freebsd-skq/sys/kern/kern_clocksource.c

967 lines
23 KiB
C
Raw Normal View History

Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
/*-
* Copyright (c) 2010-2013 Alexander Motin <mav@FreeBSD.org>
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer,
* without modification, immediately at the beginning of the file.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
* NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
* THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
/*
* Common routines to manage event timers hardware.
*/
#include "opt_device_polling.h"
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/bus.h>
#include <sys/limits.h>
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#include <sys/lock.h>
#include <sys/kdb.h>
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
#include <sys/ktr.h>
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/kernel.h>
#include <sys/sched.h>
#include <sys/smp.h>
#include <sys/sysctl.h>
#include <sys/timeet.h>
#include <sys/timetc.h>
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#include <machine/atomic.h>
#include <machine/clock.h>
#include <machine/cpu.h>
#include <machine/smp.h>
int cpu_deepest_sleep = 0; /* Deepest Cx state available. */
int cpu_disable_c2_sleep = 0; /* Timer dies in C2. */
int cpu_disable_c3_sleep = 0; /* Timer dies in C3. */
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
static void setuptimer(void);
static void loadtimer(sbintime_t now, int first);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
static int doconfigtimer(void);
static void configtimer(int start);
static int round_freq(struct eventtimer *et, int freq);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
static sbintime_t getnextcpuevent(int idle);
static sbintime_t getnextevent(void);
static int handleevents(sbintime_t now, int fake);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
static struct mtx et_hw_mtx;
#define ET_HW_LOCK(state) \
{ \
if (timer->et_flags & ET_FLAGS_PERCPU) \
mtx_lock_spin(&(state)->et_hw_mtx); \
else \
mtx_lock_spin(&et_hw_mtx); \
}
#define ET_HW_UNLOCK(state) \
{ \
if (timer->et_flags & ET_FLAGS_PERCPU) \
mtx_unlock_spin(&(state)->et_hw_mtx); \
else \
mtx_unlock_spin(&et_hw_mtx); \
}
static struct eventtimer *timer = NULL;
static sbintime_t timerperiod; /* Timer period for periodic mode. */
static sbintime_t statperiod; /* statclock() events period. */
static sbintime_t profperiod; /* profclock() events period. */
static sbintime_t nexttick; /* Next global timer tick time. */
static u_int busy = 1; /* Reconfiguration is in progress. */
static int profiling; /* Profiling events enabled. */
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
static char timername[32]; /* Wanted timer. */
TUNABLE_STR("kern.eventtimer.timer", timername, sizeof(timername));
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
static int singlemul; /* Multiplier for periodic mode. */
SYSCTL_INT(_kern_eventtimer, OID_AUTO, singlemul, CTLFLAG_RWTUN, &singlemul,
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
0, "Multiplier for periodic mode");
static u_int idletick; /* Run periodic events when idle. */
SYSCTL_UINT(_kern_eventtimer, OID_AUTO, idletick, CTLFLAG_RWTUN, &idletick,
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
0, "Run periodic events when idle");
static int periodic; /* Periodic or one-shot mode. */
static int want_periodic; /* What mode to prefer. */
TUNABLE_INT("kern.eventtimer.periodic", &want_periodic);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state {
struct mtx et_hw_mtx; /* Per-CPU timer mutex. */
u_int action; /* Reconfiguration requests. */
u_int handle; /* Immediate handle resuests. */
sbintime_t now; /* Last tick time. */
sbintime_t nextevent; /* Next scheduled event on this CPU. */
sbintime_t nexttick; /* Next timer tick time. */
sbintime_t nexthard; /* Next hardclock() event. */
sbintime_t nextstat; /* Next statclock() event. */
sbintime_t nextprof; /* Next profclock() event. */
sbintime_t nextcall; /* Next callout event. */
sbintime_t nextcallopt; /* Next optional callout event. */
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
int ipi; /* This CPU needs IPI. */
int idle; /* This CPU is in idle mode. */
};
static DPCPU_DEFINE(struct pcpu_state, timerstate);
DPCPU_DEFINE(sbintime_t, hardclocktime);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Timer broadcast IPI handler.
*/
int
hardclockintr(void)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
sbintime_t now;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
int done;
if (doconfigtimer() || busy)
return (FILTER_HANDLED);
state = DPCPU_PTR(timerstate);
now = state->now;
CTR3(KTR_SPARE2, "ipi at %d: now %d.%08x",
curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
done = handleevents(now, 0);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
return (done ? FILTER_HANDLED : FILTER_STRAY);
}
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Handle all events for specified time on this CPU
*/
static int
handleevents(sbintime_t now, int fake)
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
{
sbintime_t t, *hct;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct trapframe *frame;
struct pcpu_state *state;
int usermode;
int done, runs;
CTR3(KTR_SPARE2, "handle at %d: now %d.%08x",
curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
done = 0;
if (fake) {
frame = NULL;
usermode = 0;
} else {
frame = curthread->td_intr_frame;
usermode = TRAPF_USERMODE(frame);
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_PTR(timerstate);
runs = 0;
while (now >= state->nexthard) {
state->nexthard += tick_sbt;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
runs++;
}
if (runs) {
hct = DPCPU_PTR(hardclocktime);
*hct = state->nexthard - tick_sbt;
if (fake < 2) {
hardclock_cnt(runs, usermode);
done = 1;
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
runs = 0;
while (now >= state->nextstat) {
state->nextstat += statperiod;
runs++;
}
if (runs && fake < 2) {
statclock_cnt(runs, usermode);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
done = 1;
}
if (profiling) {
runs = 0;
while (now >= state->nextprof) {
state->nextprof += profperiod;
runs++;
}
if (runs && !fake) {
profclock_cnt(runs, usermode, TRAPF_PC(frame));
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
done = 1;
}
} else
state->nextprof = state->nextstat;
if (now >= state->nextcallopt) {
state->nextcall = state->nextcallopt = SBT_MAX;
callout_process(now);
}
t = getnextcpuevent(0);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_HW_LOCK(state);
if (!busy) {
state->idle = 0;
state->nextevent = t;
loadtimer(now, (fake == 2) &&
(timer->et_flags & ET_FLAGS_PERCPU));
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
ET_HW_UNLOCK(state);
return (done);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Schedule binuptime of the next event on current CPU.
*/
static sbintime_t
getnextcpuevent(int idle)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
sbintime_t event;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
u_int hardfreq;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_PTR(timerstate);
/* Handle hardclock() events, skipping some if CPU is idle. */
event = state->nexthard;
if (idle) {
hardfreq = (u_int)hz / 2;
if (tc_min_ticktock_freq > 2
#ifdef SMP
&& curcpu == CPU_FIRST()
#endif
)
hardfreq = hz / tc_min_ticktock_freq;
if (hardfreq > 1)
event += tick_sbt * (hardfreq - 1);
}
/* Handle callout events. */
if (event > state->nextcall)
event = state->nextcall;
if (!idle) { /* If CPU is active - handle other types of events. */
if (event > state->nextstat)
event = state->nextstat;
if (profiling && event > state->nextprof)
event = state->nextprof;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
return (event);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Schedule binuptime of the next event on all CPUs.
*/
static sbintime_t
getnextevent(void)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
sbintime_t event;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#ifdef SMP
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
int cpu;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#endif
int c;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_PTR(timerstate);
event = state->nextevent;
c = -1;
#ifdef SMP
if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
CPU_FOREACH(cpu) {
state = DPCPU_ID_PTR(cpu, timerstate);
if (event > state->nextevent) {
event = state->nextevent;
c = cpu;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
}
}
#endif
CTR4(KTR_SPARE2, "next at %d: next %d.%08x by %d",
curcpu, (int)(event >> 32), (u_int)(event & 0xffffffff), c);
return (event);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/* Hardware timer callback function. */
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
static void
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
timercb(struct eventtimer *et, void *arg)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
sbintime_t now;
sbintime_t *next;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
#ifdef SMP
int cpu, bcast;
#endif
/* Do not touch anything if somebody reconfiguring timers. */
if (busy)
return;
/* Update present and next tick times. */
state = DPCPU_PTR(timerstate);
if (et->et_flags & ET_FLAGS_PERCPU) {
next = &state->nexttick;
} else
next = &nexttick;
now = sbinuptime();
if (periodic)
*next = now + timerperiod;
else
*next = -1; /* Next tick is not scheduled yet. */
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state->now = now;
CTR3(KTR_SPARE2, "intr at %d: now %d.%08x",
curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#ifdef SMP
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#ifdef EARLY_AP_STARTUP
MPASS(mp_ncpus == 1 || smp_started);
#endif
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/* Prepare broadcasting to other CPUs for non-per-CPU timers. */
bcast = 0;
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#ifdef EARLY_AP_STARTUP
if ((et->et_flags & ET_FLAGS_PERCPU) == 0) {
#else
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if ((et->et_flags & ET_FLAGS_PERCPU) == 0 && smp_started) {
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#endif
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
CPU_FOREACH(cpu) {
state = DPCPU_ID_PTR(cpu, timerstate);
ET_HW_LOCK(state);
state->now = now;
if (now >= state->nextevent) {
state->nextevent += SBT_1S;
if (curcpu != cpu) {
state->ipi = 1;
bcast = 1;
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
ET_HW_UNLOCK(state);
}
}
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#endif
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/* Handle events for this time on this CPU. */
handleevents(now, 0);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
#ifdef SMP
/* Broadcast interrupt to other CPUs for non-per-CPU timers. */
if (bcast) {
CPU_FOREACH(cpu) {
if (curcpu == cpu)
continue;
state = DPCPU_ID_PTR(cpu, timerstate);
if (state->ipi) {
state->ipi = 0;
ipi_cpu(cpu, IPI_HARDCLOCK);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
}
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
#endif
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
/*
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
* Load new value into hardware timer.
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
*/
static void
loadtimer(sbintime_t now, int start)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
sbintime_t new;
sbintime_t *next;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
uint64_t tmp;
int eq;
if (timer->et_flags & ET_FLAGS_PERCPU) {
state = DPCPU_PTR(timerstate);
next = &state->nexttick;
} else
next = &nexttick;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (periodic) {
if (start) {
/*
* Try to start all periodic timers aligned
* to period to make events synchronous.
*/
tmp = now % timerperiod;
new = timerperiod - tmp;
if (new < tmp) /* Left less then passed. */
new += timerperiod;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
CTR5(KTR_SPARE2, "load p at %d: now %d.%08x first in %d.%08x",
curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff),
(int)(new >> 32), (u_int)(new & 0xffffffff));
*next = new + now;
et_start(timer, new, timerperiod);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
} else {
new = getnextevent();
eq = (new == *next);
CTR4(KTR_SPARE2, "load at %d: next %d.%08x eq %d",
curcpu, (int)(new >> 32), (u_int)(new & 0xffffffff), eq);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (!eq) {
*next = new;
et_start(timer, new - now, 0);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
}
}
/*
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
* Prepare event timer parameters after configuration changes.
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
*/
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
static void
setuptimer(void)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
int freq;
if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
periodic = 0;
else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
periodic = 1;
singlemul = MIN(MAX(singlemul, 1), 20);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
freq = hz * singlemul;
while (freq < (profiling ? profhz : stathz))
freq += hz;
freq = round_freq(timer, freq);
timerperiod = SBT_1S / freq;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Reconfigure specified per-CPU timer on other CPU. Called from IPI handler.
*/
static int
doconfigtimer(void)
{
sbintime_t now;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
state = DPCPU_PTR(timerstate);
switch (atomic_load_acq_int(&state->action)) {
case 1:
now = sbinuptime();
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_HW_LOCK(state);
loadtimer(now, 1);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_HW_UNLOCK(state);
state->handle = 0;
atomic_store_rel_int(&state->action, 0);
return (1);
case 2:
ET_HW_LOCK(state);
et_stop(timer);
ET_HW_UNLOCK(state);
state->handle = 0;
atomic_store_rel_int(&state->action, 0);
return (1);
}
if (atomic_readandclear_int(&state->handle) && !busy) {
now = sbinuptime();
handleevents(now, 0);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
return (1);
}
return (0);
}
/*
* Reconfigure specified timer.
* For per-CPU timers use IPI to make other CPUs to reconfigure.
*/
static void
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
configtimer(int start)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
sbintime_t now, next;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
int cpu;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (start) {
setuptimer();
now = sbinuptime();
} else
now = 0;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
critical_enter();
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_HW_LOCK(DPCPU_PTR(timerstate));
if (start) {
/* Initialize time machine parameters. */
next = now + timerperiod;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (periodic)
nexttick = next;
else
nexttick = -1;
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#ifdef EARLY_AP_STARTUP
MPASS(mp_ncpus == 1 || smp_started);
#endif
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
CPU_FOREACH(cpu) {
state = DPCPU_ID_PTR(cpu, timerstate);
state->now = now;
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#ifndef EARLY_AP_STARTUP
if (!smp_started && cpu != CPU_FIRST())
state->nextevent = SBT_MAX;
else
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#endif
state->nextevent = next;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (periodic)
state->nexttick = next;
else
state->nexttick = -1;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state->nexthard = next;
state->nextstat = next;
state->nextprof = next;
state->nextcall = next;
state->nextcallopt = next;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
hardclock_sync(cpu);
}
busy = 0;
/* Start global timer or per-CPU timer of this CPU. */
loadtimer(now, 1);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
} else {
busy = 1;
/* Stop global timer or per-CPU timer of this CPU. */
et_stop(timer);
}
ET_HW_UNLOCK(DPCPU_PTR(timerstate));
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
#ifdef SMP
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#ifdef EARLY_AP_STARTUP
/* If timer is global we are done. */
if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
#else
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/* If timer is global or there is no other CPUs yet - we are done. */
if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || !smp_started) {
Add an EARLY_AP_STARTUP option to start APs earlier during boot. Currently, Application Processors (non-boot CPUs) are started by MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until SI_SUB_SMP at which point they are released to run kernel threads. SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter the scheduler and start running threads until fairly late in the boot. This change moves SI_SUB_SMP up to just before software interrupt threads are created allowing the APs to start executing kernel threads much sooner (before any devices are probed). This allows several initialization routines that need to perform initialization on all CPUs to now perform that initialization in one step rather than having to defer the AP initialization to a second SYSINIT run at SI_SUB_SMP. It also permits all CPUs to be available for handling interrupts before any devices are probed. This last feature fixes a problem on with interrupt vector exhaustion. Specifically, in the old model all device interrupts were routed onto the boot CPU during boot. Later after the APs were released at SI_SUB_SMP, interrupts were redistributed across all CPUs. However, several drivers for multiqueue hardware allocate N interrupts per CPU in the system. In a system with many CPUs, just a few drivers doing this could exhaust the available pool of interrupt vectors on the boot CPU as each driver was allocating N * mp_ncpu vectors on the boot CPU. Now, drivers will allocate interrupts on their desired CPUs during boot meaning that only N interrupts are allocated from the boot CPU instead of N * mp_ncpu. Some other bits of code can also be simplified as smp_started is now true much earlier and will now always be true for these bits of code. This removes the need to treat the single-CPU boot environment as a special case. As a transition aid, the new behavior is available under a new kernel option (EARLY_AP_STARTUP). This will allow the option to be turned off if need be during initial testing. I plan to enable this on x86 by default in a followup commit in the next few days and to have all platforms moved over before 11.0. Once the transition is complete, the option will be removed along with the !EARLY_AP_STARTUP code. These changes have only been tested on x86. Other platform maintainers are encouraged to port their architectures over as well. The main things to check for are any uses of smp_started in MD code that can be simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in the EARLY_AP_STARTUP case (e.g. the interrupt shuffling). PR: kern/199321 Reviewed by: markj, gnn, kib Sponsored by: Netflix
2016-05-14 18:22:52 +00:00
#endif
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
critical_exit();
return;
}
/* Set reconfigure flags for other CPUs. */
CPU_FOREACH(cpu) {
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_ID_PTR(cpu, timerstate);
atomic_store_rel_int(&state->action,
(cpu == curcpu) ? 0 : ( start ? 1 : 2));
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/* Broadcast reconfigure IPI. */
ipi_all_but_self(IPI_HARDCLOCK);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
/* Wait for reconfiguration completed. */
restart:
cpu_spinwait();
CPU_FOREACH(cpu) {
if (cpu == curcpu)
continue;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_ID_PTR(cpu, timerstate);
if (atomic_load_acq_int(&state->action))
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
goto restart;
}
#endif
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
critical_exit();
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Calculate nearest frequency supported by hardware timer.
*/
static int
round_freq(struct eventtimer *et, int freq)
{
uint64_t div;
if (et->et_frequency != 0) {
div = lmax((et->et_frequency + freq / 2) / freq, 1);
if (et->et_flags & ET_FLAGS_POW2DIV)
div = 1 << (flsl(div + div / 2) - 1);
freq = (et->et_frequency + div / 2) / div;
}
if (et->et_min_period > SBT_1S)
panic("Event timer \"%s\" doesn't support sub-second periods!",
et->et_name);
else if (et->et_min_period != 0)
freq = min(freq, SBT2FREQ(et->et_min_period));
if (et->et_max_period < SBT_1S && et->et_max_period != 0)
freq = max(freq, SBT2FREQ(et->et_max_period));
return (freq);
}
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
/*
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
* Configure and start event timers (BSP part).
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
*/
void
cpu_initclocks_bsp(void)
{
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
int base, div, cpu;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
mtx_init(&et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
CPU_FOREACH(cpu) {
state = DPCPU_ID_PTR(cpu, timerstate);
mtx_init(&state->et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
state->nextcall = SBT_MAX;
state->nextcallopt = SBT_MAX;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
periodic = want_periodic;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/* Grab requested timer or the best of present. */
if (timername[0])
timer = et_find(timername, 0, 0);
if (timer == NULL && periodic) {
timer = et_find(NULL,
ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
}
if (timer == NULL) {
timer = et_find(NULL,
ET_FLAGS_ONESHOT, ET_FLAGS_ONESHOT);
}
if (timer == NULL && !periodic) {
timer = et_find(NULL,
ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
}
if (timer == NULL)
panic("No usable event timer found!");
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
et_init(timer, timercb, NULL, NULL);
/* Adapt to timer capabilities. */
if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
periodic = 0;
else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
periodic = 1;
if (timer->et_flags & ET_FLAGS_C3STOP)
cpu_disable_c3_sleep++;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
/*
* We honor the requested 'hz' value.
* We want to run stathz in the neighborhood of 128hz.
* We would like profhz to run as often as possible.
*/
if (singlemul <= 0 || singlemul > 20) {
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
if (hz >= 1500 || (hz % 128) == 0)
singlemul = 1;
else if (hz >= 750)
singlemul = 2;
else
singlemul = 4;
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (periodic) {
base = round_freq(timer, hz * singlemul);
singlemul = max((base + hz / 2) / hz, 1);
hz = (base + singlemul / 2) / singlemul;
if (base <= 128)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
stathz = base;
else {
div = base / 128;
if (div >= singlemul && (div % singlemul) == 0)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
div++;
stathz = base / div;
}
profhz = stathz;
while ((profhz + stathz) <= 128 * 64)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
profhz += stathz;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
profhz = round_freq(timer, profhz);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
} else {
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
hz = round_freq(timer, hz);
stathz = round_freq(timer, 127);
profhz = round_freq(timer, stathz * 64);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
tick = 1000000 / hz;
tick_sbt = SBT_1S / hz;
tick_bt = sbttobt(tick_sbt);
statperiod = SBT_1S / stathz;
profperiod = SBT_1S / profhz;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
ET_LOCK();
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
configtimer(1);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
ET_UNLOCK();
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Start per-CPU event timers on APs.
*/
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
void
cpu_initclocks_ap(void)
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
{
sbintime_t now;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
struct thread *td;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_PTR(timerstate);
now = sbinuptime();
ET_HW_LOCK(state);
state->now = now;
hardclock_sync(curcpu);
spinlock_enter();
ET_HW_UNLOCK(state);
td = curthread;
td->td_intr_nesting_level++;
handleevents(state->now, 2);
td->td_intr_nesting_level--;
spinlock_exit();
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
/*
* Switch to profiling clock rates.
*/
void
cpu_startprofclock(void)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
ET_LOCK();
if (profiling == 0) {
if (periodic) {
configtimer(0);
profiling = 1;
configtimer(1);
} else
profiling = 1;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
} else
profiling++;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
ET_UNLOCK();
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Switch to regular clock rates.
*/
void
cpu_stopprofclock(void)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_LOCK();
if (profiling == 1) {
if (periodic) {
configtimer(0);
profiling = 0;
configtimer(1);
} else
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
profiling = 0;
} else
profiling--;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_UNLOCK();
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Switch to idle mode (all ticks handled).
*/
sbintime_t
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
cpu_idleclock(void)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
sbintime_t now, t;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (idletick || busy ||
(periodic && (timer->et_flags & ET_FLAGS_PERCPU))
#ifdef DEVICE_POLLING
|| curcpu == CPU_FIRST()
#endif
)
return (-1);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_PTR(timerstate);
if (periodic)
now = state->now;
else
now = sbinuptime();
CTR3(KTR_SPARE2, "idle at %d: now %d.%08x",
curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
t = getnextcpuevent(1);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_HW_LOCK(state);
state->idle = 1;
state->nextevent = t;
if (!periodic)
loadtimer(now, 0);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_HW_UNLOCK(state);
return (MAX(t - now, 0));
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Switch to active mode (skip empty ticks).
*/
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
void
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
cpu_activeclock(void)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
sbintime_t now;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
struct pcpu_state *state;
struct thread *td;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_PTR(timerstate);
if (state->idle == 0 || busy)
return;
if (periodic)
now = state->now;
else
now = sbinuptime();
CTR3(KTR_SPARE2, "active at %d: now %d.%08x",
curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
spinlock_enter();
td = curthread;
td->td_intr_nesting_level++;
handleevents(now, 1);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
td->td_intr_nesting_level--;
spinlock_exit();
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
}
/*
* Change the frequency of the given timer. This changes et->et_frequency and
* if et is the active timer it reconfigures the timer on all CPUs. This is
* intended to be a private interface for the use of et_change_frequency() only.
*/
void
cpu_et_frequency(struct eventtimer *et, uint64_t newfreq)
{
ET_LOCK();
if (et == timer) {
configtimer(0);
et->et_frequency = newfreq;
configtimer(1);
} else
et->et_frequency = newfreq;
ET_UNLOCK();
}
void
cpu_new_callout(int cpu, sbintime_t bt, sbintime_t bt_opt)
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
{
struct pcpu_state *state;
/* Do not touch anything if somebody reconfiguring timers. */
if (busy)
return;
CTR6(KTR_SPARE2, "new co at %d: on %d at %d.%08x - %d.%08x",
curcpu, cpu, (int)(bt_opt >> 32), (u_int)(bt_opt & 0xffffffff),
(int)(bt >> 32), (u_int)(bt & 0xffffffff));
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
state = DPCPU_ID_PTR(cpu, timerstate);
ET_HW_LOCK(state);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* If there is callout time already set earlier -- do nothing.
* This check may appear redundant because we check already in
* callout_process() but this double check guarantees we're safe
* with respect to race conditions between interrupts execution
* and scheduling.
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
*/
state->nextcallopt = bt_opt;
if (bt >= state->nextcall)
goto done;
state->nextcall = bt;
/* If there is some other event set earlier -- do nothing. */
if (bt >= state->nextevent)
goto done;
state->nextevent = bt;
/* If timer is periodic -- there is nothing to reprogram. */
if (periodic)
goto done;
/* If timer is global or of the current CPU -- reprogram it. */
if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || cpu == curcpu) {
loadtimer(sbinuptime(), 0);
done:
ET_HW_UNLOCK(state);
return;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
}
/* Otherwise make other CPU to reprogram it. */
state->handle = 1;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_HW_UNLOCK(state);
#ifdef SMP
ipi_cpu(cpu, IPI_HARDCLOCK);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
#endif
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Report or change the active event timers hardware.
*/
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
static int
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
sysctl_kern_eventtimer_timer(SYSCTL_HANDLER_ARGS)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
char buf[32];
struct eventtimer *et;
int error;
ET_LOCK();
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
et = timer;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
snprintf(buf, sizeof(buf), "%s", et->et_name);
ET_UNLOCK();
error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
ET_LOCK();
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
et = timer;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
if (error != 0 || req->newptr == NULL ||
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
strcasecmp(buf, et->et_name) == 0) {
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
ET_UNLOCK();
return (error);
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
et = et_find(buf, 0, 0);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
if (et == NULL) {
ET_UNLOCK();
return (ENOENT);
}
configtimer(0);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
et_free(timer);
if (et->et_flags & ET_FLAGS_C3STOP)
cpu_disable_c3_sleep++;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
if (timer->et_flags & ET_FLAGS_C3STOP)
cpu_disable_c3_sleep--;
periodic = want_periodic;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
timer = et;
et_init(timer, timercb, NULL, NULL);
configtimer(1);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
ET_UNLOCK();
return (error);
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
SYSCTL_PROC(_kern_eventtimer, OID_AUTO, timer,
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
0, 0, sysctl_kern_eventtimer_timer, "A", "Chosen event timer");
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
/*
* Report or change the active event timer periodicity.
*/
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
static int
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
sysctl_kern_eventtimer_periodic(SYSCTL_HANDLER_ARGS)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
{
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
int error, val;
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
val = periodic;
error = sysctl_handle_int(oidp, &val, 0, req);
if (error != 0 || req->newptr == NULL)
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
return (error);
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
ET_LOCK();
configtimer(0);
periodic = want_periodic = val;
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
configtimer(1);
Implement new event timers infrastructure. It provides unified APIs for writing event timer drivers, for choosing best possible drivers by machine independent code and for operating them to supply kernel with hardclock(), statclock() and profclock() events in unified fashion on various hardware. Infrastructure provides support for both per-CPU (independent for every CPU core) and global timers in periodic and one-shot modes. MI management code at this moment uses only periodic mode, but one-shot mode use planned for later, as part of tickless kernel project. For this moment infrastructure used on i386 and amd64 architectures. Other archs are welcome to follow, while their current operation should not be affected. This patch updates existing drivers (i8254, RTC and LAPIC) for the new order, and adds event timers support into the HPET driver. These drivers have different capabilities: LAPIC - per-CPU timer, supports periodic and one-shot operation, may freeze in C3 state, calibrated on first use, so may be not exactly precise. HPET - depending on hardware can work as per-CPU or global, supports periodic and one-shot operation, usually provides several event timers. i8254 - global, limited to periodic mode, because same hardware used also as time counter. RTC - global, supports only periodic mode, set of frequencies in Hz limited by powers of 2. Depending on hardware capabilities, drivers preferred in following orders, either LAPIC, HPETs, i8254, RTC or HPETs, LAPIC, i8254, RTC. User may explicitly specify wanted timers via loader tunables or sysctls: kern.eventtimer.timer1 and kern.eventtimer.timer2. If requested driver is unavailable or unoperational, system will try to replace it. If no more timers available or "NONE" specified for second, system will operate using only one timer, multiplying it's frequency by few times and uing respective dividers to honor hz, stathz and profhz values, set during initial setup.
2010-06-20 21:33:29 +00:00
ET_UNLOCK();
return (error);
}
Refactor timer management code with priority to one-shot operation mode. The main goal of this is to generate timer interrupts only when there is some work to do. When CPU is busy interrupts are generating at full rate of hz + stathz to fullfill scheduler and timekeeping requirements. But when CPU is idle, only minimum set of interrupts (down to 8 interrupts per second per CPU now), needed to handle scheduled callouts is executed. This allows significantly increase idle CPU sleep time, increasing effect of static power-saving technologies. Also it should reduce host CPU load on virtualized systems, when guest system is idle. There is set of tunables, also available as writable sysctls, allowing to control wanted event timer subsystem behavior: kern.eventtimer.timer - allows to choose event timer hardware to use. On x86 there is up to 4 different kinds of timers. Depending on whether chosen timer is per-CPU, behavior of other options slightly differs. kern.eventtimer.periodic - allows to choose periodic and one-shot operation mode. In periodic mode, current timer hardware taken as the only source of time for time events. This mode is quite alike to previous kernel behavior. One-shot mode instead uses currently selected time counter hardware to schedule all needed events one by one and program timer to generate interrupt exactly in specified time. Default value depends of chosen timer capabilities, but one-shot mode is preferred, until other is forced by user or hardware. kern.eventtimer.singlemul - in periodic mode specifies how much times higher timer frequency should be, to not strictly alias hardclock() and statclock() events. Default values are 2 and 4, but could be reduced to 1 if extra interrupts are unwanted. kern.eventtimer.idletick - makes each CPU to receive every timer interrupt independently of whether they busy or not. By default this options is disabled. If chosen timer is per-CPU and runs in periodic mode, this option has no effect - all interrupts are generating. As soon as this patch modifies cpu_idle() on some platforms, I have also refactored one on x86. Now it makes use of MONITOR/MWAIT instrunctions (if supported) under high sleep/wakeup rate, as fast alternative to other methods. It allows SMP scheduler to wake up sleeping CPUs much faster without using IPI, significantly increasing performance on some highly task-switching loads. Tested by: many (on i386, amd64, sparc64 and powerc) H/W donated by: Gheorghe Ardelean Sponsored by: iXsystems, Inc.
2010-09-13 07:25:35 +00:00
SYSCTL_PROC(_kern_eventtimer, OID_AUTO, periodic,
CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
0, 0, sysctl_kern_eventtimer_periodic, "I", "Enable event timer periodic mode");
#include "opt_ddb.h"
#ifdef DDB
#include <ddb/ddb.h>
DB_SHOW_COMMAND(clocksource, db_show_clocksource)
{
struct pcpu_state *st;
int c;
CPU_FOREACH(c) {
st = DPCPU_ID_PTR(c, timerstate);
db_printf(
"CPU %2d: action %d handle %d ipi %d idle %d\n"
" now %#jx nevent %#jx (%jd)\n"
" ntick %#jx (%jd) nhard %#jx (%jd)\n"
" nstat %#jx (%jd) nprof %#jx (%jd)\n"
" ncall %#jx (%jd) ncallopt %#jx (%jd)\n",
c, st->action, st->handle, st->ipi, st->idle,
(uintmax_t)st->now,
(uintmax_t)st->nextevent,
(uintmax_t)(st->nextevent - st->now) / tick_sbt,
(uintmax_t)st->nexttick,
(uintmax_t)(st->nexttick - st->now) / tick_sbt,
(uintmax_t)st->nexthard,
(uintmax_t)(st->nexthard - st->now) / tick_sbt,
(uintmax_t)st->nextstat,
(uintmax_t)(st->nextstat - st->now) / tick_sbt,
(uintmax_t)st->nextprof,
(uintmax_t)(st->nextprof - st->now) / tick_sbt,
(uintmax_t)st->nextcall,
(uintmax_t)(st->nextcall - st->now) / tick_sbt,
(uintmax_t)st->nextcallopt,
(uintmax_t)(st->nextcallopt - st->now) / tick_sbt);
}
}
#endif