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 1.88 2007/12/06 15:03:25 corecode 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/signal2.h>
51 #include <sys/resourcevar.h>
52 #include <sys/vmmeter.h>
53 #include <sys/sysctl.h>
57 #include <sys/ktrace.h>
59 #include <sys/xwait.h>
62 #include <sys/thread2.h>
63 #include <sys/spinlock2.h>
65 #include <machine/cpu.h>
66 #include <machine/smp.h>
68 TAILQ_HEAD(tslpque
, thread
);
70 static void sched_setup (void *dummy
);
71 SYSINIT(sched_setup
, SI_SUB_KICK_SCHEDULER
, SI_ORDER_FIRST
, sched_setup
, NULL
)
76 int sched_quantum
; /* Roundrobin scheduling quantum in ticks. */
78 int ncpus2
, ncpus2_shift
, ncpus2_mask
;
79 int ncpus_fit
, ncpus_fit_mask
;
83 static struct callout loadav_callout
;
84 static struct callout schedcpu_callout
;
85 MALLOC_DEFINE(M_TSLEEP
, "tslpque", "tsleep queues");
87 #if !defined(KTR_TSLEEP)
88 #define KTR_TSLEEP KTR_ALL
90 KTR_INFO_MASTER(tsleep
);
91 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_beg
, 0, "tsleep enter %p", sizeof(void *));
92 KTR_INFO(KTR_TSLEEP
, tsleep
, tsleep_end
, 1, "tsleep exit", 0);
93 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_beg
, 2, "wakeup enter %p", sizeof(void *));
94 KTR_INFO(KTR_TSLEEP
, tsleep
, wakeup_end
, 3, "wakeup exit", 0);
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 unsleep_and_wakeup_thread(struct thread
*td
);
113 static void loadav (void *arg
);
114 static void schedcpu (void *arg
);
117 * Adjust the scheduler quantum. The quantum is specified in microseconds.
118 * Note that 'tick' is in microseconds per tick.
121 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS
)
125 new_val
= sched_quantum
* tick
;
126 error
= sysctl_handle_int(oidp
, &new_val
, 0, req
);
127 if (error
!= 0 || req
->newptr
== NULL
)
131 sched_quantum
= new_val
/ tick
;
132 hogticks
= 2 * sched_quantum
;
136 SYSCTL_PROC(_kern
, OID_AUTO
, quantum
, CTLTYPE_INT
|CTLFLAG_RW
,
137 0, sizeof sched_quantum
, sysctl_kern_quantum
, "I", "");
140 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
141 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
142 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
144 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
145 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
147 * If you don't want to bother with the faster/more-accurate formula, you
148 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
149 * (more general) method of calculating the %age of CPU used by a process.
151 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
153 #define CCPU_SHIFT 11
155 static fixpt_t ccpu
= 0.95122942450071400909 * FSCALE
; /* exp(-1/20) */
156 SYSCTL_INT(_kern
, OID_AUTO
, ccpu
, CTLFLAG_RD
, &ccpu
, 0, "");
159 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
161 int fscale __unused
= FSCALE
; /* exported to systat */
162 SYSCTL_INT(_kern
, OID_AUTO
, fscale
, CTLFLAG_RD
, 0, FSCALE
, "");
165 * Recompute process priorities, once a second.
167 * Since the userland schedulers are typically event oriented, if the
168 * estcpu calculation at wakeup() time is not sufficient to make a
169 * process runnable relative to other processes in the system we have
170 * a 1-second recalc to help out.
172 * This code also allows us to store sysclock_t data in the process structure
173 * without fear of an overrun, since sysclock_t are guarenteed to hold
174 * several seconds worth of count.
176 * WARNING! callouts can preempt normal threads. However, they will not
177 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
179 static int schedcpu_stats(struct proc
*p
, void *data __unused
);
180 static int schedcpu_resource(struct proc
*p
, void *data __unused
);
185 allproc_scan(schedcpu_stats
, NULL
);
186 allproc_scan(schedcpu_resource
, NULL
);
187 wakeup((caddr_t
)&lbolt
);
188 wakeup((caddr_t
)&lbolt_syncer
);
189 callout_reset(&schedcpu_callout
, hz
, schedcpu
, NULL
);
193 * General process statistics once a second
196 schedcpu_stats(struct proc
*p
, void *data __unused
)
202 FOREACH_LWP_IN_PROC(lp
, p
) {
203 if (lp
->lwp_stat
== LSSLEEP
)
207 * Only recalculate processes that are active or have slept
208 * less then 2 seconds. The schedulers understand this.
210 if (lp
->lwp_slptime
<= 1) {
211 p
->p_usched
->recalculate(lp
);
213 lp
->lwp_pctcpu
= (lp
->lwp_pctcpu
* ccpu
) >> FSHIFT
;
221 * Resource checks. XXX break out since ksignal/killproc can block,
222 * limiting us to one process killed per second. There is probably
226 schedcpu_resource(struct proc
*p
, void *data __unused
)
232 if (p
->p_stat
== SIDL
||
233 p
->p_stat
== SZOMB
||
241 FOREACH_LWP_IN_PROC(lp
, p
) {
243 * We may have caught an lp in the middle of being
244 * created, lwp_thread can be NULL.
246 if (lp
->lwp_thread
) {
247 ttime
+= lp
->lwp_thread
->td_sticks
;
248 ttime
+= lp
->lwp_thread
->td_uticks
;
252 switch(plimit_testcpulimit(p
->p_limit
, ttime
)) {
253 case PLIMIT_TESTCPU_KILL
:
254 killproc(p
, "exceeded maximum CPU limit");
256 case PLIMIT_TESTCPU_XCPU
:
257 if ((p
->p_flag
& P_XCPU
) == 0) {
270 * This is only used by ps. Generate a cpu percentage use over
271 * a period of one second.
276 updatepcpu(struct lwp
*lp
, int cpticks
, int ttlticks
)
281 acc
= (cpticks
<< FSHIFT
) / ttlticks
;
282 if (ttlticks
>= ESTCPUFREQ
) {
283 lp
->lwp_pctcpu
= acc
;
285 remticks
= ESTCPUFREQ
- ttlticks
;
286 lp
->lwp_pctcpu
= (acc
* ttlticks
+ lp
->lwp_pctcpu
* remticks
) /
292 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
293 * like addresses being slept on.
295 #define TABLESIZE 1024
296 #define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1))
298 static cpumask_t slpque_cpumasks
[TABLESIZE
];
301 * General scheduler initialization. We force a reschedule 25 times
302 * a second by default. Note that cpu0 is initialized in early boot and
303 * cannot make any high level calls.
305 * Each cpu has its own sleep queue.
308 sleep_gdinit(globaldata_t gd
)
310 static struct tslpque slpque_cpu0
[TABLESIZE
];
313 if (gd
->gd_cpuid
== 0) {
314 sched_quantum
= (hz
+ 24) / 25;
315 hogticks
= 2 * sched_quantum
;
317 gd
->gd_tsleep_hash
= slpque_cpu0
;
319 gd
->gd_tsleep_hash
= kmalloc(sizeof(slpque_cpu0
),
320 M_TSLEEP
, M_WAITOK
| M_ZERO
);
322 for (i
= 0; i
< TABLESIZE
; ++i
)
323 TAILQ_INIT(&gd
->gd_tsleep_hash
[i
]);
327 * General sleep call. Suspends the current process until a wakeup is
328 * performed on the specified identifier. The process will then be made
329 * runnable with the specified priority. Sleeps at most timo/hz seconds
330 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
331 * before and after sleeping, else signals are not checked. Returns 0 if
332 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
333 * signal needs to be delivered, ERESTART is returned if the current system
334 * call should be restarted if possible, and EINTR is returned if the system
335 * call should be interrupted by the signal (return EINTR).
337 * Note that if we are a process, we release_curproc() before messing with
338 * the LWKT scheduler.
340 * During autoconfiguration or after a panic, a sleep will simply
341 * lower the priority briefly to allow interrupts, then return.
344 tsleep(void *ident
, int flags
, const char *wmesg
, int timo
)
346 struct thread
*td
= curthread
;
347 struct lwp
*lp
= td
->td_lwp
;
348 struct proc
*p
= td
->td_proc
; /* may be NULL */
355 struct callout thandle
;
358 * NOTE: removed KTRPOINT, it could cause races due to blocking
359 * even in stable. Just scrap it for now.
361 if (tsleep_now_works
== 0 || panicstr
) {
363 * After a panic, or before we actually have an operational
364 * softclock, just give interrupts a chance, then just return;
366 * don't run any other procs or panic below,
367 * in case this is the idle process and already asleep.
370 oldpri
= td
->td_pri
& TDPRI_MASK
;
371 lwkt_setpri_self(safepri
);
373 lwkt_setpri_self(oldpri
);
376 logtsleep2(tsleep_beg
, ident
);
378 KKASSERT(td
!= &gd
->gd_idlethread
); /* you must be kidding! */
381 * NOTE: all of this occurs on the current cpu, including any
382 * callout-based wakeups, so a critical section is a sufficient
385 * The entire sequence through to where we actually sleep must
386 * run without breaking the critical section.
389 catch = flags
& PCATCH
;
393 crit_enter_quick(td
);
395 KASSERT(ident
!= NULL
, ("tsleep: no ident"));
396 KASSERT(lp
== NULL
||
397 lp
->lwp_stat
== LSRUN
|| /* Obvious */
398 lp
->lwp_stat
== LSSTOP
, /* Set in tstop */
400 ident
, wmesg
, lp
->lwp_stat
));
403 * Setup for the current process (if this is a process).
408 * Early termination if PCATCH was set and a
409 * signal is pending, interlocked with the
412 * Early termination only occurs when tsleep() is
413 * entered while in a normal LSRUN state.
415 if ((sig
= CURSIG(lp
)) != 0)
419 * Early termination if PCATCH was set and a
420 * mailbox signal was possibly delivered prior to
421 * the system call even being made, in order to
422 * allow the user to interlock without having to
423 * make additional system calls.
425 if (p
->p_flag
& P_MAILBOX
)
429 * Causes ksignal to wake us up when.
431 lp
->lwp_flag
|= LWP_SINTR
;
435 * Make sure the current process has been untangled from
436 * the userland scheduler and initialize slptime to start
439 if (flags
& PNORESCHED
)
440 td
->td_flags
|= TDF_NORESCHED
;
441 p
->p_usched
->release_curproc(lp
);
446 * Move our thread to the correct queue and setup our wchan, etc.
448 lwkt_deschedule_self(td
);
449 td
->td_flags
|= TDF_TSLEEPQ
;
450 TAILQ_INSERT_TAIL(&gd
->gd_tsleep_hash
[id
], td
, td_threadq
);
451 atomic_set_int(&slpque_cpumasks
[id
], gd
->gd_cpumask
);
453 td
->td_wchan
= ident
;
454 td
->td_wmesg
= wmesg
;
455 td
->td_wdomain
= flags
& PDOMAIN_MASK
;
458 * Setup the timeout, if any
461 callout_init(&thandle
);
462 callout_reset(&thandle
, timo
, endtsleep
, td
);
470 * Ok, we are sleeping. Place us in the SSLEEP state.
472 KKASSERT((lp
->lwp_flag
& LWP_ONRUNQ
) == 0);
474 * tstop() sets LSSTOP, so don't fiddle with that.
476 if (lp
->lwp_stat
!= LSSTOP
)
477 lp
->lwp_stat
= LSSLEEP
;
478 lp
->lwp_ru
.ru_nvcsw
++;
482 * And when we are woken up, put us back in LSRUN. If we
483 * slept for over a second, recalculate our estcpu.
485 lp
->lwp_stat
= LSRUN
;
487 p
->p_usched
->recalculate(lp
);
494 * Make sure we haven't switched cpus while we were asleep. It's
495 * not supposed to happen. Cleanup our temporary flags.
497 KKASSERT(gd
== td
->td_gd
);
498 td
->td_flags
&= ~TDF_NORESCHED
;
501 * Cleanup the timeout.
504 if (td
->td_flags
& TDF_TIMEOUT
) {
505 td
->td_flags
&= ~TDF_TIMEOUT
;
508 callout_stop(&thandle
);
513 * Since td_threadq is used both for our run queue AND for the
514 * tsleep hash queue, we can't still be on it at this point because
515 * we've gotten cpu back.
517 KASSERT((td
->td_flags
& TDF_TSLEEPQ
) == 0, ("tsleep: impossible thread flags %08x", td
->td_flags
));
523 * Figure out the correct error return. If interrupted by a
524 * signal we want to return EINTR or ERESTART.
526 * If P_MAILBOX is set no automatic system call restart occurs
527 * and we return EINTR. P_MAILBOX is meant to be used as an
528 * interlock, the user must poll it prior to any system call
529 * that it wishes to interlock a mailbox signal against since
530 * the flag is cleared on *any* system call that sleeps.
534 if (catch && error
== 0) {
535 if ((p
->p_flag
& P_MAILBOX
) && sig
== 0) {
537 } else if (sig
!= 0 || (sig
= CURSIG(lp
))) {
538 if (SIGISMEMBER(p
->p_sigacts
->ps_sigintr
, sig
))
544 lp
->lwp_flag
&= ~(LWP_BREAKTSLEEP
| LWP_SINTR
);
545 p
->p_flag
&= ~P_MAILBOX
;
547 logtsleep1(tsleep_end
);
553 * This is a dandy function that allows us to interlock tsleep/wakeup
554 * operations with unspecified upper level locks, such as lockmgr locks,
555 * simply by holding a critical section. The sequence is:
557 * (enter critical section)
558 * (acquire upper level lock)
559 * tsleep_interlock(blah)
560 * (release upper level lock)
562 * (exit critical section)
564 * Basically this function sets our cpumask for the ident which informs
565 * other cpus that our cpu 'might' be waiting (or about to wait on) the
566 * hash index related to the ident. The critical section prevents another
567 * cpu's wakeup() from being processed on our cpu until we are actually
568 * able to enter the tsleep(). Thus, no race occurs between our attempt
569 * to release a resource and sleep, and another cpu's attempt to acquire
570 * a resource and call wakeup.
572 * There isn't much of a point to this function unless you call it while
573 * holding a critical section.
576 _tsleep_interlock(globaldata_t gd
, void *ident
)
578 int id
= LOOKUP(ident
);
580 atomic_set_int(&slpque_cpumasks
[id
], gd
->gd_cpumask
);
584 tsleep_interlock(void *ident
)
586 _tsleep_interlock(mycpu
, ident
);
590 * Interlocked spinlock sleep. An exclusively held spinlock must
591 * be passed to msleep(). The function will atomically release the
592 * spinlock and tsleep on the ident, then reacquire the spinlock and
595 * This routine is fairly important along the critical path, so optimize it
599 msleep(void *ident
, struct spinlock
*spin
, int flags
,
600 const char *wmesg
, int timo
)
602 globaldata_t gd
= mycpu
;
606 _tsleep_interlock(gd
, ident
);
607 spin_unlock_wr_quick(gd
, spin
);
608 error
= tsleep(ident
, flags
, wmesg
, timo
);
609 spin_lock_wr_quick(gd
, spin
);
616 * Directly block on the LWKT thread by descheduling it. This
617 * is much faster then tsleep(), but the only legal way to wake
618 * us up is to directly schedule the thread.
620 * Setting TDF_SINTR will cause new signals to directly schedule us.
622 * This routine is typically called while in a critical section.
625 lwkt_sleep(const char *wmesg
, int flags
)
627 thread_t td
= curthread
;
630 if ((flags
& PCATCH
) == 0 || td
->td_lwp
== NULL
) {
631 td
->td_flags
|= TDF_BLOCKED
;
632 td
->td_wmesg
= wmesg
;
633 lwkt_deschedule_self(td
);
636 td
->td_flags
&= ~TDF_BLOCKED
;
639 if ((sig
= CURSIG(td
->td_lwp
)) != 0) {
640 if (SIGISMEMBER(td
->td_proc
->p_sigacts
->ps_sigintr
, sig
))
646 td
->td_flags
|= TDF_BLOCKED
| TDF_SINTR
;
647 td
->td_wmesg
= wmesg
;
648 lwkt_deschedule_self(td
);
650 td
->td_flags
&= ~(TDF_BLOCKED
| TDF_SINTR
);
656 * Implement the timeout for tsleep.
658 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
659 * we only call setrunnable if the process is not stopped.
661 * This type of callout timeout is scheduled on the same cpu the process
662 * is sleeping on. Also, at the moment, the MP lock is held.
670 ASSERT_MP_LOCK_HELD(curthread
);
674 * cpu interlock. Thread flags are only manipulated on
675 * the cpu owning the thread. proc flags are only manipulated
676 * by the older of the MP lock. We have both.
678 if (td
->td_flags
& TDF_TSLEEPQ
) {
679 td
->td_flags
|= TDF_TIMEOUT
;
681 if ((lp
= td
->td_lwp
) != NULL
) {
682 lp
->lwp_flag
|= LWP_BREAKTSLEEP
;
683 if (lp
->lwp_proc
->p_stat
!= SSTOP
)
686 unsleep_and_wakeup_thread(td
);
693 * Unsleep and wakeup a thread. This function runs without the MP lock
694 * which means that it can only manipulate thread state on the owning cpu,
695 * and cannot touch the process state at all.
699 unsleep_and_wakeup_thread(struct thread
*td
)
701 globaldata_t gd
= mycpu
;
705 if (td
->td_gd
!= gd
) {
706 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)unsleep_and_wakeup_thread
, td
);
711 if (td
->td_flags
& TDF_TSLEEPQ
) {
712 td
->td_flags
&= ~TDF_TSLEEPQ
;
713 id
= LOOKUP(td
->td_wchan
);
714 TAILQ_REMOVE(&gd
->gd_tsleep_hash
[id
], td
, td_threadq
);
715 if (TAILQ_FIRST(&gd
->gd_tsleep_hash
[id
]) == NULL
)
716 atomic_clear_int(&slpque_cpumasks
[id
], gd
->gd_cpumask
);
723 * Make all processes sleeping on the specified identifier runnable.
724 * count may be zero or one only.
726 * The domain encodes the sleep/wakeup domain AND the first cpu to check
727 * (which is always the current cpu). As we iterate across cpus
729 * This call may run without the MP lock held. We can only manipulate thread
730 * state on the cpu owning the thread. We CANNOT manipulate process state
734 _wakeup(void *ident
, int domain
)
749 logtsleep2(wakeup_beg
, ident
);
752 qp
= &gd
->gd_tsleep_hash
[id
];
754 for (td
= TAILQ_FIRST(qp
); td
!= NULL
; td
= ntd
) {
755 ntd
= TAILQ_NEXT(td
, td_threadq
);
756 if (td
->td_wchan
== ident
&&
757 td
->td_wdomain
== (domain
& PDOMAIN_MASK
)
759 KKASSERT(td
->td_flags
& TDF_TSLEEPQ
);
760 td
->td_flags
&= ~TDF_TSLEEPQ
;
761 TAILQ_REMOVE(qp
, td
, td_threadq
);
762 if (TAILQ_FIRST(qp
) == NULL
) {
763 atomic_clear_int(&slpque_cpumasks
[id
],
767 if (domain
& PWAKEUP_ONE
)
775 * We finished checking the current cpu but there still may be
776 * more work to do. Either wakeup_one was requested and no matching
777 * thread was found, or a normal wakeup was requested and we have
778 * to continue checking cpus.
780 * The cpu that started the wakeup sequence is encoded in the domain.
781 * We use this information to determine which cpus still need to be
782 * checked, locate a candidate cpu, and chain the wakeup
783 * asynchronously with an IPI message.
785 * It should be noted that this scheme is actually less expensive then
786 * the old scheme when waking up multiple threads, since we send
787 * only one IPI message per target candidate which may then schedule
788 * multiple threads. Before we could have wound up sending an IPI
789 * message for each thread on the target cpu (!= current cpu) that
790 * needed to be woken up.
792 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
793 * should be ok since we are passing idents in the IPI rather then
796 if ((domain
& PWAKEUP_MYCPU
) == 0 &&
797 (mask
= slpque_cpumasks
[id
]) != 0
800 * Look for a cpu that might have work to do. Mask out cpus
801 * which have already been processed.
803 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
805 * start currentcpu start
808 * 11111111111111110000000000000111 case1
809 * 00000000111111110000000000000000 case2
811 * case1: We started at start_case1 and processed through
812 * to the current cpu. We have to check any bits
813 * after the current cpu, then check bits before
816 * case2: We have already checked all the bits from
817 * start_case2 to the end, and from 0 to the current
818 * cpu. We just have the bits from the current cpu
819 * to start_case2 left to check.
821 startcpu
= PWAKEUP_DECODE(domain
);
822 if (gd
->gd_cpuid
>= startcpu
) {
826 tmask
= mask
& ~((gd
->gd_cpumask
<< 1) - 1);
828 nextcpu
= bsfl(mask
& tmask
);
829 lwkt_send_ipiq2(globaldata_find(nextcpu
),
830 _wakeup
, ident
, domain
);
832 tmask
= (1 << startcpu
) - 1;
834 nextcpu
= bsfl(mask
& tmask
);
836 globaldata_find(nextcpu
),
837 _wakeup
, ident
, domain
);
844 tmask
= ~((gd
->gd_cpumask
<< 1) - 1) &
845 ((1 << startcpu
) - 1);
847 nextcpu
= bsfl(mask
& tmask
);
848 lwkt_send_ipiq2(globaldata_find(nextcpu
),
849 _wakeup
, ident
, domain
);
855 logtsleep1(wakeup_end
);
860 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
865 _wakeup(ident
, PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
));
869 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
872 wakeup_one(void *ident
)
874 /* XXX potentially round-robin the first responding cpu */
875 _wakeup(ident
, PWAKEUP_ENCODE(0, mycpu
->gd_cpuid
) | PWAKEUP_ONE
);
879 * Wakeup threads tsleep()ing on the specified ident on the current cpu
883 wakeup_mycpu(void *ident
)
885 _wakeup(ident
, PWAKEUP_MYCPU
);
889 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
893 wakeup_mycpu_one(void *ident
)
895 /* XXX potentially round-robin the first responding cpu */
896 _wakeup(ident
, PWAKEUP_MYCPU
|PWAKEUP_ONE
);
900 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
904 wakeup_oncpu(globaldata_t gd
, void *ident
)
908 _wakeup(ident
, PWAKEUP_MYCPU
);
910 lwkt_send_ipiq2(gd
, _wakeup
, ident
, PWAKEUP_MYCPU
);
913 _wakeup(ident
, PWAKEUP_MYCPU
);
918 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
922 wakeup_oncpu_one(globaldata_t gd
, void *ident
)
926 _wakeup(ident
, PWAKEUP_MYCPU
| PWAKEUP_ONE
);
928 lwkt_send_ipiq2(gd
, _wakeup
, ident
, PWAKEUP_MYCPU
| PWAKEUP_ONE
);
931 _wakeup(ident
, PWAKEUP_MYCPU
| PWAKEUP_ONE
);
936 * Wakeup all threads waiting on the specified ident that slept using
937 * the specified domain, on all cpus.
940 wakeup_domain(void *ident
, int domain
)
942 _wakeup(ident
, PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
));
946 * Wakeup one thread waiting on the specified ident that slept using
947 * the specified domain, on any cpu.
950 wakeup_domain_one(void *ident
, int domain
)
952 /* XXX potentially round-robin the first responding cpu */
953 _wakeup(ident
, PWAKEUP_ENCODE(domain
, mycpu
->gd_cpuid
) | PWAKEUP_ONE
);
959 * Make a process runnable. The MP lock must be held on call. This only
960 * has an effect if we are in SSLEEP. We only break out of the
961 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
963 * NOTE: With the MP lock held we can only safely manipulate the process
964 * structure. We cannot safely manipulate the thread structure.
967 setrunnable(struct lwp
*lp
)
970 ASSERT_MP_LOCK_HELD(curthread
);
971 if (lp
->lwp_stat
== LSSTOP
)
972 lp
->lwp_stat
= LSSLEEP
;
973 if (lp
->lwp_stat
== LSSLEEP
&& (lp
->lwp_flag
& LWP_BREAKTSLEEP
))
974 unsleep_and_wakeup_thread(lp
->lwp_thread
);
979 * The process is stopped due to some condition, usually because p_stat is
980 * set to SSTOP, but also possibly due to being traced.
982 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
983 * because the parent may check the child's status before the child actually
984 * gets to this routine.
986 * This routine is called with the current lwp only, typically just
987 * before returning to userland.
989 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
990 * SIGCONT to break out of the tsleep.
995 struct lwp
*lp
= curthread
->td_lwp
;
996 struct proc
*p
= lp
->lwp_proc
;
998 lp
->lwp_flag
|= LWP_BREAKTSLEEP
;
999 lp
->lwp_stat
= LSSTOP
;
1002 * If LWP_WSTOP is set, we were sleeping
1003 * while our process was stopped. At this point
1004 * we were already counted as stopped.
1006 if ((lp
->lwp_flag
& LWP_WSTOP
) == 0) {
1008 * If we're the last thread to stop, signal
1012 lp
->lwp_flag
|= LWP_WSTOP
;
1013 if (p
->p_nstopped
== p
->p_nthreads
) {
1014 p
->p_flag
&= ~P_WAITED
;
1016 if ((p
->p_pptr
->p_sigacts
->ps_flag
& PS_NOCLDSTOP
) == 0)
1017 ksignal(p
->p_pptr
, SIGCHLD
);
1020 tsleep(lp
->lwp_proc
, 0, "stop", 0);
1022 lp
->lwp_flag
&= ~LWP_WSTOP
;
1027 * Yield / synchronous reschedule. This is a bit tricky because the trap
1028 * code might have set a lazy release on the switch function. Setting
1029 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
1030 * switch, and that we are given a greater chance of affinity with our
1033 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
1034 * run queue. lwkt_switch() will also execute any assigned passive release
1035 * (which usually calls release_curproc()), allowing a same/higher priority
1036 * process to be designated as the current process.
1038 * While it is possible for a lower priority process to be designated,
1039 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
1040 * round-robin back to us and we will be able to re-acquire the current
1041 * process designation.
1046 struct thread
*td
= curthread
;
1047 struct proc
*p
= td
->td_proc
;
1049 lwkt_setpri_self(td
->td_pri
& TDPRI_MASK
);
1051 p
->p_flag
|= P_PASSIVE_ACQ
;
1053 p
->p_flag
&= ~P_PASSIVE_ACQ
;
1060 * Compute a tenex style load average of a quantity on
1061 * 1, 5 and 15 minute intervals.
1063 static int loadav_count_runnable(struct lwp
*p
, void *data
);
1068 struct loadavg
*avg
;
1072 alllwp_scan(loadav_count_runnable
, &nrun
);
1074 for (i
= 0; i
< 3; i
++) {
1075 avg
->ldavg
[i
] = (cexp
[i
] * avg
->ldavg
[i
] +
1076 nrun
* FSCALE
* (FSCALE
- cexp
[i
])) >> FSHIFT
;
1080 * Schedule the next update to occur after 5 seconds, but add a
1081 * random variation to avoid synchronisation with processes that
1082 * run at regular intervals.
1084 callout_reset(&loadav_callout
, hz
* 4 + (int)(krandom() % (hz
* 2 + 1)),
1089 loadav_count_runnable(struct lwp
*lp
, void *data
)
1094 switch (lp
->lwp_stat
) {
1096 if ((td
= lp
->lwp_thread
) == NULL
)
1098 if (td
->td_flags
& TDF_BLOCKED
)
1110 sched_setup(void *dummy
)
1112 callout_init(&loadav_callout
);
1113 callout_init(&schedcpu_callout
);
1115 /* Kick off timeout driven events by calling first time. */