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. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
39 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.91 2008/09/09 04:06:13 dillon Exp $
43 #include "opt_ktrace.h"
45 #include <sys/param.h>
46 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/signalvar.h>
50 #include <sys/resourcevar.h>
51 #include <sys/vmmeter.h>
52 #include <sys/sysctl.h>
56 #include <sys/ktrace.h>
58 #include <sys/xwait.h>
60 #include <sys/serialize.h>
62 #include <sys/signal2.h>
63 #include <sys/thread2.h>
64 #include <sys/spinlock2.h>
65 #include <sys/mutex2.h>
66 #include <sys/mplock2.h>
68 #include <machine/cpu.h>
69 #include <machine/smp.h>
71 TAILQ_HEAD(tslpque
, thread
);
73 static void sched_setup (void *dummy
);
74 SYSINIT(sched_setup
, SI_SUB_KICK_SCHEDULER
, SI_ORDER_FIRST
, sched_setup
, NULL
)
79 int sched_quantum
; /* Roundrobin scheduling quantum in ticks. */
81 int ncpus2
, ncpus2_shift
, ncpus2_mask
;
82 int ncpus_fit
, ncpus_fit_mask
;
86 static struct callout loadav_callout
;
87 static struct callout schedcpu_callout
;
88 MALLOC_DEFINE(M_TSLEEP
, "tslpque", "tsleep queues");
90 #if !defined(KTR_TSLEEP)
91 #define KTR_TSLEEP KTR_ALL
93 KTR_INFO_MASTER(tsleep
);
94 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_beg
, 0, "tsleep enter %p", sizeof(void *));
95 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_end
, 1, "tsleep exit", 0);
96 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_beg
, 2, "wakeup enter %p", sizeof(void *));
97 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_end
, 3, "wakeup exit", 0);
98 KTR_INFO(KTR_TSLEEP
, tsleep
, ilockfail
, 4, "interlock failed %p", sizeof(void *));
100 #define logtsleep1(name) KTR_LOG(tsleep_ ## name)
101 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
103 struct loadavg averunnable
=
104 { {0, 0, 0}, FSCALE
}; /* load average, of runnable procs */
106 * Constants for averages over 1, 5, and 15 minutes
107 * when sampling at 5 second intervals.
109 static fixpt_t cexp
[3] = {
110 0.9200444146293232 * FSCALE
, /* exp(-1/12) */
111 0.9834714538216174 * FSCALE
, /* exp(-1/60) */
112 0.9944598480048967 * FSCALE
, /* exp(-1/180) */
115 static void endtsleep (void *);
116 static void loadav (void *arg
);
117 static void schedcpu (void *arg
);
119 static void tsleep_wakeup(struct thread
*td
);
123 * Adjust the scheduler quantum. The quantum is specified in microseconds.
124 * Note that 'tick' is in microseconds per tick.
127 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS
)
131 new_val
= sched_quantum
* ustick
;
132 error
= sysctl_handle_int(oidp
, &new_val
, 0, req
);
133 if (error
!= 0 || req
->newptr
== NULL
)
135 if (new_val
< ustick
)
137 sched_quantum
= new_val
/ ustick
;
138 hogticks
= 2 * sched_quantum
;
142 SYSCTL_PROC(_kern
, OID_AUTO
, quantum
, CTLTYPE_INT
|CTLFLAG_RW
,
143 0, sizeof sched_quantum
, sysctl_kern_quantum
, "I", "");
146 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
147 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
148 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
150 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
151 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
153 * If you don't want to bother with the faster/more-accurate formula, you
154 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
155 * (more general) method of calculating the %age of CPU used by a process.
157 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
159 #define CCPU_SHIFT 11
161 static fixpt_t ccpu
= 0.95122942450071400909 * FSCALE
; /* exp(-1/20) */
162 SYSCTL_INT(_kern
, OID_AUTO
, ccpu
, CTLFLAG_RD
, &ccpu
, 0, "");
165 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
167 int fscale __unused
= FSCALE
; /* exported to systat */
168 SYSCTL_INT(_kern
, OID_AUTO
, fscale
, CTLFLAG_RD
, 0, FSCALE
, "");
171 * Recompute process priorities, once a second.
173 * Since the userland schedulers are typically event oriented, if the
174 * estcpu calculation at wakeup() time is not sufficient to make a
175 * process runnable relative to other processes in the system we have
176 * a 1-second recalc to help out.
178 * This code also allows us to store sysclock_t data in the process structure
179 * without fear of an overrun, since sysclock_t are guarenteed to hold
180 * several seconds worth of count.
182 * WARNING! callouts can preempt normal threads. However, they will not
183 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
185 static int schedcpu_stats(struct proc
*p
, void *data __unused
);
186 static int schedcpu_resource(struct proc
*p
, void *data __unused
);
191 allproc_scan(schedcpu_stats
, NULL
);
192 allproc_scan(schedcpu_resource
, NULL
);
193 wakeup((caddr_t
)&lbolt
);
194 wakeup((caddr_t
)&lbolt_syncer
);
195 callout_reset(&schedcpu_callout
, hz
, schedcpu
, NULL
);
199 * General process statistics once a second
202 schedcpu_stats(struct proc
*p
, void *data __unused
)
208 FOREACH_LWP_IN_PROC(lp
, p
) {
209 if (lp
->lwp_stat
== LSSLEEP
)
213 * Only recalculate processes that are active or have slept
214 * less then 2 seconds. The schedulers understand this.
216 if (lp
->lwp_slptime
<= 1) {
217 p
->p_usched
->recalculate(lp
);
219 lp
->lwp_pctcpu
= (lp
->lwp_pctcpu
* ccpu
) >> FSHIFT
;
227 * Resource checks. XXX break out since ksignal/killproc can block,
228 * limiting us to one process killed per second. There is probably
232 schedcpu_resource(struct proc
*p
, void *data __unused
)
238 if (p
->p_stat
== SIDL
||
239 p
->p_stat
== SZOMB
||
247 FOREACH_LWP_IN_PROC(lp
, p
) {
249 * We may have caught an lp in the middle of being
250 * created, lwp_thread can be NULL.
252 if (lp
->lwp_thread
) {
253 ttime
+= lp
->lwp_thread
->td_sticks
;
254 ttime
+= lp
->lwp_thread
->td_uticks
;
258 switch(plimit_testcpulimit(p
->p_limit
, ttime
)) {
259 case PLIMIT_TESTCPU_KILL
:
260 killproc(p
, "exceeded maximum CPU limit");
262 case PLIMIT_TESTCPU_XCPU
:
263 if ((p
->p_flag
& P_XCPU
) == 0) {
276 * This is only used by ps. Generate a cpu percentage use over
277 * a period of one second.
282 updatepcpu(struct lwp
*lp
, int cpticks
, int ttlticks
)
287 acc
= (cpticks
<< FSHIFT
) / ttlticks
;
288 if (ttlticks
>= ESTCPUFREQ
) {
289 lp
->lwp_pctcpu
= acc
;
291 remticks
= ESTCPUFREQ
- ttlticks
;
292 lp
->lwp_pctcpu
= (acc
* ttlticks
+ lp
->lwp_pctcpu
* remticks
) /
298 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
299 * like addresses being slept on.
301 #define TABLESIZE 1024
302 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1))
304 static cpumask_t slpque_cpumasks
[TABLESIZE
];
307 * General scheduler initialization. We force a reschedule 25 times
308 * a second by default. Note that cpu0 is initialized in early boot and
309 * cannot make any high level calls.
311 * Each cpu has its own sleep queue.
314 sleep_gdinit(globaldata_t gd
)
316 static struct tslpque slpque_cpu0
[TABLESIZE
];
319 if (gd
->gd_cpuid
== 0) {
320 sched_quantum
= (hz
+ 24) / 25;
321 hogticks
= 2 * sched_quantum
;
323 gd
->gd_tsleep_hash
= slpque_cpu0
;
325 gd
->gd_tsleep_hash
= kmalloc(sizeof(slpque_cpu0
),
326 M_TSLEEP
, M_WAITOK
| M_ZERO
);
328 for (i
= 0; i
< TABLESIZE
; ++i
)
329 TAILQ_INIT(&gd
->gd_tsleep_hash
[i
]);
333 * This is a dandy function that allows us to interlock tsleep/wakeup
334 * operations with unspecified upper level locks, such as lockmgr locks,
335 * simply by holding a critical section. The sequence is:
337 * (acquire upper level lock)
338 * tsleep_interlock(blah)
339 * (release upper level lock)
342 * Basically this functions queues us on the tsleep queue without actually
343 * descheduling us. When tsleep() is later called with PINTERLOCK it
344 * assumes the thread was already queued, otherwise it queues it there.
346 * Thus it is possible to receive the wakeup prior to going to sleep and
347 * the race conditions are covered.
350 _tsleep_interlock(globaldata_t gd
, void *ident
, int flags
)
352 thread_t td
= gd
->gd_curthread
;
355 crit_enter_quick(td
);
356 if (td
->td_flags
& TDF_TSLEEPQ
) {
357 id
= LOOKUP(td
->td_wchan
);
358 TAILQ_REMOVE(&gd
->gd_tsleep_hash
[id
], td
, td_sleepq
);
359 if (TAILQ_FIRST(&gd
->gd_tsleep_hash
[id
]) == NULL
)
360 atomic_clear_int(&slpque_cpumasks
[id
], gd
->gd_cpumask
);
362 td
->td_flags
|= TDF_TSLEEPQ
;
365 TAILQ_INSERT_TAIL(&gd
->gd_tsleep_hash
[id
], td
, td_sleepq
);
366 atomic_set_int(&slpque_cpumasks
[id
], gd
->gd_cpumask
);
367 td
->td_wchan
= ident
;
368 td
->td_wdomain
= flags
& PDOMAIN_MASK
;
373 tsleep_interlock(void *ident
, int flags
)
375 _tsleep_interlock(mycpu
, ident
, flags
);
379 * Remove thread from sleepq. Must be called with a critical section held.
382 _tsleep_remove(thread_t td
)
384 globaldata_t gd
= mycpu
;
387 KKASSERT(td
->td_gd
== gd
);
388 if (td
->td_flags
& TDF_TSLEEPQ
) {
389 td
->td_flags
&= ~TDF_TSLEEPQ
;
390 id
= LOOKUP(td
->td_wchan
);
391 TAILQ_REMOVE(&gd
->gd_tsleep_hash
[id
], td
, td_sleepq
);
392 if (TAILQ_FIRST(&gd
->gd_tsleep_hash
[id
]) == NULL
)
393 atomic_clear_int(&slpque_cpumasks
[id
], gd
->gd_cpumask
);
400 tsleep_remove(thread_t td
)
406 * This function removes a thread from the tsleep queue and schedules
407 * it. This function may act asynchronously. The target thread may be
408 * sleeping on a different cpu.
410 * This function mus be called while in a critical section but if the
411 * target thread is sleeping on a different cpu we cannot safely probe
416 _tsleep_wakeup(struct thread
*td
)
419 globaldata_t gd
= mycpu
;
421 if (td
->td_gd
!= gd
) {
422 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)tsleep_wakeup
, td
);
427 if (td
->td_flags
& TDF_TSLEEP_DESCHEDULED
) {
428 td
->td_flags
&= ~TDF_TSLEEP_DESCHEDULED
;
436 tsleep_wakeup(struct thread
*td
)
444 * General sleep call. Suspends the current process until a wakeup is
445 * performed on the specified identifier. The process will then be made
446 * runnable with the specified priority. Sleeps at most timo/hz seconds
447 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
448 * before and after sleeping, else signals are not checked. Returns 0 if
449 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
450 * signal needs to be delivered, ERESTART is returned if the current system
451 * call should be restarted if possible, and EINTR is returned if the system
452 * call should be interrupted by the signal (return EINTR).
454 * Note that if we are a process, we release_curproc() before messing with
455 * the LWKT scheduler.
457 * During autoconfiguration or after a panic, a sleep will simply
458 * lower the priority briefly to allow interrupts, then return.
461 tsleep(void *ident
, int flags
, const char *wmesg
, int timo
)
463 struct thread
*td
= curthread
;
464 struct lwp
*lp
= td
->td_lwp
;
465 struct proc
*p
= td
->td_proc
; /* may be NULL */
472 struct callout thandle
;
475 * NOTE: removed KTRPOINT, it could cause races due to blocking
476 * even in stable. Just scrap it for now.
478 if (tsleep_now_works
== 0 || panicstr
) {
480 * After a panic, or before we actually have an operational
481 * softclock, just give interrupts a chance, then just return;
483 * don't run any other procs or panic below,
484 * in case this is the idle process and already asleep.
487 oldpri
= td
->td_pri
& TDPRI_MASK
;
488 lwkt_setpri_self(safepri
);
490 lwkt_setpri_self(oldpri
);
493 logtsleep2(tsleep_beg
, ident
);
495 KKASSERT(td
!= &gd
->gd_idlethread
); /* you must be kidding! */
498 * NOTE: all of this occurs on the current cpu, including any
499 * callout-based wakeups, so a critical section is a sufficient
502 * The entire sequence through to where we actually sleep must
503 * run without breaking the critical section.
505 catch = flags
& PCATCH
;
509 crit_enter_quick(td
);
511 KASSERT(ident
!= NULL
, ("tsleep: no ident"));
512 KASSERT(lp
== NULL
||
513 lp
->lwp_stat
== LSRUN
|| /* Obvious */
514 lp
->lwp_stat
== LSSTOP
, /* Set in tstop */
516 ident
, wmesg
, lp
->lwp_stat
));
519 * Setup for the current process (if this is a process).
524 * Early termination if PCATCH was set and a
525 * signal is pending, interlocked with the
528 * Early termination only occurs when tsleep() is
529 * entered while in a normal LSRUN state.
531 if ((sig
= CURSIG(lp
)) != 0)
535 * Early termination if PCATCH was set and a
536 * mailbox signal was possibly delivered prior to
537 * the system call even being made, in order to
538 * allow the user to interlock without having to
539 * make additional system calls.
541 if (p
->p_flag
& P_MAILBOX
)
545 * Causes ksignal to wake us up when.
547 lp
->lwp_flag
|= LWP_SINTR
;
552 * We interlock the sleep queue if the caller has not already done
555 if ((flags
& PINTERLOCKED
) == 0) {
557 _tsleep_interlock(gd
, ident
, flags
);
562 * If no interlock was set we do an integrated interlock here.
563 * Make sure the current process has been untangled from
564 * the userland scheduler and initialize slptime to start
565 * counting. We must interlock the sleep queue before doing
566 * this to avoid wakeup/process-ipi races which can occur under
570 p
->p_usched
->release_curproc(lp
);
575 * If the interlocked flag is set but our cpu bit in the slpqueue
576 * is no longer set, then a wakeup was processed inbetween the
577 * tsleep_interlock() (ours or the callers), and here. This can
578 * occur under numerous circumstances including when we release the
581 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
582 * to process incoming IPIs, thus draining incoming wakeups.
584 if ((td
->td_flags
& TDF_TSLEEPQ
) == 0) {
585 logtsleep2(ilockfail
, ident
);
590 * scheduling is blocked while in a critical section. Coincide
591 * the descheduled-by-tsleep flag with the descheduling of the
594 lwkt_deschedule_self(td
);
595 td
->td_flags
|= TDF_TSLEEP_DESCHEDULED
;
596 td
->td_wmesg
= wmesg
;
599 * Setup the timeout, if any
602 callout_init(&thandle
);
603 callout_reset(&thandle
, timo
, endtsleep
, td
);
611 * Ok, we are sleeping. Place us in the SSLEEP state.
613 KKASSERT((lp
->lwp_flag
& LWP_ONRUNQ
) == 0);
615 * tstop() sets LSSTOP, so don't fiddle with that.
617 if (lp
->lwp_stat
!= LSSTOP
)
618 lp
->lwp_stat
= LSSLEEP
;
619 lp
->lwp_ru
.ru_nvcsw
++;
623 * And when we are woken up, put us back in LSRUN. If we
624 * slept for over a second, recalculate our estcpu.
626 lp
->lwp_stat
= LSRUN
;
628 p
->p_usched
->recalculate(lp
);
635 * Make sure we haven't switched cpus while we were asleep. It's
636 * not supposed to happen. Cleanup our temporary flags.
638 KKASSERT(gd
== td
->td_gd
);
641 * Cleanup the timeout.
644 if (td
->td_flags
& TDF_TIMEOUT
) {
645 td
->td_flags
&= ~TDF_TIMEOUT
;
648 callout_stop(&thandle
);
653 * Make sure we have been removed from the sleepq. This should
654 * have been done for us already.
658 if (td
->td_flags
& TDF_TSLEEP_DESCHEDULED
) {
659 td
->td_flags
&= ~TDF_TSLEEP_DESCHEDULED
;
660 kprintf("td %p (%s) unexpectedly rescheduled\n",
665 * Figure out the correct error return. If interrupted by a
666 * signal we want to return EINTR or ERESTART.
668 * If P_MAILBOX is set no automatic system call restart occurs
669 * and we return EINTR. P_MAILBOX is meant to be used as an
670 * interlock, the user must poll it prior to any system call
671 * that it wishes to interlock a mailbox signal against since
672 * the flag is cleared on *any* system call that sleeps.
676 if (catch && error
== 0) {
677 if ((p
->p_flag
& P_MAILBOX
) && sig
== 0) {
679 } else if (sig
!= 0 || (sig
= CURSIG(lp
))) {
680 if (SIGISMEMBER(p
->p_sigacts
->ps_sigintr
, sig
))
686 lp
->lwp_flag
&= ~(LWP_BREAKTSLEEP
| LWP_SINTR
);
687 p
->p_flag
&= ~P_MAILBOX
;
689 logtsleep1(tsleep_end
);
695 * Interlocked spinlock sleep. An exclusively held spinlock must
696 * be passed to ssleep(). The function will atomically release the
697 * spinlock and tsleep on the ident, then reacquire the spinlock and
700 * This routine is fairly important along the critical path, so optimize it
704 ssleep(void *ident
, struct spinlock
*spin
, int flags
,
705 const char *wmesg
, int timo
)
707 globaldata_t gd
= mycpu
;
710 _tsleep_interlock(gd
, ident
, flags
);
711 spin_unlock_wr_quick(gd
, spin
);
712 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
713 spin_lock_wr_quick(gd
, spin
);
719 lksleep(void *ident
, struct lock
*lock
, int flags
,
720 const char *wmesg
, int timo
)
722 globaldata_t gd
= mycpu
;
725 _tsleep_interlock(gd
, ident
, flags
);
726 lockmgr(lock
, LK_RELEASE
);
727 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
728 lockmgr(lock
, LK_EXCLUSIVE
);
734 * Interlocked mutex sleep. An exclusively held mutex must be passed
735 * to mtxsleep(). The function will atomically release the mutex
736 * and tsleep on the ident, then reacquire the mutex and return.
739 mtxsleep(void *ident
, struct mtx
*mtx
, int flags
,
740 const char *wmesg
, int timo
)
742 globaldata_t gd
= mycpu
;
745 _tsleep_interlock(gd
, ident
, flags
);
747 error
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
748 mtx_lock_ex_quick(mtx
, wmesg
);
754 * Interlocked serializer sleep. An exclusively held serializer must
755 * be passed to zsleep(). The function will atomically release
756 * the serializer and tsleep on the ident, then reacquire the serializer
760 zsleep(void *ident
, struct lwkt_serialize
*slz
, int flags
,
761 const char *wmesg
, int timo
)
763 globaldata_t gd
= mycpu
;
766 ASSERT_SERIALIZED(slz
);
768 _tsleep_interlock(gd
, ident
, flags
);
769 lwkt_serialize_exit(slz
);
770 ret
= tsleep(ident
, flags
| PINTERLOCKED
, wmesg
, timo
);
771 lwkt_serialize_enter(slz
);
777 * Directly block on the LWKT thread by descheduling it. This
778 * is much faster then tsleep(), but the only legal way to wake
779 * us up is to directly schedule the thread.
781 * Setting TDF_SINTR will cause new signals to directly schedule us.
783 * This routine must be called while in a critical section.
786 lwkt_sleep(const char *wmesg
, int flags
)
788 thread_t td
= curthread
;
791 if ((flags
& PCATCH
) == 0 || td
->td_lwp
== NULL
) {
792 td
->td_flags
|= TDF_BLOCKED
;
793 td
->td_wmesg
= wmesg
;
794 lwkt_deschedule_self(td
);
797 td
->td_flags
&= ~TDF_BLOCKED
;
800 if ((sig
= CURSIG(td
->td_lwp
)) != 0) {
801 if (SIGISMEMBER(td
->td_proc
->p_sigacts
->ps_sigintr
, sig
))
807 td
->td_flags
|= TDF_BLOCKED
| TDF_SINTR
;
808 td
->td_wmesg
= wmesg
;
809 lwkt_deschedule_self(td
);
811 td
->td_flags
&= ~(TDF_BLOCKED
| TDF_SINTR
);
817 * Implement the timeout for tsleep.
819 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
820 * we only call setrunnable if the process is not stopped.
822 * This type of callout timeout is scheduled on the same cpu the process
823 * is sleeping on. Also, at the moment, the MP lock is held.
831 ASSERT_MP_LOCK_HELD(curthread
);
835 * cpu interlock. Thread flags are only manipulated on
836 * the cpu owning the thread. proc flags are only manipulated
837 * by the older of the MP lock. We have both.
839 if (td
->td_flags
& TDF_TSLEEP_DESCHEDULED
) {
840 td
->td_flags
|= TDF_TIMEOUT
;
842 if ((lp
= td
->td_lwp
) != NULL
) {
843 lp
->lwp_flag
|= LWP_BREAKTSLEEP
;
844 if (lp
->lwp_proc
->p_stat
!= SSTOP
)
854 * Make all processes sleeping on the specified identifier runnable.
855 * count may be zero or one only.
857 * The domain encodes the sleep/wakeup domain AND the first cpu to check
858 * (which is always the current cpu). As we iterate across cpus
860 * This call may run without the MP lock held. We can only manipulate thread
861 * state on the cpu owning the thread. We CANNOT manipulate process state
865 _wakeup(void *ident
, int domain
)
877 logtsleep2(wakeup_beg
, ident
);
880 qp
= &gd
->gd_tsleep_hash
[id
];
882 for (td
= TAILQ_FIRST(qp
); td
!= NULL
; td
= ntd
) {
883 ntd
= TAILQ_NEXT(td
, td_sleepq
);
884 if (td
->td_wchan
== ident
&&
885 td
->td_wdomain
== (domain
& PDOMAIN_MASK
)
887 KKASSERT(td
->td_gd
== gd
);
889 if (td
->td_flags
& TDF_TSLEEP_DESCHEDULED
) {
890 td
->td_flags
&= ~TDF_TSLEEP_DESCHEDULED
;
892 if (domain
& PWAKEUP_ONE
)
901 * We finished checking the current cpu but there still may be
902 * more work to do. Either wakeup_one was requested and no matching
903 * thread was found, or a normal wakeup was requested and we have
904 * to continue checking cpus.
906 * It should be noted that this scheme is actually less expensive then
907 * the old scheme when waking up multiple threads, since we send
908 * only one IPI message per target candidate which may then schedule
909 * multiple threads. Before we could have wound up sending an IPI
910 * message for each thread on the target cpu (!= current cpu) that
911 * needed to be woken up.
913 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
914 * should be ok since we are passing idents in the IPI rather then
917 if ((domain
& PWAKEUP_MYCPU
) == 0 &&
918 (mask
= slpque_cpumasks
[id
] & gd
->gd_other_cpus
) != 0) {
919 lwkt_send_ipiq2_mask(mask
, _wakeup
, ident
,
920 domain
| PWAKEUP_MYCPU
);
924 logtsleep1(wakeup_end
);
929 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
934 _wakeup(ident
, PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
));
938 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
941 wakeup_one(void *ident
)
943 /* XXX potentially round-robin the first responding cpu */
944 _wakeup(ident
, PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) | PWAKEUP_ONE
);
948 * Wakeup threads tsleep()ing on the specified ident on the current cpu
952 wakeup_mycpu(void *ident
)
954 _wakeup(ident
, PWAKEUP_MYCPU
);
958 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
962 wakeup_mycpu_one(void *ident
)
964 /* XXX potentially round-robin the first responding cpu */
965 _wakeup(ident
, PWAKEUP_MYCPU
|PWAKEUP_ONE
);
969 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
973 wakeup_oncpu(globaldata_t gd
, void *ident
)
977 _wakeup(ident
, PWAKEUP_MYCPU
);
979 lwkt_send_ipiq2(gd
, _wakeup
, ident
, PWAKEUP_MYCPU
);
982 _wakeup(ident
, PWAKEUP_MYCPU
);
987 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
991 wakeup_oncpu_one(globaldata_t gd
, void *ident
)
995 _wakeup(ident
, PWAKEUP_MYCPU
| PWAKEUP_ONE
);
997 lwkt_send_ipiq2(gd
, _wakeup
, ident
, PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1000 _wakeup(ident
, PWAKEUP_MYCPU
| PWAKEUP_ONE
);
1005 * Wakeup all threads waiting on the specified ident that slept using
1006 * the specified domain, on all cpus.
1009 wakeup_domain(void *ident
, int domain
)
1011 _wakeup(ident
, PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
));
1015 * Wakeup one thread waiting on the specified ident that slept using
1016 * the specified domain, on any cpu.
1019 wakeup_domain_one(void *ident
, int domain
)
1021 /* XXX potentially round-robin the first responding cpu */
1022 _wakeup(ident
, PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
) | PWAKEUP_ONE
);
1028 * Make a process runnable. The MP lock must be held on call. This only
1029 * has an effect if we are in SSLEEP. We only break out of the
1030 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
1032 * NOTE: With the MP lock held we can only safely manipulate the process
1033 * structure. We cannot safely manipulate the thread structure.
1036 setrunnable(struct lwp
*lp
)
1039 ASSERT_MP_LOCK_HELD(curthread
);
1040 if (lp
->lwp_stat
== LSSTOP
)
1041 lp
->lwp_stat
= LSSLEEP
;
1042 if (lp
->lwp_stat
== LSSLEEP
&& (lp
->lwp_flag
& LWP_BREAKTSLEEP
))
1043 _tsleep_wakeup(lp
->lwp_thread
);
1048 * The process is stopped due to some condition, usually because p_stat is
1049 * set to SSTOP, but also possibly due to being traced.
1051 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
1052 * because the parent may check the child's status before the child actually
1053 * gets to this routine.
1055 * This routine is called with the current lwp only, typically just
1056 * before returning to userland.
1058 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
1059 * SIGCONT to break out of the tsleep.
1064 struct lwp
*lp
= curthread
->td_lwp
;
1065 struct proc
*p
= lp
->lwp_proc
;
1069 * If LWP_WSTOP is set, we were sleeping
1070 * while our process was stopped. At this point
1071 * we were already counted as stopped.
1073 if ((lp
->lwp_flag
& LWP_WSTOP
) == 0) {
1075 * If we're the last thread to stop, signal
1079 lp
->lwp_flag
|= LWP_WSTOP
;
1080 wakeup(&p
->p_nstopped
);
1081 if (p
->p_nstopped
== p
->p_nthreads
) {
1082 p
->p_flag
&= ~P_WAITED
;
1084 if ((p
->p_pptr
->p_sigacts
->ps_flag
& PS_NOCLDSTOP
) == 0)
1085 ksignal(p
->p_pptr
, SIGCHLD
);
1088 while (p
->p_stat
== SSTOP
) {
1089 lp
->lwp_flag
|= LWP_BREAKTSLEEP
;
1090 lp
->lwp_stat
= LSSTOP
;
1091 tsleep(p
, 0, "stop", 0);
1094 lp
->lwp_flag
&= ~LWP_WSTOP
;
1099 * Yield / synchronous reschedule. This is a bit tricky because the trap
1100 * code might have set a lazy release on the switch function. Setting
1101 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
1102 * switch, and that we are given a greater chance of affinity with our
1105 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
1106 * run queue. lwkt_switch() will also execute any assigned passive release
1107 * (which usually calls release_curproc()), allowing a same/higher priority
1108 * process to be designated as the current process.
1110 * While it is possible for a lower priority process to be designated,
1111 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
1112 * round-robin back to us and we will be able to re-acquire the current
1113 * process designation.
1120 struct thread
*td
= curthread
;
1121 struct proc
*p
= td
->td_proc
;
1123 lwkt_setpri_self(td
->td_pri
& TDPRI_MASK
);
1125 p
->p_flag
|= P_PASSIVE_ACQ
;
1127 p
->p_flag
&= ~P_PASSIVE_ACQ
;
1134 * Compute a tenex style load average of a quantity on
1135 * 1, 5 and 15 minute intervals.
1137 static int loadav_count_runnable(struct lwp
*p
, void *data
);
1142 struct loadavg
*avg
;
1146 alllwp_scan(loadav_count_runnable
, &nrun
);
1148 for (i
= 0; i
< 3; i
++) {
1149 avg
->ldavg
[i
] = (cexp
[i
] * avg
->ldavg
[i
] +
1150 nrun
* FSCALE
* (FSCALE
- cexp
[i
])) >> FSHIFT
;
1154 * Schedule the next update to occur after 5 seconds, but add a
1155 * random variation to avoid synchronisation with processes that
1156 * run at regular intervals.
1158 callout_reset(&loadav_callout
, hz
* 4 + (int)(krandom() % (hz
* 2 + 1)),
1163 loadav_count_runnable(struct lwp
*lp
, void *data
)
1168 switch (lp
->lwp_stat
) {
1170 if ((td
= lp
->lwp_thread
) == NULL
)
1172 if (td
->td_flags
& TDF_BLOCKED
)
1184 sched_setup(void *dummy
)
1186 callout_init(&loadav_callout
);
1187 callout_init(&schedcpu_callout
);
1189 /* Kick off timeout driven events by calling first time. */