2 * Copyright (c) 1982, 1986, 1990, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
35 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
38 #include "opt_ktrace.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/signalvar.h>
45 #include <sys/resourcevar.h>
46 #include <sys/vmmeter.h>
47 #include <sys/sysctl.h>
51 #include <sys/kcollect.h>
53 #include <sys/ktrace.h>
56 #include <sys/serialize.h>
58 #include <sys/signal2.h>
59 #include <sys/thread2.h>
60 #include <sys/spinlock2.h>
61 #include <sys/mutex2.h>
63 #include <machine/cpu.h>
64 #include <machine/smp.h>
66 #include <vm/vm_extern.h>
69 TAILQ_HEAD(, thread
) queue
;
70 const volatile void *ident0
;
71 const volatile void *ident1
;
72 const volatile void *ident2
;
73 const volatile void *ident3
;
76 static void sched_setup (void *dummy
);
77 SYSINIT(sched_setup
, SI_SUB_KICK_SCHEDULER
, SI_ORDER_FIRST
, sched_setup
, NULL
);
78 static void sched_dyninit (void *dummy
);
79 SYSINIT(sched_dyninit
, SI_BOOT1_DYNALLOC
, SI_ORDER_FIRST
, sched_dyninit
, NULL
);
84 int ncpus_fit
, ncpus_fit_mask
; /* note: mask not cpumask_t */
87 int tsleep_crypto_dump
= 0;
89 MALLOC_DEFINE(M_TSLEEP
, "tslpque", "tsleep queues");
91 #define __DEALL(ident) __DEQUALIFY(void *, ident)
93 #if !defined(KTR_TSLEEP)
94 #define KTR_TSLEEP KTR_ALL
96 KTR_INFO_MASTER(tsleep
);
97 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_beg
, 0, "tsleep enter %p", const volatile void *ident
);
98 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_end
, 1, "tsleep exit");
99 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_beg
, 2, "wakeup enter %p", const volatile void *ident
);
100 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_end
, 3, "wakeup exit");
101 KTR_INFO(KTR_TSLEEP
, tsleep
, ilockfail
, 4, "interlock failed %p", const volatile void *ident
);
103 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
104 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
106 struct loadavg averunnable
=
107 { {0, 0, 0}, FSCALE
}; /* load average, of runnable procs */
109 * Constants for averages over 1, 5, and 15 minutes
110 * when sampling at 5 second intervals.
112 static fixpt_t cexp
[3] = {
113 0.9200444146293232 * FSCALE
, /* exp(-1/12) */
114 0.9834714538216174 * FSCALE
, /* exp(-1/60) */
115 0.9944598480048967 * FSCALE
, /* exp(-1/180) */
118 static void endtsleep (void *);
119 static void loadav (void *arg
);
120 static void schedcpu (void *arg
);
122 static int pctcpu_decay
= 10;
123 SYSCTL_INT(_kern
, OID_AUTO
, pctcpu_decay
, CTLFLAG_RW
,
124 &pctcpu_decay
, 0, "");
127 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
129 int fscale __unused
= FSCALE
; /* exported to systat */
130 SYSCTL_INT(_kern
, OID_AUTO
, fscale
, CTLFLAG_RD
, 0, FSCALE
, "");
133 * Issue a wakeup() from userland (debugging)
136 sysctl_wakeup(SYSCTL_HANDLER_ARGS
)
141 if (req
->newptr
!= NULL
) {
142 if (priv_check(curthread
, PRIV_ROOT
))
144 error
= SYSCTL_IN(req
, &ident
, sizeof(ident
));
147 kprintf("issue wakeup %016jx\n", ident
);
148 wakeup((void *)(intptr_t)ident
);
150 if (req
->oldptr
!= NULL
) {
151 error
= SYSCTL_OUT(req
, &ident
, sizeof(ident
));
156 SYSCTL_PROC(_debug
, OID_AUTO
, wakeup
, CTLTYPE_UQUAD
|CTLFLAG_RW
, 0, 0,
157 sysctl_wakeup
, "Q", "issue wakeup(addr)");
160 * Recompute process priorities, once a second.
162 * Since the userland schedulers are typically event oriented, if the
163 * estcpu calculation at wakeup() time is not sufficient to make a
164 * process runnable relative to other processes in the system we have
165 * a 1-second recalc to help out.
167 * This code also allows us to store sysclock_t data in the process structure
168 * without fear of an overrun, since sysclock_t are guarenteed to hold
169 * several seconds worth of count.
171 * WARNING! callouts can preempt normal threads. However, they will not
172 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
174 static int schedcpu_stats(struct proc
*p
, void *data __unused
);
175 static int schedcpu_resource(struct proc
*p
, void *data __unused
);
180 allproc_scan(schedcpu_stats
, NULL
, 1);
181 allproc_scan(schedcpu_resource
, NULL
, 1);
182 if (mycpu
->gd_cpuid
== 0) {
183 wakeup((caddr_t
)&lbolt
);
184 wakeup(lbolt_syncer
);
186 callout_reset(&mycpu
->gd_schedcpu_callout
, hz
, schedcpu
, NULL
);
190 * General process statistics once a second
193 schedcpu_stats(struct proc
*p
, void *data __unused
)
198 * Threads may not be completely set up if process in SIDL state.
200 if (p
->p_stat
== SIDL
)
204 if (lwkt_trytoken(&p
->p_token
) == FALSE
) {
210 FOREACH_LWP_IN_PROC(lp
, p
) {
211 if (lp
->lwp_stat
== LSSLEEP
) {
213 if (lp
->lwp_slptime
== 1)
214 p
->p_usched
->uload_update(lp
);
218 * Only recalculate processes that are active or have slept
219 * less then 2 seconds. The schedulers understand this.
220 * Otherwise decay by 50% per second.
222 if (lp
->lwp_slptime
<= 1) {
223 p
->p_usched
->recalculate(lp
);
227 decay
= pctcpu_decay
;
233 lp
->lwp_pctcpu
= (lp
->lwp_pctcpu
* (decay
- 1)) / decay
;
236 lwkt_reltoken(&p
->p_token
);
243 * Resource checks. XXX break out since ksignal/killproc can block,
244 * limiting us to one process killed per second. There is probably
248 schedcpu_resource(struct proc
*p
, void *data __unused
)
253 if (p
->p_stat
== SIDL
)
257 if (lwkt_trytoken(&p
->p_token
) == FALSE
) {
262 if (p
->p_stat
== SZOMB
|| p
->p_limit
== NULL
) {
263 lwkt_reltoken(&p
->p_token
);
269 FOREACH_LWP_IN_PROC(lp
, p
) {
271 * We may have caught an lp in the middle of being
272 * created, lwp_thread can be NULL.
274 if (lp
->lwp_thread
) {
275 ttime
+= lp
->lwp_thread
->td_sticks
;
276 ttime
+= lp
->lwp_thread
->td_uticks
;
280 switch(plimit_testcpulimit(p
->p_limit
, ttime
)) {
281 case PLIMIT_TESTCPU_KILL
:
282 killproc(p
, "exceeded maximum CPU limit");
284 case PLIMIT_TESTCPU_XCPU
:
285 if ((p
->p_flags
& P_XCPU
) == 0) {
286 p
->p_flags
|= P_XCPU
;
293 lwkt_reltoken(&p
->p_token
);
300 * This is only used by ps. Generate a cpu percentage use over
301 * a period of one second.
304 updatepcpu(struct lwp
*lp
, int cpticks
, int ttlticks
)
309 acc
= (cpticks
<< FSHIFT
) / ttlticks
;
310 if (ttlticks
>= ESTCPUFREQ
) {
311 lp
->lwp_pctcpu
= acc
;
313 remticks
= ESTCPUFREQ
- ttlticks
;
314 lp
->lwp_pctcpu
= (acc
* ttlticks
+ lp
->lwp_pctcpu
* remticks
) /
320 * Handy macros to calculate hash indices. LOOKUP() calculates the
321 * global cpumask hash index, TCHASHSHIFT() converts that into the
324 * By making the pcpu hash arrays smaller we save a significant amount
325 * of memory at very low cost. The real cost is in IPIs, which are handled
326 * by the much larger global cpumask hash table.
328 #define LOOKUP_PRIME 66555444443333333ULL
329 #define LOOKUP(x) ((((uintptr_t)(x) + ((uintptr_t)(x) >> 18)) ^ \
330 LOOKUP_PRIME) % slpque_tablesize)
331 #define TCHASHSHIFT(x) ((x) >> 4)
333 static uint32_t slpque_tablesize
;
334 static cpumask_t
*slpque_cpumasks
;
336 SYSCTL_UINT(_kern
, OID_AUTO
, slpque_tablesize
, CTLFLAG_RD
, &slpque_tablesize
,
340 * This is a dandy function that allows us to interlock tsleep/wakeup
341 * operations with unspecified upper level locks, such as lockmgr locks,
342 * simply by holding a critical section. The sequence is:
344 * (acquire upper level lock)
345 * tsleep_interlock(blah)
346 * (release upper level lock)
349 * Basically this functions queues us on the tsleep queue without actually
350 * descheduling us. When tsleep() is later called with PINTERLOCK it
351 * assumes the thread was already queued, otherwise it queues it there.
353 * Thus it is possible to receive the wakeup prior to going to sleep and
354 * the race conditions are covered.
357 _tsleep_interlock(globaldata_t gd
, const volatile void *ident
, int flags
)
359 thread_t td
= gd
->gd_curthread
;
365 kprintf("tsleep_interlock: NULL ident %s\n", td
->td_comm
);
369 crit_enter_quick(td
);
370 if (td
->td_flags
& TDF_TSLEEPQ
) {
371 cid
= LOOKUP(td
->td_wchan
);
372 gid
= TCHASHSHIFT(cid
);
373 qp
= &gd
->gd_tsleep_hash
[gid
];
374 TAILQ_REMOVE(&qp
->queue
, td
, td_sleepq
);
375 if (TAILQ_FIRST(&qp
->queue
) == NULL
) {
380 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[cid
],
384 td
->td_flags
|= TDF_TSLEEPQ
;
387 gid
= TCHASHSHIFT(cid
);
388 qp
= &gd
->gd_tsleep_hash
[gid
];
389 TAILQ_INSERT_TAIL(&qp
->queue
, td
, td_sleepq
);
390 if (qp
->ident0
!= ident
&& qp
->ident1
!= ident
&&
391 qp
->ident2
!= ident
&& qp
->ident3
!= ident
) {
392 if (qp
->ident0
== NULL
)
394 else if (qp
->ident1
== NULL
)
396 else if (qp
->ident2
== NULL
)
398 else if (qp
->ident3
== NULL
)
401 qp
->ident0
= (void *)(intptr_t)-1;
403 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks
[cid
], gd
->gd_cpuid
);
404 td
->td_wchan
= ident
;
405 td
->td_wdomain
= flags
& PDOMAIN_MASK
;
410 tsleep_interlock(const volatile void *ident
, int flags
)
412 _tsleep_interlock(mycpu
, ident
, flags
);
416 * Remove thread from sleepq. Must be called with a critical section held.
417 * The thread must not be migrating.
420 _tsleep_remove(thread_t td
)
422 globaldata_t gd
= mycpu
;
427 KKASSERT(td
->td_gd
== gd
&& IN_CRITICAL_SECT(td
));
428 KKASSERT((td
->td_flags
& TDF_MIGRATING
) == 0);
429 if (td
->td_flags
& TDF_TSLEEPQ
) {
430 td
->td_flags
&= ~TDF_TSLEEPQ
;
431 cid
= LOOKUP(td
->td_wchan
);
432 gid
= TCHASHSHIFT(cid
);
433 qp
= &gd
->gd_tsleep_hash
[gid
];
434 TAILQ_REMOVE(&qp
->queue
, td
, td_sleepq
);
435 if (TAILQ_FIRST(&qp
->queue
) == NULL
) {
436 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[cid
],
445 tsleep_remove(thread_t td
)
451 * General sleep call. Suspends the current process until a wakeup is
452 * performed on the specified identifier. The process will then be made
453 * runnable with the specified priority. Sleeps at most timo/hz seconds
454 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
455 * before and after sleeping, else signals are not checked. Returns 0 if
456 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
457 * signal needs to be delivered, ERESTART is returned if the current system
458 * call should be restarted if possible, and EINTR is returned if the system
459 * call should be interrupted by the signal (return EINTR).
461 * Note that if we are a process, we release_curproc() before messing with
462 * the LWKT scheduler.
464 * During autoconfiguration or after a panic, a sleep will simply
465 * lower the priority briefly to allow interrupts, then return.
467 * WARNING! This code can't block (short of switching away), or bad things
468 * will happen. No getting tokens, no blocking locks, etc.
471 tsleep(const volatile void *ident
, int flags
, const char *wmesg
, int timo
)
473 struct thread
*td
= curthread
;
474 struct lwp
*lp
= td
->td_lwp
;
475 struct proc
*p
= td
->td_proc
; /* may be NULL */
481 struct callout thandle
;
484 * Currently a severe hack. Make sure any delayed wakeups
485 * are flushed before we sleep or we might deadlock on whatever
486 * event we are sleeping on.
488 if (td
->td_flags
& TDF_DELAYED_WAKEUP
)
489 wakeup_end_delayed();
492 * NOTE: removed KTRPOINT, it could cause races due to blocking
493 * even in stable. Just scrap it for now.
495 if (!tsleep_crypto_dump
&& (tsleep_now_works
== 0 || panicstr
)) {
497 * After a panic, or before we actually have an operational
498 * softclock, just give interrupts a chance, then just return;
500 * don't run any other procs or panic below,
501 * in case this is the idle process and already asleep.
505 lwkt_setpri_self(safepri
);
507 lwkt_setpri_self(oldpri
);
510 logtsleep2(tsleep_beg
, ident
);
512 KKASSERT(td
!= &gd
->gd_idlethread
); /* you must be kidding! */
513 td
->td_wakefromcpu
= -1; /* overwritten by _wakeup */
516 * NOTE: all of this occurs on the current cpu, including any
517 * callout-based wakeups, so a critical section is a sufficient
520 * The entire sequence through to where we actually sleep must
521 * run without breaking the critical section.
523 catch = flags
& PCATCH
;
527 crit_enter_quick(td
);
529 KASSERT(ident
!= NULL
, ("tsleep: no ident"));
530 KASSERT(lp
== NULL
||
531 lp
->lwp_stat
== LSRUN
|| /* Obvious */
532 lp
->lwp_stat
== LSSTOP
, /* Set in tstop */
534 ident
, wmesg
, lp
->lwp_stat
));
537 * We interlock the sleep queue if the caller has not already done
538 * it for us. This must be done before we potentially acquire any
539 * tokens or we can loose the wakeup.
541 if ((flags
& PINTERLOCKED
) == 0) {
542 _tsleep_interlock(gd
, ident
, flags
);
546 * Setup for the current process (if this is a process). We must
547 * interlock with lwp_token to avoid remote wakeup races via
551 lwkt_gettoken(&lp
->lwp_token
);
554 * If the umbrella process is in the SCORE state then
555 * make sure that the thread is flagged going into a
556 * normal sleep to allow the core dump to proceed, otherwise
557 * the coredump can end up waiting forever. If the normal
558 * sleep is woken up, the thread will enter a stopped state
559 * upon return to userland.
561 * We do not want to interrupt or cause a thread exist at
562 * this juncture because that will mess-up the state the
563 * coredump is trying to save.
565 if (p
->p_stat
== SCORE
&&
566 (lp
->lwp_mpflags
& LWP_MP_WSTOP
) == 0) {
567 atomic_set_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
576 * Early termination if PCATCH was set and a
577 * signal is pending, interlocked with the
580 * Early termination only occurs when tsleep() is
581 * entered while in a normal LSRUN state.
583 if ((sig
= CURSIG(lp
)) != 0)
587 * Causes ksignal to wake us up if a signal is
588 * received (interlocked with lp->lwp_token).
590 lp
->lwp_flags
|= LWP_SINTR
;
597 * Make sure the current process has been untangled from
598 * the userland scheduler and initialize slptime to start
601 * NOTE: td->td_wakefromcpu is pre-set by the release function
602 * for the dfly scheduler, and then adjusted by _wakeup()
605 p
->p_usched
->release_curproc(lp
);
610 * If the interlocked flag is set but our cpu bit in the slpqueue
611 * is no longer set, then a wakeup was processed inbetween the
612 * tsleep_interlock() (ours or the callers), and here. This can
613 * occur under numerous circumstances including when we release the
616 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
617 * to process incoming IPIs, thus draining incoming wakeups.
619 if ((td
->td_flags
& TDF_TSLEEPQ
) == 0) {
620 logtsleep2(ilockfail
, ident
);
625 * scheduling is blocked while in a critical section. Coincide
626 * the descheduled-by-tsleep flag with the descheduling of the
629 * The timer callout is localized on our cpu and interlocked by
630 * our critical section.
632 lwkt_deschedule_self(td
);
633 td
->td_flags
|= TDF_TSLEEP_DESCHEDULED
;
634 td
->td_wmesg
= wmesg
;
637 * Setup the timeout, if any. The timeout is only operable while
638 * the thread is flagged descheduled.
640 KKASSERT((td
->td_flags
& TDF_TIMEOUT
) == 0);
642 callout_init_mp(&thandle
);
643 callout_reset(&thandle
, timo
, endtsleep
, td
);
651 * Ok, we are sleeping. Place us in the SSLEEP state.
653 KKASSERT((lp
->lwp_mpflags
& LWP_MP_ONRUNQ
) == 0);
656 * tstop() sets LSSTOP, so don't fiddle with that.
658 if (lp
->lwp_stat
!= LSSTOP
)
659 lp
->lwp_stat
= LSSLEEP
;
660 lp
->lwp_ru
.ru_nvcsw
++;
661 p
->p_usched
->uload_update(lp
);
665 * And when we are woken up, put us back in LSRUN. If we
666 * slept for over a second, recalculate our estcpu.
668 lp
->lwp_stat
= LSRUN
;
669 if (lp
->lwp_slptime
) {
670 p
->p_usched
->uload_update(lp
);
671 p
->p_usched
->recalculate(lp
);
679 * Make sure we haven't switched cpus while we were asleep. It's
680 * not supposed to happen. Cleanup our temporary flags.
682 KKASSERT(gd
== td
->td_gd
);
685 * Cleanup the timeout. If the timeout has already occured thandle
686 * has already been stopped, otherwise stop thandle. If the timeout
687 * is running (the callout thread must be blocked trying to get
688 * lwp_token) then wait for us to get scheduled.
691 while (td
->td_flags
& TDF_TIMEOUT_RUNNING
) {
692 /* else we won't get rescheduled! */
693 if (lp
->lwp_stat
!= LSSTOP
)
694 lp
->lwp_stat
= LSSLEEP
;
695 lwkt_deschedule_self(td
);
696 td
->td_wmesg
= "tsrace";
698 kprintf("td %p %s: timeout race\n", td
, td
->td_comm
);
700 if (td
->td_flags
& TDF_TIMEOUT
) {
701 td
->td_flags
&= ~TDF_TIMEOUT
;
704 /* does not block when on same cpu */
705 callout_stop(&thandle
);
708 td
->td_flags
&= ~TDF_TSLEEP_DESCHEDULED
;
711 * Make sure we have been removed from the sleepq. In most
712 * cases this will have been done for us already but it is
713 * possible for a scheduling IPI to be in-flight from a
714 * previous tsleep/tsleep_interlock() or due to a straight-out
715 * call to lwkt_schedule() (in the case of an interrupt thread),
716 * causing a spurious wakeup.
722 * Figure out the correct error return. If interrupted by a
723 * signal we want to return EINTR or ERESTART.
727 if (catch && error
== 0) {
728 if (sig
!= 0 || (sig
= CURSIG(lp
))) {
729 if (SIGISMEMBER(p
->p_sigacts
->ps_sigintr
, sig
))
736 lp
->lwp_flags
&= ~LWP_SINTR
;
739 * Unconditionally set us to LSRUN on resume. lwp_stat could
740 * be in a weird state due to the goto resume, particularly
741 * when tsleep() is called from tstop().
743 lp
->lwp_stat
= LSRUN
;
744 lwkt_reltoken(&lp
->lwp_token
);
746 logtsleep1(tsleep_end
);
752 * Interlocked spinlock sleep. An exclusively held spinlock must
753 * be passed to ssleep(). The function will atomically release the
754 * spinlock and tsleep on the ident, then reacquire the spinlock and
757 * This routine is fairly important along the critical path, so optimize it
761 ssleep(const volatile void *ident
, struct spinlock
*spin
, int flags
,
762 const char *wmesg
, int timo
)
764 globaldata_t gd
= mycpu
;
767 _tsleep_interlock(gd
, ident
, flags
);
768 spin_unlock_quick(gd
, spin
);
769 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
770 KKASSERT(gd
== mycpu
);
771 _spin_lock_quick(gd
, spin
, wmesg
);
777 lksleep(const volatile void *ident
, struct lock
*lock
, int flags
,
778 const char *wmesg
, int timo
)
780 globaldata_t gd
= mycpu
;
783 _tsleep_interlock(gd
, ident
, flags
);
784 lockmgr(lock
, LK_RELEASE
);
785 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
786 lockmgr(lock
, LK_EXCLUSIVE
);
792 * Interlocked mutex sleep. An exclusively held mutex must be passed
793 * to mtxsleep(). The function will atomically release the mutex
794 * and tsleep on the ident, then reacquire the mutex and return.
797 mtxsleep(const volatile void *ident
, struct mtx
*mtx
, int flags
,
798 const char *wmesg
, int timo
)
800 globaldata_t gd
= mycpu
;
803 _tsleep_interlock(gd
, ident
, flags
);
805 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
806 mtx_lock_ex_quick(mtx
);
812 * Interlocked serializer sleep. An exclusively held serializer must
813 * be passed to zsleep(). The function will atomically release
814 * the serializer and tsleep on the ident, then reacquire the serializer
818 zsleep(const volatile void *ident
, struct lwkt_serialize
*slz
, int flags
,
819 const char *wmesg
, int timo
)
821 globaldata_t gd
= mycpu
;
824 ASSERT_SERIALIZED(slz
);
826 _tsleep_interlock(gd
, ident
, flags
);
827 lwkt_serialize_exit(slz
);
828 ret
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
829 lwkt_serialize_enter(slz
);
835 * Directly block on the LWKT thread by descheduling it. This
836 * is much faster then tsleep(), but the only legal way to wake
837 * us up is to directly schedule the thread.
839 * Setting TDF_SINTR will cause new signals to directly schedule us.
841 * This routine must be called while in a critical section.
844 lwkt_sleep(const char *wmesg
, int flags
)
846 thread_t td
= curthread
;
849 if ((flags
& PCATCH
) == 0 || td
->td_lwp
== NULL
) {
850 td
->td_flags
|= TDF_BLOCKED
;
851 td
->td_wmesg
= wmesg
;
852 lwkt_deschedule_self(td
);
855 td
->td_flags
&= ~TDF_BLOCKED
;
858 if ((sig
= CURSIG(td
->td_lwp
)) != 0) {
859 if (SIGISMEMBER(td
->td_proc
->p_sigacts
->ps_sigintr
, sig
))
865 td
->td_flags
|= TDF_BLOCKED
| TDF_SINTR
;
866 td
->td_wmesg
= wmesg
;
867 lwkt_deschedule_self(td
);
869 td
->td_flags
&= ~(TDF_BLOCKED
| TDF_SINTR
);
875 * Implement the timeout for tsleep.
877 * This type of callout timeout is scheduled on the same cpu the process
878 * is sleeping on. Also, at the moment, the MP lock is held.
887 * We are going to have to get the lwp_token, which means we might
888 * block. This can race a tsleep getting woken up by other means
889 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
890 * processing to complete (sorry tsleep!).
892 * We can safely set td_flags because td MUST be on the same cpu
895 KKASSERT(td
->td_gd
== mycpu
);
897 td
->td_flags
|= TDF_TIMEOUT_RUNNING
| TDF_TIMEOUT
;
900 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
901 * from exiting the tsleep on us. The flag is interlocked by virtue
902 * of lp being on the same cpu as we are.
904 if ((lp
= td
->td_lwp
) != NULL
)
905 lwkt_gettoken(&lp
->lwp_token
);
907 KKASSERT(td
->td_flags
& TDF_TSLEEP_DESCHEDULED
);
911 * callout timer should normally never be set in tstop()
912 * because it passes a timeout of 0. However, there is a
913 * case during thread exit (which SSTOP's all the threads)
914 * for which tstop() must break out and can (properly) leave
915 * the thread in LSSTOP.
917 KKASSERT(lp
->lwp_stat
!= LSSTOP
||
918 (lp
->lwp_mpflags
& LWP_MP_WEXIT
));
920 lwkt_reltoken(&lp
->lwp_token
);
925 KKASSERT(td
->td_gd
== mycpu
);
926 td
->td_flags
&= ~TDF_TIMEOUT_RUNNING
;
931 * Make all processes sleeping on the specified identifier runnable.
932 * count may be zero or one only.
934 * The domain encodes the sleep/wakeup domain, flags, plus the originating
937 * This call may run without the MP lock held. We can only manipulate thread
938 * state on the cpu owning the thread. We CANNOT manipulate process state
941 * _wakeup() can be passed to an IPI so we can't use (const volatile
945 _wakeup(void *ident
, int domain
)
957 logtsleep2(wakeup_beg
, ident
);
960 gid
= TCHASHSHIFT(cid
);
961 qp
= &gd
->gd_tsleep_hash
[gid
];
963 for (td
= TAILQ_FIRST(&qp
->queue
); td
!= NULL
; td
= ntd
) {
964 ntd
= TAILQ_NEXT(td
, td_sleepq
);
965 if (td
->td_wchan
== ident
&&
966 td
->td_wdomain
== (domain
& PDOMAIN_MASK
)
968 KKASSERT(td
->td_gd
== gd
);
970 td
->td_wakefromcpu
= PWAKEUP_DECODE(domain
);
971 if (td
->td_flags
& TDF_TSLEEP_DESCHEDULED
) {
973 if (domain
& PWAKEUP_ONE
)
978 if (td
->td_wchan
== qp
->ident0
)
980 else if (td
->td_wchan
== qp
->ident1
)
982 else if (td
->td_wchan
== qp
->ident2
)
984 else if (td
->td_wchan
== qp
->ident3
)
987 wids
|= 16; /* force ident0 to be retained (-1) */
991 * Because a bunch of cpumask array entries cover the same queue, it
992 * is possible for our bit to remain set in some of them and cause
993 * spurious wakeup IPIs later on. Make sure that the bit is cleared
994 * when a spurious IPI occurs to prevent further spurious IPIs.
996 if (TAILQ_FIRST(&qp
->queue
) == NULL
) {
997 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[cid
], gd
->gd_cpuid
);
1003 if ((wids
& 1) == 0) {
1004 if ((wids
& 16) == 0) {
1007 KKASSERT(qp
->ident0
== (void *)(intptr_t)-1);
1010 if ((wids
& 2) == 0)
1012 if ((wids
& 4) == 0)
1014 if ((wids
& 8) == 0)
1019 * We finished checking the current cpu but there still may be
1020 * more work to do. Either wakeup_one was requested and no matching
1021 * thread was found, or a normal wakeup was requested and we have
1022 * to continue checking cpus.
1024 * It should be noted that this scheme is actually less expensive then
1025 * the old scheme when waking up multiple threads, since we send
1026 * only one IPI message per target candidate which may then schedule
1027 * multiple threads. Before we could have wound up sending an IPI
1028 * message for each thread on the target cpu (!= current cpu) that
1029 * needed to be woken up.
1031 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
1032 * should be ok since we are passing idents in the IPI rather
1033 * then thread pointers.
1035 * NOTE: We MUST mfence (or use an atomic op) prior to reading
1036 * the cpumask, as another cpu may have written to it in
1037 * a fashion interlocked with whatever the caller did before
1038 * calling wakeup(). Otherwise we might miss the interaction
1039 * (kern_mutex.c can cause this problem).
1041 * lfence is insufficient as it may allow a written state to
1042 * reorder around the cpumask load.
1044 if ((domain
& PWAKEUP_MYCPU
) == 0) {
1046 const volatile void *id0
;
1051 mask
= slpque_cpumasks
[cid
];
1052 CPUMASK_ANDMASK(mask
, gd
->gd_other_cpus
);
1053 while (CPUMASK_TESTNZERO(mask
)) {
1054 n
= BSRCPUMASK(mask
);
1055 CPUMASK_NANDBIT(mask
, n
);
1056 tgd
= globaldata_find(n
);
1059 * Both ident0 compares must from a single load
1060 * to avoid ident0 update races crossing the two
1063 qp
= &tgd
->gd_tsleep_hash
[gid
];
1066 if (id0
== (void *)(intptr_t)-1) {
1067 lwkt_send_ipiq2(tgd
, _wakeup
, ident
,
1068 domain
| PWAKEUP_MYCPU
);
1069 ++tgd
->gd_cnt
.v_wakeup_colls
;
1070 } else if (id0
== ident
||
1071 qp
->ident1
== ident
||
1072 qp
->ident2
== ident
||
1073 qp
->ident3
== ident
) {
1074 lwkt_send_ipiq2(tgd
, _wakeup
, ident
,
1075 domain
| PWAKEUP_MYCPU
);
1079 if (CPUMASK_TESTNZERO(mask
)) {
1080 lwkt_send_ipiq2_mask(mask
, _wakeup
, ident
,
1081 domain
| PWAKEUP_MYCPU
);
1086 logtsleep1(wakeup_end
);
1091 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
1094 wakeup(const volatile void *ident
)
1096 globaldata_t gd
= mycpu
;
1097 thread_t td
= gd
->gd_curthread
;
1099 if (td
&& (td
->td_flags
& TDF_DELAYED_WAKEUP
)) {
1101 * If we are in a delayed wakeup section, record up to two wakeups in
1102 * a per-CPU queue and issue them when we block or exit the delayed
1105 if (atomic_cmpset_ptr(&gd
->gd_delayed_wakeup
[0], NULL
, ident
))
1107 if (atomic_cmpset_ptr(&gd
->gd_delayed_wakeup
[1], NULL
, ident
))
1110 ident
= atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd
->gd_delayed_wakeup
[1]),
1112 ident
= atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd
->gd_delayed_wakeup
[0]),
1116 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, gd
->gd_cpuid
));
1120 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
1123 wakeup_one(const volatile void *ident
)
1125 /* XXX potentially round-robin the first responding cpu */
1126 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1131 * Wakeup threads tsleep()ing on the specified ident on the current cpu
1135 wakeup_mycpu(const volatile void *ident
)
1137 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1142 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
1146 wakeup_mycpu_one(const volatile void *ident
)
1148 /* XXX potentially round-robin the first responding cpu */
1149 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1150 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1154 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1158 wakeup_oncpu(globaldata_t gd
, const volatile void *ident
)
1160 globaldata_t mygd
= mycpu
;
1162 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1165 lwkt_send_ipiq2(gd
, _wakeup
, __DEALL(ident
),
1166 PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1172 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1176 wakeup_oncpu_one(globaldata_t gd
, const volatile void *ident
)
1178 globaldata_t mygd
= mycpu
;
1180 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1181 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1183 lwkt_send_ipiq2(gd
, _wakeup
, __DEALL(ident
),
1184 PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1185 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1190 * Wakeup all threads waiting on the specified ident that slept using
1191 * the specified domain, on all cpus.
1194 wakeup_domain(const volatile void *ident
, int domain
)
1196 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
));
1200 * Wakeup one thread waiting on the specified ident that slept using
1201 * the specified domain, on any cpu.
1204 wakeup_domain_one(const volatile void *ident
, int domain
)
1206 /* XXX potentially round-robin the first responding cpu */
1207 _wakeup(__DEALL(ident
),
1208 PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
) | PWAKEUP_ONE
);
1212 wakeup_start_delayed(void)
1214 globaldata_t gd
= mycpu
;
1217 gd
->gd_curthread
->td_flags
|= TDF_DELAYED_WAKEUP
;
1222 wakeup_end_delayed(void)
1224 globaldata_t gd
= mycpu
;
1226 if (gd
->gd_curthread
->td_flags
& TDF_DELAYED_WAKEUP
) {
1228 gd
->gd_curthread
->td_flags
&= ~TDF_DELAYED_WAKEUP
;
1229 if (gd
->gd_delayed_wakeup
[0] || gd
->gd_delayed_wakeup
[1]) {
1230 if (gd
->gd_delayed_wakeup
[0]) {
1231 wakeup(gd
->gd_delayed_wakeup
[0]);
1232 gd
->gd_delayed_wakeup
[0] = NULL
;
1234 if (gd
->gd_delayed_wakeup
[1]) {
1235 wakeup(gd
->gd_delayed_wakeup
[1]);
1236 gd
->gd_delayed_wakeup
[1] = NULL
;
1246 * Make a process runnable. lp->lwp_token must be held on call and this
1247 * function must be called from the cpu owning lp.
1249 * This only has an effect if we are in LSSTOP or LSSLEEP.
1252 setrunnable(struct lwp
*lp
)
1254 thread_t td
= lp
->lwp_thread
;
1256 ASSERT_LWKT_TOKEN_HELD(&lp
->lwp_token
);
1257 KKASSERT(td
->td_gd
== mycpu
);
1259 if (lp
->lwp_stat
== LSSTOP
)
1260 lp
->lwp_stat
= LSSLEEP
;
1261 if (lp
->lwp_stat
== LSSLEEP
) {
1264 } else if (td
->td_flags
& TDF_SINTR
) {
1271 * The process is stopped due to some condition, usually because p_stat is
1272 * set to SSTOP, but also possibly due to being traced.
1274 * Caller must hold p->p_token
1276 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1277 * because the parent may check the child's status before the child actually
1278 * gets to this routine.
1280 * This routine is called with the current lwp only, typically just
1281 * before returning to userland if the process state is detected as
1282 * possibly being in a stopped state.
1287 struct lwp
*lp
= curthread
->td_lwp
;
1288 struct proc
*p
= lp
->lwp_proc
;
1291 lwkt_gettoken(&lp
->lwp_token
);
1295 * If LWP_MP_WSTOP is set, we were sleeping
1296 * while our process was stopped. At this point
1297 * we were already counted as stopped.
1299 if ((lp
->lwp_mpflags
& LWP_MP_WSTOP
) == 0) {
1301 * If we're the last thread to stop, signal
1305 atomic_set_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
1306 wakeup(&p
->p_nstopped
);
1307 if (p
->p_nstopped
== p
->p_nthreads
) {
1309 * Token required to interlock kern_wait()
1313 lwkt_gettoken(&q
->p_token
);
1314 p
->p_flags
&= ~P_WAITED
;
1316 if ((q
->p_sigacts
->ps_flag
& PS_NOCLDSTOP
) == 0)
1317 ksignal(q
, SIGCHLD
);
1318 lwkt_reltoken(&q
->p_token
);
1324 * Wait here while in a stopped state, interlocked with lwp_token.
1325 * We must break-out if the whole process is trying to exit.
1327 while (STOPLWP(p
, lp
)) {
1328 lp
->lwp_stat
= LSSTOP
;
1329 tsleep(p
, 0, "stop", 0);
1332 atomic_clear_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
1334 lwkt_reltoken(&lp
->lwp_token
);
1338 * Compute a tenex style load average of a quantity on
1339 * 1, 5 and 15 minute intervals. This is a pcpu callout.
1341 * We segment the lwp scan on a pcpu basis. This does NOT
1342 * mean the associated lwps are on this cpu, it is done
1343 * just to break the work up.
1345 * The callout on cpu0 rolls up the stats from the other
1348 static int loadav_count_runnable(struct lwp
*p
, void *data
);
1353 globaldata_t gd
= mycpu
;
1354 struct loadavg
*avg
;
1358 alllwp_scan(loadav_count_runnable
, &nrun
, 1);
1359 gd
->gd_loadav_nrunnable
= nrun
;
1360 if (gd
->gd_cpuid
== 0) {
1363 for (i
= 0; i
< ncpus
; ++i
)
1364 nrun
+= globaldata_find(i
)->gd_loadav_nrunnable
;
1365 for (i
= 0; i
< 3; i
++) {
1366 avg
->ldavg
[i
] = (cexp
[i
] * avg
->ldavg
[i
] +
1367 (long)nrun
* FSCALE
* (FSCALE
- cexp
[i
])) >> FSHIFT
;
1372 * Schedule the next update to occur after 5 seconds, but add a
1373 * random variation to avoid synchronisation with processes that
1374 * run at regular intervals.
1376 callout_reset(&gd
->gd_loadav_callout
,
1377 hz
* 4 + (int)(krandom() % (hz
* 2 + 1)),
1382 loadav_count_runnable(struct lwp
*lp
, void *data
)
1387 switch (lp
->lwp_stat
) {
1389 if ((td
= lp
->lwp_thread
) == NULL
)
1391 if (td
->td_flags
& TDF_BLOCKED
)
1403 * Regular data collection
1406 collect_load_callback(int n
)
1408 int fscale
= averunnable
.fscale
;
1410 return ((averunnable
.ldavg
[0] * 100 + (fscale
>> 1)) / fscale
);
1414 sched_setup(void *dummy __unused
)
1416 globaldata_t save_gd
= mycpu
;
1420 kcollect_register(KCOLLECT_LOAD
, "load", collect_load_callback
,
1421 KCOLLECT_SCALE(KCOLLECT_LOAD_FORMAT
, 0));
1424 * Kick off timeout driven events by calling first time. We
1425 * split the work across available cpus to help scale it,
1426 * it can eat a lot of cpu when there are a lot of processes
1429 for (n
= 0; n
< ncpus
; ++n
) {
1430 gd
= globaldata_find(n
);
1431 lwkt_setcpu_self(gd
);
1432 callout_init_mp(&gd
->gd_loadav_callout
);
1433 callout_init_mp(&gd
->gd_schedcpu_callout
);
1437 lwkt_setcpu_self(save_gd
);
1441 * Extremely early initialization, dummy-up the tables so we don't have
1442 * to conditionalize for NULL in _wakeup() and tsleep_interlock(). Even
1443 * though the system isn't blocking this early, these functions still
1444 * try to access the hash table.
1446 * This setup will be overridden once sched_dyninit() -> sleep_gdinit()
1450 sleep_early_gdinit(globaldata_t gd
)
1452 static struct tslpque dummy_slpque
;
1453 static cpumask_t dummy_cpumasks
;
1455 slpque_tablesize
= 1;
1456 gd
->gd_tsleep_hash
= &dummy_slpque
;
1457 slpque_cpumasks
= &dummy_cpumasks
;
1458 TAILQ_INIT(&dummy_slpque
.queue
);
1462 * PCPU initialization. Called after KMALLOC is operational, by
1463 * sched_dyninit() for cpu 0, and by mi_gdinit() for other cpus later.
1465 * WARNING! The pcpu hash table is smaller than the global cpumask
1466 * hash table, which can save us a lot of memory when maxproc
1470 sleep_gdinit(globaldata_t gd
)
1478 * This shouldn't happen, that is there shouldn't be any threads
1479 * waiting on the dummy tsleep queue this early in the boot.
1481 if (gd
->gd_cpuid
== 0) {
1482 struct tslpque
*qp
= &gd
->gd_tsleep_hash
[0];
1483 TAILQ_FOREACH(td
, &qp
->queue
, td_sleepq
) {
1484 kprintf("SLEEP_GDINIT SWITCH %s\n", td
->td_comm
);
1489 * Note that we have to allocate one extra slot because we are
1490 * shifting a modulo value. TCHASHSHIFT(slpque_tablesize - 1) can
1491 * return the same value as TCHASHSHIFT(slpque_tablesize).
1493 n
= TCHASHSHIFT(slpque_tablesize
) + 1;
1495 hash_size
= sizeof(struct tslpque
) * n
;
1496 gd
->gd_tsleep_hash
= (void *)kmem_alloc3(&kernel_map
, hash_size
,
1498 KM_CPU(gd
->gd_cpuid
));
1499 memset(gd
->gd_tsleep_hash
, 0, hash_size
);
1500 for (i
= 0; i
< n
; ++i
)
1501 TAILQ_INIT(&gd
->gd_tsleep_hash
[i
].queue
);
1505 * Dynamic initialization after the memory system is operational.
1508 sched_dyninit(void *dummy __unused
)
1515 * Calculate table size for slpque hash. We want a prime number
1516 * large enough to avoid overloading slpque_cpumasks when the
1517 * system has a large number of sleeping processes, which will
1518 * spam IPIs on wakeup().
1520 * While it is true this is really a per-lwp factor, generally
1521 * speaking the maxproc limit is a good metric to go by.
1523 for (tblsize
= maxproc
| 1; ; tblsize
+= 2) {
1524 if (tblsize
% 3 == 0)
1526 if (tblsize
% 5 == 0)
1528 tblsize2
= (tblsize
/ 2) | 1;
1529 for (n
= 7; n
< tblsize2
; n
+= 2) {
1530 if (tblsize
% n
== 0)
1538 * PIDs are currently limited to 6 digits. Cap the table size
1541 if (tblsize
> 2000003)
1544 slpque_tablesize
= tblsize
;
1545 slpque_cpumasks
= kmalloc(sizeof(*slpque_cpumasks
) * slpque_tablesize
,
1546 M_TSLEEP
, M_WAITOK
| M_ZERO
);
1547 sleep_gdinit(mycpu
);