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 1.34 2004/07/24 20:21:35 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>
53 #include <sys/thread2.h>
56 #include <sys/ktrace.h>
58 #include <sys/xwait.h>
60 #include <machine/cpu.h>
61 #include <machine/ipl.h>
62 #include <machine/smp.h>
64 static void sched_setup (void *dummy);
65 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
69 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
71 int ncpus2, ncpus2_shift, ncpus2_mask;
73 static struct callout loadav_callout;
75 struct loadavg averunnable =
76 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
78 * Constants for averages over 1, 5, and 15 minutes
79 * when sampling at 5 second intervals.
81 static fixpt_t cexp[3] = {
82 0.9200444146293232 * FSCALE, /* exp(-1/12) */
83 0.9834714538216174 * FSCALE, /* exp(-1/60) */
84 0.9944598480048967 * FSCALE, /* exp(-1/180) */
87 static void endtsleep (void *);
88 static void loadav (void *arg);
89 static void roundrobin (void *arg);
90 static void schedcpu (void *arg);
91 static void updatepri (struct proc *p);
92 static void crit_panicints(void);
95 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
99 new_val = sched_quantum * tick;
100 error = sysctl_handle_int(oidp, &new_val, 0, req);
101 if (error != 0 || req->newptr == NULL)
105 sched_quantum = new_val / tick;
106 hogticks = 2 * sched_quantum;
110 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
111 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
114 roundrobin_interval(void)
116 return (sched_quantum);
120 * Force switch among equal priority processes every 100ms.
122 * WARNING! The MP lock is not held on ipi message remotes.
127 roundrobin_remote(void *arg)
129 struct proc *p = lwkt_preempted_proc();
130 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
137 roundrobin(void *arg)
139 struct proc *p = lwkt_preempted_proc();
140 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
143 lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL);
145 timeout(roundrobin, NULL, sched_quantum);
151 resched_cpus(u_int32_t mask)
153 lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL);
159 * The load average is scaled by FSCALE (2048 typ). The estimated cpu is
160 * incremented at a rate of ESTCPUVFREQ per second (40hz typ), but this is
161 * divided up across all cpu bound processes running in the system so an
162 * individual process will get less under load. ESTCPULIM typicaly caps
163 * out at ESTCPUMAX (around 376, or 11 nice levels).
165 * Generally speaking the decay equation needs to break-even on growth
166 * at the limit at all load levels >= 1.0, so if the estimated cpu for
167 * a process increases by (ESTVCPUFREQ / load) per second, then the decay
168 * should reach this value when estcpu reaches ESTCPUMAX. That calculation
171 * ESTCPUMAX * decay = ESTCPUVFREQ / load
172 * decay = ESTCPUVFREQ / (load * ESTCPUMAX)
173 * decay = estcpu * 0.053 / load
175 * If the load is less then 1.0 we assume a load of 1.0.
178 #define cload(loadav) ((loadav) < FSCALE ? FSCALE : (loadav))
179 #define decay_cpu(loadav,estcpu) \
180 ((estcpu) * (FSCALE * ESTCPUVFREQ / ESTCPUMAX) / cload(loadav))
182 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
183 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
184 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
186 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
187 static int fscale __unused = FSCALE;
188 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
191 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
192 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
193 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
195 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
196 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
198 * If you don't want to bother with the faster/more-accurate formula, you
199 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
200 * (more general) method of calculating the %age of CPU used by a process.
202 #define CCPU_SHIFT 11
205 * Recompute process priorities, once a second.
211 fixpt_t loadfac = averunnable.ldavg[0];
216 FOREACH_PROC_IN_SYSTEM(p) {
218 * Increment time in/out of memory and sleep time
219 * (if sleeping). We ignore overflow; with 16-bit int's
220 * (remember them?) overflow takes 45 days.
223 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
225 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
228 * If the process has slept the entire second,
229 * stop recalculating its priority until it wakes up.
231 * Note that interactive calculations do not occur for
232 * long sleeps (because that isn't necessarily indicative
233 * of an interactive process).
235 if (p->p_slptime > 1)
237 /* prevent state changes and protect run queue */
240 * p_cpticks runs at ESTCPUFREQ but must be divided by the
241 * load average for par-100% use. Higher p_interactive
242 * values mean less interactive, lower values mean more
245 if ((((fixpt_t)p->p_cpticks * cload(loadfac)) >> FSHIFT) >
247 if (p->p_interactive < 127)
250 if (p->p_interactive > -127)
254 * p_pctcpu is only for ps.
256 #if (FSHIFT >= CCPU_SHIFT)
257 p->p_pctcpu += (ESTCPUFREQ == 100)?
258 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
259 100 * (((fixpt_t) p->p_cpticks)
260 << (FSHIFT - CCPU_SHIFT)) / ESTCPUFREQ;
262 p->p_pctcpu += ((FSCALE - ccpu) *
263 (p->p_cpticks * FSCALE / ESTCPUFREQ)) >> FSHIFT;
266 ndecay = decay_cpu(loadfac, p->p_estcpu);
267 if (p->p_estcpu > ndecay)
268 p->p_estcpu -= ndecay;
274 wakeup((caddr_t)&lbolt);
275 timeout(schedcpu, (void *)0, hz);
279 * Recalculate the priority of a process after it has slept for a while.
280 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
281 * least six times the loadfactor will decay p_estcpu to zero.
284 updatepri(struct proc *p)
288 ndecay = decay_cpu(averunnable.ldavg[0], p->p_estcpu) * p->p_slptime;
289 if (p->p_estcpu > ndecay)
290 p->p_estcpu -= ndecay;
297 * We're only looking at 7 bits of the address; everything is
298 * aligned to 4, lots of things are aligned to greater powers
299 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
301 #define TABLESIZE 128
302 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
303 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
306 * During autoconfiguration or after a panic, a sleep will simply
307 * lower the priority briefly to allow interrupts, then return.
308 * The priority to be used (safepri) is machine-dependent, thus this
309 * value is initialized and maintained in the machine-dependent layers.
310 * This priority will typically be 0, or the lowest priority
311 * that is safe for use on the interrupt stack; it can be made
312 * higher to block network software interrupts after panics.
321 sched_quantum = hz/10;
322 hogticks = 2 * sched_quantum;
323 for (i = 0; i < TABLESIZE; i++)
324 TAILQ_INIT(&slpque[i]);
328 * General sleep call. Suspends the current process until a wakeup is
329 * performed on the specified identifier. The process will then be made
330 * runnable with the specified priority. Sleeps at most timo/hz seconds
331 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
332 * before and after sleeping, else signals are not checked. Returns 0 if
333 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
334 * signal needs to be delivered, ERESTART is returned if the current system
335 * call should be restarted if possible, and EINTR is returned if the system
336 * call should be interrupted by the signal (return EINTR).
338 * Note that if we are a process, we release_curproc() before messing with
339 * the LWKT scheduler.
342 tsleep(void *ident, int flags, const char *wmesg, int timo)
344 struct thread *td = curthread;
345 struct proc *p = td->td_proc; /* may be NULL */
346 int sig = 0, catch = flags & PCATCH;
347 int id = LOOKUP(ident);
348 struct callout_handle thandle;
351 * NOTE: removed KTRPOINT, it could cause races due to blocking
352 * even in stable. Just scrap it for now.
354 if (cold || panicstr) {
356 * After a panic, or during autoconfiguration,
357 * just give interrupts a chance, then just return;
358 * don't run any other procs or panic below,
359 * in case this is the idle process and already asleep.
364 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */
365 crit_enter_quick(td);
366 KASSERT(ident != NULL, ("tsleep: no ident"));
367 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
368 ident, wmesg, p->p_stat));
370 td->td_wchan = ident;
371 td->td_wmesg = wmesg;
373 if (flags & PNORESCHED)
374 td->td_flags |= TDF_NORESCHED;
378 lwkt_deschedule_self(td);
379 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
381 thandle = timeout(endtsleep, (void *)td, timo);
383 * We put ourselves on the sleep queue and start our timeout
384 * before calling CURSIG, as we could stop there, and a wakeup
385 * or a SIGCONT (or both) could occur while we were stopped.
386 * A SIGCONT would cause us to be marked as SSLEEP
387 * without resuming us, thus we must be ready for sleep
388 * when CURSIG is called. If the wakeup happens while we're
389 * stopped, td->td_wchan will be 0 upon return from CURSIG.
393 p->p_flag |= P_SINTR;
394 if ((sig = CURSIG(p))) {
397 lwkt_schedule_self(td);
402 if (td->td_wchan == NULL) {
411 * If we are not the current process we have to remove ourself
412 * from the run queue.
414 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
416 * If this is the current 'user' process schedule another one.
418 clrrunnable(p, SSLEEP);
419 p->p_stats->p_ru.ru_nvcsw++;
421 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
427 p->p_flag &= ~P_SINTR;
429 td->td_flags &= ~TDF_NORESCHED;
430 if (td->td_flags & TDF_TIMEOUT) {
431 td->td_flags &= ~TDF_TIMEOUT;
433 return (EWOULDBLOCK);
435 untimeout(endtsleep, (void *)td, thandle);
436 } else if (td->td_wmesg) {
438 * This can happen if a thread is woken up directly. Clear
439 * wmesg to avoid debugging confusion.
443 /* inline of iscaught() */
445 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
446 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
455 * Implement the timeout for tsleep. We interlock against
456 * wchan when setting TDF_TIMEOUT. For processes we remove
457 * the sleep if the process is stopped rather then sleeping,
458 * so it remains stopped.
468 td->td_flags |= TDF_TIMEOUT;
469 if ((p = td->td_proc) != NULL) {
470 if (p->p_stat == SSLEEP)
483 * Remove a process from its wait queue
486 unsleep(struct thread *td)
491 if (p->p_flag & P_XSLEEP) {
492 struct xwait *w = p->p_wchan;
493 TAILQ_REMOVE(&w->waitq, p, p_procq);
494 p->p_flag &= ~P_XSLEEP;
497 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
505 * Make all processes sleeping on the explicit lock structure runnable.
508 xwakeup(struct xwait *w)
514 while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
515 TAILQ_REMOVE(&w->waitq, p, p_procq);
516 KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
517 ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
519 p->p_flag &= ~P_XSLEEP;
520 if (p->p_stat == SSLEEP) {
521 /* OPTIMIZED EXPANSION OF setrunnable(p); */
522 if (p->p_slptime > 1)
526 if (p->p_flag & P_INMEM) {
529 p->p_flag |= P_SWAPINREQ;
530 wakeup((caddr_t)&proc0);
539 * Make all processes sleeping on the specified identifier runnable.
542 _wakeup(void *ident, int count)
544 struct slpquehead *qp;
548 int id = LOOKUP(ident);
553 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
554 ntd = TAILQ_NEXT(td, td_threadq);
555 if (td->td_wchan == ident) {
556 TAILQ_REMOVE(qp, td, td_threadq);
558 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
559 /* OPTIMIZED EXPANSION OF setrunnable(p); */
560 if (p->p_slptime > 1)
564 if (p->p_flag & P_INMEM) {
566 * LWKT scheduled now, there is no
567 * userland runq interaction until
568 * the thread tries to return to user
575 p->p_flag |= P_SWAPINREQ;
576 wakeup((caddr_t)&proc0);
578 /* END INLINE EXPANSION */
579 } else if (p == NULL) {
597 wakeup_one(void *ident)
603 * The machine independent parts of mi_switch().
605 * 'p' must be the current process.
608 mi_switch(struct proc *p)
610 thread_t td = p->p_thread;
614 KKASSERT(td == mycpu->gd_curthread);
616 crit_enter_quick(td);
619 * Check if the process exceeds its cpu resource allocation.
620 * If over max, kill it. Time spent in interrupts is not
621 * included. YYY 64 bit match is expensive. Ick.
623 ttime = td->td_sticks + td->td_uticks;
624 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
625 ttime > p->p_limit->p_cpulimit) {
626 rlim = &p->p_rlimit[RLIMIT_CPU];
627 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
628 killproc(p, "exceeded maximum CPU limit");
631 if (rlim->rlim_cur < rlim->rlim_max) {
632 /* XXX: we should make a private copy */
639 * If we are in a SSTOPped state we deschedule ourselves.
640 * YYY this needs to be cleaned up, remember that LWKTs stay on
641 * their run queue which works differently then the user scheduler
642 * which removes the process from the runq when it runs it.
644 mycpu->gd_cnt.v_swtch++;
645 if (p->p_stat == SSTOP)
646 lwkt_deschedule_self(td);
652 * Change process state to be runnable,
653 * placing it on the run queue if it is in memory,
654 * and awakening the swapper if it isn't in memory.
657 setrunnable(struct proc *p)
667 panic("setrunnable");
670 unsleep(p->p_thread); /* e.g. when sending signals */
679 * The process is controlled by LWKT at this point, we do not mess
680 * around with the userland scheduler until the thread tries to
681 * return to user mode.
684 if (p->p_flag & P_INMEM)
687 if (p->p_flag & P_INMEM)
688 lwkt_schedule(p->p_thread);
690 if (p->p_slptime > 1)
693 if ((p->p_flag & P_INMEM) == 0) {
694 p->p_flag |= P_SWAPINREQ;
695 wakeup((caddr_t)&proc0);
700 * Change the process state to NOT be runnable, removing it from the run
704 clrrunnable(struct proc *p, int stat)
706 crit_enter_quick(p->p_thread);
707 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ))
710 crit_exit_quick(p->p_thread);
714 * Compute the priority of a process when running in user mode.
715 * Arrange to reschedule if the resulting priority is better
716 * than that of the current process.
719 resetpriority(struct proc *p)
727 * Set p_priority for general process comparisons
729 switch(p->p_rtprio.type) {
730 case RTP_PRIO_REALTIME:
731 p->p_priority = PRIBASE_REALTIME + p->p_rtprio.prio;
733 case RTP_PRIO_NORMAL:
736 p->p_priority = PRIBASE_IDLE + p->p_rtprio.prio;
738 case RTP_PRIO_THREAD:
739 p->p_priority = PRIBASE_THREAD + p->p_rtprio.prio;
744 * NORMAL priorities fall through. These are based on niceness
745 * and cpu use. Lower numbers == higher priorities.
747 newpriority = (int)(NICE_ADJUST(p->p_nice - PRIO_MIN) +
748 p->p_estcpu / ESTCPURAMP);
751 * p_interactive is -128 to +127 and represents very long term
752 * interactivity or batch (whereas estcpu is a much faster variable).
753 * Interactivity can modify the priority by up to 8 units either way.
754 * (8 units == approximately 4 nice levels).
756 interactive = p->p_interactive / 10;
757 newpriority += interactive;
759 newpriority = min(newpriority, MAXPRI);
760 newpriority = max(newpriority, 0);
761 npq = newpriority / PPQ;
763 opq = (p->p_priority & PRIMASK) / PPQ;
764 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) {
766 * We have to move the process to another queue
769 p->p_priority = PRIBASE_NORMAL + newpriority;
773 * We can just adjust the priority and it will be picked
776 KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0);
777 p->p_priority = PRIBASE_NORMAL + newpriority;
783 * Compute a tenex style load average of a quantity on
784 * 1, 5 and 15 minute intervals.
796 FOREACH_PROC_IN_SYSTEM(p) {
799 if ((td = p->p_thread) == NULL)
801 if (td->td_flags & TDF_BLOCKED)
811 for (i = 0; i < 3; i++)
812 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
813 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
816 * Schedule the next update to occur after 5 seconds, but add a
817 * random variation to avoid synchronisation with processes that
818 * run at regular intervals.
820 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
826 sched_setup(void *dummy)
829 callout_init(&loadav_callout);
831 /* Kick off timeout driven events by calling first time. */
838 * We adjust the priority of the current process. The priority of
839 * a process gets worse as it accumulates CPU time. The cpu usage
840 * estimator (p_estcpu) is increased here. resetpriority() will
841 * compute a different priority each time p_estcpu increases by
842 * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached).
844 * The cpu usage estimator ramps up quite quickly when the process is
845 * running (linearly), and decays away exponentially, at a rate which
846 * is proportionally slower when the system is busy. The basic principle
847 * is that the system will 90% forget that the process used a lot of CPU
848 * time in 5 * loadav seconds. This causes the system to favor processes
849 * which haven't run much recently, and to round-robin among other processes.
851 * The actual schedulerclock interrupt rate is ESTCPUFREQ, but we generally
852 * want to ramp-up at a faster rate, ESTCPUVFREQ, so p_estcpu is scaled
853 * by (ESTCPUVFREQ / ESTCPUFREQ). You can control the ramp-up/ramp-down
854 * rate by adjusting ESTCPUVFREQ in sys/proc.h in integer multiples
857 * WARNING! called from a fast-int or an IPI, the MP lock MIGHT NOT BE HELD
858 * and we cannot block.
861 schedulerclock(void *dummy)
867 if ((p = td->td_proc) != NULL) {
868 p->p_cpticks++; /* cpticks runs at ESTCPUFREQ */
869 p->p_estcpu = ESTCPULIM(p->p_estcpu + ESTCPUVFREQ / ESTCPUFREQ);
885 cpri = crit_panic_save();
887 crit_panic_restore(cpri);