freebsd-dev/sys/kern/kern_synch.c
Alexander Motin f91aa773be Add wakeup_any(), cheaper wakeup_one() for taskqueue(9).
wakeup_one() and underlying sleepq_signal() spend additional time trying
to be fair, waking thread with highest priority, sleeping longest time.
But in case of taskqueue there are many absolutely identical threads, and
any fairness between them is quite pointless.  It makes even worse, since
round-robin wakeups not only make previous CPU affinity in scheduler quite
useless, but also hide from user chance to see CPU bottlenecks, when
sequential workload with one request at a time looks evenly distributed
between multiple threads.

This change adds new SLEEPQ_UNFAIR flag to sleepq_signal(), making it wakeup
thread that went to sleep last, but no longer in context switch (to avoid
immediate spinning on the thread lock).  On top of that new wakeup_any()
function is added, equivalent to wakeup_one(), but setting the flag.
On top of that taskqueue(9) is switchied to wakeup_any() to wakeup its
threads.

As result, on 72-core Xeon v4 machine sequential ZFS write to 12 ZVOLs
with 16KB block size spend 34% less time in wakeup_any() and descendants
then it was spending in wakeup_one(), and total write throughput increased
by ~10% with the same as before CPU usage.

Reviewed by:	markj, mmacy
MFC after:	2 weeks
Sponsored by:	iXsystems, Inc.
Differential Revision:	https://reviews.freebsd.org/D20669
2019-06-20 01:15:33 +00:00

589 lines
16 KiB
C

/*-
* SPDX-License-Identifier: BSD-3-Clause
*
* Copyright (c) 1982, 1986, 1990, 1991, 1993
* The Regents of the University of California. All rights reserved.
* (c) UNIX System Laboratories, Inc.
* All or some portions of this file are derived from material licensed
* to the University of California by American Telephone and Telegraph
* Co. or Unix System Laboratories, Inc. and are reproduced herein with
* the permission of UNIX System Laboratories, Inc.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 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.
* 3. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``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 REGENTS OR CONTRIBUTORS 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.
*
* @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_ktrace.h"
#include "opt_sched.h"
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/condvar.h>
#include <sys/kdb.h>
#include <sys/kernel.h>
#include <sys/ktr.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/resourcevar.h>
#include <sys/sched.h>
#include <sys/sdt.h>
#include <sys/signalvar.h>
#include <sys/sleepqueue.h>
#include <sys/smp.h>
#include <sys/sx.h>
#include <sys/sysctl.h>
#include <sys/sysproto.h>
#include <sys/vmmeter.h>
#ifdef KTRACE
#include <sys/uio.h>
#include <sys/ktrace.h>
#endif
#include <machine/cpu.h>
static void synch_setup(void *dummy);
SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup,
NULL);
int hogticks;
static uint8_t pause_wchan[MAXCPU];
static struct callout loadav_callout;
struct loadavg averunnable =
{ {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
/*
* Constants for averages over 1, 5, and 15 minutes
* when sampling at 5 second intervals.
*/
static fixpt_t cexp[3] = {
0.9200444146293232 * FSCALE, /* exp(-1/12) */
0.9834714538216174 * FSCALE, /* exp(-1/60) */
0.9944598480048967 * FSCALE, /* exp(-1/180) */
};
/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, FSCALE, "");
static void loadav(void *arg);
SDT_PROVIDER_DECLARE(sched);
SDT_PROBE_DEFINE(sched, , , preempt);
static void
sleepinit(void *unused)
{
hogticks = (hz / 10) * 2; /* Default only. */
init_sleepqueues();
}
/*
* vmem tries to lock the sleepq mutexes when free'ing kva, so make sure
* it is available.
*/
SYSINIT(sleepinit, SI_SUB_KMEM, SI_ORDER_ANY, sleepinit, NULL);
/*
* General sleep call. Suspends the current thread until a wakeup is
* performed on the specified identifier. The thread will then be made
* runnable with the specified priority. Sleeps at most sbt units of time
* (0 means no timeout). If pri includes the PCATCH flag, let signals
* interrupt the sleep, otherwise ignore them while sleeping. Returns 0 if
* awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
* signal becomes pending, ERESTART is returned if the current system
* call should be restarted if possible, and EINTR is returned if the system
* call should be interrupted by the signal (return EINTR).
*
* The lock argument is unlocked before the caller is suspended, and
* re-locked before _sleep() returns. If priority includes the PDROP
* flag the lock is not re-locked before returning.
*/
int
_sleep(void *ident, struct lock_object *lock, int priority,
const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
{
struct thread *td;
struct lock_class *class;
uintptr_t lock_state;
int catch, pri, rval, sleepq_flags;
WITNESS_SAVE_DECL(lock_witness);
td = curthread;
#ifdef KTRACE
if (KTRPOINT(td, KTR_CSW))
ktrcsw(1, 0, wmesg);
#endif
WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock,
"Sleeping on \"%s\"", wmesg);
KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL,
("sleeping without a lock"));
KASSERT(ident != NULL, ("_sleep: NULL ident"));
KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running"));
KASSERT(td->td_epochnest == 0, ("sleeping in an epoch section"));
if (priority & PDROP)
KASSERT(lock != NULL && lock != &Giant.lock_object,
("PDROP requires a non-Giant lock"));
if (lock != NULL)
class = LOCK_CLASS(lock);
else
class = NULL;
if (SCHEDULER_STOPPED_TD(td)) {
if (lock != NULL && priority & PDROP)
class->lc_unlock(lock);
return (0);
}
catch = priority & PCATCH;
pri = priority & PRIMASK;
KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep"));
if ((uint8_t *)ident >= &pause_wchan[0] &&
(uint8_t *)ident <= &pause_wchan[MAXCPU - 1])
sleepq_flags = SLEEPQ_PAUSE;
else
sleepq_flags = SLEEPQ_SLEEP;
if (catch)
sleepq_flags |= SLEEPQ_INTERRUPTIBLE;
sleepq_lock(ident);
CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)",
td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
if (lock == &Giant.lock_object)
mtx_assert(&Giant, MA_OWNED);
DROP_GIANT();
if (lock != NULL && lock != &Giant.lock_object &&
!(class->lc_flags & LC_SLEEPABLE)) {
WITNESS_SAVE(lock, lock_witness);
lock_state = class->lc_unlock(lock);
} else
/* GCC needs to follow the Yellow Brick Road */
lock_state = -1;
/*
* We put ourselves on the sleep queue and start our timeout
* before calling thread_suspend_check, as we could stop there,
* and a wakeup or a SIGCONT (or both) could occur while we were
* stopped without resuming us. Thus, we must be ready for sleep
* when cursig() is called. If the wakeup happens while we're
* stopped, then td will no longer be on a sleep queue upon
* return from cursig().
*/
sleepq_add(ident, lock, wmesg, sleepq_flags, 0);
if (sbt != 0)
sleepq_set_timeout_sbt(ident, sbt, pr, flags);
if (lock != NULL && class->lc_flags & LC_SLEEPABLE) {
sleepq_release(ident);
WITNESS_SAVE(lock, lock_witness);
lock_state = class->lc_unlock(lock);
sleepq_lock(ident);
}
if (sbt != 0 && catch)
rval = sleepq_timedwait_sig(ident, pri);
else if (sbt != 0)
rval = sleepq_timedwait(ident, pri);
else if (catch)
rval = sleepq_wait_sig(ident, pri);
else {
sleepq_wait(ident, pri);
rval = 0;
}
#ifdef KTRACE
if (KTRPOINT(td, KTR_CSW))
ktrcsw(0, 0, wmesg);
#endif
PICKUP_GIANT();
if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) {
class->lc_lock(lock, lock_state);
WITNESS_RESTORE(lock, lock_witness);
}
return (rval);
}
int
msleep_spin_sbt(void *ident, struct mtx *mtx, const char *wmesg,
sbintime_t sbt, sbintime_t pr, int flags)
{
struct thread *td;
int rval;
WITNESS_SAVE_DECL(mtx);
td = curthread;
KASSERT(mtx != NULL, ("sleeping without a mutex"));
KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident"));
KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running"));
if (SCHEDULER_STOPPED_TD(td))
return (0);
sleepq_lock(ident);
CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)",
td->td_tid, td->td_proc->p_pid, td->td_name, wmesg, ident);
DROP_GIANT();
mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED);
WITNESS_SAVE(&mtx->lock_object, mtx);
mtx_unlock_spin(mtx);
/*
* We put ourselves on the sleep queue and start our timeout.
*/
sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0);
if (sbt != 0)
sleepq_set_timeout_sbt(ident, sbt, pr, flags);
/*
* Can't call ktrace with any spin locks held so it can lock the
* ktrace_mtx lock, and WITNESS_WARN considers it an error to hold
* any spin lock. Thus, we have to drop the sleepq spin lock while
* we handle those requests. This is safe since we have placed our
* thread on the sleep queue already.
*/
#ifdef KTRACE
if (KTRPOINT(td, KTR_CSW)) {
sleepq_release(ident);
ktrcsw(1, 0, wmesg);
sleepq_lock(ident);
}
#endif
#ifdef WITNESS
sleepq_release(ident);
WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"",
wmesg);
sleepq_lock(ident);
#endif
if (sbt != 0)
rval = sleepq_timedwait(ident, 0);
else {
sleepq_wait(ident, 0);
rval = 0;
}
#ifdef KTRACE
if (KTRPOINT(td, KTR_CSW))
ktrcsw(0, 0, wmesg);
#endif
PICKUP_GIANT();
mtx_lock_spin(mtx);
WITNESS_RESTORE(&mtx->lock_object, mtx);
return (rval);
}
/*
* pause_sbt() delays the calling thread by the given signed binary
* time. During cold bootup, pause_sbt() uses the DELAY() function
* instead of the _sleep() function to do the waiting. The "sbt"
* argument must be greater than or equal to zero. A "sbt" value of
* zero is equivalent to a "sbt" value of one tick.
*/
int
pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags)
{
KASSERT(sbt >= 0, ("pause_sbt: timeout must be >= 0"));
/* silently convert invalid timeouts */
if (sbt == 0)
sbt = tick_sbt;
if ((cold && curthread == &thread0) || kdb_active ||
SCHEDULER_STOPPED()) {
/*
* We delay one second at a time to avoid overflowing the
* system specific DELAY() function(s):
*/
while (sbt >= SBT_1S) {
DELAY(1000000);
sbt -= SBT_1S;
}
/* Do the delay remainder, if any */
sbt = howmany(sbt, SBT_1US);
if (sbt > 0)
DELAY(sbt);
return (EWOULDBLOCK);
}
return (_sleep(&pause_wchan[curcpu], NULL,
(flags & C_CATCH) ? PCATCH : 0, wmesg, sbt, pr, flags));
}
/*
* Make all threads sleeping on the specified identifier runnable.
*/
void
wakeup(void *ident)
{
int wakeup_swapper;
sleepq_lock(ident);
wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0);
sleepq_release(ident);
if (wakeup_swapper) {
KASSERT(ident != &proc0,
("wakeup and wakeup_swapper and proc0"));
kick_proc0();
}
}
/*
* Make a thread sleeping on the specified identifier runnable.
* May wake more than one thread if a target thread is currently
* swapped out.
*/
void
wakeup_one(void *ident)
{
int wakeup_swapper;
sleepq_lock(ident);
wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0);
sleepq_release(ident);
if (wakeup_swapper)
kick_proc0();
}
void
wakeup_any(void *ident)
{
int wakeup_swapper;
sleepq_lock(ident);
wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP | SLEEPQ_UNFAIR,
0, 0);
sleepq_release(ident);
if (wakeup_swapper)
kick_proc0();
}
static void
kdb_switch(void)
{
thread_unlock(curthread);
kdb_backtrace();
kdb_reenter();
panic("%s: did not reenter debugger", __func__);
}
/*
* The machine independent parts of context switching.
*/
void
mi_switch(int flags, struct thread *newtd)
{
uint64_t runtime, new_switchtime;
struct thread *td;
td = curthread; /* XXX */
THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code"));
#ifdef INVARIANTS
if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td))
mtx_assert(&Giant, MA_NOTOWNED);
#endif
KASSERT(td->td_critnest == 1 || panicstr,
("mi_switch: switch in a critical section"));
KASSERT((flags & (SW_INVOL | SW_VOL)) != 0,
("mi_switch: switch must be voluntary or involuntary"));
KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself"));
/*
* Don't perform context switches from the debugger.
*/
if (kdb_active)
kdb_switch();
if (SCHEDULER_STOPPED_TD(td))
return;
if (flags & SW_VOL) {
td->td_ru.ru_nvcsw++;
td->td_swvoltick = ticks;
} else {
td->td_ru.ru_nivcsw++;
td->td_swinvoltick = ticks;
}
#ifdef SCHED_STATS
SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]);
#endif
/*
* Compute the amount of time during which the current
* thread was running, and add that to its total so far.
*/
new_switchtime = cpu_ticks();
runtime = new_switchtime - PCPU_GET(switchtime);
td->td_runtime += runtime;
td->td_incruntime += runtime;
PCPU_SET(switchtime, new_switchtime);
td->td_generation++; /* bump preempt-detect counter */
VM_CNT_INC(v_swtch);
PCPU_SET(switchticks, ticks);
CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)",
td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
#ifdef KDTRACE_HOOKS
if (SDT_PROBES_ENABLED() &&
((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 &&
(flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)))
SDT_PROBE0(sched, , , preempt);
#endif
sched_switch(td, newtd, flags);
CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)",
td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name);
/*
* If the last thread was exiting, finish cleaning it up.
*/
if ((td = PCPU_GET(deadthread))) {
PCPU_SET(deadthread, NULL);
thread_stash(td);
}
}
/*
* Change thread state to be runnable, placing it on the run queue if
* it is in memory. If it is swapped out, return true so our caller
* will know to awaken the swapper.
*/
int
setrunnable(struct thread *td)
{
THREAD_LOCK_ASSERT(td, MA_OWNED);
KASSERT(td->td_proc->p_state != PRS_ZOMBIE,
("setrunnable: pid %d is a zombie", td->td_proc->p_pid));
switch (td->td_state) {
case TDS_RUNNING:
case TDS_RUNQ:
return (0);
case TDS_INHIBITED:
/*
* If we are only inhibited because we are swapped out
* then arange to swap in this process. Otherwise just return.
*/
if (td->td_inhibitors != TDI_SWAPPED)
return (0);
/* FALLTHROUGH */
case TDS_CAN_RUN:
break;
default:
printf("state is 0x%x", td->td_state);
panic("setrunnable(2)");
}
if ((td->td_flags & TDF_INMEM) == 0) {
if ((td->td_flags & TDF_SWAPINREQ) == 0) {
td->td_flags |= TDF_SWAPINREQ;
return (1);
}
} else
sched_wakeup(td);
return (0);
}
/*
* Compute a tenex style load average of a quantity on
* 1, 5 and 15 minute intervals.
*/
static void
loadav(void *arg)
{
int i, nrun;
struct loadavg *avg;
nrun = sched_load();
avg = &averunnable;
for (i = 0; i < 3; i++)
avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
/*
* Schedule the next update to occur after 5 seconds, but add a
* random variation to avoid synchronisation with processes that
* run at regular intervals.
*/
callout_reset_sbt(&loadav_callout,
SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US,
loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
}
/* ARGSUSED */
static void
synch_setup(void *dummy)
{
callout_init(&loadav_callout, 1);
/* Kick off timeout driven events by calling first time. */
loadav(NULL);
}
int
should_yield(void)
{
return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks);
}
void
maybe_yield(void)
{
if (should_yield())
kern_yield(PRI_USER);
}
void
kern_yield(int prio)
{
struct thread *td;
td = curthread;
DROP_GIANT();
thread_lock(td);
if (prio == PRI_USER)
prio = td->td_user_pri;
if (prio >= 0)
sched_prio(td, prio);
mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
thread_unlock(td);
PICKUP_GIANT();
}
/*
* General purpose yield system call.
*/
int
sys_yield(struct thread *td, struct yield_args *uap)
{
thread_lock(td);
if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
sched_prio(td, PRI_MAX_TIMESHARE);
mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
thread_unlock(td);
td->td_retval[0] = 0;
return (0);
}