freebsd-dev/sys/kern/sched_ule.c

2926 lines
77 KiB
C
Raw Normal View History

/*-
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
* 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 unmodified, this list of conditions, and the following
* disclaimer.
* 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.
*/
/*
* This file implements the ULE scheduler. ULE supports independent CPU
* run queues and fine grain locking. It has superior interactive
* performance under load even on uni-processor systems.
*
* etymology:
* ULE is the last three letters in schedule. It owes its name to a
* generic user created for a scheduling system by Paul Mikesell at
* Isilon Systems and a general lack of creativity on the part of the author.
*/
2003-06-11 00:56:59 +00:00
#include <sys/cdefs.h>
2009-04-29 03:26:30 +00:00
__FBSDID("$FreeBSD$");
2003-06-11 00:56:59 +00:00
#include "opt_hwpmc_hooks.h"
#include "opt_sched.h"
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kdb.h>
#include <sys/kernel.h>
#include <sys/ktr.h>
2014-03-19 00:38:27 +00:00
#include <sys/limits.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/resource.h>
#include <sys/resourcevar.h>
#include <sys/sched.h>
#include <sys/sdt.h>
#include <sys/smp.h>
#include <sys/sx.h>
#include <sys/sysctl.h>
#include <sys/sysproto.h>
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
#include <sys/turnstile.h>
#include <sys/umtx.h>
#include <sys/vmmeter.h>
#include <sys/cpuset.h>
#include <sys/sbuf.h>
#ifdef HWPMC_HOOKS
#include <sys/pmckern.h>
#endif
#ifdef KDTRACE_HOOKS
#include <sys/dtrace_bsd.h>
int dtrace_vtime_active;
dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
#endif
#include <machine/cpu.h>
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#include <machine/smp.h>
#define KTR_ULE 0
#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
/*
* Thread scheduler specific section. All fields are protected
* by the thread lock.
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
*/
struct td_sched {
struct runq *ts_runq; /* Run-queue we're queued on. */
short ts_flags; /* TSF_* flags. */
int ts_cpu; /* CPU that we have affinity for. */
int ts_rltick; /* Real last tick, for affinity. */
int ts_slice; /* Ticks of slice remaining. */
u_int ts_slptime; /* Number of ticks we vol. slept */
u_int ts_runtime; /* Number of ticks we were running */
int ts_ltick; /* Last tick that we were running on */
int ts_ftick; /* First tick that we were running on */
int ts_ticks; /* Tick count */
#ifdef KTR
char ts_name[TS_NAME_LEN];
#endif
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
};
/* flags kept in ts_flags */
#define TSF_BOUND 0x0001 /* Thread can not migrate. */
#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
#define THREAD_CAN_SCHED(td, cpu) \
CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
_Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
sizeof(struct thread0_storage),
"increase struct thread0_storage.t0st_sched size");
/*
* Priority ranges used for interactive and non-interactive timeshare
* threads. The timeshare priorities are split up into four ranges.
* The first range handles interactive threads. The last three ranges
* (NHALF, x, and NHALF) handle non-interactive threads with the outer
* ranges supporting nice values.
*/
#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
#define PRI_MAX_BATCH PRI_MAX_TIMESHARE
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* Cpu percentage computation macros and defines.
*
* SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
* SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
* SCHED_TICK_MAX: Maximum number of ticks before scaling back.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
* SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
* SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
*/
#define SCHED_TICK_SECS 10
#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_TICK_SHIFT 10
#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* These macros determine priorities for non-interactive threads. They are
* assigned a priority based on their recent cpu utilization as expressed
* by the ratio of ticks to the tick total. NHALF priorities at the start
* and end of the MIN to MAX timeshare range are only reachable with negative
* or positive nice respectively.
*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* PRI_RANGE: Priority range for utilization dependent priorities.
* PRI_NRESV: Number of nice values.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
* PRI_NICE: Determines the part of the priority inherited from nice.
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_PRI_TICKS(ts) \
(SCHED_TICK_HZ((ts)) / \
(roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_PRI_NICE(nice) (nice)
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* These determine the interactivity of a process. Interactivity differs from
* cpu utilization in that it expresses the voluntary time slept vs time ran
* while cpu utilization includes all time not running. This more accurately
* models the intent of the thread.
*
* SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
* before throttling back.
* SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
* INTERACT_MAX: Maximum interactivity value. Smaller is better.
* INTERACT_THRESH: Threshold for placement on the current runq.
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
#define SCHED_INTERACT_MAX (100)
#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
#define SCHED_INTERACT_THRESH (30)
/*
* These parameters determine the slice behavior for batch work.
*/
#define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
#define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
/* Flags kept in td_flags. */
#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* tickincr: Converts a stathz tick into a hz domain scaled by
* the shift factor. Without the shift the error rate
* due to rounding would be unacceptably high.
* realstathz: stathz is sometimes 0 and run off of hz.
* sched_slice: Runtime of each thread before rescheduling.
* preempt_thresh: Priority threshold for preemption and remote IPIs.
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static int sched_interact = SCHED_INTERACT_THRESH;
static int tickincr = 8 << SCHED_TICK_SHIFT;
static int realstathz = 127; /* reset during boot. */
static int sched_slice = 10; /* reset during boot. */
static int sched_slice_min = 1; /* reset during boot. */
#ifdef PREEMPTION
#ifdef FULL_PREEMPTION
static int preempt_thresh = PRI_MAX_IDLE;
#else
static int preempt_thresh = PRI_MIN_KERN;
#endif
#else
static int preempt_thresh = 0;
#endif
static int static_boost = PRI_MIN_BATCH;
static int sched_idlespins = 10000;
static int sched_idlespinthresh = -1;
/*
* tdq - per processor runqs and statistics. All fields are protected by the
* tdq_lock. The load and lowpri may be accessed without to avoid excess
* locking in sched_pickcpu();
*/
struct tdq {
/*
* Ordered to improve efficiency of cpu_search() and switch().
* tdq_lock is padded to avoid false sharing with tdq_load and
* tdq_cpu_idle.
*/
struct mtx_padalign tdq_lock; /* run queue lock. */
struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
volatile int tdq_load; /* Aggregate load. */
volatile int tdq_cpu_idle; /* cpu_idle() is active. */
int tdq_sysload; /* For loadavg, !ITHD load. */
int tdq_transferable; /* Transferable thread count. */
short tdq_switchcnt; /* Switches this tick. */
short tdq_oldswitchcnt; /* Switches last tick. */
u_char tdq_lowpri; /* Lowest priority thread. */
u_char tdq_ipipending; /* IPI pending. */
u_char tdq_idx; /* Current insert index. */
u_char tdq_ridx; /* Current removal index. */
struct runq tdq_realtime; /* real-time run queue. */
struct runq tdq_timeshare; /* timeshare run queue. */
struct runq tdq_idle; /* Queue of IDLE threads. */
char tdq_name[TDQ_NAME_LEN];
#ifdef KTR
char tdq_loadname[TDQ_LOADNAME_LEN];
#endif
} __aligned(64);
/* Idle thread states and config. */
#define TDQ_RUNNING 1
#define TDQ_IDLE 2
#ifdef SMP
struct cpu_group *cpu_top; /* CPU topology */
#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
/*
* Run-time tunables.
*/
static int rebalance = 1;
static int balance_interval = 128; /* Default set in sched_initticks(). */
static int affinity;
static int steal_idle = 1;
static int steal_thresh = 2;
/*
* One thread queue per processor.
*/
static struct tdq tdq_cpu[MAXCPU];
static struct tdq *balance_tdq;
static int balance_ticks;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
static DPCPU_DEFINE(uint32_t, randomval);
#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
#define TDQ_CPU(x) (&tdq_cpu[(x)])
#define TDQ_ID(x) ((int)((x) - tdq_cpu))
#else /* !SMP */
static struct tdq tdq_cpu;
#define TDQ_ID(x) (0)
#define TDQ_SELF() (&tdq_cpu)
#define TDQ_CPU(x) (&tdq_cpu)
#endif
#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
#define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
static void sched_priority(struct thread *);
static void sched_thread_priority(struct thread *, u_char);
static int sched_interact_score(struct thread *);
static void sched_interact_update(struct thread *);
static void sched_interact_fork(struct thread *);
static void sched_pctcpu_update(struct td_sched *, int);
/* Operations on per processor queues */
static struct thread *tdq_choose(struct tdq *);
static void tdq_setup(struct tdq *);
static void tdq_load_add(struct tdq *, struct thread *);
static void tdq_load_rem(struct tdq *, struct thread *);
static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
static __inline void tdq_runq_rem(struct tdq *, struct thread *);
static inline int sched_shouldpreempt(int, int, int);
void tdq_print(int cpu);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void runq_print(struct runq *rq);
static void tdq_add(struct tdq *, struct thread *, int);
#ifdef SMP
static int tdq_move(struct tdq *, struct tdq *);
static int tdq_idled(struct tdq *);
static void tdq_notify(struct tdq *, struct thread *);
static struct thread *tdq_steal(struct tdq *, int);
static struct thread *runq_steal(struct runq *, int);
static int sched_pickcpu(struct thread *, int);
static void sched_balance(void);
static int sched_balance_pair(struct tdq *, struct tdq *);
static inline struct tdq *sched_setcpu(struct thread *, int, int);
static inline void thread_unblock_switch(struct thread *, struct mtx *);
static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
struct cpu_group *cg, int indent);
#endif
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void sched_setup(void *dummy);
SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void sched_initticks(void *dummy);
SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
NULL);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
SDT_PROVIDER_DEFINE(sched);
SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
"struct proc *", "uint8_t");
SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
"struct proc *", "void *");
SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
"struct proc *", "void *", "int");
SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
"struct proc *", "uint8_t", "struct thread *");
SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
"struct proc *");
SDT_PROBE_DEFINE(sched, , , on__cpu);
SDT_PROBE_DEFINE(sched, , , remain__cpu);
SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
"struct proc *");
/*
* Print the threads waiting on a run-queue.
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void
runq_print(struct runq *rq)
{
struct rqhead *rqh;
struct thread *td;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
int pri;
int j;
int i;
for (i = 0; i < RQB_LEN; i++) {
printf("\t\trunq bits %d 0x%zx\n",
i, rq->rq_status.rqb_bits[i]);
for (j = 0; j < RQB_BPW; j++)
if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
pri = j + (i << RQB_L2BPW);
rqh = &rq->rq_queues[pri];
TAILQ_FOREACH(td, rqh, td_runq) {
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
td, td->td_name, td->td_priority,
td->td_rqindex, pri);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
}
}
}
}
/*
* Print the status of a per-cpu thread queue. Should be a ddb show cmd.
*/
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
void
tdq_print(int cpu)
{
struct tdq *tdq;
tdq = TDQ_CPU(cpu);
printf("tdq %d:\n", TDQ_ID(tdq));
printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
printf("\tLock name: %s\n", tdq->tdq_name);
printf("\tload: %d\n", tdq->tdq_load);
printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
printf("\tload transferable: %d\n", tdq->tdq_transferable);
printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
printf("\trealtime runq:\n");
runq_print(&tdq->tdq_realtime);
printf("\ttimeshare runq:\n");
runq_print(&tdq->tdq_timeshare);
printf("\tidle runq:\n");
runq_print(&tdq->tdq_idle);
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
}
static inline int
sched_shouldpreempt(int pri, int cpri, int remote)
{
/*
* If the new priority is not better than the current priority there is
* nothing to do.
*/
if (pri >= cpri)
return (0);
/*
* Always preempt idle.
*/
if (cpri >= PRI_MIN_IDLE)
return (1);
/*
* If preemption is disabled don't preempt others.
*/
if (preempt_thresh == 0)
return (0);
/*
* Preempt if we exceed the threshold.
*/
if (pri <= preempt_thresh)
return (1);
/*
* If we're interactive or better and there is non-interactive
* or worse running preempt only remote processors.
*/
if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
return (1);
return (0);
}
/*
* Add a thread to the actual run-queue. Keeps transferable counts up to
* date with what is actually on the run-queue. Selects the correct
* queue position for timeshare threads.
*/
static __inline void
tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
{
struct td_sched *ts;
u_char pri;
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
THREAD_LOCK_ASSERT(td, MA_OWNED);
pri = td->td_priority;
ts = td_get_sched(td);
TD_SET_RUNQ(td);
if (THREAD_CAN_MIGRATE(td)) {
tdq->tdq_transferable++;
ts->ts_flags |= TSF_XFERABLE;
}
if (pri < PRI_MIN_BATCH) {
ts->ts_runq = &tdq->tdq_realtime;
} else if (pri <= PRI_MAX_BATCH) {
ts->ts_runq = &tdq->tdq_timeshare;
KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
("Invalid priority %d on timeshare runq", pri));
/*
* This queue contains only priorities between MIN and MAX
* realtime. Use the whole queue to represent these values.
*/
if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
pri = (pri + tdq->tdq_idx) % RQ_NQS;
/*
* This effectively shortens the queue by one so we
* can have a one slot difference between idx and
* ridx while we wait for threads to drain.
*/
if (tdq->tdq_ridx != tdq->tdq_idx &&
pri == tdq->tdq_ridx)
pri = (unsigned char)(pri - 1) % RQ_NQS;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
} else
pri = tdq->tdq_ridx;
runq_add_pri(ts->ts_runq, td, pri, flags);
return;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
} else
ts->ts_runq = &tdq->tdq_idle;
runq_add(ts->ts_runq, td, flags);
}
/*
* Remove a thread from a run-queue. This typically happens when a thread
* is selected to run. Running threads are not on the queue and the
* transferable count does not reflect them.
*/
static __inline void
tdq_runq_rem(struct tdq *tdq, struct thread *td)
{
struct td_sched *ts;
ts = td_get_sched(td);
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
KASSERT(ts->ts_runq != NULL,
("tdq_runq_remove: thread %p null ts_runq", td));
if (ts->ts_flags & TSF_XFERABLE) {
tdq->tdq_transferable--;
ts->ts_flags &= ~TSF_XFERABLE;
}
if (ts->ts_runq == &tdq->tdq_timeshare) {
if (tdq->tdq_idx != tdq->tdq_ridx)
runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
else
runq_remove_idx(ts->ts_runq, td, NULL);
} else
runq_remove(ts->ts_runq, td);
}
/*
* Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
* for this thread to the referenced thread queue.
*/
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
static void
tdq_load_add(struct tdq *tdq, struct thread *td)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
{
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
THREAD_LOCK_ASSERT(td, MA_OWNED);
tdq->tdq_load++;
if ((td->td_flags & TDF_NOLOAD) == 0)
tdq->tdq_sysload++;
KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
/*
* Remove the load from a thread that is transitioning to a sleep state or
* exiting.
*/
static void
tdq_load_rem(struct tdq *tdq, struct thread *td)
{
THREAD_LOCK_ASSERT(td, MA_OWNED);
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
KASSERT(tdq->tdq_load != 0,
("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
tdq->tdq_load--;
if ((td->td_flags & TDF_NOLOAD) == 0)
tdq->tdq_sysload--;
KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
}
/*
* Bound timeshare latency by decreasing slice size as load increases. We
* consider the maximum latency as the sum of the threads waiting to run
* aside from curthread and target no more than sched_slice latency but
* no less than sched_slice_min runtime.
*/
static inline int
tdq_slice(struct tdq *tdq)
{
int load;
/*
* It is safe to use sys_load here because this is called from
* contexts where timeshare threads are running and so there
* cannot be higher priority load in the system.
*/
load = tdq->tdq_sysload - 1;
if (load >= SCHED_SLICE_MIN_DIVISOR)
return (sched_slice_min);
if (load <= 1)
return (sched_slice);
return (sched_slice / load);
}
/*
* Set lowpri to its exact value by searching the run-queue and
* evaluating curthread. curthread may be passed as an optimization.
*/
static void
tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
{
struct thread *td;
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
if (ctd == NULL)
ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
td = tdq_choose(tdq);
if (td == NULL || td->td_priority > ctd->td_priority)
tdq->tdq_lowpri = ctd->td_priority;
else
tdq->tdq_lowpri = td->td_priority;
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
#ifdef SMP
/*
* We need some randomness. Implement a classic Linear Congruential
* Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
* m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
* of the random state (in the low bits of our answer) to keep
* the maximum randomness.
*/
static uint32_t
sched_random(void)
{
uint32_t *rndptr;
rndptr = DPCPU_PTR(randomval);
*rndptr = *rndptr * 69069 + 5;
return (*rndptr >> 16);
}
struct cpu_search {
cpuset_t cs_mask;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
u_int cs_prefer;
int cs_pri; /* Min priority for low. */
int cs_limit; /* Max load for low, min load for high. */
int cs_cpu;
int cs_load;
};
#define CPU_SEARCH_LOWEST 0x1
#define CPU_SEARCH_HIGHEST 0x2
#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
#define CPUSET_FOREACH(cpu, mask) \
for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
Commit the support for removing cpumask_t and replacing it directly with cpuset_t objects. That is going to offer the underlying support for a simple bump of MAXCPU and then support for number of cpus > 32 (as it is today). Right now, cpumask_t is an int, 32 bits on all our supported architecture. cpumask_t on the other side is implemented as an array of longs, and easilly extendible by definition. The architectures touched by this commit are the following: - amd64 - i386 - pc98 - arm - ia64 - XEN while the others are still missing. Userland is believed to be fully converted with the changes contained here. Some technical notes: - This commit may be considered an ABI nop for all the architectures different from amd64 and ia64 (and sparc64 in the future) - per-cpu members, which are now converted to cpuset_t, needs to be accessed avoiding migration, because the size of cpuset_t should be considered unknown - size of cpuset_t objects is different from kernel and userland (this is primirally done in order to leave some more space in userland to cope with KBI extensions). If you need to access kernel cpuset_t from the userland please refer to example in this patch on how to do that correctly (kgdb may be a good source, for example). - Support for other architectures is going to be added soon - Only MAXCPU for amd64 is bumped now The patch has been tested by sbruno and Nicholas Esborn on opteron 4 x 12 pack CPUs. More testing on big SMP is expected to came soon. pluknet tested the patch with his 8-ways on both amd64 and i386. Tested by: pluknet, sbruno, gianni, Nicholas Esborn Reviewed by: jeff, jhb, sbruno
2011-05-05 14:39:14 +00:00
if (CPU_ISSET(cpu, &mask))
static __always_inline int cpu_search(const struct cpu_group *cg,
struct cpu_search *low, struct cpu_search *high, const int match);
int __noinline cpu_search_lowest(const struct cpu_group *cg,
struct cpu_search *low);
int __noinline cpu_search_highest(const struct cpu_group *cg,
struct cpu_search *high);
int __noinline cpu_search_both(const struct cpu_group *cg,
struct cpu_search *low, struct cpu_search *high);
/*
* Search the tree of cpu_groups for the lowest or highest loaded cpu
* according to the match argument. This routine actually compares the
* load on all paths through the tree and finds the least loaded cpu on
* the least loaded path, which may differ from the least loaded cpu in
* the system. This balances work among caches and busses.
*
* This inline is instantiated in three forms below using constants for the
* match argument. It is reduced to the minimum set for each case. It is
* also recursive to the depth of the tree.
*/
static __always_inline int
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cpu_search(const struct cpu_group *cg, struct cpu_search *low,
struct cpu_search *high, const int match)
{
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
struct cpu_search lgroup;
struct cpu_search hgroup;
cpuset_t cpumask;
struct cpu_group *child;
struct tdq *tdq;
int cpu, i, hload, lload, load, total, rnd;
total = 0;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cpumask = cg->cg_mask;
if (match & CPU_SEARCH_LOWEST) {
lload = INT_MAX;
lgroup = *low;
}
if (match & CPU_SEARCH_HIGHEST) {
hload = INT_MIN;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
hgroup = *high;
}
/* Iterate through the child CPU groups and then remaining CPUs. */
for (i = cg->cg_children, cpu = mp_maxid; ; ) {
if (i == 0) {
#ifdef HAVE_INLINE_FFSL
cpu = CPU_FFS(&cpumask) - 1;
#else
while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
cpu--;
#endif
if (cpu < 0)
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
break;
child = NULL;
} else
child = &cg->cg_child[i - 1];
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
if (match & CPU_SEARCH_LOWEST)
lgroup.cs_cpu = -1;
if (match & CPU_SEARCH_HIGHEST)
hgroup.cs_cpu = -1;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
if (child) { /* Handle child CPU group. */
CPU_NAND(&cpumask, &child->cg_mask);
switch (match) {
case CPU_SEARCH_LOWEST:
load = cpu_search_lowest(child, &lgroup);
break;
case CPU_SEARCH_HIGHEST:
load = cpu_search_highest(child, &hgroup);
break;
case CPU_SEARCH_BOTH:
load = cpu_search_both(child, &lgroup, &hgroup);
break;
}
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
} else { /* Handle child CPU. */
CPU_CLR(cpu, &cpumask);
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
tdq = TDQ_CPU(cpu);
load = tdq->tdq_load * 256;
rnd = sched_random() % 32;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
if (match & CPU_SEARCH_LOWEST) {
if (cpu == low->cs_prefer)
load -= 64;
/* If that CPU is allowed and get data. */
if (tdq->tdq_lowpri > lgroup.cs_pri &&
tdq->tdq_load <= lgroup.cs_limit &&
CPU_ISSET(cpu, &lgroup.cs_mask)) {
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
lgroup.cs_cpu = cpu;
lgroup.cs_load = load - rnd;
}
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
}
if (match & CPU_SEARCH_HIGHEST)
if (tdq->tdq_load >= hgroup.cs_limit &&
tdq->tdq_transferable &&
CPU_ISSET(cpu, &hgroup.cs_mask)) {
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
hgroup.cs_cpu = cpu;
hgroup.cs_load = load - rnd;
}
}
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
total += load;
/* We have info about child item. Compare it. */
if (match & CPU_SEARCH_LOWEST) {
if (lgroup.cs_cpu >= 0 &&
(load < lload ||
(load == lload && lgroup.cs_load < low->cs_load))) {
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
lload = load;
low->cs_cpu = lgroup.cs_cpu;
low->cs_load = lgroup.cs_load;
}
}
if (match & CPU_SEARCH_HIGHEST)
if (hgroup.cs_cpu >= 0 &&
(load > hload ||
(load == hload && hgroup.cs_load > high->cs_load))) {
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
hload = load;
high->cs_cpu = hgroup.cs_cpu;
high->cs_load = hgroup.cs_load;
}
if (child) {
i--;
if (i == 0 && CPU_EMPTY(&cpumask))
break;
}
#ifndef HAVE_INLINE_FFSL
else
cpu--;
#endif
}
return (total);
}
/*
* cpu_search instantiations must pass constants to maintain the inline
* optimization.
*/
int
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
{
return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
}
int
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
{
return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
}
int
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
struct cpu_search *high)
{
return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
}
/*
* Find the cpu with the least load via the least loaded path that has a
* lowpri greater than pri pri. A pri of -1 indicates any priority is
* acceptable.
*/
static inline int
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
int prefer)
{
struct cpu_search low;
low.cs_cpu = -1;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
low.cs_prefer = prefer;
low.cs_mask = mask;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
low.cs_pri = pri;
low.cs_limit = maxload;
cpu_search_lowest(cg, &low);
return low.cs_cpu;
}
/*
* Find the cpu with the highest load via the highest loaded path.
*/
static inline int
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
{
struct cpu_search high;
high.cs_cpu = -1;
high.cs_mask = mask;
high.cs_limit = minload;
cpu_search_highest(cg, &high);
return high.cs_cpu;
}
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
static void
sched_balance_group(struct cpu_group *cg)
{
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cpuset_t hmask, lmask;
int high, low, anylow;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
CPU_FILL(&hmask);
for (;;) {
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
high = sched_highest(cg, hmask, 1);
/* Stop if there is no more CPU with transferrable threads. */
if (high == -1)
break;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
CPU_CLR(high, &hmask);
CPU_COPY(&hmask, &lmask);
/* Stop if there is no more CPU left for low. */
if (CPU_EMPTY(&lmask))
break;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
anylow = 1;
nextlow:
low = sched_lowest(cg, lmask, -1,
TDQ_CPU(high)->tdq_load - 1, high);
/* Stop if we looked well and found no less loaded CPU. */
if (anylow && low == -1)
break;
/* Go to next high if we found no less loaded CPU. */
if (low == -1)
continue;
/* Transfer thread from high to low. */
if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
/* CPU that got thread can no longer be a donor. */
CPU_CLR(low, &hmask);
} else {
/*
* If failed, then there is no threads on high
* that can run on this low. Drop low from low
* mask and look for different one.
*/
CPU_CLR(low, &lmask);
anylow = 0;
goto nextlow;
}
}
}
static void
sched_balance(void)
{
struct tdq *tdq;
if (smp_started == 0 || rebalance == 0)
return;
balance_ticks = max(balance_interval / 2, 1) +
(sched_random() % balance_interval);
tdq = TDQ_SELF();
TDQ_UNLOCK(tdq);
sched_balance_group(cpu_top);
TDQ_LOCK(tdq);
}
/*
* Lock two thread queues using their address to maintain lock order.
*/
static void
tdq_lock_pair(struct tdq *one, struct tdq *two)
{
if (one < two) {
TDQ_LOCK(one);
TDQ_LOCK_FLAGS(two, MTX_DUPOK);
} else {
TDQ_LOCK(two);
TDQ_LOCK_FLAGS(one, MTX_DUPOK);
}
}
/*
* Unlock two thread queues. Order is not important here.
*/
static void
tdq_unlock_pair(struct tdq *one, struct tdq *two)
{
TDQ_UNLOCK(one);
TDQ_UNLOCK(two);
}
/*
* Transfer load between two imbalanced thread queues.
*/
static int
sched_balance_pair(struct tdq *high, struct tdq *low)
{
int moved;
int cpu;
tdq_lock_pair(high, low);
moved = 0;
/*
* Determine what the imbalance is and then adjust that to how many
* threads we actually have to give up (transferable).
*/
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
(moved = tdq_move(high, low)) > 0) {
/*
* In case the target isn't the current cpu IPI it to force a
* reschedule with the new workload.
*/
cpu = TDQ_ID(low);
if (cpu != PCPU_GET(cpuid))
ipi_cpu(cpu, IPI_PREEMPT);
}
tdq_unlock_pair(high, low);
return (moved);
}
/*
* Move a thread from one thread queue to another.
*/
static int
tdq_move(struct tdq *from, struct tdq *to)
{
struct td_sched *ts;
struct thread *td;
struct tdq *tdq;
int cpu;
TDQ_LOCK_ASSERT(from, MA_OWNED);
TDQ_LOCK_ASSERT(to, MA_OWNED);
tdq = from;
cpu = TDQ_ID(to);
td = tdq_steal(tdq, cpu);
if (td == NULL)
return (0);
ts = td_get_sched(td);
/*
* Although the run queue is locked the thread may be blocked. Lock
* it to clear this and acquire the run-queue lock.
*/
thread_lock(td);
/* Drop recursive lock on from acquired via thread_lock(). */
TDQ_UNLOCK(from);
sched_rem(td);
ts->ts_cpu = cpu;
td->td_lock = TDQ_LOCKPTR(to);
tdq_add(to, td, SRQ_YIELDING);
return (1);
}
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
/*
* This tdq has idled. Try to steal a thread from another cpu and switch
* to it.
*/
static int
tdq_idled(struct tdq *tdq)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
{
struct cpu_group *cg;
struct tdq *steal;
cpuset_t mask;
int thresh;
int cpu;
if (smp_started == 0 || steal_idle == 0)
return (1);
CPU_FILL(&mask);
CPU_CLR(PCPU_GET(cpuid), &mask);
/* We don't want to be preempted while we're iterating. */
spinlock_enter();
for (cg = tdq->tdq_cg; cg != NULL; ) {
if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
thresh = steal_thresh;
else
thresh = 1;
cpu = sched_highest(cg, mask, thresh);
if (cpu == -1) {
cg = cg->cg_parent;
continue;
}
steal = TDQ_CPU(cpu);
CPU_CLR(cpu, &mask);
tdq_lock_pair(tdq, steal);
if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
tdq_unlock_pair(tdq, steal);
continue;
}
/*
* If a thread was added while interrupts were disabled don't
* steal one here. If we fail to acquire one due to affinity
* restrictions loop again with this cpu removed from the
* set.
*/
if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
tdq_unlock_pair(tdq, steal);
continue;
}
spinlock_exit();
TDQ_UNLOCK(steal);
mi_switch(SW_VOL | SWT_IDLE, NULL);
thread_unlock(curthread);
return (0);
}
spinlock_exit();
return (1);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
/*
* Notify a remote cpu of new work. Sends an IPI if criteria are met.
*/
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
static void
tdq_notify(struct tdq *tdq, struct thread *td)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
{
struct thread *ctd;
int pri;
int cpu;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
if (tdq->tdq_ipipending)
return;
cpu = td_get_sched(td)->ts_cpu;
pri = td->td_priority;
ctd = pcpu_find(cpu)->pc_curthread;
if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
return;
/*
* Make sure that our caller's earlier update to tdq_load is
* globally visible before we read tdq_cpu_idle. Idle thread
* accesses both of them without locks, and the order is important.
*/
atomic_thread_fence_seq_cst();
if (TD_IS_IDLETHREAD(ctd)) {
/*
* If the MD code has an idle wakeup routine try that before
* falling back to IPI.
*/
if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
return;
}
tdq->tdq_ipipending = 1;
ipi_cpu(cpu, IPI_PREEMPT);
}
/*
* Steals load from a timeshare queue. Honors the rotating queue head
* index.
*/
static struct thread *
runq_steal_from(struct runq *rq, int cpu, u_char start)
{
struct rqbits *rqb;
struct rqhead *rqh;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
struct thread *td, *first;
int bit;
int i;
rqb = &rq->rq_status;
bit = start & (RQB_BPW -1);
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
first = NULL;
again:
for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
if (rqb->rqb_bits[i] == 0)
continue;
if (bit == 0)
bit = RQB_FFS(rqb->rqb_bits[i]);
for (; bit < RQB_BPW; bit++) {
if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
continue;
rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
TAILQ_FOREACH(td, rqh, td_runq) {
if (first && THREAD_CAN_MIGRATE(td) &&
THREAD_CAN_SCHED(td, cpu))
return (td);
first = td;
}
}
}
if (start != 0) {
start = 0;
goto again;
}
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
if (first && THREAD_CAN_MIGRATE(first) &&
THREAD_CAN_SCHED(first, cpu))
return (first);
return (NULL);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
/*
* Steals load from a standard linear queue.
*/
static struct thread *
runq_steal(struct runq *rq, int cpu)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
{
struct rqhead *rqh;
struct rqbits *rqb;
struct thread *td;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
int word;
int bit;
rqb = &rq->rq_status;
for (word = 0; word < RQB_LEN; word++) {
if (rqb->rqb_bits[word] == 0)
continue;
for (bit = 0; bit < RQB_BPW; bit++) {
if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
continue;
rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
TAILQ_FOREACH(td, rqh, td_runq)
if (THREAD_CAN_MIGRATE(td) &&
THREAD_CAN_SCHED(td, cpu))
return (td);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
}
return (NULL);
}
/*
* Attempt to steal a thread in priority order from a thread queue.
*/
static struct thread *
tdq_steal(struct tdq *tdq, int cpu)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
{
struct thread *td;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
return (td);
if ((td = runq_steal_from(&tdq->tdq_timeshare,
cpu, tdq->tdq_ridx)) != NULL)
return (td);
return (runq_steal(&tdq->tdq_idle, cpu));
}
/*
* Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
* current lock and returns with the assigned queue locked.
*/
static inline struct tdq *
sched_setcpu(struct thread *td, int cpu, int flags)
{
struct tdq *tdq;
THREAD_LOCK_ASSERT(td, MA_OWNED);
tdq = TDQ_CPU(cpu);
td_get_sched(td)->ts_cpu = cpu;
/*
* If the lock matches just return the queue.
*/
if (td->td_lock == TDQ_LOCKPTR(tdq))
return (tdq);
#ifdef notyet
/*
* If the thread isn't running its lockptr is a
* turnstile or a sleepqueue. We can just lock_set without
* blocking.
*/
if (TD_CAN_RUN(td)) {
TDQ_LOCK(tdq);
thread_lock_set(td, TDQ_LOCKPTR(tdq));
return (tdq);
}
#endif
/*
* The hard case, migration, we need to block the thread first to
* prevent order reversals with other cpus locks.
*/
spinlock_enter();
thread_lock_block(td);
TDQ_LOCK(tdq);
thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
spinlock_exit();
return (tdq);
}
SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
static int
sched_pickcpu(struct thread *td, int flags)
{
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
struct cpu_group *cg, *ccg;
struct td_sched *ts;
struct tdq *tdq;
cpuset_t mask;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
int cpu, pri, self;
self = PCPU_GET(cpuid);
ts = td_get_sched(td);
if (smp_started == 0)
return (self);
/*
* Don't migrate a running thread from sched_switch().
*/
if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
return (ts->ts_cpu);
/*
* Prefer to run interrupt threads on the processors that generate
* the interrupt.
*/
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
pri = td->td_priority;
if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
curthread->td_intr_nesting_level && ts->ts_cpu != self) {
SCHED_STAT_INC(pickcpu_intrbind);
ts->ts_cpu = self;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
if (TDQ_CPU(self)->tdq_lowpri > pri) {
SCHED_STAT_INC(pickcpu_affinity);
return (ts->ts_cpu);
}
}
/*
* If the thread can run on the last cpu and the affinity has not
* expired or it is idle run it there.
*/
tdq = TDQ_CPU(ts->ts_cpu);
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cg = tdq->tdq_cg;
if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
tdq->tdq_lowpri >= PRI_MIN_IDLE &&
SCHED_AFFINITY(ts, CG_SHARE_L2)) {
if (cg->cg_flags & CG_FLAG_THREAD) {
CPUSET_FOREACH(cpu, cg->cg_mask) {
if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
break;
}
} else
cpu = INT_MAX;
if (cpu > mp_maxid) {
SCHED_STAT_INC(pickcpu_idle_affinity);
return (ts->ts_cpu);
}
}
/*
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
* Search for the last level cache CPU group in the tree.
* Skip caches with expired affinity time and SMT groups.
* Affinity to higher level caches will be handled less aggressively.
*/
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
if (cg->cg_flags & CG_FLAG_THREAD)
continue;
if (!SCHED_AFFINITY(ts, cg->cg_level))
continue;
ccg = cg;
}
if (ccg != NULL)
cg = ccg;
cpu = -1;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
/* Search the group for the less loaded idle CPU we can run now. */
mask = td->td_cpuset->cs_mask;
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
if (cg != NULL && cg != cpu_top &&
CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
INT_MAX, ts->ts_cpu);
/* Search globally for the less loaded CPU we can run now. */
if (cpu == -1)
cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
/* Search globally for the less loaded CPU. */
if (cpu == -1)
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
/*
* Compare the lowest loaded cpu to current cpu.
*/
if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
Rework CPU load balancing in SCHED_ULE: - In sched_pickcpu() be more careful taking previous CPU on SMT systems. Do it only if all other logical CPUs of that physical one are idle to avoid extra resource sharing. - In sched_pickcpu() change general logic of CPU selection. First look for idle CPU, sharing last level cache with previously used one, skipping SMT CPU groups. If none found, search all CPUs for the least loaded one, where the thread with its priority can run now. If none found, search just for the least loaded CPU. - Make cpu_search() compare lowest/highest CPU load when comparing CPU groups with equal load. That allows to differentiate 1+1 and 2+0 loads. - Make cpu_search() to prefer specified (previous) CPU or group if load is equal. This improves cache affinity for more complicated topologies. - Randomize CPU selection if above factors are equal. Previous code tend to prefer CPUs with lower IDs, causing unneeded collisions. - Rework periodic balancer in sched_balance_group(). With cpu_search() more intelligent now, make balansing process flat, removing recursion over the topology tree. That fixes double swap problem and makes load distribution more even and predictable. All together this gives 10-15% performance improvement in many tests on CPUs with SMT, such as Core i7, for number of threads is less then number of logical CPUs. In some tests it also gives positive effect to systems without SMT. Reviewed by: jeff Tested by: flo, hackers@ MFC after: 1 month Sponsored by: iXsystems, Inc.
2012-02-27 10:31:54 +00:00
TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
SCHED_STAT_INC(pickcpu_local);
cpu = self;
} else
SCHED_STAT_INC(pickcpu_lowest);
if (cpu != ts->ts_cpu)
SCHED_STAT_INC(pickcpu_migration);
return (cpu);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
#endif
/*
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
* Pick the highest priority task we have and return it.
*/
static struct thread *
tdq_choose(struct tdq *tdq)
{
struct thread *td;
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
td = runq_choose(&tdq->tdq_realtime);
if (td != NULL)
return (td);
td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
if (td != NULL) {
KASSERT(td->td_priority >= PRI_MIN_BATCH,
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
("tdq_choose: Invalid priority on timeshare queue %d",
td->td_priority));
return (td);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
}
td = runq_choose(&tdq->tdq_idle);
if (td != NULL) {
KASSERT(td->td_priority >= PRI_MIN_IDLE,
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
("tdq_choose: Invalid priority on idle queue %d",
td->td_priority));
return (td);
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
return (NULL);
}
/*
* Initialize a thread queue.
*/
static void
tdq_setup(struct tdq *tdq)
{
if (bootverbose)
printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
runq_init(&tdq->tdq_realtime);
runq_init(&tdq->tdq_timeshare);
runq_init(&tdq->tdq_idle);
snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
"sched lock %d", (int)TDQ_ID(tdq));
mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
MTX_SPIN | MTX_RECURSE);
#ifdef KTR
snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
"CPU %d load", (int)TDQ_ID(tdq));
#endif
}
#ifdef SMP
static void
sched_setup_smp(void)
{
struct tdq *tdq;
int i;
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
cpu_top = smp_topo();
CPU_FOREACH(i) {
tdq = TDQ_CPU(i);
tdq_setup(tdq);
tdq->tdq_cg = smp_topo_find(cpu_top, i);
if (tdq->tdq_cg == NULL)
panic("Can't find cpu group for %d\n", i);
}
balance_tdq = TDQ_SELF();
sched_balance();
}
#endif
/*
* Setup the thread queues and initialize the topology based on MD
* information.
*/
static void
sched_setup(void *dummy)
{
struct tdq *tdq;
tdq = TDQ_SELF();
#ifdef SMP
sched_setup_smp();
#else
tdq_setup(tdq);
#endif
/* Add thread0's load since it's running. */
TDQ_LOCK(tdq);
thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
tdq_load_add(tdq, &thread0);
tdq->tdq_lowpri = thread0.td_priority;
TDQ_UNLOCK(tdq);
}
/*
* This routine determines time constants after stathz and hz are setup.
*/
/* ARGSUSED */
static void
sched_initticks(void *dummy)
{
int incr;
realstathz = stathz ? stathz : hz;
sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
realstathz);
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* tickincr is shifted out by 10 to avoid rounding errors due to
* hz not being evenly divisible by stathz on all platforms.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
*/
incr = (hz << SCHED_TICK_SHIFT) / realstathz;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* This does not work for values of stathz that are more than
* 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
*/
if (incr == 0)
incr = 1;
tickincr = incr;
#ifdef SMP
/*
* Set the default balance interval now that we know
* what realstathz is.
*/
balance_interval = realstathz;
affinity = SCHED_AFFINITY_DEFAULT;
#endif
if (sched_idlespinthresh < 0)
sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
}
/*
* This is the core of the interactivity algorithm. Determines a score based
* on past behavior. It is the ratio of sleep time to run time scaled to
* a [0, 100] integer. This is the voluntary sleep time of a process, which
* differs from the cpu usage because it does not account for time spent
* waiting on a run-queue. Would be prettier if we had floating point.
*
* When a thread's sleep time is greater than its run time the
* calculation is:
*
* scaling factor
* interactivity score = ---------------------
* sleep time / run time
*
*
* When a thread's run time is greater than its sleep time the
* calculation is:
*
* scaling factor
* interactivity score = --------------------- + scaling factor
* run time / sleep time
*/
static int
sched_interact_score(struct thread *td)
{
struct td_sched *ts;
int div;
ts = td_get_sched(td);
/*
* The score is only needed if this is likely to be an interactive
* task. Don't go through the expense of computing it if there's
* no chance.
*/
if (sched_interact <= SCHED_INTERACT_HALF &&
ts->ts_runtime >= ts->ts_slptime)
return (SCHED_INTERACT_HALF);
if (ts->ts_runtime > ts->ts_slptime) {
div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
return (SCHED_INTERACT_HALF +
(SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
}
if (ts->ts_slptime > ts->ts_runtime) {
div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
return (ts->ts_runtime / div);
}
/* runtime == slptime */
if (ts->ts_runtime)
return (SCHED_INTERACT_HALF);
/*
* This can happen if slptime and runtime are 0.
*/
return (0);
}
/*
* Scale the scheduling priority according to the "interactivity" of this
* process.
*/
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
static void
sched_priority(struct thread *td)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
int score;
int pri;
if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
return;
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* If the score is interactive we place the thread in the realtime
* queue with a priority that is less than kernel and interrupt
* priorities. These threads are not subject to nice restrictions.
*
* Scores greater than this are placed on the normal timeshare queue
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* where the priority is partially decided by the most recent cpu
* utilization and the rest is decided by nice value.
*
* The nice value of the process has a linear effect on the calculated
* score. Negative nice values make it easier for a thread to be
* considered interactive.
*/
score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (score < sched_interact) {
pri = PRI_MIN_INTERACT;
pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
sched_interact) * score;
KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
("sched_priority: invalid interactive priority %d score %d",
pri, score));
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
} else {
pri = SCHED_PRI_MIN;
if (td_get_sched(td)->ts_ticks)
pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
SCHED_PRI_RANGE - 1);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
pri += SCHED_PRI_NICE(td->td_proc->p_nice);
KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
("sched_priority: invalid priority %d: nice %d, "
"ticks %d ftick %d ltick %d tick pri %d",
pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
SCHED_PRI_TICKS(td_get_sched(td))));
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
}
sched_user_prio(td, pri);
return;
}
/*
* This routine enforces a maximum limit on the amount of scheduling history
* kept. It is called after either the slptime or runtime is adjusted. This
* function is ugly due to integer math.
*/
static void
sched_interact_update(struct thread *td)
{
struct td_sched *ts;
u_int sum;
ts = td_get_sched(td);
sum = ts->ts_runtime + ts->ts_slptime;
if (sum < SCHED_SLP_RUN_MAX)
return;
/*
* This only happens from two places:
* 1) We have added an unusual amount of run time from fork_exit.
* 2) We have added an unusual amount of sleep time from sched_sleep().
*/
if (sum > SCHED_SLP_RUN_MAX * 2) {
if (ts->ts_runtime > ts->ts_slptime) {
ts->ts_runtime = SCHED_SLP_RUN_MAX;
ts->ts_slptime = 1;
} else {
ts->ts_slptime = SCHED_SLP_RUN_MAX;
ts->ts_runtime = 1;
}
return;
}
/*
* If we have exceeded by more than 1/5th then the algorithm below
* will not bring us back into range. Dividing by two here forces
* us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
*/
if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
ts->ts_runtime /= 2;
ts->ts_slptime /= 2;
return;
}
ts->ts_runtime = (ts->ts_runtime / 5) * 4;
ts->ts_slptime = (ts->ts_slptime / 5) * 4;
}
/*
* Scale back the interactivity history when a child thread is created. The
* history is inherited from the parent but the thread may behave totally
* differently. For example, a shell spawning a compiler process. We want
* to learn that the compiler is behaving badly very quickly.
*/
static void
sched_interact_fork(struct thread *td)
{
struct td_sched *ts;
int ratio;
int sum;
ts = td_get_sched(td);
sum = ts->ts_runtime + ts->ts_slptime;
if (sum > SCHED_SLP_RUN_FORK) {
ratio = sum / SCHED_SLP_RUN_FORK;
ts->ts_runtime /= ratio;
ts->ts_slptime /= ratio;
}
}
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
/*
* Called from proc0_init() to setup the scheduler fields.
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
*/
void
schedinit(void)
{
struct td_sched *ts0;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
/*
* Set up the scheduler specific parts of thread0.
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
*/
ts0 = td_get_sched(&thread0);
ts0->ts_ltick = ticks;
ts0->ts_ftick = ticks;
ts0->ts_slice = 0;
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
/*
* This is only somewhat accurate since given many processes of the same
* priority they will switch when their slices run out, which will be
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* at most sched_slice stathz ticks.
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
*/
int
sched_rr_interval(void)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/* Convert sched_slice from stathz to hz. */
return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
}
/*
* Update the percent cpu tracking information when it is requested or
* the total history exceeds the maximum. We keep a sliding history of
* tick counts that slowly decays. This is less precise than the 4BSD
* mechanism since it happens with less regular and frequent events.
*/
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
static void
sched_pctcpu_update(struct td_sched *ts, int run)
{
int t = ticks;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
ts->ts_ticks = 0;
ts->ts_ftick = t - SCHED_TICK_TARG;
} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
(ts->ts_ltick - (t - SCHED_TICK_TARG));
ts->ts_ftick = t - SCHED_TICK_TARG;
}
if (run)
ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
ts->ts_ltick = t;
}
/*
* Adjust the priority of a thread. Move it to the appropriate run-queue
* if necessary. This is the back-end for several priority related
* functions.
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
sched_thread_priority(struct thread *td, u_char prio)
{
struct td_sched *ts;
struct tdq *tdq;
int oldpri;
KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
"prio:%d", td->td_priority, "new prio:%d", prio,
KTR_ATTR_LINKED, sched_tdname(curthread));
SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
if (td != curthread && prio < td->td_priority) {
KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
"lend prio", "prio:%d", td->td_priority, "new prio:%d",
prio, KTR_ATTR_LINKED, sched_tdname(td));
SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
curthread);
}
ts = td_get_sched(td);
THREAD_LOCK_ASSERT(td, MA_OWNED);
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
if (td->td_priority == prio)
return;
/*
* If the priority has been elevated due to priority
* propagation, we may have to move ourselves to a new
* queue. This could be optimized to not re-add in some
* cases.
*/
if (TD_ON_RUNQ(td) && prio < td->td_priority) {
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
sched_rem(td);
td->td_priority = prio;
sched_add(td, SRQ_BORROWING);
return;
}
/*
* If the thread is currently running we may have to adjust the lowpri
* information so other cpus are aware of our current priority.
*/
if (TD_IS_RUNNING(td)) {
tdq = TDQ_CPU(ts->ts_cpu);
oldpri = td->td_priority;
td->td_priority = prio;
if (prio < tdq->tdq_lowpri)
tdq->tdq_lowpri = prio;
else if (tdq->tdq_lowpri == oldpri)
tdq_setlowpri(tdq, td);
return;
}
td->td_priority = prio;
}
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
/*
* Update a thread's priority when it is lent another thread's
* priority.
*/
void
sched_lend_prio(struct thread *td, u_char prio)
{
td->td_flags |= TDF_BORROWING;
sched_thread_priority(td, prio);
}
/*
* Restore a thread's priority when priority propagation is
* over. The prio argument is the minimum priority the thread
* needs to have to satisfy other possible priority lending
* requests. If the thread's regular priority is less
* important than prio, the thread will keep a priority boost
* of prio.
*/
void
sched_unlend_prio(struct thread *td, u_char prio)
{
u_char base_pri;
if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
td->td_base_pri <= PRI_MAX_TIMESHARE)
base_pri = td->td_user_pri;
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
else
base_pri = td->td_base_pri;
if (prio >= base_pri) {
2004-12-30 22:17:00 +00:00
td->td_flags &= ~TDF_BORROWING;
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
sched_thread_priority(td, base_pri);
} else
sched_lend_prio(td, prio);
}
/*
* Standard entry for setting the priority to an absolute value.
*/
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
void
sched_prio(struct thread *td, u_char prio)
{
u_char oldprio;
/* First, update the base priority. */
td->td_base_pri = prio;
/*
2004-12-30 22:17:00 +00:00
* If the thread is borrowing another thread's priority, don't
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
* ever lower the priority.
*/
if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
return;
/* Change the real priority. */
oldprio = td->td_priority;
sched_thread_priority(td, prio);
/*
* If the thread is on a turnstile, then let the turnstile update
* its state.
*/
if (TD_ON_LOCK(td) && oldprio != prio)
turnstile_adjust(td, oldprio);
}
2004-12-30 22:17:00 +00:00
/*
* Set the base user priority, does not effect current running priority.
*/
void
sched_user_prio(struct thread *td, u_char prio)
{
td->td_base_user_pri = prio;
if (td->td_lend_user_pri <= prio)
return;
td->td_user_pri = prio;
}
void
sched_lend_user_prio(struct thread *td, u_char prio)
{
THREAD_LOCK_ASSERT(td, MA_OWNED);
td->td_lend_user_pri = prio;
td->td_user_pri = min(prio, td->td_base_user_pri);
if (td->td_priority > td->td_user_pri)
sched_prio(td, td->td_user_pri);
else if (td->td_priority != td->td_user_pri)
td->td_flags |= TDF_NEEDRESCHED;
}
/*
* Handle migration from sched_switch(). This happens only for
* cpu binding.
*/
static struct mtx *
sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
{
struct tdq *tdn;
tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
#ifdef SMP
tdq_load_rem(tdq, td);
/*
* Do the lock dance required to avoid LOR. We grab an extra
* spinlock nesting to prevent preemption while we're
* not holding either run-queue lock.
*/
spinlock_enter();
thread_lock_block(td); /* This releases the lock on tdq. */
Fix sched_switch_migrate(): - In 8.x and above the run-queue locks are nomore shared even in the HTT case, so remove the special case. - The deadlock explained in the removed comment here is still possible even with different locks, with the contribution of tdq_lock_pair(). An explanation is here: (hypotesis: a thread needs to migrate on another CPU, thread1 is doing sched_switch_migrate() and thread2 is the one handling the sched_switch() request or in other words, thread1 is the thread that needs to migrate and thread2 is a thread that is going to be preempted, most likely an idle thread. Also, 'old' is referred to the context (in terms of run-queue and CPU) thread1 is leaving and 'new' is referred to the context thread1 is going into. Finally, thread3 is doing tdq_idletd() or sched_balance() and definitively doing tdq_lock_pair()) * thread1 blocks its td_lock. Now td_lock is 'blocked' * thread1 drops its old runqueue lock * thread1 acquires the new runqueue lock * thread1 adds itself to the new runqueue and sends an IPI_PREEMPT through tdq_notify() to the new CPU * thread1 drops the new lock * thread3, scanning the runqueues, locks the old lock * thread2 received the IPI_PREEMPT and does thread_lock() with td_lock pointing to the new runqueue * thread3 wants to acquire the new runqueue lock, but it can't because it is held by thread2 so it spins * thread1 wants to acquire old lock, but as long as it is held by thread3 it can't * thread2 going further, at some point wants to switchin in thread1, but it will wait forever because thread1->td_lock is in blocked state This deadlock has been manifested mostly on 7.x and reported several time on mailing lists under the voice 'spinlock held too long'. Many thanks to des@ for having worked hard on producing suitable textdumps and Jeff for help on the comment wording. Reviewed by: jeff Reported by: des, others Tested by: des, Giovanni Trematerra <giovanni dot trematerra at gmail dot com> (STABLE_7 based version)
2009-09-15 16:56:17 +00:00
/*
Fix sched_switch_migrate(): - In 8.x and above the run-queue locks are nomore shared even in the HTT case, so remove the special case. - The deadlock explained in the removed comment here is still possible even with different locks, with the contribution of tdq_lock_pair(). An explanation is here: (hypotesis: a thread needs to migrate on another CPU, thread1 is doing sched_switch_migrate() and thread2 is the one handling the sched_switch() request or in other words, thread1 is the thread that needs to migrate and thread2 is a thread that is going to be preempted, most likely an idle thread. Also, 'old' is referred to the context (in terms of run-queue and CPU) thread1 is leaving and 'new' is referred to the context thread1 is going into. Finally, thread3 is doing tdq_idletd() or sched_balance() and definitively doing tdq_lock_pair()) * thread1 blocks its td_lock. Now td_lock is 'blocked' * thread1 drops its old runqueue lock * thread1 acquires the new runqueue lock * thread1 adds itself to the new runqueue and sends an IPI_PREEMPT through tdq_notify() to the new CPU * thread1 drops the new lock * thread3, scanning the runqueues, locks the old lock * thread2 received the IPI_PREEMPT and does thread_lock() with td_lock pointing to the new runqueue * thread3 wants to acquire the new runqueue lock, but it can't because it is held by thread2 so it spins * thread1 wants to acquire old lock, but as long as it is held by thread3 it can't * thread2 going further, at some point wants to switchin in thread1, but it will wait forever because thread1->td_lock is in blocked state This deadlock has been manifested mostly on 7.x and reported several time on mailing lists under the voice 'spinlock held too long'. Many thanks to des@ for having worked hard on producing suitable textdumps and Jeff for help on the comment wording. Reviewed by: jeff Reported by: des, others Tested by: des, Giovanni Trematerra <giovanni dot trematerra at gmail dot com> (STABLE_7 based version)
2009-09-15 16:56:17 +00:00
* Acquire both run-queue locks before placing the thread on the new
* run-queue to avoid deadlocks created by placing a thread with a
* blocked lock on the run-queue of a remote processor. The deadlock
* occurs when a third processor attempts to lock the two queues in
* question while the target processor is spinning with its own
* run-queue lock held while waiting for the blocked lock to clear.
*/
Fix sched_switch_migrate(): - In 8.x and above the run-queue locks are nomore shared even in the HTT case, so remove the special case. - The deadlock explained in the removed comment here is still possible even with different locks, with the contribution of tdq_lock_pair(). An explanation is here: (hypotesis: a thread needs to migrate on another CPU, thread1 is doing sched_switch_migrate() and thread2 is the one handling the sched_switch() request or in other words, thread1 is the thread that needs to migrate and thread2 is a thread that is going to be preempted, most likely an idle thread. Also, 'old' is referred to the context (in terms of run-queue and CPU) thread1 is leaving and 'new' is referred to the context thread1 is going into. Finally, thread3 is doing tdq_idletd() or sched_balance() and definitively doing tdq_lock_pair()) * thread1 blocks its td_lock. Now td_lock is 'blocked' * thread1 drops its old runqueue lock * thread1 acquires the new runqueue lock * thread1 adds itself to the new runqueue and sends an IPI_PREEMPT through tdq_notify() to the new CPU * thread1 drops the new lock * thread3, scanning the runqueues, locks the old lock * thread2 received the IPI_PREEMPT and does thread_lock() with td_lock pointing to the new runqueue * thread3 wants to acquire the new runqueue lock, but it can't because it is held by thread2 so it spins * thread1 wants to acquire old lock, but as long as it is held by thread3 it can't * thread2 going further, at some point wants to switchin in thread1, but it will wait forever because thread1->td_lock is in blocked state This deadlock has been manifested mostly on 7.x and reported several time on mailing lists under the voice 'spinlock held too long'. Many thanks to des@ for having worked hard on producing suitable textdumps and Jeff for help on the comment wording. Reviewed by: jeff Reported by: des, others Tested by: des, Giovanni Trematerra <giovanni dot trematerra at gmail dot com> (STABLE_7 based version)
2009-09-15 16:56:17 +00:00
tdq_lock_pair(tdn, tdq);
tdq_add(tdn, td, flags);
tdq_notify(tdn, td);
TDQ_UNLOCK(tdn);
spinlock_exit();
#endif
return (TDQ_LOCKPTR(tdn));
}
/*
* Variadic version of thread_lock_unblock() that does not assume td_lock
* is blocked.
*/
static inline void
thread_unblock_switch(struct thread *td, struct mtx *mtx)
{
atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
(uintptr_t)mtx);
}
/*
* Switch threads. This function has to handle threads coming in while
* blocked for some reason, running, or idle. It also must deal with
* migrating a thread from one queue to another as running threads may
* be assigned elsewhere via binding.
*/
void
sched_switch(struct thread *td, struct thread *newtd, int flags)
{
2006-12-29 12:55:32 +00:00
struct tdq *tdq;
struct td_sched *ts;
struct mtx *mtx;
int srqflag;
int cpuid, preempted;
THREAD_LOCK_ASSERT(td, MA_OWNED);
KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
cpuid = PCPU_GET(cpuid);
tdq = TDQ_CPU(cpuid);
ts = td_get_sched(td);
mtx = td->td_lock;
sched_pctcpu_update(ts, 1);
ts->ts_rltick = ticks;
td->td_lastcpu = td->td_oncpu;
td->td_oncpu = NOCPU;
preempted = !((td->td_flags & TDF_SLICEEND) ||
(flags & SWT_RELINQUISH));
td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
td->td_owepreempt = 0;
if (!TD_IS_IDLETHREAD(td))
tdq->tdq_switchcnt++;
/*
* The lock pointer in an idle thread should never change. Reset it
* to CAN_RUN as well.
*/
if (TD_IS_IDLETHREAD(td)) {
MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
TD_SET_CAN_RUN(td);
} else if (TD_IS_RUNNING(td)) {
MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
srqflag = preempted ?
SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
SRQ_OURSELF|SRQ_YIELDING;
2010-09-02 16:23:05 +00:00
#ifdef SMP
if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
ts->ts_cpu = sched_pickcpu(td, 0);
2010-09-02 16:23:05 +00:00
#endif
if (ts->ts_cpu == cpuid)
tdq_runq_add(tdq, td, srqflag);
else {
KASSERT(THREAD_CAN_MIGRATE(td) ||
(ts->ts_flags & TSF_BOUND) != 0,
("Thread %p shouldn't migrate", td));
mtx = sched_switch_migrate(tdq, td, srqflag);
}
} else {
/* This thread must be going to sleep. */
TDQ_LOCK(tdq);
mtx = thread_lock_block(td);
tdq_load_rem(tdq, td);
}
/*
* We enter here with the thread blocked and assigned to the
* appropriate cpu run-queue or sleep-queue and with the current
* thread-queue locked.
*/
TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
newtd = choosethread();
/*
* Call the MD code to switch contexts if necessary.
*/
if (td != newtd) {
#ifdef HWPMC_HOOKS
if (PMC_PROC_IS_USING_PMCS(td->td_proc))
PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
#endif
SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
sched_pctcpu_update(td_get_sched(newtd), 0);
#ifdef KDTRACE_HOOKS
/*
* If DTrace has set the active vtime enum to anything
* other than INACTIVE (0), then it should have set the
* function to call.
*/
if (dtrace_vtime_active)
(*dtrace_vtime_switch_func)(newtd);
#endif
cpu_switch(td, newtd, mtx);
/*
* We may return from cpu_switch on a different cpu. However,
* we always return with td_lock pointing to the current cpu's
* run queue lock.
*/
cpuid = PCPU_GET(cpuid);
tdq = TDQ_CPU(cpuid);
lock_profile_obtain_lock_success(
&TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
SDT_PROBE0(sched, , , on__cpu);
#ifdef HWPMC_HOOKS
if (PMC_PROC_IS_USING_PMCS(td->td_proc))
PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
#endif
} else {
thread_unblock_switch(td, mtx);
SDT_PROBE0(sched, , , remain__cpu);
}
/*
* Assert that all went well and return.
*/
TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
td->td_oncpu = cpuid;
}
/*
* Adjust thread priorities as a result of a nice request.
*/
void
sched_nice(struct proc *p, int nice)
{
struct thread *td;
PROC_LOCK_ASSERT(p, MA_OWNED);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
p->p_nice = nice;
FOREACH_THREAD_IN_PROC(p, td) {
thread_lock(td);
sched_priority(td);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
sched_prio(td, td->td_base_user_pri);
thread_unlock(td);
}
}
/*
* Record the sleep time for the interactivity scorer.
*/
void
sched_sleep(struct thread *td, int prio)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
td->td_slptick = ticks;
if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
td->td_flags |= TDF_CANSWAP;
if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
return;
if (static_boost == 1 && prio)
sched_prio(td, prio);
else if (static_boost && td->td_priority > static_boost)
sched_prio(td, static_boost);
}
/*
* Schedule a thread to resume execution and record how long it voluntarily
* slept. We also update the pctcpu, interactivity, and priority.
*/
void
sched_wakeup(struct thread *td)
{
struct td_sched *ts;
int slptick;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
ts = td_get_sched(td);
td->td_flags &= ~TDF_CANSWAP;
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* If we slept for more than a tick update our interactivity and
* priority.
*/
slptick = td->td_slptick;
td->td_slptick = 0;
if (slptick && slptick != ticks) {
ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
sched_interact_update(td);
sched_pctcpu_update(ts, 0);
}
/*
* Reset the slice value since we slept and advanced the round-robin.
*/
ts->ts_slice = 0;
sched_add(td, SRQ_BORING);
}
/*
* Penalize the parent for creating a new child and initialize the child's
* priority.
*/
void
sched_fork(struct thread *td, struct thread *child)
{
THREAD_LOCK_ASSERT(td, MA_OWNED);
sched_pctcpu_update(td_get_sched(td), 1);
sched_fork_thread(td, child);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* Penalize the parent and child for forking.
*/
sched_interact_fork(child);
sched_priority(child);
td_get_sched(td)->ts_runtime += tickincr;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
sched_interact_update(td);
sched_priority(td);
}
/*
* Fork a new thread, may be within the same process.
*/
void
sched_fork_thread(struct thread *td, struct thread *child)
{
struct td_sched *ts;
struct td_sched *ts2;
struct tdq *tdq;
tdq = TDQ_SELF();
THREAD_LOCK_ASSERT(td, MA_OWNED);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* Initialize child.
*/
ts = td_get_sched(td);
ts2 = td_get_sched(child);
child->td_oncpu = NOCPU;
child->td_lastcpu = NOCPU;
child->td_lock = TDQ_LOCKPTR(tdq);
child->td_cpuset = cpuset_ref(td->td_cpuset);
ts2->ts_cpu = ts->ts_cpu;
ts2->ts_flags = 0;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* Grab our parents cpu estimation information.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
*/
ts2->ts_ticks = ts->ts_ticks;
ts2->ts_ltick = ts->ts_ltick;
ts2->ts_ftick = ts->ts_ftick;
/*
* Do not inherit any borrowed priority from the parent.
*/
child->td_priority = child->td_base_pri;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* And update interactivity score.
*/
ts2->ts_slptime = ts->ts_slptime;
ts2->ts_runtime = ts->ts_runtime;
/* Attempt to quickly learn interactivity. */
ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
#ifdef KTR
bzero(ts2->ts_name, sizeof(ts2->ts_name));
#endif
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
}
/*
* Adjust the priority class of a thread.
*/
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
void
sched_class(struct thread *td, int class)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
{
THREAD_LOCK_ASSERT(td, MA_OWNED);
if (td->td_pri_class == class)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
return;
td->td_pri_class = class;
}
/*
* Return some of the child's priority and interactivity to the parent.
*/
void
sched_exit(struct proc *p, struct thread *child)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
struct thread *td;
KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
"prio:%d", child->td_priority);
PROC_LOCK_ASSERT(p, MA_OWNED);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
td = FIRST_THREAD_IN_PROC(p);
sched_exit_thread(td, child);
}
/*
* Penalize another thread for the time spent on this one. This helps to
* worsen the priority and interactivity of processes which schedule batch
* jobs such as make. This has little effect on the make process itself but
* causes new processes spawned by it to receive worse scores immediately.
*/
void
sched_exit_thread(struct thread *td, struct thread *child)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
"prio:%d", child->td_priority);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* Give the child's runtime to the parent without returning the
* sleep time as a penalty to the parent. This causes shells that
* launch expensive things to mark their children as expensive.
*/
thread_lock(td);
td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
sched_interact_update(td);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
sched_priority(td);
thread_unlock(td);
}
void
sched_preempt(struct thread *td)
{
struct tdq *tdq;
SDT_PROBE2(sched, , , surrender, td, td->td_proc);
thread_lock(td);
tdq = TDQ_SELF();
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
tdq->tdq_ipipending = 0;
if (td->td_priority > tdq->tdq_lowpri) {
int flags;
flags = SW_INVOL | SW_PREEMPT;
if (td->td_critnest > 1)
td->td_owepreempt = 1;
else if (TD_IS_IDLETHREAD(td))
mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
else
mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
}
thread_unlock(td);
}
/*
* Fix priorities on return to user-space. Priorities may be elevated due
* to static priorities in msleep() or similar.
*/
void
sched_userret(struct thread *td)
{
/*
* XXX we cheat slightly on the locking here to avoid locking in
* the usual case. Setting td_priority here is essentially an
* incomplete workaround for not setting it properly elsewhere.
* Now that some interrupt handlers are threads, not setting it
* properly elsewhere can clobber it in the window between setting
* it here and returning to user mode, so don't waste time setting
* it perfectly here.
*/
KASSERT((td->td_flags & TDF_BORROWING) == 0,
("thread with borrowed priority returning to userland"));
if (td->td_priority != td->td_user_pri) {
thread_lock(td);
td->td_priority = td->td_user_pri;
td->td_base_pri = td->td_user_pri;
tdq_setlowpri(TDQ_SELF(), td);
thread_unlock(td);
}
}
/*
* Handle a stathz tick. This is really only relevant for timeshare
* threads.
*/
void
sched_clock(struct thread *td)
{
struct tdq *tdq;
struct td_sched *ts;
THREAD_LOCK_ASSERT(td, MA_OWNED);
tdq = TDQ_SELF();
#ifdef SMP
/*
* We run the long term load balancer infrequently on the first cpu.
*/
if (balance_tdq == tdq) {
if (balance_ticks && --balance_ticks == 0)
sched_balance();
}
#endif
/*
* Save the old switch count so we have a record of the last ticks
* activity. Initialize the new switch count based on our load.
* If there is some activity seed it to reflect that.
*/
tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
tdq->tdq_switchcnt = tdq->tdq_load;
/*
* Advance the insert index once for each tick to ensure that all
* threads get a chance to run.
*/
if (tdq->tdq_idx == tdq->tdq_ridx) {
tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
tdq->tdq_ridx = tdq->tdq_idx;
}
ts = td_get_sched(td);
sched_pctcpu_update(ts, 1);
if (td->td_pri_class & PRI_FIFO_BIT)
return;
if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
/*
* We used a tick; charge it to the thread so
* that we can compute our interactivity.
*/
td_get_sched(td)->ts_runtime += tickincr;
sched_interact_update(td);
sched_priority(td);
}
/*
* Force a context switch if the current thread has used up a full
* time slice (default is 100ms).
*/
if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
ts->ts_slice = 0;
td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
}
}
u_int
sched_estcpu(struct thread *td __unused)
{
return (0);
}
/*
* Return whether the current CPU has runnable tasks. Used for in-kernel
* cooperative idle threads.
*/
int
sched_runnable(void)
{
struct tdq *tdq;
int load;
load = 1;
tdq = TDQ_SELF();
if ((curthread->td_flags & TDF_IDLETD) != 0) {
if (tdq->tdq_load > 0)
goto out;
} else
if (tdq->tdq_load - 1 > 0)
goto out;
load = 0;
out:
return (load);
}
/*
* Choose the highest priority thread to run. The thread is removed from
* the run-queue while running however the load remains. For SMP we set
* the tdq in the global idle bitmask if it idles here.
*/
struct thread *
sched_choose(void)
{
struct thread *td;
struct tdq *tdq;
tdq = TDQ_SELF();
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
td = tdq_choose(tdq);
if (td) {
tdq_runq_rem(tdq, td);
tdq->tdq_lowpri = td->td_priority;
return (td);
}
tdq->tdq_lowpri = PRI_MAX_IDLE;
return (PCPU_GET(idlethread));
}
/*
* Set owepreempt if necessary. Preemption never happens directly in ULE,
* we always request it once we exit a critical section.
*/
static inline void
sched_setpreempt(struct thread *td)
{
struct thread *ctd;
int cpri;
int pri;
THREAD_LOCK_ASSERT(curthread, MA_OWNED);
ctd = curthread;
pri = td->td_priority;
cpri = ctd->td_priority;
if (pri < cpri)
ctd->td_flags |= TDF_NEEDRESCHED;
if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
return;
if (!sched_shouldpreempt(pri, cpri, 0))
return;
ctd->td_owepreempt = 1;
}
/*
* Add a thread to a thread queue. Select the appropriate runq and add the
* thread to it. This is the internal function called when the tdq is
* predetermined.
*/
void
tdq_add(struct tdq *tdq, struct thread *td, int flags)
{
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
KASSERT((td->td_inhibitors == 0),
("sched_add: trying to run inhibited thread"));
KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
("sched_add: bad thread state"));
KASSERT(td->td_flags & TDF_INMEM,
("sched_add: thread swapped out"));
if (td->td_priority < tdq->tdq_lowpri)
tdq->tdq_lowpri = td->td_priority;
tdq_runq_add(tdq, td, flags);
tdq_load_add(tdq, td);
}
/*
* Select the target thread queue and add a thread to it. Request
* preemption or IPI a remote processor if required.
*/
void
sched_add(struct thread *td, int flags)
{
struct tdq *tdq;
#ifdef SMP
int cpu;
#endif
KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
"prio:%d", td->td_priority, KTR_ATTR_LINKED,
sched_tdname(curthread));
KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
KTR_ATTR_LINKED, sched_tdname(td));
SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
flags & SRQ_PREEMPTED);
THREAD_LOCK_ASSERT(td, MA_OWNED);
/*
* Recalculate the priority before we select the target cpu or
* run-queue.
*/
if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
sched_priority(td);
#ifdef SMP
/*
* Pick the destination cpu and if it isn't ours transfer to the
* target cpu.
*/
cpu = sched_pickcpu(td, flags);
tdq = sched_setcpu(td, cpu, flags);
tdq_add(tdq, td, flags);
if (cpu != PCPU_GET(cpuid)) {
tdq_notify(tdq, td);
return;
}
#else
tdq = TDQ_SELF();
TDQ_LOCK(tdq);
/*
* Now that the thread is moving to the run-queue, set the lock
* to the scheduler's lock.
*/
thread_lock_set(td, TDQ_LOCKPTR(tdq));
tdq_add(tdq, td, flags);
#endif
if (!(flags & SRQ_YIELDING))
sched_setpreempt(td);
}
/*
* Remove a thread from a run-queue without running it. This is used
* when we're stealing a thread from a remote queue. Otherwise all threads
* exit by calling sched_exit_thread() and sched_throw() themselves.
*/
void
sched_rem(struct thread *td)
{
struct tdq *tdq;
KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
"prio:%d", td->td_priority);
SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
TDQ_LOCK_ASSERT(tdq, MA_OWNED);
MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
KASSERT(TD_ON_RUNQ(td),
("sched_rem: thread not on run queue"));
tdq_runq_rem(tdq, td);
tdq_load_rem(tdq, td);
TD_SET_CAN_RUN(td);
if (td->td_priority == tdq->tdq_lowpri)
tdq_setlowpri(tdq, NULL);
}
/*
* Fetch cpu utilization information. Updates on demand.
*/
fixpt_t
sched_pctcpu(struct thread *td)
{
fixpt_t pctcpu;
struct td_sched *ts;
pctcpu = 0;
ts = td_get_sched(td);
THREAD_LOCK_ASSERT(td, MA_OWNED);
sched_pctcpu_update(ts, TD_IS_RUNNING(td));
if (ts->ts_ticks) {
int rtick;
/* How many rtick per second ? */
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
}
return (pctcpu);
}
/*
* Enforce affinity settings for a thread. Called after adjustments to
* cpumask.
*/
void
sched_affinity(struct thread *td)
{
#ifdef SMP
struct td_sched *ts;
THREAD_LOCK_ASSERT(td, MA_OWNED);
ts = td_get_sched(td);
if (THREAD_CAN_SCHED(td, ts->ts_cpu))
return;
if (TD_ON_RUNQ(td)) {
sched_rem(td);
sched_add(td, SRQ_BORING);
return;
}
if (!TD_IS_RUNNING(td))
return;
/*
* Force a switch before returning to userspace. If the
* target thread is not running locally send an ipi to force
* the issue.
*/
td->td_flags |= TDF_NEEDRESCHED;
if (td != curthread)
ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
#endif
}
/*
* Bind a thread to a target cpu.
*/
void
sched_bind(struct thread *td, int cpu)
{
struct td_sched *ts;
THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
ts = td_get_sched(td);
if (ts->ts_flags & TSF_BOUND)
sched_unbind(td);
KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
ts->ts_flags |= TSF_BOUND;
sched_pin();
if (PCPU_GET(cpuid) == cpu)
return;
ts->ts_cpu = cpu;
/* When we return from mi_switch we'll be on the correct cpu. */
mi_switch(SW_VOL, NULL);
}
/*
* Release a bound thread.
*/
void
sched_unbind(struct thread *td)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
struct td_sched *ts;
THREAD_LOCK_ASSERT(td, MA_OWNED);
KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
ts = td_get_sched(td);
if ((ts->ts_flags & TSF_BOUND) == 0)
return;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
ts->ts_flags &= ~TSF_BOUND;
sched_unpin();
}
int
sched_is_bound(struct thread *td)
{
THREAD_LOCK_ASSERT(td, MA_OWNED);
return (td_get_sched(td)->ts_flags & TSF_BOUND);
}
/*
* Basic yield call.
*/
void
sched_relinquish(struct thread *td)
{
thread_lock(td);
mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
thread_unlock(td);
}
/*
* Return the total system load.
*/
int
sched_load(void)
{
#ifdef SMP
int total;
int i;
total = 0;
CPU_FOREACH(i)
total += TDQ_CPU(i)->tdq_sysload;
return (total);
#else
return (TDQ_SELF()->tdq_sysload);
#endif
}
int
sched_sizeof_proc(void)
{
return (sizeof(struct proc));
}
int
sched_sizeof_thread(void)
{
return (sizeof(struct thread) + sizeof(struct td_sched));
}
Add scheduler CORE, the work I have done half a year ago, recent, I picked it up again. The scheduler is forked from ULE, but the algorithm to detect an interactive process is almost completely different with ULE, it comes from Linux paper "Understanding the Linux 2.6.8.1 CPU Scheduler", although I still use same word "score" as a priority boost in ULE scheduler. Briefly, the scheduler has following characteristic: 1. Timesharing process's nice value is seriously respected, timeslice and interaction detecting algorithm are based on nice value. 2. per-cpu scheduling queue and load balancing. 3. O(1) scheduling. 4. Some cpu affinity code in wakeup path. 5. Support POSIX SCHED_FIFO and SCHED_RR. Unlike scheduler 4BSD and ULE which using fuzzy RQ_PPQ, the scheduler uses 256 priority queues. Unlike ULE which using pull and push, the scheduelr uses pull method, the main reason is to let relative idle cpu do the work, but current the whole scheduler is protected by the big sched_lock, so the benefit is not visible, it really can be worse than nothing because all other cpu are locked out when we are doing balancing work, which the 4BSD scheduelr does not have this problem. The scheduler does not support hyperthreading very well, in fact, the scheduler does not make the difference between physical CPU and logical CPU, this should be improved in feature. The scheduler has priority inversion problem on MP machine, it is not good for realtime scheduling, it can cause realtime process starving. As a result, it seems the MySQL super-smack runs better on my Pentium-D machine when using libthr, despite on UP or SMP kernel.
2006-06-13 13:12:56 +00:00
#ifdef SMP
#define TDQ_IDLESPIN(tdq) \
((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
#else
#define TDQ_IDLESPIN(tdq) 1
#endif
/*
* The actual idle process.
*/
void
sched_idletd(void *dummy)
{
struct thread *td;
struct tdq *tdq;
int oldswitchcnt, switchcnt;
int i;
mtx_assert(&Giant, MA_NOTOWNED);
td = curthread;
tdq = TDQ_SELF();
THREAD_NO_SLEEPING();
oldswitchcnt = -1;
for (;;) {
if (tdq->tdq_load) {
thread_lock(td);
mi_switch(SW_VOL | SWT_IDLE, NULL);
thread_unlock(td);
}
switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
#ifdef SMP
if (switchcnt != oldswitchcnt) {
oldswitchcnt = switchcnt;
if (tdq_idled(tdq) == 0)
continue;
}
switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
#else
oldswitchcnt = switchcnt;
#endif
/*
* If we're switching very frequently, spin while checking
* for load rather than entering a low power state that
* may require an IPI. However, don't do any busy
* loops while on SMT machines as this simply steals
* cycles from cores doing useful work.
*/
if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
for (i = 0; i < sched_idlespins; i++) {
if (tdq->tdq_load)
break;
cpu_spinwait();
}
}
/* If there was context switch during spin, restart it. */
switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
continue;
/* Run main MD idle handler. */
tdq->tdq_cpu_idle = 1;
/*
* Make sure that tdq_cpu_idle update is globally visible
* before cpu_idle() read tdq_load. The order is important
* to avoid race with tdq_notify.
*/
atomic_thread_fence_seq_cst();
cpu_idle(switchcnt * 4 > sched_idlespinthresh);
tdq->tdq_cpu_idle = 0;
/*
* Account thread-less hardware interrupts and
* other wakeup reasons equal to context switches.
*/
switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
if (switchcnt != oldswitchcnt)
continue;
tdq->tdq_switchcnt++;
oldswitchcnt++;
}
Add scheduler CORE, the work I have done half a year ago, recent, I picked it up again. The scheduler is forked from ULE, but the algorithm to detect an interactive process is almost completely different with ULE, it comes from Linux paper "Understanding the Linux 2.6.8.1 CPU Scheduler", although I still use same word "score" as a priority boost in ULE scheduler. Briefly, the scheduler has following characteristic: 1. Timesharing process's nice value is seriously respected, timeslice and interaction detecting algorithm are based on nice value. 2. per-cpu scheduling queue and load balancing. 3. O(1) scheduling. 4. Some cpu affinity code in wakeup path. 5. Support POSIX SCHED_FIFO and SCHED_RR. Unlike scheduler 4BSD and ULE which using fuzzy RQ_PPQ, the scheduler uses 256 priority queues. Unlike ULE which using pull and push, the scheduelr uses pull method, the main reason is to let relative idle cpu do the work, but current the whole scheduler is protected by the big sched_lock, so the benefit is not visible, it really can be worse than nothing because all other cpu are locked out when we are doing balancing work, which the 4BSD scheduelr does not have this problem. The scheduler does not support hyperthreading very well, in fact, the scheduler does not make the difference between physical CPU and logical CPU, this should be improved in feature. The scheduler has priority inversion problem on MP machine, it is not good for realtime scheduling, it can cause realtime process starving. As a result, it seems the MySQL super-smack runs better on my Pentium-D machine when using libthr, despite on UP or SMP kernel.
2006-06-13 13:12:56 +00:00
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* A CPU is entering for the first time or a thread is exiting.
*/
void
sched_throw(struct thread *td)
{
struct thread *newtd;
struct tdq *tdq;
tdq = TDQ_SELF();
if (td == NULL) {
/* Correct spinlock nesting and acquire the correct lock. */
TDQ_LOCK(tdq);
spinlock_exit();
PCPU_SET(switchtime, cpu_ticks());
PCPU_SET(switchticks, ticks);
} else {
MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
tdq_load_rem(tdq, td);
lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
td->td_lastcpu = td->td_oncpu;
td->td_oncpu = NOCPU;
}
KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
newtd = choosethread();
TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
cpu_throw(td, newtd); /* doesn't return */
}
/*
* This is called from fork_exit(). Just acquire the correct locks and
* let fork do the rest of the work.
*/
void
sched_fork_exit(struct thread *td)
{
struct tdq *tdq;
int cpuid;
/*
* Finish setting up thread glue so that it begins execution in a
* non-nested critical section with the scheduler lock held.
*/
cpuid = PCPU_GET(cpuid);
tdq = TDQ_CPU(cpuid);
if (TD_IS_IDLETHREAD(td))
td->td_lock = TDQ_LOCKPTR(tdq);
MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
td->td_oncpu = cpuid;
TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
lock_profile_obtain_lock_success(
&TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
}
/*
* Create on first use to catch odd startup conditons.
*/
char *
sched_tdname(struct thread *td)
{
#ifdef KTR
struct td_sched *ts;
ts = td_get_sched(td);
if (ts->ts_name[0] == '\0')
snprintf(ts->ts_name, sizeof(ts->ts_name),
"%s tid %d", td->td_name, td->td_tid);
return (ts->ts_name);
#else
return (td->td_name);
#endif
}
#ifdef KTR
void
sched_clear_tdname(struct thread *td)
{
struct td_sched *ts;
ts = td_get_sched(td);
ts->ts_name[0] = '\0';
}
#endif
#ifdef SMP
/*
* Build the CPU topology dump string. Is recursively called to collect
* the topology tree.
*/
static int
sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
int indent)
{
Commit the support for removing cpumask_t and replacing it directly with cpuset_t objects. That is going to offer the underlying support for a simple bump of MAXCPU and then support for number of cpus > 32 (as it is today). Right now, cpumask_t is an int, 32 bits on all our supported architecture. cpumask_t on the other side is implemented as an array of longs, and easilly extendible by definition. The architectures touched by this commit are the following: - amd64 - i386 - pc98 - arm - ia64 - XEN while the others are still missing. Userland is believed to be fully converted with the changes contained here. Some technical notes: - This commit may be considered an ABI nop for all the architectures different from amd64 and ia64 (and sparc64 in the future) - per-cpu members, which are now converted to cpuset_t, needs to be accessed avoiding migration, because the size of cpuset_t should be considered unknown - size of cpuset_t objects is different from kernel and userland (this is primirally done in order to leave some more space in userland to cope with KBI extensions). If you need to access kernel cpuset_t from the userland please refer to example in this patch on how to do that correctly (kgdb may be a good source, for example). - Support for other architectures is going to be added soon - Only MAXCPU for amd64 is bumped now The patch has been tested by sbruno and Nicholas Esborn on opteron 4 x 12 pack CPUs. More testing on big SMP is expected to came soon. pluknet tested the patch with his 8-ways on both amd64 and i386. Tested by: pluknet, sbruno, gianni, Nicholas Esborn Reviewed by: jeff, jhb, sbruno
2011-05-05 14:39:14 +00:00
char cpusetbuf[CPUSETBUFSIZ];
int i, first;
sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
"", 1 + indent / 2, cg->cg_level);
Commit the support for removing cpumask_t and replacing it directly with cpuset_t objects. That is going to offer the underlying support for a simple bump of MAXCPU and then support for number of cpus > 32 (as it is today). Right now, cpumask_t is an int, 32 bits on all our supported architecture. cpumask_t on the other side is implemented as an array of longs, and easilly extendible by definition. The architectures touched by this commit are the following: - amd64 - i386 - pc98 - arm - ia64 - XEN while the others are still missing. Userland is believed to be fully converted with the changes contained here. Some technical notes: - This commit may be considered an ABI nop for all the architectures different from amd64 and ia64 (and sparc64 in the future) - per-cpu members, which are now converted to cpuset_t, needs to be accessed avoiding migration, because the size of cpuset_t should be considered unknown - size of cpuset_t objects is different from kernel and userland (this is primirally done in order to leave some more space in userland to cope with KBI extensions). If you need to access kernel cpuset_t from the userland please refer to example in this patch on how to do that correctly (kgdb may be a good source, for example). - Support for other architectures is going to be added soon - Only MAXCPU for amd64 is bumped now The patch has been tested by sbruno and Nicholas Esborn on opteron 4 x 12 pack CPUs. More testing on big SMP is expected to came soon. pluknet tested the patch with his 8-ways on both amd64 and i386. Tested by: pluknet, sbruno, gianni, Nicholas Esborn Reviewed by: jeff, jhb, sbruno
2011-05-05 14:39:14 +00:00
sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
first = TRUE;
for (i = 0; i < MAXCPU; i++) {
Commit the support for removing cpumask_t and replacing it directly with cpuset_t objects. That is going to offer the underlying support for a simple bump of MAXCPU and then support for number of cpus > 32 (as it is today). Right now, cpumask_t is an int, 32 bits on all our supported architecture. cpumask_t on the other side is implemented as an array of longs, and easilly extendible by definition. The architectures touched by this commit are the following: - amd64 - i386 - pc98 - arm - ia64 - XEN while the others are still missing. Userland is believed to be fully converted with the changes contained here. Some technical notes: - This commit may be considered an ABI nop for all the architectures different from amd64 and ia64 (and sparc64 in the future) - per-cpu members, which are now converted to cpuset_t, needs to be accessed avoiding migration, because the size of cpuset_t should be considered unknown - size of cpuset_t objects is different from kernel and userland (this is primirally done in order to leave some more space in userland to cope with KBI extensions). If you need to access kernel cpuset_t from the userland please refer to example in this patch on how to do that correctly (kgdb may be a good source, for example). - Support for other architectures is going to be added soon - Only MAXCPU for amd64 is bumped now The patch has been tested by sbruno and Nicholas Esborn on opteron 4 x 12 pack CPUs. More testing on big SMP is expected to came soon. pluknet tested the patch with his 8-ways on both amd64 and i386. Tested by: pluknet, sbruno, gianni, Nicholas Esborn Reviewed by: jeff, jhb, sbruno
2011-05-05 14:39:14 +00:00
if (CPU_ISSET(i, &cg->cg_mask)) {
if (!first)
sbuf_printf(sb, ", ");
else
first = FALSE;
sbuf_printf(sb, "%d", i);
}
}
sbuf_printf(sb, "</cpu>\n");
if (cg->cg_flags != 0) {
sbuf_printf(sb, "%*s <flags>", indent, "");
if ((cg->cg_flags & CG_FLAG_HTT) != 0)
sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2010-06-10 11:48:14 +00:00
if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2010-06-10 11:48:14 +00:00
sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
sbuf_printf(sb, "</flags>\n");
}
if (cg->cg_children > 0) {
sbuf_printf(sb, "%*s <children>\n", indent, "");
for (i = 0; i < cg->cg_children; i++)
sysctl_kern_sched_topology_spec_internal(sb,
&cg->cg_child[i], indent+2);
sbuf_printf(sb, "%*s </children>\n", indent, "");
}
sbuf_printf(sb, "%*s</group>\n", indent, "");
return (0);
}
/*
* Sysctl handler for retrieving topology dump. It's a wrapper for
* the recursive sysctl_kern_smp_topology_spec_internal().
*/
static int
sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
{
struct sbuf *topo;
int err;
KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
if (topo == NULL)
return (ENOMEM);
sbuf_printf(topo, "<groups>\n");
err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
sbuf_printf(topo, "</groups>\n");
if (err == 0) {
err = sbuf_finish(topo);
}
sbuf_delete(topo);
return (err);
}
#endif
static int
sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
{
int error, new_val, period;
period = 1000000 / realstathz;
new_val = period * sched_slice;
error = sysctl_handle_int(oidp, &new_val, 0, req);
if (error != 0 || req->newptr == NULL)
return (error);
if (new_val <= 0)
return (EINVAL);
sched_slice = imax(1, (new_val + period / 2) / period);
sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
realstathz);
return (0);
}
SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
"Scheduler name");
SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
NULL, 0, sysctl_kern_quantum, "I",
"Quantum for timeshare threads in microseconds");
SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
"Quantum for timeshare threads in stathz ticks");
SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
"Interactivity score threshold");
SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
&preempt_thresh, 0,
"Maximal (lowest) priority for preemption");
SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
"Assign static kernel priorities to sleeping threads");
SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
"Number of times idle thread will spin waiting for new work");
SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
&sched_idlespinthresh, 0,
"Threshold before we will permit idle thread spinning");
#ifdef SMP
SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
"Number of hz ticks to keep thread affinity for");
SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
"Enables the long-term load balancer");
SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
&balance_interval, 0,
"Average period in stathz ticks to run the long-term balancer");
SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
"Attempts to steal work from other cores before idling");
SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
"Minimum load on remote CPU before we'll steal");
SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
"XML dump of detected CPU topology");
#endif
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/* ps compat. All cpu percentages from ULE are weighted. */
static int ccpu = 0;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");