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 1.38 2004/11/10 08:27:54 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;
74 static struct callout roundrobin_callout;
75 static struct callout schedcpu_callout;
77 struct loadavg averunnable =
78 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
80 * Constants for averages over 1, 5, and 15 minutes
81 * when sampling at 5 second intervals.
83 static fixpt_t cexp[3] = {
84 0.9200444146293232 * FSCALE, /* exp(-1/12) */
85 0.9834714538216174 * FSCALE, /* exp(-1/60) */
86 0.9944598480048967 * FSCALE, /* exp(-1/180) */
89 static void endtsleep (void *);
90 static void loadav (void *arg);
91 static void roundrobin (void *arg);
92 static void schedcpu (void *arg);
93 static void updatepri (struct proc *p);
94 static void crit_panicints(void);
97 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
101 new_val = sched_quantum * tick;
102 error = sysctl_handle_int(oidp, &new_val, 0, req);
103 if (error != 0 || req->newptr == NULL)
107 sched_quantum = new_val / tick;
108 hogticks = 2 * sched_quantum;
112 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
113 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
116 roundrobin_interval(void)
118 return (sched_quantum);
122 * Force switch among equal priority processes every 100ms.
124 * WARNING! The MP lock is not held on ipi message remotes.
129 roundrobin_remote(void *arg)
131 struct proc *p = lwkt_preempted_proc();
132 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
139 roundrobin(void *arg)
141 struct proc *p = lwkt_preempted_proc();
142 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
145 lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL);
147 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
153 resched_cpus(u_int32_t mask)
155 lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL);
161 * The load average is scaled by FSCALE (2048 typ). The estimated cpu is
162 * incremented at a rate of ESTCPUVFREQ per second (40hz typ), but this is
163 * divided up across all cpu bound processes running in the system so an
164 * individual process will get less under load. ESTCPULIM typicaly caps
165 * out at ESTCPUMAX (around 376, or 11 nice levels).
167 * Generally speaking the decay equation needs to break-even on growth
168 * at the limit at all load levels >= 1.0, so if the estimated cpu for
169 * a process increases by (ESTVCPUFREQ / load) per second, then the decay
170 * should reach this value when estcpu reaches ESTCPUMAX. That calculation
173 * ESTCPUMAX * decay = ESTCPUVFREQ / load
174 * decay = ESTCPUVFREQ / (load * ESTCPUMAX)
175 * decay = estcpu * 0.053 / load
177 * If the load is less then 1.0 we assume a load of 1.0.
180 #define cload(loadav) ((loadav) < FSCALE ? FSCALE : (loadav))
181 #define decay_cpu(loadav,estcpu) \
182 ((estcpu) * (FSCALE * ESTCPUVFREQ / ESTCPUMAX) / cload(loadav))
184 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
185 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
186 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
188 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
189 static int fscale __unused = FSCALE;
190 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
193 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
194 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
195 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
197 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
198 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
200 * If you don't want to bother with the faster/more-accurate formula, you
201 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
202 * (more general) method of calculating the %age of CPU used by a process.
204 #define CCPU_SHIFT 11
207 * Recompute process priorities, once a second.
213 fixpt_t loadfac = averunnable.ldavg[0];
218 FOREACH_PROC_IN_SYSTEM(p) {
220 * Increment time in/out of memory and sleep time
221 * (if sleeping). We ignore overflow; with 16-bit int's
222 * (remember them?) overflow takes 45 days.
225 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
227 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
230 * If the process has slept the entire second,
231 * stop recalculating its priority until it wakes up.
233 * Note that interactive calculations do not occur for
234 * long sleeps (because that isn't necessarily indicative
235 * of an interactive process).
237 if (p->p_slptime > 1)
239 /* prevent state changes and protect run queue */
242 * p_cpticks runs at ESTCPUFREQ but must be divided by the
243 * load average for par-100% use. Higher p_interactive
244 * values mean less interactive, lower values mean more
247 if ((((fixpt_t)p->p_cpticks * cload(loadfac)) >> FSHIFT) >
249 if (p->p_interactive < 127)
252 if (p->p_interactive > -127)
256 * p_pctcpu is only for ps.
258 #if (FSHIFT >= CCPU_SHIFT)
259 p->p_pctcpu += (ESTCPUFREQ == 100)?
260 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
261 100 * (((fixpt_t) p->p_cpticks)
262 << (FSHIFT - CCPU_SHIFT)) / ESTCPUFREQ;
264 p->p_pctcpu += ((FSCALE - ccpu) *
265 (p->p_cpticks * FSCALE / ESTCPUFREQ)) >> FSHIFT;
268 ndecay = decay_cpu(loadfac, p->p_estcpu);
269 if (p->p_estcpu > ndecay)
270 p->p_estcpu -= ndecay;
276 wakeup((caddr_t)&lbolt);
277 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
281 * Recalculate the priority of a process after it has slept for a while.
282 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
283 * least six times the loadfactor will decay p_estcpu to zero.
286 updatepri(struct proc *p)
290 ndecay = decay_cpu(averunnable.ldavg[0], p->p_estcpu) * p->p_slptime;
291 if (p->p_estcpu > ndecay)
292 p->p_estcpu -= ndecay;
299 * We're only looking at 7 bits of the address; everything is
300 * aligned to 4, lots of things are aligned to greater powers
301 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
303 #define TABLESIZE 128
304 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
305 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
308 * During autoconfiguration or after a panic, a sleep will simply
309 * lower the priority briefly to allow interrupts, then return.
310 * The priority to be used (safepri) is machine-dependent, thus this
311 * value is initialized and maintained in the machine-dependent layers.
312 * This priority will typically be 0, or the lowest priority
313 * that is safe for use on the interrupt stack; it can be made
314 * higher to block network software interrupts after panics.
323 sched_quantum = hz/10;
324 hogticks = 2 * sched_quantum;
325 for (i = 0; i < TABLESIZE; i++)
326 TAILQ_INIT(&slpque[i]);
330 * General sleep call. Suspends the current process until a wakeup is
331 * performed on the specified identifier. The process will then be made
332 * runnable with the specified priority. Sleeps at most timo/hz seconds
333 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
334 * before and after sleeping, else signals are not checked. Returns 0 if
335 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
336 * signal needs to be delivered, ERESTART is returned if the current system
337 * call should be restarted if possible, and EINTR is returned if the system
338 * call should be interrupted by the signal (return EINTR).
340 * Note that if we are a process, we release_curproc() before messing with
341 * the LWKT scheduler.
344 tsleep(void *ident, int flags, const char *wmesg, int timo)
346 struct thread *td = curthread;
347 struct proc *p = td->td_proc; /* may be NULL */
348 int sig = 0, catch = flags & PCATCH;
349 int id = LOOKUP(ident);
350 struct callout thandle;
353 * NOTE: removed KTRPOINT, it could cause races due to blocking
354 * even in stable. Just scrap it for now.
356 if (cold || panicstr) {
358 * After a panic, or during autoconfiguration,
359 * just give interrupts a chance, then just return;
360 * don't run any other procs or panic below,
361 * in case this is the idle process and already asleep.
366 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */
367 crit_enter_quick(td);
368 KASSERT(ident != NULL, ("tsleep: no ident"));
369 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
370 ident, wmesg, p->p_stat));
372 td->td_wchan = ident;
373 td->td_wmesg = wmesg;
375 if (flags & PNORESCHED)
376 td->td_flags |= TDF_NORESCHED;
380 lwkt_deschedule_self(td);
381 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
383 callout_init(&thandle);
384 callout_reset(&thandle, timo, endtsleep, td);
387 * We put ourselves on the sleep queue and start our timeout
388 * before calling CURSIG, as we could stop there, and a wakeup
389 * or a SIGCONT (or both) could occur while we were stopped.
390 * A SIGCONT would cause us to be marked as SSLEEP
391 * without resuming us, thus we must be ready for sleep
392 * when CURSIG is called. If the wakeup happens while we're
393 * stopped, td->td_wchan will be 0 upon return from CURSIG.
397 p->p_flag |= P_SINTR;
398 if ((sig = CURSIG(p))) {
401 lwkt_schedule_self(td);
406 if (td->td_wchan == NULL) {
415 * If we are not the current process we have to remove ourself
416 * from the run queue.
418 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
420 * If this is the current 'user' process schedule another one.
422 clrrunnable(p, SSLEEP);
423 p->p_stats->p_ru.ru_nvcsw++;
425 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
431 p->p_flag &= ~P_SINTR;
433 td->td_flags &= ~TDF_NORESCHED;
434 if (td->td_flags & TDF_TIMEOUT) {
435 td->td_flags &= ~TDF_TIMEOUT;
437 return (EWOULDBLOCK);
439 callout_stop(&thandle);
440 } else if (td->td_wmesg) {
442 * This can happen if a thread is woken up directly. Clear
443 * wmesg to avoid debugging confusion.
447 /* inline of iscaught() */
449 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
450 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
459 * Implement the timeout for tsleep. We interlock against
460 * wchan when setting TDF_TIMEOUT. For processes we remove
461 * the sleep if the process is stopped rather then sleeping,
462 * so it remains stopped.
472 td->td_flags |= TDF_TIMEOUT;
473 if ((p = td->td_proc) != NULL) {
474 if (p->p_stat == SSLEEP)
487 * Remove a process from its wait queue
490 unsleep(struct thread *td)
495 if (p->p_flag & P_XSLEEP) {
496 struct xwait *w = p->p_wchan;
497 TAILQ_REMOVE(&w->waitq, p, p_procq);
498 p->p_flag &= ~P_XSLEEP;
501 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
509 * Make all processes sleeping on the explicit lock structure runnable.
512 xwakeup(struct xwait *w)
518 while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
519 TAILQ_REMOVE(&w->waitq, p, p_procq);
520 KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
521 ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
523 p->p_flag &= ~P_XSLEEP;
524 if (p->p_stat == SSLEEP) {
525 /* OPTIMIZED EXPANSION OF setrunnable(p); */
526 if (p->p_slptime > 1)
530 if (p->p_flag & P_INMEM) {
533 p->p_flag |= P_SWAPINREQ;
534 wakeup((caddr_t)&proc0);
543 * Make all processes sleeping on the specified identifier runnable.
546 _wakeup(void *ident, int count)
548 struct slpquehead *qp;
552 int id = LOOKUP(ident);
557 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
558 ntd = TAILQ_NEXT(td, td_threadq);
559 if (td->td_wchan == ident) {
560 TAILQ_REMOVE(qp, td, td_threadq);
562 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
563 /* OPTIMIZED EXPANSION OF setrunnable(p); */
564 if (p->p_slptime > 1)
568 if (p->p_flag & P_INMEM) {
570 * LWKT scheduled now, there is no
571 * userland runq interaction until
572 * the thread tries to return to user
579 p->p_flag |= P_SWAPINREQ;
580 wakeup((caddr_t)&proc0);
582 /* END INLINE EXPANSION */
583 } else if (p == NULL) {
601 wakeup_one(void *ident)
607 * The machine independent parts of mi_switch().
609 * 'p' must be the current process.
612 mi_switch(struct proc *p)
614 thread_t td = p->p_thread;
618 KKASSERT(td == mycpu->gd_curthread);
620 crit_enter_quick(td);
623 * Check if the process exceeds its cpu resource allocation.
624 * If over max, kill it. Time spent in interrupts is not
625 * included. YYY 64 bit match is expensive. Ick.
627 ttime = td->td_sticks + td->td_uticks;
628 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
629 ttime > p->p_limit->p_cpulimit) {
630 rlim = &p->p_rlimit[RLIMIT_CPU];
631 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
632 killproc(p, "exceeded maximum CPU limit");
635 if (rlim->rlim_cur < rlim->rlim_max) {
636 /* XXX: we should make a private copy */
643 * If we are in a SSTOPped state we deschedule ourselves.
644 * YYY this needs to be cleaned up, remember that LWKTs stay on
645 * their run queue which works differently then the user scheduler
646 * which removes the process from the runq when it runs it.
648 mycpu->gd_cnt.v_swtch++;
649 if (p->p_stat == SSTOP)
650 lwkt_deschedule_self(td);
656 * Change process state to be runnable,
657 * placing it on the run queue if it is in memory,
658 * and awakening the swapper if it isn't in memory.
661 setrunnable(struct proc *p)
671 panic("setrunnable");
674 unsleep(p->p_thread); /* e.g. when sending signals */
683 * The process is controlled by LWKT at this point, we do not mess
684 * around with the userland scheduler until the thread tries to
685 * return to user mode.
688 if (p->p_flag & P_INMEM)
691 if (p->p_flag & P_INMEM)
692 lwkt_schedule(p->p_thread);
694 if (p->p_slptime > 1)
697 if ((p->p_flag & P_INMEM) == 0) {
698 p->p_flag |= P_SWAPINREQ;
699 wakeup((caddr_t)&proc0);
704 * Change the process state to NOT be runnable, removing it from the run
708 clrrunnable(struct proc *p, int stat)
710 crit_enter_quick(p->p_thread);
711 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ))
714 crit_exit_quick(p->p_thread);
718 * Compute the priority of a process when running in user mode.
719 * Arrange to reschedule if the resulting priority is better
720 * than that of the current process.
723 resetpriority(struct proc *p)
731 * Set p_priority for general process comparisons
733 switch(p->p_rtprio.type) {
734 case RTP_PRIO_REALTIME:
735 p->p_priority = PRIBASE_REALTIME + p->p_rtprio.prio;
737 case RTP_PRIO_NORMAL:
740 p->p_priority = PRIBASE_IDLE + p->p_rtprio.prio;
742 case RTP_PRIO_THREAD:
743 p->p_priority = PRIBASE_THREAD + p->p_rtprio.prio;
748 * NORMAL priorities fall through. These are based on niceness
749 * and cpu use. Lower numbers == higher priorities.
751 newpriority = (int)(NICE_ADJUST(p->p_nice - PRIO_MIN) +
752 p->p_estcpu / ESTCPURAMP);
755 * p_interactive is -128 to +127 and represents very long term
756 * interactivity or batch (whereas estcpu is a much faster variable).
757 * Interactivity can modify the priority by up to 8 units either way.
758 * (8 units == approximately 4 nice levels).
760 interactive = p->p_interactive / 10;
761 newpriority += interactive;
763 newpriority = MIN(newpriority, MAXPRI);
764 newpriority = MAX(newpriority, 0);
765 npq = newpriority / PPQ;
767 opq = (p->p_priority & PRIMASK) / PPQ;
768 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) {
770 * We have to move the process to another queue
773 p->p_priority = PRIBASE_NORMAL + newpriority;
777 * We can just adjust the priority and it will be picked
780 KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0);
781 p->p_priority = PRIBASE_NORMAL + newpriority;
787 * Compute a tenex style load average of a quantity on
788 * 1, 5 and 15 minute intervals.
800 FOREACH_PROC_IN_SYSTEM(p) {
803 if ((td = p->p_thread) == NULL)
805 if (td->td_flags & TDF_BLOCKED)
815 for (i = 0; i < 3; i++)
816 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
817 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
820 * Schedule the next update to occur after 5 seconds, but add a
821 * random variation to avoid synchronisation with processes that
822 * run at regular intervals.
824 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
830 sched_setup(void *dummy)
832 callout_init(&loadav_callout);
833 callout_init(&roundrobin_callout);
834 callout_init(&schedcpu_callout);
836 /* Kick off timeout driven events by calling first time. */
843 * We adjust the priority of the current process. The priority of
844 * a process gets worse as it accumulates CPU time. The cpu usage
845 * estimator (p_estcpu) is increased here. resetpriority() will
846 * compute a different priority each time p_estcpu increases by
847 * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached).
849 * The cpu usage estimator ramps up quite quickly when the process is
850 * running (linearly), and decays away exponentially, at a rate which
851 * is proportionally slower when the system is busy. The basic principle
852 * is that the system will 90% forget that the process used a lot of CPU
853 * time in 5 * loadav seconds. This causes the system to favor processes
854 * which haven't run much recently, and to round-robin among other processes.
856 * The actual schedulerclock interrupt rate is ESTCPUFREQ, but we generally
857 * want to ramp-up at a faster rate, ESTCPUVFREQ, so p_estcpu is scaled
858 * by (ESTCPUVFREQ / ESTCPUFREQ). You can control the ramp-up/ramp-down
859 * rate by adjusting ESTCPUVFREQ in sys/proc.h in integer multiples
862 * WARNING! called from a fast-int or an IPI, the MP lock MIGHT NOT BE HELD
863 * and we cannot block.
866 schedulerclock(void *dummy)
872 if ((p = td->td_proc) != NULL) {
873 p->p_cpticks++; /* cpticks runs at ESTCPUFREQ */
874 p->p_estcpu = ESTCPULIM(p->p_estcpu + ESTCPUVFREQ / ESTCPUFREQ);
890 cpri = crit_panic_save();
892 crit_panic_restore(cpri);