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 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include "opt_hwpmc_hooks.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
45 #include <sys/kthread.h>
46 #include <sys/mutex.h>
48 #include <sys/resourcevar.h>
49 #include <sys/sched.h>
51 #include <sys/sysctl.h>
53 #include <sys/turnstile.h>
55 #include <machine/pcb.h>
56 #include <machine/smp.h>
59 #include <sys/pmckern.h>
63 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
64 * the range 100-256 Hz (approximately).
66 #define ESTCPULIM(e) \
67 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
68 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
70 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
72 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
74 #define NICE_WEIGHT 1 /* Priorities per nice level. */
77 * The schedulable entity that runs a context.
78 * This is an extension to the thread structure and is tailored to
79 * the requirements of this scheduler
82 TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */
83 struct thread *ts_thread; /* (*) Active associated thread. */
84 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
85 u_char ts_rqindex; /* (j) Run queue index. */
86 int ts_cpticks; /* (j) Ticks of cpu time. */
87 int ts_slptime; /* (j) Seconds !RUNNING. */
88 struct runq *ts_runq; /* runq the thread is currently on */
91 /* flags kept in td_flags */
92 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
93 #define TDF_EXIT TDF_SCHED1 /* thread is being killed. */
94 #define TDF_BOUND TDF_SCHED2
96 #define ts_flags ts_thread->td_flags
97 #define TSF_DIDRUN TDF_DIDRUN /* thread actually ran. */
98 #define TSF_EXIT TDF_EXIT /* thread is being killed. */
99 #define TSF_BOUND TDF_BOUND /* stuck to one CPU */
101 #define SKE_RUNQ_PCPU(ts) \
102 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
104 static struct td_sched td_sched0;
105 struct mtx sched_lock;
107 static int sched_tdcnt; /* Total runnable threads in the system. */
108 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
109 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
111 static struct callout roundrobin_callout;
113 static void setup_runqs(void);
114 static void roundrobin(void *arg);
115 static void schedcpu(void);
116 static void schedcpu_thread(void);
117 static void sched_priority(struct thread *td, u_char prio);
118 static void sched_setup(void *dummy);
119 static void maybe_resched(struct thread *td);
120 static void updatepri(struct thread *td);
121 static void resetpriority(struct thread *td);
122 static void resetpriority_thread(struct thread *td);
124 static int forward_wakeup(int cpunum);
127 static struct kproc_desc sched_kp = {
132 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp)
133 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
138 static struct runq runq;
144 static struct runq runq_pcpu[MAXCPU];
153 for (i = 0; i < MAXCPU; ++i)
154 runq_init(&runq_pcpu[i]);
161 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
165 new_val = sched_quantum * tick;
166 error = sysctl_handle_int(oidp, &new_val, 0, req);
167 if (error != 0 || req->newptr == NULL)
171 sched_quantum = new_val / tick;
172 hogticks = 2 * sched_quantum;
176 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
178 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
181 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
182 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
183 "Roundrobin scheduling quantum in microseconds");
186 /* Enable forwarding of wakeups to all other cpus */
187 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
189 static int forward_wakeup_enabled = 1;
190 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
191 &forward_wakeup_enabled, 0,
192 "Forwarding of wakeup to idle CPUs");
194 static int forward_wakeups_requested = 0;
195 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
196 &forward_wakeups_requested, 0,
197 "Requests for Forwarding of wakeup to idle CPUs");
199 static int forward_wakeups_delivered = 0;
200 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
201 &forward_wakeups_delivered, 0,
202 "Completed Forwarding of wakeup to idle CPUs");
204 static int forward_wakeup_use_mask = 1;
205 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
206 &forward_wakeup_use_mask, 0,
207 "Use the mask of idle cpus");
209 static int forward_wakeup_use_loop = 0;
210 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
211 &forward_wakeup_use_loop, 0,
212 "Use a loop to find idle cpus");
214 static int forward_wakeup_use_single = 0;
215 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
216 &forward_wakeup_use_single, 0,
217 "Only signal one idle cpu");
219 static int forward_wakeup_use_htt = 0;
220 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
221 &forward_wakeup_use_htt, 0,
226 static int sched_followon = 0;
227 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
229 "allow threads to share a quantum");
236 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
243 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
246 * Arrange to reschedule if necessary, taking the priorities and
247 * schedulers into account.
250 maybe_resched(struct thread *td)
253 THREAD_LOCK_ASSERT(td, MA_OWNED);
254 if (td->td_priority < curthread->td_priority)
255 curthread->td_flags |= TDF_NEEDRESCHED;
259 * Force switch among equal priority processes every 100ms.
260 * We don't actually need to force a context switch of the current process.
261 * The act of firing the event triggers a context switch to softclock() and
262 * then switching back out again which is equivalent to a preemption, thus
263 * no further work is needed on the local CPU.
267 roundrobin(void *arg)
271 mtx_lock_spin(&sched_lock);
272 forward_roundrobin();
273 mtx_unlock_spin(&sched_lock);
276 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
280 * Constants for digital decay and forget:
281 * 90% of (td_estcpu) usage in 5 * loadav time
282 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
283 * Note that, as ps(1) mentions, this can let percentages
284 * total over 100% (I've seen 137.9% for 3 processes).
286 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
288 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
289 * That is, the system wants to compute a value of decay such
290 * that the following for loop:
291 * for (i = 0; i < (5 * loadavg); i++)
292 * td_estcpu *= decay;
295 * for all values of loadavg:
297 * Mathematically this loop can be expressed by saying:
298 * decay ** (5 * loadavg) ~= .1
300 * The system computes decay as:
301 * decay = (2 * loadavg) / (2 * loadavg + 1)
303 * We wish to prove that the system's computation of decay
304 * will always fulfill the equation:
305 * decay ** (5 * loadavg) ~= .1
307 * If we compute b as:
310 * decay = b / (b + 1)
312 * We now need to prove two things:
313 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
314 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
317 * For x close to zero, exp(x) =~ 1 + x, since
318 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
319 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
320 * For x close to zero, ln(1+x) =~ x, since
321 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
322 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
326 * Solve (factor)**(power) =~ .1 given power (5*loadav):
327 * solving for factor,
328 * ln(factor) =~ (-2.30/5*loadav), or
329 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
330 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
333 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
335 * power*ln(b/(b+1)) =~ -2.30, or
336 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
338 * Actual power values for the implemented algorithm are as follows:
340 * power: 5.68 10.32 14.94 19.55
343 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
344 #define loadfactor(loadav) (2 * (loadav))
345 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
347 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
348 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
349 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
352 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
353 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
354 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
356 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
357 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
359 * If you don't want to bother with the faster/more-accurate formula, you
360 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
361 * (more general) method of calculating the %age of CPU used by a process.
363 #define CCPU_SHIFT 11
366 * Recompute process priorities, every hz ticks.
367 * MP-safe, called without the Giant mutex.
373 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
377 int awake, realstathz;
379 realstathz = stathz ? stathz : hz;
380 sx_slock(&allproc_lock);
381 FOREACH_PROC_IN_SYSTEM(p) {
383 FOREACH_THREAD_IN_PROC(p, td) {
388 * Increment sleep time (if sleeping). We
389 * ignore overflow, as above.
392 * The td_sched slptimes are not touched in wakeup
393 * because the thread may not HAVE everything in
394 * memory? XXX I think this is out of date.
396 if (TD_ON_RUNQ(td)) {
398 ts->ts_flags &= ~TSF_DIDRUN;
399 } else if (TD_IS_RUNNING(td)) {
401 /* Do not clear TSF_DIDRUN */
402 } else if (ts->ts_flags & TSF_DIDRUN) {
404 ts->ts_flags &= ~TSF_DIDRUN;
408 * ts_pctcpu is only for ps and ttyinfo().
409 * Do it per td_sched, and add them up at the end?
412 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
414 * If the td_sched has been idle the entire second,
415 * stop recalculating its priority until
418 if (ts->ts_cpticks != 0) {
419 #if (FSHIFT >= CCPU_SHIFT)
420 ts->ts_pctcpu += (realstathz == 100)
421 ? ((fixpt_t) ts->ts_cpticks) <<
422 (FSHIFT - CCPU_SHIFT) :
423 100 * (((fixpt_t) ts->ts_cpticks)
424 << (FSHIFT - CCPU_SHIFT)) / realstathz;
426 ts->ts_pctcpu += ((FSCALE - ccpu) *
428 FSCALE / realstathz)) >> FSHIFT;
433 * If there are ANY running threads in this process,
434 * then don't count it as sleeping.
439 if (ts->ts_slptime > 1) {
441 * In an ideal world, this should not
442 * happen, because whoever woke us
443 * up from the long sleep should have
444 * unwound the slptime and reset our
445 * priority before we run at the stale
446 * priority. Should KASSERT at some
447 * point when all the cases are fixed.
454 if (ts->ts_slptime > 1) {
458 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
460 resetpriority_thread(td);
462 } /* end of thread loop */
464 } /* end of process loop */
465 sx_sunlock(&allproc_lock);
469 * Main loop for a kthread that executes schedcpu once a second.
472 schedcpu_thread(void)
482 * Recalculate the priority of a process after it has slept for a while.
483 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
484 * least six times the loadfactor will decay td_estcpu to zero.
487 updatepri(struct thread *td)
494 loadfac = loadfactor(averunnable.ldavg[0]);
495 if (ts->ts_slptime > 5 * loadfac)
498 newcpu = td->td_estcpu;
499 ts->ts_slptime--; /* was incremented in schedcpu() */
500 while (newcpu && --ts->ts_slptime)
501 newcpu = decay_cpu(loadfac, newcpu);
502 td->td_estcpu = newcpu;
507 * Compute the priority of a process when running in user mode.
508 * Arrange to reschedule if the resulting priority is better
509 * than that of the current process.
512 resetpriority(struct thread *td)
514 register unsigned int newpriority;
516 if (td->td_pri_class == PRI_TIMESHARE) {
517 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
518 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
519 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
521 sched_user_prio(td, newpriority);
526 * Update the thread's priority when the associated process's user
530 resetpriority_thread(struct thread *td)
533 /* Only change threads with a time sharing user priority. */
534 if (td->td_priority < PRI_MIN_TIMESHARE ||
535 td->td_priority > PRI_MAX_TIMESHARE)
538 /* XXX the whole needresched thing is broken, but not silly. */
541 sched_prio(td, td->td_user_pri);
546 sched_setup(void *dummy)
550 if (sched_quantum == 0)
551 sched_quantum = SCHED_QUANTUM;
552 hogticks = 2 * sched_quantum;
554 callout_init(&roundrobin_callout, CALLOUT_MPSAFE);
556 /* Kick off timeout driven events by calling first time. */
559 /* Account for thread0. */
563 /* External interfaces start here */
565 * Very early in the boot some setup of scheduler-specific
566 * parts of proc0 and of some scheduler resources needs to be done.
574 * Set up the scheduler specific parts of proc0.
576 proc0.p_sched = NULL; /* XXX */
577 thread0.td_sched = &td_sched0;
578 thread0.td_lock = &sched_lock;
579 td_sched0.ts_thread = &thread0;
580 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
587 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
589 return runq_check(&runq);
594 sched_rr_interval(void)
596 if (sched_quantum == 0)
597 sched_quantum = SCHED_QUANTUM;
598 return (sched_quantum);
602 * We adjust the priority of the current process. The priority of
603 * a process gets worse as it accumulates CPU time. The cpu usage
604 * estimator (td_estcpu) is increased here. resetpriority() will
605 * compute a different priority each time td_estcpu increases by
606 * INVERSE_ESTCPU_WEIGHT
607 * (until MAXPRI is reached). The cpu usage estimator ramps up
608 * quite quickly when the process is running (linearly), and decays
609 * away exponentially, at a rate which is proportionally slower when
610 * the system is busy. The basic principle is that the system will
611 * 90% forget that the process used a lot of CPU time in 5 * loadav
612 * seconds. This causes the system to favor processes which haven't
613 * run much recently, and to round-robin among other processes.
616 sched_clock(struct thread *td)
620 THREAD_LOCK_ASSERT(td, MA_OWNED);
624 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
625 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
627 resetpriority_thread(td);
632 * charge childs scheduling cpu usage to parent.
635 sched_exit(struct proc *p, struct thread *td)
638 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
639 td, td->td_proc->p_comm, td->td_priority);
640 PROC_SLOCK_ASSERT(p, MA_OWNED);
641 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
645 sched_exit_thread(struct thread *td, struct thread *child)
648 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
649 child, child->td_proc->p_comm, child->td_priority);
651 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
653 mtx_lock_spin(&sched_lock);
654 if ((child->td_proc->p_flag & P_NOLOAD) == 0)
656 mtx_unlock_spin(&sched_lock);
660 sched_fork(struct thread *td, struct thread *childtd)
662 sched_fork_thread(td, childtd);
666 sched_fork_thread(struct thread *td, struct thread *childtd)
668 childtd->td_estcpu = td->td_estcpu;
669 childtd->td_lock = &sched_lock;
670 sched_newthread(childtd);
674 sched_nice(struct proc *p, int nice)
678 PROC_LOCK_ASSERT(p, MA_OWNED);
679 PROC_SLOCK_ASSERT(p, MA_OWNED);
681 FOREACH_THREAD_IN_PROC(p, td) {
684 resetpriority_thread(td);
690 sched_class(struct thread *td, int class)
692 THREAD_LOCK_ASSERT(td, MA_OWNED);
693 td->td_pri_class = class;
697 * Adjust the priority of a thread.
700 sched_priority(struct thread *td, u_char prio)
702 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
703 td, td->td_proc->p_comm, td->td_priority, prio, curthread,
704 curthread->td_proc->p_comm);
706 THREAD_LOCK_ASSERT(td, MA_OWNED);
707 if (td->td_priority == prio)
709 td->td_priority = prio;
710 if (TD_ON_RUNQ(td) &&
711 td->td_sched->ts_rqindex != (prio / RQ_PPQ)) {
713 sched_add(td, SRQ_BORING);
718 * Update a thread's priority when it is lent another thread's
722 sched_lend_prio(struct thread *td, u_char prio)
725 td->td_flags |= TDF_BORROWING;
726 sched_priority(td, prio);
730 * Restore a thread's priority when priority propagation is
731 * over. The prio argument is the minimum priority the thread
732 * needs to have to satisfy other possible priority lending
733 * requests. If the thread's regulary priority is less
734 * important than prio the thread will keep a priority boost
738 sched_unlend_prio(struct thread *td, u_char prio)
742 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
743 td->td_base_pri <= PRI_MAX_TIMESHARE)
744 base_pri = td->td_user_pri;
746 base_pri = td->td_base_pri;
747 if (prio >= base_pri) {
748 td->td_flags &= ~TDF_BORROWING;
749 sched_prio(td, base_pri);
751 sched_lend_prio(td, prio);
755 sched_prio(struct thread *td, u_char prio)
759 /* First, update the base priority. */
760 td->td_base_pri = prio;
763 * If the thread is borrowing another thread's priority, don't ever
764 * lower the priority.
766 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
769 /* Change the real priority. */
770 oldprio = td->td_priority;
771 sched_priority(td, prio);
774 * If the thread is on a turnstile, then let the turnstile update
777 if (TD_ON_LOCK(td) && oldprio != prio)
778 turnstile_adjust(td, oldprio);
782 sched_user_prio(struct thread *td, u_char prio)
786 td->td_base_user_pri = prio;
787 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
789 oldprio = td->td_user_pri;
790 td->td_user_pri = prio;
792 if (TD_ON_UPILOCK(td) && oldprio != prio)
793 umtx_pi_adjust(td, oldprio);
797 sched_lend_user_prio(struct thread *td, u_char prio)
801 td->td_flags |= TDF_UBORROWING;
803 oldprio = td->td_user_pri;
804 td->td_user_pri = prio;
806 if (TD_ON_UPILOCK(td) && oldprio != prio)
807 umtx_pi_adjust(td, oldprio);
811 sched_unlend_user_prio(struct thread *td, u_char prio)
815 base_pri = td->td_base_user_pri;
816 if (prio >= base_pri) {
817 td->td_flags &= ~TDF_UBORROWING;
818 sched_user_prio(td, base_pri);
820 sched_lend_user_prio(td, prio);
824 sched_sleep(struct thread *td)
827 THREAD_LOCK_ASSERT(td, MA_OWNED);
828 td->td_slptick = ticks;
829 td->td_sched->ts_slptime = 0;
833 sched_switch(struct thread *td, struct thread *newtd, int flags)
841 THREAD_LOCK_ASSERT(td, MA_OWNED);
843 * Switch to the sched lock to fix things up and pick
846 if (td->td_lock != &sched_lock) {
847 mtx_lock_spin(&sched_lock);
851 if ((p->p_flag & P_NOLOAD) == 0)
855 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
857 td->td_lastcpu = td->td_oncpu;
858 td->td_flags &= ~TDF_NEEDRESCHED;
859 td->td_owepreempt = 0;
860 td->td_oncpu = NOCPU;
862 * At the last moment, if this thread is still marked RUNNING,
863 * then put it back on the run queue as it has not been suspended
864 * or stopped or any thing else similar. We never put the idle
865 * threads on the run queue, however.
867 if (td->td_flags & TDF_IDLETD) {
870 idle_cpus_mask &= ~PCPU_GET(cpumask);
873 if (TD_IS_RUNNING(td)) {
874 /* Put us back on the run queue. */
875 sched_add(td, (flags & SW_PREEMPT) ?
876 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
877 SRQ_OURSELF|SRQ_YIELDING);
882 * The thread we are about to run needs to be counted
883 * as if it had been added to the run queue and selected.
889 KASSERT((newtd->td_inhibitors == 0),
890 ("trying to run inhibited thread"));
891 newtd->td_sched->ts_flags |= TSF_DIDRUN;
892 TD_SET_RUNNING(newtd);
893 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0)
896 newtd = choosethread();
898 MPASS(newtd->td_lock == &sched_lock);
902 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
903 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
907 cpu_switch(td, newtd, td->td_lock);
909 * Where am I? What year is it?
910 * We are in the same thread that went to sleep above,
911 * but any amount of time may have passed. All out context
912 * will still be available as will local variables.
913 * PCPU values however may have changed as we may have
914 * changed CPU so don't trust cached values of them.
915 * New threads will go to fork_exit() instead of here
916 * so if you change things here you may need to change
918 * If the thread above was exiting it will never wake
919 * up again here, so either it has saved everything it
920 * needed to, or the thread_wait() or wait() will
924 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
925 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
930 if (td->td_flags & TDF_IDLETD)
931 idle_cpus_mask |= PCPU_GET(cpumask);
933 sched_lock.mtx_lock = (uintptr_t)td;
934 td->td_oncpu = PCPU_GET(cpuid);
935 MPASS(td->td_lock == &sched_lock);
939 sched_wakeup(struct thread *td)
943 THREAD_LOCK_ASSERT(td, MA_OWNED);
945 if (ts->ts_slptime > 1) {
949 td->td_slptick = ticks;
951 sched_add(td, SRQ_BORING);
955 /* enable HTT_2 if you have a 2-way HTT cpu.*/
957 forward_wakeup(int cpunum)
959 cpumask_t map, me, dontuse;
964 mtx_assert(&sched_lock, MA_OWNED);
966 CTR0(KTR_RUNQ, "forward_wakeup()");
968 if ((!forward_wakeup_enabled) ||
969 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
971 if (!smp_started || cold || panicstr)
974 forward_wakeups_requested++;
977 * check the idle mask we received against what we calculated before
978 * in the old version.
980 me = PCPU_GET(cpumask);
982 * don't bother if we should be doing it ourself..
984 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
987 dontuse = me | stopped_cpus | hlt_cpus_mask;
989 if (forward_wakeup_use_loop) {
990 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
992 if ( (id & dontuse) == 0 &&
993 pc->pc_curthread == pc->pc_idlethread) {
999 if (forward_wakeup_use_mask) {
1001 map = idle_cpus_mask & ~dontuse;
1003 /* If they are both on, compare and use loop if different */
1004 if (forward_wakeup_use_loop) {
1006 printf("map (%02X) != map3 (%02X)\n",
1014 /* If we only allow a specific CPU, then mask off all the others */
1015 if (cpunum != NOCPU) {
1016 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1017 map &= (1 << cpunum);
1019 /* Try choose an idle die. */
1020 if (forward_wakeup_use_htt) {
1021 map2 = (map & (map >> 1)) & 0x5555;
1027 /* set only one bit */
1028 if (forward_wakeup_use_single) {
1029 map = map & ((~map) + 1);
1033 forward_wakeups_delivered++;
1034 ipi_selected(map, IPI_AST);
1037 if (cpunum == NOCPU)
1038 printf("forward_wakeup: Idle processor not found\n");
1044 static void kick_other_cpu(int pri,int cpuid);
1047 kick_other_cpu(int pri,int cpuid)
1049 struct pcpu * pcpu = pcpu_find(cpuid);
1050 int cpri = pcpu->pc_curthread->td_priority;
1052 if (idle_cpus_mask & pcpu->pc_cpumask) {
1053 forward_wakeups_delivered++;
1054 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1061 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1062 #if !defined(FULL_PREEMPTION)
1063 if (pri <= PRI_MAX_ITHD)
1064 #endif /* ! FULL_PREEMPTION */
1066 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1069 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1071 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1072 ipi_selected( pcpu->pc_cpumask , IPI_AST);
1078 sched_add(struct thread *td, int flags)
1081 struct td_sched *ts;
1087 THREAD_LOCK_ASSERT(td, MA_OWNED);
1088 KASSERT((td->td_inhibitors == 0),
1089 ("sched_add: trying to run inhibited thread"));
1090 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1091 ("sched_add: bad thread state"));
1092 KASSERT(td->td_flags & TDF_INMEM,
1093 ("sched_add: thread swapped out"));
1094 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1095 td, td->td_proc->p_comm, td->td_priority, curthread,
1096 curthread->td_proc->p_comm);
1098 * Now that the thread is moving to the run-queue, set the lock
1099 * to the scheduler's lock.
1101 if (td->td_lock != &sched_lock) {
1102 mtx_lock_spin(&sched_lock);
1103 thread_lock_set(td, &sched_lock);
1107 if (td->td_pinned != 0) {
1108 cpu = td->td_lastcpu;
1109 ts->ts_runq = &runq_pcpu[cpu];
1112 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1113 } else if ((ts)->ts_flags & TSF_BOUND) {
1114 /* Find CPU from bound runq */
1115 KASSERT(SKE_RUNQ_PCPU(ts),("sched_add: bound td_sched not on cpu runq"));
1116 cpu = ts->ts_runq - &runq_pcpu[0];
1119 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1122 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, td);
1124 ts->ts_runq = &runq;
1127 if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1128 kick_other_cpu(td->td_priority,cpu);
1132 cpumask_t me = PCPU_GET(cpumask);
1133 int idle = idle_cpus_mask & me;
1135 if (!idle && ((flags & SRQ_INTR) == 0) &&
1136 (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1137 forwarded = forward_wakeup(cpu);
1141 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1148 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1150 runq_add(ts->ts_runq, ts, flags);
1154 struct td_sched *ts;
1156 THREAD_LOCK_ASSERT(td, MA_OWNED);
1157 KASSERT((td->td_inhibitors == 0),
1158 ("sched_add: trying to run inhibited thread"));
1159 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1160 ("sched_add: bad thread state"));
1161 KASSERT(td->td_flags & TDF_INMEM,
1162 ("sched_add: thread swapped out"));
1163 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1164 td, td->td_proc->p_comm, td->td_priority, curthread,
1165 curthread->td_proc->p_comm);
1167 * Now that the thread is moving to the run-queue, set the lock
1168 * to the scheduler's lock.
1170 if (td->td_lock != &sched_lock) {
1171 mtx_lock_spin(&sched_lock);
1172 thread_lock_set(td, &sched_lock);
1175 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1176 ts->ts_runq = &runq;
1179 * If we are yielding (on the way out anyhow)
1180 * or the thread being saved is US,
1181 * then don't try be smart about preemption
1182 * or kicking off another CPU
1183 * as it won't help and may hinder.
1184 * In the YIEDLING case, we are about to run whoever is
1185 * being put in the queue anyhow, and in the
1186 * OURSELF case, we are puting ourself on the run queue
1187 * which also only happens when we are about to yield.
1189 if((flags & SRQ_YIELDING) == 0) {
1190 if (maybe_preempt(td))
1193 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1195 runq_add(ts->ts_runq, ts, flags);
1201 sched_rem(struct thread *td)
1203 struct td_sched *ts;
1206 KASSERT(td->td_flags & TDF_INMEM,
1207 ("sched_rem: thread swapped out"));
1208 KASSERT(TD_ON_RUNQ(td),
1209 ("sched_rem: thread not on run queue"));
1210 mtx_assert(&sched_lock, MA_OWNED);
1211 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1212 td, td->td_proc->p_comm, td->td_priority, curthread,
1213 curthread->td_proc->p_comm);
1215 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1217 runq_remove(ts->ts_runq, ts);
1222 * Select threads to run.
1223 * Notice that the running threads still consume a slot.
1228 struct td_sched *ts;
1231 mtx_assert(&sched_lock, MA_OWNED);
1233 struct td_sched *kecpu;
1236 ts = runq_choose(&runq);
1237 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1241 kecpu->ts_thread->td_priority < ts->ts_thread->td_priority)) {
1242 CTR2(KTR_RUNQ, "choosing td_sched %p from pcpu runq %d", kecpu,
1245 rq = &runq_pcpu[PCPU_GET(cpuid)];
1247 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", ts);
1252 ts = runq_choose(&runq);
1256 runq_remove(rq, ts);
1257 ts->ts_flags |= TSF_DIDRUN;
1259 KASSERT(ts->ts_thread->td_flags & TDF_INMEM,
1260 ("sched_choose: thread swapped out"));
1261 return (ts->ts_thread);
1263 return (PCPU_GET(idlethread));
1267 sched_userret(struct thread *td)
1270 * XXX we cheat slightly on the locking here to avoid locking in
1271 * the usual case. Setting td_priority here is essentially an
1272 * incomplete workaround for not setting it properly elsewhere.
1273 * Now that some interrupt handlers are threads, not setting it
1274 * properly elsewhere can clobber it in the window between setting
1275 * it here and returning to user mode, so don't waste time setting
1276 * it perfectly here.
1278 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1279 ("thread with borrowed priority returning to userland"));
1280 if (td->td_priority != td->td_user_pri) {
1282 td->td_priority = td->td_user_pri;
1283 td->td_base_pri = td->td_user_pri;
1289 sched_bind(struct thread *td, int cpu)
1291 struct td_sched *ts;
1293 THREAD_LOCK_ASSERT(td, MA_OWNED);
1294 KASSERT(TD_IS_RUNNING(td),
1295 ("sched_bind: cannot bind non-running thread"));
1299 ts->ts_flags |= TSF_BOUND;
1301 ts->ts_runq = &runq_pcpu[cpu];
1302 if (PCPU_GET(cpuid) == cpu)
1305 mi_switch(SW_VOL, NULL);
1310 sched_unbind(struct thread* td)
1312 THREAD_LOCK_ASSERT(td, MA_OWNED);
1313 td->td_sched->ts_flags &= ~TSF_BOUND;
1317 sched_is_bound(struct thread *td)
1319 THREAD_LOCK_ASSERT(td, MA_OWNED);
1320 return (td->td_sched->ts_flags & TSF_BOUND);
1324 sched_relinquish(struct thread *td)
1327 if (td->td_pri_class == PRI_TIMESHARE)
1328 sched_prio(td, PRI_MAX_TIMESHARE);
1329 SCHED_STAT_INC(switch_relinquish);
1330 mi_switch(SW_VOL, NULL);
1337 return (sched_tdcnt);
1341 sched_sizeof_proc(void)
1343 return (sizeof(struct proc));
1347 sched_sizeof_thread(void)
1349 return (sizeof(struct thread) + sizeof(struct td_sched));
1353 sched_pctcpu(struct thread *td)
1355 struct td_sched *ts;
1358 return (ts->ts_pctcpu);
1367 * The actual idle process.
1370 sched_idletd(void *dummy)
1378 mtx_assert(&Giant, MA_NOTOWNED);
1380 while (sched_runnable() == 0)
1383 mtx_lock_spin(&sched_lock);
1384 mi_switch(SW_VOL, NULL);
1385 mtx_unlock_spin(&sched_lock);
1390 * A CPU is entering for the first time or a thread is exiting.
1393 sched_throw(struct thread *td)
1396 * Correct spinlock nesting. The idle thread context that we are
1397 * borrowing was created so that it would start out with a single
1398 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1399 * explicitly acquired locks in this function, the nesting count
1400 * is now 2 rather than 1. Since we are nested, calling
1401 * spinlock_exit() will simply adjust the counts without allowing
1402 * spin lock using code to interrupt us.
1405 mtx_lock_spin(&sched_lock);
1408 MPASS(td->td_lock == &sched_lock);
1410 mtx_assert(&sched_lock, MA_OWNED);
1411 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1412 PCPU_SET(switchtime, cpu_ticks());
1413 PCPU_SET(switchticks, ticks);
1414 cpu_throw(td, choosethread()); /* doesn't return */
1418 sched_fork_exit(struct thread *td)
1422 * Finish setting up thread glue so that it begins execution in a
1423 * non-nested critical section with sched_lock held but not recursed.
1425 td->td_oncpu = PCPU_GET(cpuid);
1426 sched_lock.mtx_lock = (uintptr_t)td;
1427 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1430 #define KERN_SWITCH_INCLUDE 1
1431 #include "kern/kern_switch.c"