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>
50 #include <sys/kcollect.h>
51 #include <sys/malloc.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
);
83 __read_mostly
int tsleep_crypto_dump
= 0;
84 __read_mostly
int ncpus
;
85 __read_mostly
int ncpus_fit
, ncpus_fit_mask
; /* note: mask not cpumask_t */
86 __read_mostly
int safepri
;
87 __read_mostly
int tsleep_now_works
;
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 __exclusive_cache_line
107 struct loadavg averunnable
=
108 { {0, 0, 0}, FSCALE
}; /* load average, of runnable procs */
110 * Constants for averages over 1, 5, and 15 minutes
111 * when sampling at 5 second intervals.
114 static fixpt_t cexp
[3] = {
115 0.9200444146293232 * FSCALE
, /* exp(-1/12) */
116 0.9834714538216174 * FSCALE
, /* exp(-1/60) */
117 0.9944598480048967 * FSCALE
, /* exp(-1/180) */
120 static void endtsleep (void *);
121 static void loadav (void *arg
);
122 static void schedcpu (void *arg
);
124 __read_mostly
static int pctcpu_decay
= 10;
125 SYSCTL_INT(_kern
, OID_AUTO
, pctcpu_decay
, CTLFLAG_RW
,
126 &pctcpu_decay
, 0, "");
129 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
131 __read_mostly
int fscale __unused
= FSCALE
; /* exported to systat */
132 SYSCTL_INT(_kern
, OID_AUTO
, fscale
, CTLFLAG_RD
, 0, FSCALE
, "");
135 * Issue a wakeup() from userland (debugging)
138 sysctl_wakeup(SYSCTL_HANDLER_ARGS
)
143 if (req
->newptr
!= NULL
) {
144 if (caps_priv_check_self(SYSCAP_RESTRICTEDROOT
))
146 error
= SYSCTL_IN(req
, &ident
, sizeof(ident
));
149 kprintf("issue wakeup %016jx\n", ident
);
150 wakeup((void *)(intptr_t)ident
);
152 if (req
->oldptr
!= NULL
) {
153 error
= SYSCTL_OUT(req
, &ident
, sizeof(ident
));
159 sysctl_wakeup_umtx(SYSCTL_HANDLER_ARGS
)
164 if (req
->newptr
!= NULL
) {
165 if (caps_priv_check_self(SYSCAP_RESTRICTEDROOT
))
167 error
= SYSCTL_IN(req
, &ident
, sizeof(ident
));
170 kprintf("issue wakeup %016jx, PDOMAIN_UMTX\n", ident
);
171 wakeup_domain((void *)(intptr_t)ident
, PDOMAIN_UMTX
);
173 if (req
->oldptr
!= NULL
) {
174 error
= SYSCTL_OUT(req
, &ident
, sizeof(ident
));
179 SYSCTL_PROC(_debug
, OID_AUTO
, wakeup
, CTLTYPE_UQUAD
|CTLFLAG_RW
, 0, 0,
180 sysctl_wakeup
, "Q", "issue wakeup(addr)");
181 SYSCTL_PROC(_debug
, OID_AUTO
, wakeup_umtx
, CTLTYPE_UQUAD
|CTLFLAG_RW
, 0, 0,
182 sysctl_wakeup_umtx
, "Q", "issue wakeup(addr, PDOMAIN_UMTX)");
185 * Recompute process priorities, once a second.
187 * Since the userland schedulers are typically event oriented, if the
188 * estcpu calculation at wakeup() time is not sufficient to make a
189 * process runnable relative to other processes in the system we have
190 * a 1-second recalc to help out.
192 * This code also allows us to store sysclock_t data in the process structure
193 * without fear of an overrun, since sysclock_t are guarenteed to hold
194 * several seconds worth of count.
196 * WARNING! callouts can preempt normal threads. However, they will not
197 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
199 static int schedcpu_stats(struct proc
*p
, void *data __unused
);
200 static int schedcpu_resource(struct proc
*p
, void *data __unused
);
205 allproc_scan(schedcpu_stats
, NULL
, 1);
206 allproc_scan(schedcpu_resource
, NULL
, 1);
207 if (mycpu
->gd_cpuid
== 0) {
208 wakeup((caddr_t
)&lbolt
);
209 wakeup(lbolt_syncer
);
211 callout_reset(&mycpu
->gd_schedcpu_callout
, hz
, schedcpu
, NULL
);
215 * General process statistics once a second
218 schedcpu_stats(struct proc
*p
, void *data __unused
)
223 * Threads may not be completely set up if process in SIDL state.
225 if (p
->p_stat
== SIDL
)
229 if (lwkt_trytoken(&p
->p_token
) == FALSE
) {
235 FOREACH_LWP_IN_PROC(lp
, p
) {
236 if (lp
->lwp_stat
== LSSLEEP
) {
238 if (lp
->lwp_slptime
== 1)
239 p
->p_usched
->uload_update(lp
);
243 * Only recalculate processes that are active or have slept
244 * less then 2 seconds. The schedulers understand this.
245 * Otherwise decay by 50% per second.
247 * NOTE: uload_update is called separately from kern_synch.c
248 * when slptime == 1, removing the thread's
251 if (lp
->lwp_slptime
<= 1) {
252 p
->p_usched
->recalculate(lp
);
256 decay
= pctcpu_decay
;
262 lp
->lwp_pctcpu
= (lp
->lwp_pctcpu
* (decay
- 1)) / decay
;
265 lwkt_reltoken(&p
->p_token
);
272 * Resource checks. XXX break out since ksignal/killproc can block,
273 * limiting us to one process killed per second. There is probably
277 schedcpu_resource(struct proc
*p
, void *data __unused
)
282 if (p
->p_stat
== SIDL
)
286 if (lwkt_trytoken(&p
->p_token
) == FALSE
) {
291 if (p
->p_stat
== SZOMB
|| p
->p_limit
== NULL
) {
292 lwkt_reltoken(&p
->p_token
);
298 FOREACH_LWP_IN_PROC(lp
, p
) {
300 * We may have caught an lp in the middle of being
301 * created, lwp_thread can be NULL.
303 if (lp
->lwp_thread
) {
304 ttime
+= lp
->lwp_thread
->td_sticks
;
305 ttime
+= lp
->lwp_thread
->td_uticks
;
309 switch(plimit_testcpulimit(p
, ttime
)) {
310 case PLIMIT_TESTCPU_KILL
:
311 killproc(p
, "exceeded maximum CPU limit");
313 case PLIMIT_TESTCPU_XCPU
:
314 if ((p
->p_flags
& P_XCPU
) == 0) {
315 p
->p_flags
|= P_XCPU
;
322 lwkt_reltoken(&p
->p_token
);
329 * This is only used by ps. Generate a cpu percentage use over
330 * a period of one second.
333 updatepcpu(struct lwp
*lp
, int cpticks
, int ttlticks
)
338 acc
= (cpticks
<< FSHIFT
) / ttlticks
;
339 if (ttlticks
>= ESTCPUFREQ
) {
340 lp
->lwp_pctcpu
= acc
;
342 remticks
= ESTCPUFREQ
- ttlticks
;
343 lp
->lwp_pctcpu
= (acc
* ttlticks
+ lp
->lwp_pctcpu
* remticks
) /
349 * Handy macros to calculate hash indices. LOOKUP() calculates the
350 * global cpumask hash index, TCHASHSHIFT() converts that into the
353 * By making the pcpu hash arrays smaller we save a significant amount
354 * of memory at very low cost. The real cost is in IPIs, which are handled
355 * by the much larger global cpumask hash table.
357 #define LOOKUP_PRIME 66555444443333333ULL
358 #define LOOKUP(x) ((((uintptr_t)(x) + ((uintptr_t)(x) >> 18)) ^ \
359 LOOKUP_PRIME) % slpque_tablesize)
360 #define TCHASHSHIFT(x) ((x) >> 4)
362 __read_mostly
static uint32_t slpque_tablesize
;
363 __read_mostly
static cpumask_t
*slpque_cpumasks
;
365 SYSCTL_UINT(_kern
, OID_AUTO
, slpque_tablesize
, CTLFLAG_RD
, &slpque_tablesize
,
369 * This is a dandy function that allows us to interlock tsleep/wakeup
370 * operations with unspecified upper level locks, such as lockmgr locks,
371 * simply by holding a critical section. The sequence is:
373 * (acquire upper level lock)
374 * tsleep_interlock(blah)
375 * (release upper level lock)
378 * Basically this functions queues us on the tsleep queue without actually
379 * descheduling us. When tsleep() is later called with PINTERLOCK it
380 * assumes the thread was already queued, otherwise it queues it there.
382 * Thus it is possible to receive the wakeup prior to going to sleep and
383 * the race conditions are covered.
386 _tsleep_interlock(globaldata_t gd
, const volatile void *ident
, int flags
)
388 thread_t td
= gd
->gd_curthread
;
394 kprintf("tsleep_interlock: NULL ident %s\n", td
->td_comm
);
398 crit_enter_quick(td
);
399 if (td
->td_flags
& TDF_TSLEEPQ
) {
401 * Shortcut if unchanged
403 if (td
->td_wchan
== ident
&&
404 td
->td_wdomain
== (flags
& PDOMAIN_MASK
)) {
410 * Remove current sleepq
412 cid
= LOOKUP(td
->td_wchan
);
413 gid
= TCHASHSHIFT(cid
);
414 qp
= &gd
->gd_tsleep_hash
[gid
];
415 TAILQ_REMOVE(&qp
->queue
, td
, td_sleepq
);
416 if (TAILQ_FIRST(&qp
->queue
) == NULL
) {
421 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[cid
],
425 td
->td_flags
|= TDF_TSLEEPQ
;
428 gid
= TCHASHSHIFT(cid
);
429 qp
= &gd
->gd_tsleep_hash
[gid
];
430 TAILQ_INSERT_TAIL(&qp
->queue
, td
, td_sleepq
);
431 if (qp
->ident0
!= ident
&& qp
->ident1
!= ident
&&
432 qp
->ident2
!= ident
&& qp
->ident3
!= ident
) {
433 if (qp
->ident0
== NULL
)
435 else if (qp
->ident1
== NULL
)
437 else if (qp
->ident2
== NULL
)
439 else if (qp
->ident3
== NULL
)
442 qp
->ident0
= (void *)(intptr_t)-1;
444 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks
[cid
], gd
->gd_cpuid
);
445 td
->td_wchan
= ident
;
446 td
->td_wdomain
= flags
& PDOMAIN_MASK
;
451 tsleep_interlock(const volatile void *ident
, int flags
)
453 _tsleep_interlock(mycpu
, ident
, flags
);
457 * Remove thread from sleepq. Must be called with a critical section held.
458 * The thread must not be migrating.
461 _tsleep_remove(thread_t td
)
463 globaldata_t gd
= mycpu
;
468 KKASSERT(td
->td_gd
== gd
&& IN_CRITICAL_SECT(td
));
469 KKASSERT((td
->td_flags
& TDF_MIGRATING
) == 0);
470 if (td
->td_flags
& TDF_TSLEEPQ
) {
471 td
->td_flags
&= ~TDF_TSLEEPQ
;
472 cid
= LOOKUP(td
->td_wchan
);
473 gid
= TCHASHSHIFT(cid
);
474 qp
= &gd
->gd_tsleep_hash
[gid
];
475 TAILQ_REMOVE(&qp
->queue
, td
, td_sleepq
);
476 if (TAILQ_FIRST(&qp
->queue
) == NULL
) {
477 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[cid
],
486 tsleep_remove(thread_t td
)
492 * General sleep call. Suspends the current process until a wakeup is
493 * performed on the specified identifier. The process will then be made
494 * runnable with the specified priority. Sleeps at most timo/hz seconds
495 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
496 * before and after sleeping, else signals are not checked. Returns 0 if
497 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
498 * signal needs to be delivered, ERESTART is returned if the current system
499 * call should be restarted if possible, and EINTR is returned if the system
500 * call should be interrupted by the signal (return EINTR).
502 * Note that if we are a process, we release_curproc() before messing with
503 * the LWKT scheduler.
505 * During autoconfiguration or after a panic, a sleep will simply
506 * lower the priority briefly to allow interrupts, then return.
508 * WARNING! This code can't block (short of switching away), or bad things
509 * will happen. No getting tokens, no blocking locks, etc.
512 tsleep(const volatile void *ident
, int flags
, const char *wmesg
, int timo
)
514 struct thread
*td
= curthread
;
515 struct lwp
*lp
= td
->td_lwp
;
516 struct proc
*p
= td
->td_proc
; /* may be NULL */
522 struct callout thandle1
;
523 struct _callout thandle2
;
526 * Currently a severe hack. Make sure any delayed wakeups
527 * are flushed before we sleep or we might deadlock on whatever
528 * event we are sleeping on.
530 if (td
->td_flags
& TDF_DELAYED_WAKEUP
)
531 wakeup_end_delayed();
534 * NOTE: removed KTRPOINT, it could cause races due to blocking
535 * even in stable. Just scrap it for now.
537 if (!tsleep_crypto_dump
&& (tsleep_now_works
== 0 || panicstr
)) {
539 * After a panic, or before we actually have an operational
540 * softclock, just give interrupts a chance, then just return;
542 * don't run any other procs or panic below,
543 * in case this is the idle process and already asleep.
547 lwkt_setpri_self(safepri
);
549 lwkt_setpri_self(oldpri
);
552 logtsleep2(tsleep_beg
, ident
);
554 KKASSERT(td
!= &gd
->gd_idlethread
); /* you must be kidding! */
557 * NOTE: all of this occurs on the current cpu, including any
558 * callout-based wakeups, so a critical section is a sufficient
561 * The entire sequence through to where we actually sleep must
562 * run without breaking the critical section.
564 catch = flags
& PCATCH
;
568 crit_enter_quick(td
);
570 KASSERT(ident
!= NULL
, ("tsleep: no ident"));
571 KASSERT(lp
== NULL
||
572 lp
->lwp_stat
== LSRUN
|| /* Obvious */
573 lp
->lwp_stat
== LSSTOP
, /* Set in tstop */
575 ident
, wmesg
, lp
->lwp_stat
));
578 * We interlock the sleep queue if the caller has not already done
579 * it for us. This must be done before we potentially acquire any
580 * tokens or we can loose the wakeup.
582 if ((flags
& PINTERLOCKED
) == 0) {
583 _tsleep_interlock(gd
, ident
, flags
);
587 * Setup for the current process (if this is a process). We must
588 * interlock with lwp_token to avoid remote wakeup races via
592 lwkt_gettoken(&lp
->lwp_token
);
595 * If the umbrella process is in the SCORE state then
596 * make sure that the thread is flagged going into a
597 * normal sleep to allow the core dump to proceed, otherwise
598 * the coredump can end up waiting forever. If the normal
599 * sleep is woken up, the thread will enter a stopped state
600 * upon return to userland.
602 * We do not want to interrupt or cause a thread exist at
603 * this juncture because that will mess-up the state the
604 * coredump is trying to save.
606 if (p
->p_stat
== SCORE
) {
607 lwkt_gettoken(&p
->p_token
);
608 if ((lp
->lwp_mpflags
& LWP_MP_WSTOP
) == 0) {
609 atomic_set_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
612 lwkt_reltoken(&p
->p_token
);
620 * Early termination if PCATCH was set and a
621 * signal is pending, interlocked with the
624 * Early termination only occurs when tsleep() is
625 * entered while in a normal LSRUN state.
627 if ((sig
= CURSIG(lp
)) != 0)
631 * Causes ksignal to wake us up if a signal is
632 * received (interlocked with lp->lwp_token).
634 lp
->lwp_flags
|= LWP_SINTR
;
641 * Make sure the current process has been untangled from
642 * the userland scheduler and initialize slptime to start
645 * NOTE: td->td_wakefromcpu is pre-set by the release function
646 * for the dfly scheduler, and then adjusted by _wakeup()
649 p
->p_usched
->release_curproc(lp
);
654 * For PINTERLOCKED operation, TDF_TSLEEPQ might not be set if
655 * a wakeup() was processed before the thread could go to sleep.
657 * If TDF_TSLEEPQ is set, make sure the ident matches the recorded
658 * ident. If it does not then the thread slept inbetween the
659 * caller's initial tsleep_interlock() call and the caller's tsleep()
662 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
663 * to process incoming IPIs, thus draining incoming wakeups.
665 if ((td
->td_flags
& TDF_TSLEEPQ
) == 0) {
666 logtsleep2(ilockfail
, ident
);
668 } else if (td
->td_wchan
!= ident
||
669 td
->td_wdomain
!= (flags
& PDOMAIN_MASK
)) {
670 logtsleep2(ilockfail
, ident
);
675 * scheduling is blocked while in a critical section. Coincide
676 * the descheduled-by-tsleep flag with the descheduling of the
679 * The timer callout is localized on our cpu and interlocked by
680 * our critical section.
682 lwkt_deschedule_self(td
);
683 td
->td_flags
|= TDF_TSLEEP_DESCHEDULED
;
684 td
->td_wmesg
= wmesg
;
687 * Setup the timeout, if any. The timeout is only operable while
688 * the thread is flagged descheduled.
690 KKASSERT((td
->td_flags
& TDF_TIMEOUT
) == 0);
692 _callout_setup_quick(&thandle1
, &thandle2
, timo
, endtsleep
, td
);
700 * Ok, we are sleeping. Place us in the SSLEEP state.
702 KKASSERT((lp
->lwp_mpflags
& LWP_MP_ONRUNQ
) == 0);
705 * tstop() sets LSSTOP, so don't fiddle with that.
707 if (lp
->lwp_stat
!= LSSTOP
)
708 lp
->lwp_stat
= LSSLEEP
;
709 lp
->lwp_ru
.ru_nvcsw
++;
710 p
->p_usched
->uload_update(lp
);
714 * And when we are woken up, put us back in LSRUN. If we
715 * slept for over a second, recalculate our estcpu.
717 lp
->lwp_stat
= LSRUN
;
718 if (lp
->lwp_slptime
) {
719 p
->p_usched
->uload_update(lp
);
720 p
->p_usched
->recalculate(lp
);
728 * Make sure we haven't switched cpus while we were asleep. It's
729 * not supposed to happen. Cleanup our temporary flags.
731 KKASSERT(gd
== td
->td_gd
);
734 * Cleanup the timeout. If the timeout has already occured thandle
735 * has already been stopped, otherwise stop thandle.
737 * If the timeout is still running the callout thread must be blocked
738 * trying to get lwp_token, or this is a VM where cpu-cpu races are
739 * common, then wait for us to get scheduled.
742 while (td
->td_flags
& TDF_TIMEOUT_RUNNING
) {
743 /* else we won't get rescheduled! */
744 if (lp
->lwp_stat
!= LSSTOP
)
745 lp
->lwp_stat
= LSSLEEP
;
746 lwkt_deschedule_self(td
);
747 td
->td_wmesg
= "tsrace";
750 if (td
->td_flags
& TDF_TIMEOUT
) {
751 td
->td_flags
&= ~TDF_TIMEOUT
;
755 * We are on the same cpu so use the quick version
756 * which is guaranteed not to block or race.
758 _callout_cancel_quick(&thandle2
);
761 td
->td_flags
&= ~TDF_TSLEEP_DESCHEDULED
;
764 * Make sure we have been removed from the sleepq. In most
765 * cases this will have been done for us already but it is
766 * possible for a scheduling IPI to be in-flight from a
767 * previous tsleep/tsleep_interlock() or due to a straight-out
768 * call to lwkt_schedule() (in the case of an interrupt thread),
769 * causing a spurious wakeup.
775 * Figure out the correct error return. If interrupted by a
776 * signal we want to return EINTR or ERESTART.
780 if (catch && error
== 0) {
781 if (sig
!= 0 || (sig
= CURSIG(lp
))) {
782 if (SIGISMEMBER(p
->p_sigacts
->ps_sigintr
, sig
))
789 lp
->lwp_flags
&= ~LWP_SINTR
;
792 * Unconditionally set us to LSRUN on resume. lwp_stat could
793 * be in a weird state due to the goto resume, particularly
794 * when tsleep() is called from tstop().
796 lp
->lwp_stat
= LSRUN
;
797 lwkt_reltoken(&lp
->lwp_token
);
799 logtsleep1(tsleep_end
);
806 * Interlocked spinlock sleep. An exclusively held spinlock must
807 * be passed to ssleep(). The function will atomically release the
808 * spinlock and tsleep on the ident, then reacquire the spinlock and
811 * This routine is fairly important along the critical path, so optimize it
815 ssleep(const volatile void *ident
, struct spinlock
*spin
, int flags
,
816 const char *wmesg
, int timo
)
818 globaldata_t gd
= mycpu
;
821 _tsleep_interlock(gd
, ident
, flags
);
822 spin_unlock_quick(gd
, spin
);
823 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
824 KKASSERT(gd
== mycpu
);
825 _spin_lock_quick(gd
, spin
, wmesg
);
831 lksleep(const volatile void *ident
, struct lock
*lock
, int flags
,
832 const char *wmesg
, int timo
)
834 globaldata_t gd
= mycpu
;
837 _tsleep_interlock(gd
, ident
, flags
);
838 lockmgr(lock
, LK_RELEASE
);
839 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
840 lockmgr(lock
, LK_EXCLUSIVE
);
846 * Interlocked mutex sleep. An exclusively held mutex must be passed
847 * to mtxsleep(). The function will atomically release the mutex
848 * and tsleep on the ident, then reacquire the mutex and return.
851 mtxsleep(const volatile void *ident
, struct mtx
*mtx
, int flags
,
852 const char *wmesg
, int timo
)
854 globaldata_t gd
= mycpu
;
857 _tsleep_interlock(gd
, ident
, flags
);
859 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
860 mtx_lock_ex_quick(mtx
);
866 * Interlocked serializer sleep. An exclusively held serializer must
867 * be passed to zsleep(). The function will atomically release
868 * the serializer and tsleep on the ident, then reacquire the serializer
872 zsleep(const volatile void *ident
, struct lwkt_serialize
*slz
, int flags
,
873 const char *wmesg
, int timo
)
875 globaldata_t gd
= mycpu
;
878 ASSERT_SERIALIZED(slz
);
880 _tsleep_interlock(gd
, ident
, flags
);
881 lwkt_serialize_exit(slz
);
882 ret
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
883 lwkt_serialize_enter(slz
);
889 * Directly block on the LWKT thread by descheduling it. This
890 * is much faster then tsleep(), but the only legal way to wake
891 * us up is to directly schedule the thread.
893 * Setting TDF_SINTR will cause new signals to directly schedule us.
895 * This routine must be called while in a critical section.
898 lwkt_sleep(const char *wmesg
, int flags
)
900 thread_t td
= curthread
;
903 if ((flags
& PCATCH
) == 0 || td
->td_lwp
== NULL
) {
904 td
->td_flags
|= TDF_BLOCKED
;
905 td
->td_wmesg
= wmesg
;
906 lwkt_deschedule_self(td
);
909 td
->td_flags
&= ~TDF_BLOCKED
;
912 if ((sig
= CURSIG(td
->td_lwp
)) != 0) {
913 if (SIGISMEMBER(td
->td_proc
->p_sigacts
->ps_sigintr
, sig
))
919 td
->td_flags
|= TDF_BLOCKED
| TDF_SINTR
;
920 td
->td_wmesg
= wmesg
;
921 lwkt_deschedule_self(td
);
923 td
->td_flags
&= ~(TDF_BLOCKED
| TDF_SINTR
);
929 * Implement the timeout for tsleep.
931 * This type of callout timeout is scheduled on the same cpu the process
932 * is sleeping on. Also, at the moment, the MP lock is held.
941 * We are going to have to get the lwp_token, which means we might
942 * block. This can race a tsleep getting woken up by other means
943 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
944 * processing to complete (sorry tsleep!).
946 * We can safely set td_flags because td MUST be on the same cpu
949 KKASSERT(td
->td_gd
== mycpu
);
951 td
->td_flags
|= TDF_TIMEOUT_RUNNING
| TDF_TIMEOUT
;
954 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
955 * from exiting the tsleep on us. The flag is interlocked by virtue
956 * of lp being on the same cpu as we are.
958 if ((lp
= td
->td_lwp
) != NULL
)
959 lwkt_gettoken(&lp
->lwp_token
);
961 KKASSERT(td
->td_flags
& TDF_TSLEEP_DESCHEDULED
);
965 * callout timer should normally never be set in tstop()
966 * because it passes a timeout of 0. However, there is a
967 * case during thread exit (which SSTOP's all the threads)
968 * for which tstop() must break out and can (properly) leave
969 * the thread in LSSTOP.
971 KKASSERT(lp
->lwp_stat
!= LSSTOP
||
972 (lp
->lwp_mpflags
& LWP_MP_WEXIT
));
974 lwkt_reltoken(&lp
->lwp_token
);
979 KKASSERT(td
->td_gd
== mycpu
);
980 td
->td_flags
&= ~TDF_TIMEOUT_RUNNING
;
985 * Make all processes sleeping on the specified identifier runnable.
986 * count may be zero or one only.
988 * The domain encodes the sleep/wakeup domain, flags, plus the originating
991 * This call may run without the MP lock held. We can only manipulate thread
992 * state on the cpu owning the thread. We CANNOT manipulate process state
995 * _wakeup() can be passed to an IPI so we can't use (const volatile
999 _wakeup(void *ident
, int domain
)
1011 logtsleep2(wakeup_beg
, ident
);
1013 cid
= LOOKUP(ident
);
1014 gid
= TCHASHSHIFT(cid
);
1015 qp
= &gd
->gd_tsleep_hash
[gid
];
1017 for (td
= TAILQ_FIRST(&qp
->queue
); td
!= NULL
; td
= ntd
) {
1018 ntd
= TAILQ_NEXT(td
, td_sleepq
);
1019 if (td
->td_wchan
== ident
&&
1020 td
->td_wdomain
== (domain
& PDOMAIN_MASK
)
1022 KKASSERT(td
->td_gd
== gd
);
1024 td
->td_wakefromcpu
= PWAKEUP_DECODE(domain
);
1025 if (td
->td_flags
& TDF_TSLEEP_DESCHEDULED
) {
1027 if (domain
& PWAKEUP_ONE
)
1032 if (td
->td_wchan
== qp
->ident0
)
1034 else if (td
->td_wchan
== qp
->ident1
)
1036 else if (td
->td_wchan
== qp
->ident2
)
1038 else if (td
->td_wchan
== qp
->ident3
)
1041 wids
|= 16; /* force ident0 to be retained (-1) */
1045 * Because a bunch of cpumask array entries cover the same queue, it
1046 * is possible for our bit to remain set in some of them and cause
1047 * spurious wakeup IPIs later on. Make sure that the bit is cleared
1048 * when a spurious IPI occurs to prevent further spurious IPIs.
1050 if (TAILQ_FIRST(&qp
->queue
) == NULL
) {
1051 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[cid
], gd
->gd_cpuid
);
1057 if ((wids
& 1) == 0) {
1058 if ((wids
& 16) == 0) {
1061 KKASSERT(qp
->ident0
== (void *)(intptr_t)-1);
1064 if ((wids
& 2) == 0)
1066 if ((wids
& 4) == 0)
1068 if ((wids
& 8) == 0)
1073 * We finished checking the current cpu but there still may be
1074 * more work to do. Either wakeup_one was requested and no matching
1075 * thread was found, or a normal wakeup was requested and we have
1076 * to continue checking cpus.
1078 * It should be noted that this scheme is actually less expensive then
1079 * the old scheme when waking up multiple threads, since we send
1080 * only one IPI message per target candidate which may then schedule
1081 * multiple threads. Before we could have wound up sending an IPI
1082 * message for each thread on the target cpu (!= current cpu) that
1083 * needed to be woken up.
1085 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
1086 * should be ok since we are passing idents in the IPI rather
1087 * then thread pointers.
1089 * NOTE: We MUST mfence (or use an atomic op) prior to reading
1090 * the cpumask, as another cpu may have written to it in
1091 * a fashion interlocked with whatever the caller did before
1092 * calling wakeup(). Otherwise we might miss the interaction
1093 * (kern_mutex.c can cause this problem).
1095 * lfence is insufficient as it may allow a written state to
1096 * reorder around the cpumask load.
1098 if ((domain
& PWAKEUP_MYCPU
) == 0) {
1100 const volatile void *id0
;
1105 mask
= slpque_cpumasks
[cid
];
1106 CPUMASK_ANDMASK(mask
, gd
->gd_other_cpus
);
1107 while (CPUMASK_TESTNZERO(mask
)) {
1108 n
= BSRCPUMASK(mask
);
1109 CPUMASK_NANDBIT(mask
, n
);
1110 tgd
= globaldata_find(n
);
1113 * Both ident0 compares must from a single load
1114 * to avoid ident0 update races crossing the two
1117 qp
= &tgd
->gd_tsleep_hash
[gid
];
1120 if (id0
== (void *)(intptr_t)-1) {
1121 lwkt_send_ipiq2(tgd
, _wakeup
, ident
,
1122 domain
| PWAKEUP_MYCPU
);
1123 ++tgd
->gd_cnt
.v_wakeup_colls
;
1124 } else if (id0
== ident
||
1125 qp
->ident1
== ident
||
1126 qp
->ident2
== ident
||
1127 qp
->ident3
== ident
) {
1128 lwkt_send_ipiq2(tgd
, _wakeup
, ident
,
1129 domain
| PWAKEUP_MYCPU
);
1133 if (CPUMASK_TESTNZERO(mask
)) {
1134 lwkt_send_ipiq2_mask(mask
, _wakeup
, ident
,
1135 domain
| PWAKEUP_MYCPU
);
1140 logtsleep1(wakeup_end
);
1145 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
1148 wakeup(const volatile void *ident
)
1150 globaldata_t gd
= mycpu
;
1151 thread_t td
= gd
->gd_curthread
;
1153 if (td
&& (td
->td_flags
& TDF_DELAYED_WAKEUP
)) {
1155 * If we are in a delayed wakeup section, record up to two wakeups in
1156 * a per-CPU queue and issue them when we block or exit the delayed
1159 if (atomic_cmpset_ptr(&gd
->gd_delayed_wakeup
[0], NULL
, ident
))
1161 if (atomic_cmpset_ptr(&gd
->gd_delayed_wakeup
[1], NULL
, ident
))
1164 ident
= atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd
->gd_delayed_wakeup
[1]),
1166 ident
= atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd
->gd_delayed_wakeup
[0]),
1170 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, gd
->gd_cpuid
));
1174 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
1177 wakeup_one(const volatile void *ident
)
1179 /* XXX potentially round-robin the first responding cpu */
1180 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1185 * Wakeup threads tsleep()ing on the specified ident on the current cpu
1189 wakeup_mycpu(const volatile void *ident
)
1191 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1196 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
1200 wakeup_mycpu_one(const volatile void *ident
)
1202 /* XXX potentially round-robin the first responding cpu */
1203 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1204 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1208 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1212 wakeup_oncpu(globaldata_t gd
, const volatile void *ident
)
1214 globaldata_t mygd
= mycpu
;
1216 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1219 lwkt_send_ipiq2(gd
, _wakeup
, __DEALL(ident
),
1220 PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1226 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1230 wakeup_oncpu_one(globaldata_t gd
, const volatile void *ident
)
1232 globaldata_t mygd
= mycpu
;
1234 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1235 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1237 lwkt_send_ipiq2(gd
, _wakeup
, __DEALL(ident
),
1238 PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1239 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1244 * Wakeup all threads waiting on the specified ident that slept using
1245 * the specified domain, on all cpus.
1248 wakeup_domain(const volatile void *ident
, int domain
)
1250 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
));
1254 * Wakeup one thread waiting on the specified ident that slept using
1255 * the specified domain, on any cpu.
1258 wakeup_domain_one(const volatile void *ident
, int domain
)
1260 /* XXX potentially round-robin the first responding cpu */
1261 _wakeup(__DEALL(ident
),
1262 PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
) | PWAKEUP_ONE
);
1266 wakeup_start_delayed(void)
1268 globaldata_t gd
= mycpu
;
1271 gd
->gd_curthread
->td_flags
|= TDF_DELAYED_WAKEUP
;
1276 wakeup_end_delayed(void)
1278 globaldata_t gd
= mycpu
;
1280 if (gd
->gd_curthread
->td_flags
& TDF_DELAYED_WAKEUP
) {
1282 gd
->gd_curthread
->td_flags
&= ~TDF_DELAYED_WAKEUP
;
1283 if (gd
->gd_delayed_wakeup
[0] || gd
->gd_delayed_wakeup
[1]) {
1284 if (gd
->gd_delayed_wakeup
[0]) {
1285 wakeup(gd
->gd_delayed_wakeup
[0]);
1286 gd
->gd_delayed_wakeup
[0] = NULL
;
1288 if (gd
->gd_delayed_wakeup
[1]) {
1289 wakeup(gd
->gd_delayed_wakeup
[1]);
1290 gd
->gd_delayed_wakeup
[1] = NULL
;
1300 * Make a process runnable. lp->lwp_token must be held on call and this
1301 * function must be called from the cpu owning lp.
1303 * This only has an effect if we are in LSSTOP or LSSLEEP.
1306 setrunnable(struct lwp
*lp
)
1308 thread_t td
= lp
->lwp_thread
;
1310 ASSERT_LWKT_TOKEN_HELD(&lp
->lwp_token
);
1311 KKASSERT(td
->td_gd
== mycpu
);
1313 if (lp
->lwp_stat
== LSSTOP
)
1314 lp
->lwp_stat
= LSSLEEP
;
1315 if (lp
->lwp_stat
== LSSLEEP
) {
1318 } else if (td
->td_flags
& TDF_SINTR
) {
1325 * The process is stopped due to some condition, usually because p_stat is
1326 * set to SSTOP, but also possibly due to being traced.
1328 * Caller must hold p->p_token
1330 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1331 * because the parent may check the child's status before the child actually
1332 * gets to this routine.
1334 * This routine is called with the current lwp only, typically just
1335 * before returning to userland if the process state is detected as
1336 * possibly being in a stopped state.
1341 struct lwp
*lp
= curthread
->td_lwp
;
1342 struct proc
*p
= lp
->lwp_proc
;
1345 lwkt_gettoken(&lp
->lwp_token
);
1349 * If LWP_MP_WSTOP is set, we were sleeping
1350 * while our process was stopped. At this point
1351 * we were already counted as stopped.
1353 if ((lp
->lwp_mpflags
& LWP_MP_WSTOP
) == 0) {
1355 * If we're the last thread to stop, signal
1359 atomic_set_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
1360 wakeup(&p
->p_nstopped
);
1361 if (p
->p_nstopped
== p
->p_nthreads
) {
1363 * Token required to interlock kern_wait()
1367 lwkt_gettoken(&q
->p_token
);
1368 p
->p_flags
&= ~P_WAITED
;
1370 if ((q
->p_sigacts
->ps_flag
& PS_NOCLDSTOP
) == 0)
1371 ksignal(q
, SIGCHLD
);
1372 lwkt_reltoken(&q
->p_token
);
1378 * Wait here while in a stopped state, interlocked with lwp_token.
1379 * We must break-out if the whole process is trying to exit.
1381 while (STOPLWP(p
, lp
)) {
1382 lp
->lwp_stat
= LSSTOP
;
1383 tsleep(p
, 0, "stop", 0);
1386 atomic_clear_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
1388 lwkt_reltoken(&lp
->lwp_token
);
1392 * Compute a tenex style load average of a quantity on
1393 * 1, 5 and 15 minute intervals. This is a pcpu callout.
1395 * We segment the lwp scan on a pcpu basis. This does NOT
1396 * mean the associated lwps are on this cpu, it is done
1397 * just to break the work up.
1399 * The callout on cpu0 rolls up the stats from the other
1402 static int loadav_count_runnable(struct lwp
*p
, void *data
);
1407 globaldata_t gd
= mycpu
;
1408 struct loadavg
*avg
;
1412 alllwp_scan(loadav_count_runnable
, &nrun
, 1);
1413 gd
->gd_loadav_nrunnable
= nrun
;
1414 if (gd
->gd_cpuid
== 0) {
1417 for (i
= 0; i
< ncpus
; ++i
)
1418 nrun
+= globaldata_find(i
)->gd_loadav_nrunnable
;
1419 for (i
= 0; i
< 3; i
++) {
1420 avg
->ldavg
[i
] = (cexp
[i
] * avg
->ldavg
[i
] +
1421 (long)nrun
* FSCALE
* (FSCALE
- cexp
[i
])) >> FSHIFT
;
1426 * Schedule the next update to occur after 5 seconds, but add a
1427 * random variation to avoid synchronisation with processes that
1428 * run at regular intervals.
1430 callout_reset(&gd
->gd_loadav_callout
,
1431 hz
* 4 + (int)(krandom() % (hz
* 2 + 1)),
1436 loadav_count_runnable(struct lwp
*lp
, void *data
)
1441 switch (lp
->lwp_stat
) {
1443 if ((td
= lp
->lwp_thread
) == NULL
)
1445 if (td
->td_flags
& TDF_BLOCKED
)
1457 * Regular data collection
1460 collect_load_callback(int n
)
1462 int fscale
= averunnable
.fscale
;
1464 return ((averunnable
.ldavg
[0] * 100 + (fscale
>> 1)) / fscale
);
1468 sched_setup(void *dummy __unused
)
1470 globaldata_t save_gd
= mycpu
;
1474 kcollect_register(KCOLLECT_LOAD
, "load", collect_load_callback
,
1475 KCOLLECT_SCALE(KCOLLECT_LOAD_FORMAT
, 0));
1478 * Kick off timeout driven events by calling first time. We
1479 * split the work across available cpus to help scale it,
1480 * it can eat a lot of cpu when there are a lot of processes
1483 for (n
= 0; n
< ncpus
; ++n
) {
1484 gd
= globaldata_find(n
);
1485 lwkt_setcpu_self(gd
);
1486 callout_init_mp(&gd
->gd_loadav_callout
);
1487 callout_init_mp(&gd
->gd_schedcpu_callout
);
1491 lwkt_setcpu_self(save_gd
);
1495 * Extremely early initialization, dummy-up the tables so we don't have
1496 * to conditionalize for NULL in _wakeup() and tsleep_interlock(). Even
1497 * though the system isn't blocking this early, these functions still
1498 * try to access the hash table.
1500 * This setup will be overridden once sched_dyninit() -> sleep_gdinit()
1504 sleep_early_gdinit(globaldata_t gd
)
1506 static struct tslpque dummy_slpque
;
1507 static cpumask_t dummy_cpumasks
;
1509 slpque_tablesize
= 1;
1510 gd
->gd_tsleep_hash
= &dummy_slpque
;
1511 slpque_cpumasks
= &dummy_cpumasks
;
1512 TAILQ_INIT(&dummy_slpque
.queue
);
1516 * PCPU initialization. Called after KMALLOC is operational, by
1517 * sched_dyninit() for cpu 0, and by mi_gdinit() for other cpus later.
1519 * WARNING! The pcpu hash table is smaller than the global cpumask
1520 * hash table, which can save us a lot of memory when maxproc
1524 sleep_gdinit(globaldata_t gd
)
1532 * This shouldn't happen, that is there shouldn't be any threads
1533 * waiting on the dummy tsleep queue this early in the boot.
1535 if (gd
->gd_cpuid
== 0) {
1536 struct tslpque
*qp
= &gd
->gd_tsleep_hash
[0];
1537 TAILQ_FOREACH(td
, &qp
->queue
, td_sleepq
) {
1538 kprintf("SLEEP_GDINIT SWITCH %s\n", td
->td_comm
);
1543 * Note that we have to allocate one extra slot because we are
1544 * shifting a modulo value. TCHASHSHIFT(slpque_tablesize - 1) can
1545 * return the same value as TCHASHSHIFT(slpque_tablesize).
1547 n
= TCHASHSHIFT(slpque_tablesize
) + 1;
1549 hash_size
= sizeof(struct tslpque
) * n
;
1550 gd
->gd_tsleep_hash
= (void *)kmem_alloc3(kernel_map
, hash_size
,
1552 KM_CPU(gd
->gd_cpuid
));
1553 memset(gd
->gd_tsleep_hash
, 0, hash_size
);
1554 for (i
= 0; i
< n
; ++i
)
1555 TAILQ_INIT(&gd
->gd_tsleep_hash
[i
].queue
);
1559 * Dynamic initialization after the memory system is operational.
1562 sched_dyninit(void *dummy __unused
)
1569 * Calculate table size for slpque hash. We want a prime number
1570 * large enough to avoid overloading slpque_cpumasks when the
1571 * system has a large number of sleeping processes, which will
1572 * spam IPIs on wakeup().
1574 * While it is true this is really a per-lwp factor, generally
1575 * speaking the maxproc limit is a good metric to go by.
1577 for (tblsize
= maxproc
| 1; ; tblsize
+= 2) {
1578 if (tblsize
% 3 == 0)
1580 if (tblsize
% 5 == 0)
1582 tblsize2
= (tblsize
/ 2) | 1;
1583 for (n
= 7; n
< tblsize2
; n
+= 2) {
1584 if (tblsize
% n
== 0)
1592 * PIDs are currently limited to 6 digits. Cap the table size
1595 if (tblsize
> 2000003)
1598 slpque_tablesize
= tblsize
;
1599 slpque_cpumasks
= kmalloc(sizeof(*slpque_cpumasks
) * slpque_tablesize
,
1600 M_TSLEEP
, M_WAITOK
| M_ZERO
);
1601 sleep_gdinit(mycpu
);