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>
52 #include <sys/ktrace.h>
55 #include <sys/serialize.h>
57 #include <sys/signal2.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <sys/mutex2.h>
62 #include <machine/cpu.h>
63 #include <machine/smp.h>
65 TAILQ_HEAD(tslpque
, thread
);
67 static void sched_setup (void *dummy
);
68 SYSINIT(sched_setup
, SI_SUB_KICK_SCHEDULER
, SI_ORDER_FIRST
, sched_setup
, NULL
);
72 int sched_quantum
; /* Roundrobin scheduling quantum in ticks. */
74 int ncpus2
, ncpus2_shift
, ncpus2_mask
; /* note: mask not cpumask_t */
75 int ncpus_fit
, ncpus_fit_mask
; /* note: mask not cpumask_t */
78 int tsleep_crypto_dump
= 0;
80 static struct callout loadav_callout
;
81 static struct callout schedcpu_callout
;
82 MALLOC_DEFINE(M_TSLEEP
, "tslpque", "tsleep queues");
84 #define __DEALL(ident) __DEQUALIFY(void *, ident)
86 #if !defined(KTR_TSLEEP)
87 #define KTR_TSLEEP KTR_ALL
89 KTR_INFO_MASTER(tsleep
);
90 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_beg
, 0, "tsleep enter %p", const volatile void *ident
);
91 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_end
, 1, "tsleep exit");
92 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_beg
, 2, "wakeup enter %p", const volatile void *ident
);
93 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_end
, 3, "wakeup exit");
94 KTR_INFO(KTR_TSLEEP
, tsleep
, ilockfail
, 4, "interlock failed %p", const volatile void *ident
);
96 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
97 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
99 struct loadavg averunnable
=
100 { {0, 0, 0}, FSCALE
}; /* load average, of runnable procs */
102 * Constants for averages over 1, 5, and 15 minutes
103 * when sampling at 5 second intervals.
105 static fixpt_t cexp
[3] = {
106 0.9200444146293232 * FSCALE
, /* exp(-1/12) */
107 0.9834714538216174 * FSCALE
, /* exp(-1/60) */
108 0.9944598480048967 * FSCALE
, /* exp(-1/180) */
111 static void endtsleep (void *);
112 static void loadav (void *arg
);
113 static void schedcpu (void *arg
);
116 * Adjust the scheduler quantum. The quantum is specified in microseconds.
117 * Note that 'tick' is in microseconds per tick.
120 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS
)
124 new_val
= sched_quantum
* ustick
;
125 error
= sysctl_handle_int(oidp
, &new_val
, 0, req
);
126 if (error
!= 0 || req
->newptr
== NULL
)
128 if (new_val
< ustick
)
130 sched_quantum
= new_val
/ ustick
;
134 SYSCTL_PROC(_kern
, OID_AUTO
, quantum
, CTLTYPE_INT
|CTLFLAG_RW
,
135 0, sizeof sched_quantum
, sysctl_kern_quantum
, "I", "");
137 static int pctcpu_decay
= 10;
138 SYSCTL_INT(_kern
, OID_AUTO
, pctcpu_decay
, CTLFLAG_RW
, &pctcpu_decay
, 0, "");
141 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
143 int fscale __unused
= FSCALE
; /* exported to systat */
144 SYSCTL_INT(_kern
, OID_AUTO
, fscale
, CTLFLAG_RD
, 0, FSCALE
, "");
147 * Recompute process priorities, once a second.
149 * Since the userland schedulers are typically event oriented, if the
150 * estcpu calculation at wakeup() time is not sufficient to make a
151 * process runnable relative to other processes in the system we have
152 * a 1-second recalc to help out.
154 * This code also allows us to store sysclock_t data in the process structure
155 * without fear of an overrun, since sysclock_t are guarenteed to hold
156 * several seconds worth of count.
158 * WARNING! callouts can preempt normal threads. However, they will not
159 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
161 static int schedcpu_stats(struct proc
*p
, void *data __unused
);
162 static int schedcpu_resource(struct proc
*p
, void *data __unused
);
167 allproc_scan(schedcpu_stats
, NULL
);
168 allproc_scan(schedcpu_resource
, NULL
);
169 wakeup((caddr_t
)&lbolt
);
170 wakeup(lbolt_syncer
);
171 callout_reset(&schedcpu_callout
, hz
, schedcpu
, NULL
);
175 * General process statistics once a second
178 schedcpu_stats(struct proc
*p
, void *data __unused
)
183 * Threads may not be completely set up if process in SIDL state.
185 if (p
->p_stat
== SIDL
)
189 if (lwkt_trytoken(&p
->p_token
) == FALSE
) {
195 FOREACH_LWP_IN_PROC(lp
, p
) {
196 if (lp
->lwp_stat
== LSSLEEP
) {
198 if (lp
->lwp_slptime
== 1)
199 p
->p_usched
->uload_update(lp
);
203 * Only recalculate processes that are active or have slept
204 * less then 2 seconds. The schedulers understand this.
205 * Otherwise decay by 50% per second.
207 if (lp
->lwp_slptime
<= 1) {
208 p
->p_usched
->recalculate(lp
);
212 decay
= pctcpu_decay
;
218 lp
->lwp_pctcpu
= (lp
->lwp_pctcpu
* (decay
- 1)) / decay
;
221 lwkt_reltoken(&p
->p_token
);
228 * Resource checks. XXX break out since ksignal/killproc can block,
229 * limiting us to one process killed per second. There is probably
233 schedcpu_resource(struct proc
*p
, void *data __unused
)
238 if (p
->p_stat
== SIDL
)
242 if (lwkt_trytoken(&p
->p_token
) == FALSE
) {
247 if (p
->p_stat
== SZOMB
|| p
->p_limit
== NULL
) {
248 lwkt_reltoken(&p
->p_token
);
254 FOREACH_LWP_IN_PROC(lp
, p
) {
256 * We may have caught an lp in the middle of being
257 * created, lwp_thread can be NULL.
259 if (lp
->lwp_thread
) {
260 ttime
+= lp
->lwp_thread
->td_sticks
;
261 ttime
+= lp
->lwp_thread
->td_uticks
;
265 switch(plimit_testcpulimit(p
->p_limit
, ttime
)) {
266 case PLIMIT_TESTCPU_KILL
:
267 killproc(p
, "exceeded maximum CPU limit");
269 case PLIMIT_TESTCPU_XCPU
:
270 if ((p
->p_flags
& P_XCPU
) == 0) {
271 p
->p_flags
|= P_XCPU
;
278 lwkt_reltoken(&p
->p_token
);
285 * This is only used by ps. Generate a cpu percentage use over
286 * a period of one second.
289 updatepcpu(struct lwp
*lp
, int cpticks
, int ttlticks
)
294 acc
= (cpticks
<< FSHIFT
) / ttlticks
;
295 if (ttlticks
>= ESTCPUFREQ
) {
296 lp
->lwp_pctcpu
= acc
;
298 remticks
= ESTCPUFREQ
- ttlticks
;
299 lp
->lwp_pctcpu
= (acc
* ttlticks
+ lp
->lwp_pctcpu
* remticks
) /
305 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
306 * like addresses being slept on. The larger the table, the fewer
307 * unnecessary IPIs. However, larger sizes also have diminishing returns
310 #define TABLESIZE 8191 /* 4001, 8191, or 16369 */
311 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
313 static cpumask_t slpque_cpumasks
[TABLESIZE
];
316 * General scheduler initialization. We force a reschedule 25 times
317 * a second by default. Note that cpu0 is initialized in early boot and
318 * cannot make any high level calls.
320 * Each cpu has its own sleep queue.
323 sleep_gdinit(globaldata_t gd
)
325 static struct tslpque slpque_cpu0
[TABLESIZE
];
328 if (gd
->gd_cpuid
== 0) {
329 sched_quantum
= (hz
+ 24) / 25;
330 gd
->gd_tsleep_hash
= slpque_cpu0
;
332 gd
->gd_tsleep_hash
= kmalloc(sizeof(slpque_cpu0
),
333 M_TSLEEP
, M_WAITOK
| M_ZERO
);
335 for (i
= 0; i
< TABLESIZE
; ++i
)
336 TAILQ_INIT(&gd
->gd_tsleep_hash
[i
]);
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
;
362 crit_enter_quick(td
);
363 if (td
->td_flags
& TDF_TSLEEPQ
) {
364 id
= LOOKUP(td
->td_wchan
);
365 TAILQ_REMOVE(&gd
->gd_tsleep_hash
[id
], td
, td_sleepq
);
366 if (TAILQ_FIRST(&gd
->gd_tsleep_hash
[id
]) == NULL
) {
367 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[id
],
371 td
->td_flags
|= TDF_TSLEEPQ
;
374 TAILQ_INSERT_TAIL(&gd
->gd_tsleep_hash
[id
], td
, td_sleepq
);
375 ATOMIC_CPUMASK_ORBIT(slpque_cpumasks
[id
], gd
->gd_cpuid
);
376 td
->td_wchan
= ident
;
377 td
->td_wdomain
= flags
& PDOMAIN_MASK
;
382 tsleep_interlock(const volatile void *ident
, int flags
)
384 _tsleep_interlock(mycpu
, ident
, flags
);
388 * Remove thread from sleepq. Must be called with a critical section held.
389 * The thread must not be migrating.
392 _tsleep_remove(thread_t td
)
394 globaldata_t gd
= mycpu
;
397 KKASSERT(td
->td_gd
== gd
&& IN_CRITICAL_SECT(td
));
398 KKASSERT((td
->td_flags
& TDF_MIGRATING
) == 0);
399 if (td
->td_flags
& TDF_TSLEEPQ
) {
400 td
->td_flags
&= ~TDF_TSLEEPQ
;
401 id
= LOOKUP(td
->td_wchan
);
402 TAILQ_REMOVE(&gd
->gd_tsleep_hash
[id
], td
, td_sleepq
);
403 if (TAILQ_FIRST(&gd
->gd_tsleep_hash
[id
]) == NULL
) {
404 ATOMIC_CPUMASK_NANDBIT(slpque_cpumasks
[id
],
413 tsleep_remove(thread_t td
)
419 * General sleep call. Suspends the current process until a wakeup is
420 * performed on the specified identifier. The process will then be made
421 * runnable with the specified priority. Sleeps at most timo/hz seconds
422 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
423 * before and after sleeping, else signals are not checked. Returns 0 if
424 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
425 * signal needs to be delivered, ERESTART is returned if the current system
426 * call should be restarted if possible, and EINTR is returned if the system
427 * call should be interrupted by the signal (return EINTR).
429 * Note that if we are a process, we release_curproc() before messing with
430 * the LWKT scheduler.
432 * During autoconfiguration or after a panic, a sleep will simply
433 * lower the priority briefly to allow interrupts, then return.
435 * WARNING! This code can't block (short of switching away), or bad things
436 * will happen. No getting tokens, no blocking locks, etc.
439 tsleep(const volatile void *ident
, int flags
, const char *wmesg
, int timo
)
441 struct thread
*td
= curthread
;
442 struct lwp
*lp
= td
->td_lwp
;
443 struct proc
*p
= td
->td_proc
; /* may be NULL */
449 struct callout thandle
;
452 * Currently a severe hack. Make sure any delayed wakeups
453 * are flushed before we sleep or we might deadlock on whatever
454 * event we are sleeping on.
456 if (td
->td_flags
& TDF_DELAYED_WAKEUP
)
457 wakeup_end_delayed();
460 * NOTE: removed KTRPOINT, it could cause races due to blocking
461 * even in stable. Just scrap it for now.
463 if (!tsleep_crypto_dump
&& (tsleep_now_works
== 0 || panicstr
)) {
465 * After a panic, or before we actually have an operational
466 * softclock, just give interrupts a chance, then just return;
468 * don't run any other procs or panic below,
469 * in case this is the idle process and already asleep.
473 lwkt_setpri_self(safepri
);
475 lwkt_setpri_self(oldpri
);
478 logtsleep2(tsleep_beg
, ident
);
480 KKASSERT(td
!= &gd
->gd_idlethread
); /* you must be kidding! */
481 td
->td_wakefromcpu
= -1; /* overwritten by _wakeup */
484 * NOTE: all of this occurs on the current cpu, including any
485 * callout-based wakeups, so a critical section is a sufficient
488 * The entire sequence through to where we actually sleep must
489 * run without breaking the critical section.
491 catch = flags
& PCATCH
;
495 crit_enter_quick(td
);
497 KASSERT(ident
!= NULL
, ("tsleep: no ident"));
498 KASSERT(lp
== NULL
||
499 lp
->lwp_stat
== LSRUN
|| /* Obvious */
500 lp
->lwp_stat
== LSSTOP
, /* Set in tstop */
502 ident
, wmesg
, lp
->lwp_stat
));
505 * We interlock the sleep queue if the caller has not already done
506 * it for us. This must be done before we potentially acquire any
507 * tokens or we can loose the wakeup.
509 if ((flags
& PINTERLOCKED
) == 0) {
510 _tsleep_interlock(gd
, ident
, flags
);
514 * Setup for the current process (if this is a process). We must
515 * interlock with lwp_token to avoid remote wakeup races via
519 lwkt_gettoken(&lp
->lwp_token
);
522 * If the umbrella process is in the SCORE state then
523 * make sure that the thread is flagged going into a
524 * normal sleep to allow the core dump to proceed, otherwise
525 * the coredump can end up waiting forever. If the normal
526 * sleep is woken up, the thread will enter a stopped state
527 * upon return to userland.
529 * We do not want to interrupt or cause a thread exist at
530 * this juncture because that will mess-up the state the
531 * coredump is trying to save.
533 if (p
->p_stat
== SCORE
&&
534 (lp
->lwp_mpflags
& LWP_MP_WSTOP
) == 0) {
535 atomic_set_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
544 * Early termination if PCATCH was set and a
545 * signal is pending, interlocked with the
548 * Early termination only occurs when tsleep() is
549 * entered while in a normal LSRUN state.
551 if ((sig
= CURSIG(lp
)) != 0)
555 * Causes ksignal to wake us up if a signal is
556 * received (interlocked with lp->lwp_token).
558 lp
->lwp_flags
|= LWP_SINTR
;
565 * Make sure the current process has been untangled from
566 * the userland scheduler and initialize slptime to start
569 * NOTE: td->td_wakefromcpu is pre-set by the release function
570 * for the dfly scheduler, and then adjusted by _wakeup()
573 p
->p_usched
->release_curproc(lp
);
578 * If the interlocked flag is set but our cpu bit in the slpqueue
579 * is no longer set, then a wakeup was processed inbetween the
580 * tsleep_interlock() (ours or the callers), and here. This can
581 * occur under numerous circumstances including when we release the
584 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
585 * to process incoming IPIs, thus draining incoming wakeups.
587 if ((td
->td_flags
& TDF_TSLEEPQ
) == 0) {
588 logtsleep2(ilockfail
, ident
);
593 * scheduling is blocked while in a critical section. Coincide
594 * the descheduled-by-tsleep flag with the descheduling of the
597 * The timer callout is localized on our cpu and interlocked by
598 * our critical section.
600 lwkt_deschedule_self(td
);
601 td
->td_flags
|= TDF_TSLEEP_DESCHEDULED
;
602 td
->td_wmesg
= wmesg
;
605 * Setup the timeout, if any. The timeout is only operable while
606 * the thread is flagged descheduled.
608 KKASSERT((td
->td_flags
& TDF_TIMEOUT
) == 0);
610 callout_init_mp(&thandle
);
611 callout_reset(&thandle
, timo
, endtsleep
, td
);
619 * Ok, we are sleeping. Place us in the SSLEEP state.
621 KKASSERT((lp
->lwp_mpflags
& LWP_MP_ONRUNQ
) == 0);
624 * tstop() sets LSSTOP, so don't fiddle with that.
626 if (lp
->lwp_stat
!= LSSTOP
)
627 lp
->lwp_stat
= LSSLEEP
;
628 lp
->lwp_ru
.ru_nvcsw
++;
629 p
->p_usched
->uload_update(lp
);
633 * And when we are woken up, put us back in LSRUN. If we
634 * slept for over a second, recalculate our estcpu.
636 lp
->lwp_stat
= LSRUN
;
637 if (lp
->lwp_slptime
) {
638 p
->p_usched
->uload_update(lp
);
639 p
->p_usched
->recalculate(lp
);
647 * Make sure we haven't switched cpus while we were asleep. It's
648 * not supposed to happen. Cleanup our temporary flags.
650 KKASSERT(gd
== td
->td_gd
);
653 * Cleanup the timeout. If the timeout has already occured thandle
654 * has already been stopped, otherwise stop thandle. If the timeout
655 * is running (the callout thread must be blocked trying to get
656 * lwp_token) then wait for us to get scheduled.
659 while (td
->td_flags
& TDF_TIMEOUT_RUNNING
) {
660 /* else we won't get rescheduled! */
661 if (lp
->lwp_stat
!= LSSTOP
)
662 lp
->lwp_stat
= LSSLEEP
;
663 lwkt_deschedule_self(td
);
664 td
->td_wmesg
= "tsrace";
666 kprintf("td %p %s: timeout race\n", td
, td
->td_comm
);
668 if (td
->td_flags
& TDF_TIMEOUT
) {
669 td
->td_flags
&= ~TDF_TIMEOUT
;
672 /* does not block when on same cpu */
673 callout_stop(&thandle
);
676 td
->td_flags
&= ~TDF_TSLEEP_DESCHEDULED
;
679 * Make sure we have been removed from the sleepq. In most
680 * cases this will have been done for us already but it is
681 * possible for a scheduling IPI to be in-flight from a
682 * previous tsleep/tsleep_interlock() or due to a straight-out
683 * call to lwkt_schedule() (in the case of an interrupt thread),
684 * causing a spurious wakeup.
690 * Figure out the correct error return. If interrupted by a
691 * signal we want to return EINTR or ERESTART.
695 if (catch && error
== 0) {
696 if (sig
!= 0 || (sig
= CURSIG(lp
))) {
697 if (SIGISMEMBER(p
->p_sigacts
->ps_sigintr
, sig
))
704 lp
->lwp_flags
&= ~LWP_SINTR
;
707 * Unconditionally set us to LSRUN on resume. lwp_stat could
708 * be in a weird state due to the goto resume, particularly
709 * when tsleep() is called from tstop().
711 lp
->lwp_stat
= LSRUN
;
712 lwkt_reltoken(&lp
->lwp_token
);
714 logtsleep1(tsleep_end
);
720 * Interlocked spinlock sleep. An exclusively held spinlock must
721 * be passed to ssleep(). The function will atomically release the
722 * spinlock and tsleep on the ident, then reacquire the spinlock and
725 * This routine is fairly important along the critical path, so optimize it
729 ssleep(const volatile void *ident
, struct spinlock
*spin
, int flags
,
730 const char *wmesg
, int timo
)
732 globaldata_t gd
= mycpu
;
735 _tsleep_interlock(gd
, ident
, flags
);
736 spin_unlock_quick(gd
, spin
);
737 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
738 _spin_lock_quick(gd
, spin
, wmesg
);
744 lksleep(const volatile void *ident
, struct lock
*lock
, int flags
,
745 const char *wmesg
, int timo
)
747 globaldata_t gd
= mycpu
;
750 _tsleep_interlock(gd
, ident
, flags
);
751 lockmgr(lock
, LK_RELEASE
);
752 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
753 lockmgr(lock
, LK_EXCLUSIVE
);
759 * Interlocked mutex sleep. An exclusively held mutex must be passed
760 * to mtxsleep(). The function will atomically release the mutex
761 * and tsleep on the ident, then reacquire the mutex and return.
764 mtxsleep(const volatile void *ident
, struct mtx
*mtx
, int flags
,
765 const char *wmesg
, int timo
)
767 globaldata_t gd
= mycpu
;
770 _tsleep_interlock(gd
, ident
, flags
);
772 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
773 mtx_lock_ex_quick(mtx
);
779 * Interlocked serializer sleep. An exclusively held serializer must
780 * be passed to zsleep(). The function will atomically release
781 * the serializer and tsleep on the ident, then reacquire the serializer
785 zsleep(const volatile void *ident
, struct lwkt_serialize
*slz
, int flags
,
786 const char *wmesg
, int timo
)
788 globaldata_t gd
= mycpu
;
791 ASSERT_SERIALIZED(slz
);
793 _tsleep_interlock(gd
, ident
, flags
);
794 lwkt_serialize_exit(slz
);
795 ret
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
796 lwkt_serialize_enter(slz
);
802 * Directly block on the LWKT thread by descheduling it. This
803 * is much faster then tsleep(), but the only legal way to wake
804 * us up is to directly schedule the thread.
806 * Setting TDF_SINTR will cause new signals to directly schedule us.
808 * This routine must be called while in a critical section.
811 lwkt_sleep(const char *wmesg
, int flags
)
813 thread_t td
= curthread
;
816 if ((flags
& PCATCH
) == 0 || td
->td_lwp
== NULL
) {
817 td
->td_flags
|= TDF_BLOCKED
;
818 td
->td_wmesg
= wmesg
;
819 lwkt_deschedule_self(td
);
822 td
->td_flags
&= ~TDF_BLOCKED
;
825 if ((sig
= CURSIG(td
->td_lwp
)) != 0) {
826 if (SIGISMEMBER(td
->td_proc
->p_sigacts
->ps_sigintr
, sig
))
832 td
->td_flags
|= TDF_BLOCKED
| TDF_SINTR
;
833 td
->td_wmesg
= wmesg
;
834 lwkt_deschedule_self(td
);
836 td
->td_flags
&= ~(TDF_BLOCKED
| TDF_SINTR
);
842 * Implement the timeout for tsleep.
844 * This type of callout timeout is scheduled on the same cpu the process
845 * is sleeping on. Also, at the moment, the MP lock is held.
854 * We are going to have to get the lwp_token, which means we might
855 * block. This can race a tsleep getting woken up by other means
856 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
857 * processing to complete (sorry tsleep!).
859 * We can safely set td_flags because td MUST be on the same cpu
862 KKASSERT(td
->td_gd
== mycpu
);
864 td
->td_flags
|= TDF_TIMEOUT_RUNNING
| TDF_TIMEOUT
;
867 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
868 * from exiting the tsleep on us. The flag is interlocked by virtue
869 * of lp being on the same cpu as we are.
871 if ((lp
= td
->td_lwp
) != NULL
)
872 lwkt_gettoken(&lp
->lwp_token
);
874 KKASSERT(td
->td_flags
& TDF_TSLEEP_DESCHEDULED
);
878 * callout timer should normally never be set in tstop()
879 * because it passes a timeout of 0. However, there is a
880 * case during thread exit (which SSTOP's all the threads)
881 * for which tstop() must break out and can (properly) leave
882 * the thread in LSSTOP.
884 KKASSERT(lp
->lwp_stat
!= LSSTOP
||
885 (lp
->lwp_mpflags
& LWP_MP_WEXIT
));
887 lwkt_reltoken(&lp
->lwp_token
);
892 KKASSERT(td
->td_gd
== mycpu
);
893 td
->td_flags
&= ~TDF_TIMEOUT_RUNNING
;
898 * Make all processes sleeping on the specified identifier runnable.
899 * count may be zero or one only.
901 * The domain encodes the sleep/wakeup domain, flags, plus the originating
904 * This call may run without the MP lock held. We can only manipulate thread
905 * state on the cpu owning the thread. We CANNOT manipulate process state
908 * _wakeup() can be passed to an IPI so we can't use (const volatile
912 _wakeup(void *ident
, int domain
)
922 logtsleep2(wakeup_beg
, ident
);
925 qp
= &gd
->gd_tsleep_hash
[id
];
927 for (td
= TAILQ_FIRST(qp
); td
!= NULL
; td
= ntd
) {
928 ntd
= TAILQ_NEXT(td
, td_sleepq
);
929 if (td
->td_wchan
== ident
&&
930 td
->td_wdomain
== (domain
& PDOMAIN_MASK
)
932 KKASSERT(td
->td_gd
== gd
);
934 td
->td_wakefromcpu
= PWAKEUP_DECODE(domain
);
935 if (td
->td_flags
& TDF_TSLEEP_DESCHEDULED
) {
937 if (domain
& PWAKEUP_ONE
)
945 * We finished checking the current cpu but there still may be
946 * more work to do. Either wakeup_one was requested and no matching
947 * thread was found, or a normal wakeup was requested and we have
948 * to continue checking cpus.
950 * It should be noted that this scheme is actually less expensive then
951 * the old scheme when waking up multiple threads, since we send
952 * only one IPI message per target candidate which may then schedule
953 * multiple threads. Before we could have wound up sending an IPI
954 * message for each thread on the target cpu (!= current cpu) that
955 * needed to be woken up.
957 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
958 * should be ok since we are passing idents in the IPI rather then
961 if ((domain
& PWAKEUP_MYCPU
) == 0) {
962 mask
= slpque_cpumasks
[id
];
963 CPUMASK_ANDMASK(mask
, gd
->gd_other_cpus
);
964 if (CPUMASK_TESTNZERO(mask
)) {
965 lwkt_send_ipiq2_mask(mask
, _wakeup
, ident
,
966 domain
| PWAKEUP_MYCPU
);
970 logtsleep1(wakeup_end
);
975 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
978 wakeup(const volatile void *ident
)
980 globaldata_t gd
= mycpu
;
981 thread_t td
= gd
->gd_curthread
;
983 if (td
&& (td
->td_flags
& TDF_DELAYED_WAKEUP
)) {
985 * If we are in a delayed wakeup section, record up to two wakeups in
986 * a per-CPU queue and issue them when we block or exit the delayed
989 if (atomic_cmpset_ptr(&gd
->gd_delayed_wakeup
[0], NULL
, ident
))
991 if (atomic_cmpset_ptr(&gd
->gd_delayed_wakeup
[1], NULL
, ident
))
994 ident
= atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd
->gd_delayed_wakeup
[1]),
996 ident
= atomic_swap_ptr(__DEQUALIFY(volatile void **, &gd
->gd_delayed_wakeup
[0]),
1000 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, gd
->gd_cpuid
));
1004 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
1007 wakeup_one(const volatile void *ident
)
1009 /* XXX potentially round-robin the first responding cpu */
1010 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1015 * Wakeup threads tsleep()ing on the specified ident on the current cpu
1019 wakeup_mycpu(const volatile void *ident
)
1021 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1026 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
1030 wakeup_mycpu_one(const volatile void *ident
)
1032 /* XXX potentially round-robin the first responding cpu */
1033 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) |
1034 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1038 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
1042 wakeup_oncpu(globaldata_t gd
, const volatile void *ident
)
1044 globaldata_t mygd
= mycpu
;
1046 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1049 lwkt_send_ipiq2(gd
, _wakeup
, __DEALL(ident
),
1050 PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1056 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
1060 wakeup_oncpu_one(globaldata_t gd
, const volatile void *ident
)
1062 globaldata_t mygd
= mycpu
;
1064 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1065 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1067 lwkt_send_ipiq2(gd
, _wakeup
, __DEALL(ident
),
1068 PWAKEUP_ENCODE(0, mygd
->gd_cpuid
) |
1069 PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1074 * Wakeup all threads waiting on the specified ident that slept using
1075 * the specified domain, on all cpus.
1078 wakeup_domain(const volatile void *ident
, int domain
)
1080 _wakeup(__DEALL(ident
), PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
));
1084 * Wakeup one thread waiting on the specified ident that slept using
1085 * the specified domain, on any cpu.
1088 wakeup_domain_one(const volatile void *ident
, int domain
)
1090 /* XXX potentially round-robin the first responding cpu */
1091 _wakeup(__DEALL(ident
),
1092 PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
) | PWAKEUP_ONE
);
1096 wakeup_start_delayed(void)
1098 globaldata_t gd
= mycpu
;
1101 gd
->gd_curthread
->td_flags
|= TDF_DELAYED_WAKEUP
;
1106 wakeup_end_delayed(void)
1108 globaldata_t gd
= mycpu
;
1110 if (gd
->gd_curthread
->td_flags
& TDF_DELAYED_WAKEUP
) {
1112 gd
->gd_curthread
->td_flags
&= ~TDF_DELAYED_WAKEUP
;
1113 if (gd
->gd_delayed_wakeup
[0] || gd
->gd_delayed_wakeup
[1]) {
1114 if (gd
->gd_delayed_wakeup
[0]) {
1115 wakeup(gd
->gd_delayed_wakeup
[0]);
1116 gd
->gd_delayed_wakeup
[0] = NULL
;
1118 if (gd
->gd_delayed_wakeup
[1]) {
1119 wakeup(gd
->gd_delayed_wakeup
[1]);
1120 gd
->gd_delayed_wakeup
[1] = NULL
;
1130 * Make a process runnable. lp->lwp_token must be held on call and this
1131 * function must be called from the cpu owning lp.
1133 * This only has an effect if we are in LSSTOP or LSSLEEP.
1136 setrunnable(struct lwp
*lp
)
1138 thread_t td
= lp
->lwp_thread
;
1140 ASSERT_LWKT_TOKEN_HELD(&lp
->lwp_token
);
1141 KKASSERT(td
->td_gd
== mycpu
);
1143 if (lp
->lwp_stat
== LSSTOP
)
1144 lp
->lwp_stat
= LSSLEEP
;
1145 if (lp
->lwp_stat
== LSSLEEP
) {
1148 } else if (td
->td_flags
& TDF_SINTR
) {
1155 * The process is stopped due to some condition, usually because p_stat is
1156 * set to SSTOP, but also possibly due to being traced.
1158 * Caller must hold p->p_token
1160 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1161 * because the parent may check the child's status before the child actually
1162 * gets to this routine.
1164 * This routine is called with the current lwp only, typically just
1165 * before returning to userland if the process state is detected as
1166 * possibly being in a stopped state.
1171 struct lwp
*lp
= curthread
->td_lwp
;
1172 struct proc
*p
= lp
->lwp_proc
;
1175 lwkt_gettoken(&lp
->lwp_token
);
1179 * If LWP_MP_WSTOP is set, we were sleeping
1180 * while our process was stopped. At this point
1181 * we were already counted as stopped.
1183 if ((lp
->lwp_mpflags
& LWP_MP_WSTOP
) == 0) {
1185 * If we're the last thread to stop, signal
1189 atomic_set_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
1190 wakeup(&p
->p_nstopped
);
1191 if (p
->p_nstopped
== p
->p_nthreads
) {
1193 * Token required to interlock kern_wait()
1197 lwkt_gettoken(&q
->p_token
);
1198 p
->p_flags
&= ~P_WAITED
;
1200 if ((q
->p_sigacts
->ps_flag
& PS_NOCLDSTOP
) == 0)
1201 ksignal(q
, SIGCHLD
);
1202 lwkt_reltoken(&q
->p_token
);
1208 * Wait here while in a stopped state, interlocked with lwp_token.
1209 * We must break-out if the whole process is trying to exit.
1211 while (STOPLWP(p
, lp
)) {
1212 lp
->lwp_stat
= LSSTOP
;
1213 tsleep(p
, 0, "stop", 0);
1216 atomic_clear_int(&lp
->lwp_mpflags
, LWP_MP_WSTOP
);
1218 lwkt_reltoken(&lp
->lwp_token
);
1222 * Compute a tenex style load average of a quantity on
1223 * 1, 5 and 15 minute intervals.
1225 static int loadav_count_runnable(struct lwp
*p
, void *data
);
1230 struct loadavg
*avg
;
1234 alllwp_scan(loadav_count_runnable
, &nrun
);
1236 for (i
= 0; i
< 3; i
++) {
1237 avg
->ldavg
[i
] = (cexp
[i
] * avg
->ldavg
[i
] +
1238 (long)nrun
* FSCALE
* (FSCALE
- cexp
[i
])) >> FSHIFT
;
1242 * Schedule the next update to occur after 5 seconds, but add a
1243 * random variation to avoid synchronisation with processes that
1244 * run at regular intervals.
1246 callout_reset(&loadav_callout
, hz
* 4 + (int)(krandom() % (hz
* 2 + 1)),
1251 loadav_count_runnable(struct lwp
*lp
, void *data
)
1256 switch (lp
->lwp_stat
) {
1258 if ((td
= lp
->lwp_thread
) == NULL
)
1260 if (td
->td_flags
& TDF_BLOCKED
)
1272 * Regular data collection
1275 collect_load_callback(int n
)
1277 int fscale
= averunnable
.fscale
;
1279 return ((averunnable
.ldavg
[0] * 100 + (fscale
>> 1)) / fscale
);
1284 sched_setup(void *dummy
)
1286 callout_init_mp(&loadav_callout
);
1287 callout_init_mp(&schedcpu_callout
);
1288 kcollect_register(KCOLLECT_LOAD
, "load", collect_load_callback
,
1289 KCOLLECT_SCALE(KCOLLECT_LOAD_FORMAT
, 0));
1290 /* Kick off timeout driven events by calling first time. */