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. 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"
39 #include "opt_sched.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/cpuset.h>
44 #include <sys/kernel.h>
47 #include <sys/kthread.h>
48 #include <sys/mutex.h>
50 #include <sys/resourcevar.h>
51 #include <sys/sched.h>
54 #include <sys/sysctl.h>
56 #include <sys/turnstile.h>
58 #include <machine/pcb.h>
59 #include <machine/smp.h>
62 #include <sys/pmckern.h>
66 #include <sys/dtrace_bsd.h>
67 int dtrace_vtime_active;
68 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
72 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
73 * the range 100-256 Hz (approximately).
75 #define ESTCPULIM(e) \
76 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
77 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
79 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
81 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
83 #define NICE_WEIGHT 1 /* Priorities per nice level. */
85 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
88 * The schedulable entity that runs a context.
89 * This is an extension to the thread structure and is tailored to
90 * the requirements of this scheduler.
91 * All fields are protected by the scheduler lock.
94 fixpt_t ts_pctcpu; /* %cpu during p_swtime. */
95 u_int ts_estcpu; /* Estimated cpu utilization. */
96 int ts_cpticks; /* Ticks of cpu time. */
97 int ts_slptime; /* Seconds !RUNNING. */
98 int ts_slice; /* Remaining part of time slice. */
100 struct runq *ts_runq; /* runq the thread is currently on */
102 char ts_name[TS_NAME_LEN];
106 /* flags kept in td_flags */
107 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
108 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
109 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
111 /* flags kept in ts_flags */
112 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
114 #define SKE_RUNQ_PCPU(ts) \
115 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
117 #define THREAD_CAN_SCHED(td, cpu) \
118 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
120 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
121 sizeof(struct thread0_storage),
122 "increase struct thread0_storage.t0st_sched size");
124 static struct mtx sched_lock;
126 static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */
127 static int sched_tdcnt; /* Total runnable threads in the system. */
128 static int sched_slice = 12; /* Thread run time before rescheduling. */
130 static void setup_runqs(void);
131 static void schedcpu(void);
132 static void schedcpu_thread(void);
133 static void sched_priority(struct thread *td, u_char prio);
134 static void sched_setup(void *dummy);
135 static void maybe_resched(struct thread *td);
136 static void updatepri(struct thread *td);
137 static void resetpriority(struct thread *td);
138 static void resetpriority_thread(struct thread *td);
140 static int sched_pickcpu(struct thread *td);
141 static int forward_wakeup(int cpunum);
142 static void kick_other_cpu(int pri, int cpuid);
145 static struct kproc_desc sched_kp = {
150 SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start,
152 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
154 static void sched_initticks(void *dummy);
155 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
161 static struct runq runq;
167 static struct runq runq_pcpu[MAXCPU];
168 long runq_length[MAXCPU];
170 static cpuset_t idle_cpus_mask;
173 struct pcpuidlestat {
177 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
185 for (i = 0; i < MAXCPU; ++i)
186 runq_init(&runq_pcpu[i]);
193 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
195 int error, new_val, period;
197 period = 1000000 / realstathz;
198 new_val = period * sched_slice;
199 error = sysctl_handle_int(oidp, &new_val, 0, req);
200 if (error != 0 || req->newptr == NULL)
204 sched_slice = imax(1, (new_val + period / 2) / period);
205 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
210 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
212 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
214 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
215 NULL, 0, sysctl_kern_quantum, "I",
216 "Quantum for timeshare threads in microseconds");
217 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
218 "Quantum for timeshare threads in stathz ticks");
220 /* Enable forwarding of wakeups to all other cpus */
221 static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL,
224 static int runq_fuzz = 1;
225 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
227 static int forward_wakeup_enabled = 1;
228 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
229 &forward_wakeup_enabled, 0,
230 "Forwarding of wakeup to idle CPUs");
232 static int forward_wakeups_requested = 0;
233 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
234 &forward_wakeups_requested, 0,
235 "Requests for Forwarding of wakeup to idle CPUs");
237 static int forward_wakeups_delivered = 0;
238 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
239 &forward_wakeups_delivered, 0,
240 "Completed Forwarding of wakeup to idle CPUs");
242 static int forward_wakeup_use_mask = 1;
243 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
244 &forward_wakeup_use_mask, 0,
245 "Use the mask of idle cpus");
247 static int forward_wakeup_use_loop = 0;
248 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
249 &forward_wakeup_use_loop, 0,
250 "Use a loop to find idle cpus");
254 static int sched_followon = 0;
255 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
257 "allow threads to share a quantum");
260 SDT_PROVIDER_DEFINE(sched);
262 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
263 "struct proc *", "uint8_t");
264 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
265 "struct proc *", "void *");
266 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
267 "struct proc *", "void *", "int");
268 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
269 "struct proc *", "uint8_t", "struct thread *");
270 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
271 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
273 SDT_PROBE_DEFINE(sched, , , on__cpu);
274 SDT_PROBE_DEFINE(sched, , , remain__cpu);
275 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
283 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
284 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
292 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
293 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
296 * Arrange to reschedule if necessary, taking the priorities and
297 * schedulers into account.
300 maybe_resched(struct thread *td)
303 THREAD_LOCK_ASSERT(td, MA_OWNED);
304 if (td->td_priority < curthread->td_priority)
305 curthread->td_flags |= TDF_NEEDRESCHED;
309 * This function is called when a thread is about to be put on run queue
310 * because it has been made runnable or its priority has been adjusted. It
311 * determines if the new thread should preempt the current thread. If so,
312 * it sets td_owepreempt to request a preemption.
315 maybe_preempt(struct thread *td)
322 * The new thread should not preempt the current thread if any of the
323 * following conditions are true:
325 * - The kernel is in the throes of crashing (panicstr).
326 * - The current thread has a higher (numerically lower) or
327 * equivalent priority. Note that this prevents curthread from
328 * trying to preempt to itself.
329 * - The current thread has an inhibitor set or is in the process of
330 * exiting. In this case, the current thread is about to switch
331 * out anyways, so there's no point in preempting. If we did,
332 * the current thread would not be properly resumed as well, so
333 * just avoid that whole landmine.
334 * - If the new thread's priority is not a realtime priority and
335 * the current thread's priority is not an idle priority and
336 * FULL_PREEMPTION is disabled.
338 * If all of these conditions are false, but the current thread is in
339 * a nested critical section, then we have to defer the preemption
340 * until we exit the critical section. Otherwise, switch immediately
344 THREAD_LOCK_ASSERT(td, MA_OWNED);
345 KASSERT((td->td_inhibitors == 0),
346 ("maybe_preempt: trying to run inhibited thread"));
347 pri = td->td_priority;
348 cpri = ctd->td_priority;
349 if (panicstr != NULL || pri >= cpri /* || dumping */ ||
350 TD_IS_INHIBITED(ctd))
352 #ifndef FULL_PREEMPTION
353 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
357 CTR0(KTR_PROC, "maybe_preempt: scheduling preemption");
358 ctd->td_owepreempt = 1;
366 * Constants for digital decay and forget:
367 * 90% of (ts_estcpu) usage in 5 * loadav time
368 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
369 * Note that, as ps(1) mentions, this can let percentages
370 * total over 100% (I've seen 137.9% for 3 processes).
372 * Note that schedclock() updates ts_estcpu and p_cpticks asynchronously.
374 * We wish to decay away 90% of ts_estcpu in (5 * loadavg) seconds.
375 * That is, the system wants to compute a value of decay such
376 * that the following for loop:
377 * for (i = 0; i < (5 * loadavg); i++)
378 * ts_estcpu *= decay;
381 * for all values of loadavg:
383 * Mathematically this loop can be expressed by saying:
384 * decay ** (5 * loadavg) ~= .1
386 * The system computes decay as:
387 * decay = (2 * loadavg) / (2 * loadavg + 1)
389 * We wish to prove that the system's computation of decay
390 * will always fulfill the equation:
391 * decay ** (5 * loadavg) ~= .1
393 * If we compute b as:
396 * decay = b / (b + 1)
398 * We now need to prove two things:
399 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
400 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
403 * For x close to zero, exp(x) =~ 1 + x, since
404 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
405 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
406 * For x close to zero, ln(1+x) =~ x, since
407 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
408 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
412 * Solve (factor)**(power) =~ .1 given power (5*loadav):
413 * solving for factor,
414 * ln(factor) =~ (-2.30/5*loadav), or
415 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
416 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
419 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
421 * power*ln(b/(b+1)) =~ -2.30, or
422 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
424 * Actual power values for the implemented algorithm are as follows:
426 * power: 5.68 10.32 14.94 19.55
429 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
430 #define loadfactor(loadav) (2 * (loadav))
431 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
433 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
434 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
435 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
438 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
439 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
440 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
442 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
443 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
445 * If you don't want to bother with the faster/more-accurate formula, you
446 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
447 * (more general) method of calculating the %age of CPU used by a process.
449 #define CCPU_SHIFT 11
452 * Recompute process priorities, every hz ticks.
453 * MP-safe, called without the Giant mutex.
459 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
465 sx_slock(&allproc_lock);
466 FOREACH_PROC_IN_SYSTEM(p) {
468 if (p->p_state == PRS_NEW) {
472 FOREACH_THREAD_IN_PROC(p, td) {
474 ts = td_get_sched(td);
477 * Increment sleep time (if sleeping). We
478 * ignore overflow, as above.
481 * The td_sched slptimes are not touched in wakeup
482 * because the thread may not HAVE everything in
483 * memory? XXX I think this is out of date.
485 if (TD_ON_RUNQ(td)) {
487 td->td_flags &= ~TDF_DIDRUN;
488 } else if (TD_IS_RUNNING(td)) {
490 /* Do not clear TDF_DIDRUN */
491 } else if (td->td_flags & TDF_DIDRUN) {
493 td->td_flags &= ~TDF_DIDRUN;
497 * ts_pctcpu is only for ps and ttyinfo().
499 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
501 * If the td_sched has been idle the entire second,
502 * stop recalculating its priority until
505 if (ts->ts_cpticks != 0) {
506 #if (FSHIFT >= CCPU_SHIFT)
507 ts->ts_pctcpu += (realstathz == 100)
508 ? ((fixpt_t) ts->ts_cpticks) <<
509 (FSHIFT - CCPU_SHIFT) :
510 100 * (((fixpt_t) ts->ts_cpticks)
511 << (FSHIFT - CCPU_SHIFT)) / realstathz;
513 ts->ts_pctcpu += ((FSCALE - ccpu) *
515 FSCALE / realstathz)) >> FSHIFT;
520 * If there are ANY running threads in this process,
521 * then don't count it as sleeping.
522 * XXX: this is broken.
525 if (ts->ts_slptime > 1) {
527 * In an ideal world, this should not
528 * happen, because whoever woke us
529 * up from the long sleep should have
530 * unwound the slptime and reset our
531 * priority before we run at the stale
532 * priority. Should KASSERT at some
533 * point when all the cases are fixed.
540 if (ts->ts_slptime > 1) {
544 ts->ts_estcpu = decay_cpu(loadfac, ts->ts_estcpu);
546 resetpriority_thread(td);
551 sx_sunlock(&allproc_lock);
555 * Main loop for a kthread that executes schedcpu once a second.
558 schedcpu_thread(void)
568 * Recalculate the priority of a process after it has slept for a while.
569 * For all load averages >= 1 and max ts_estcpu of 255, sleeping for at
570 * least six times the loadfactor will decay ts_estcpu to zero.
573 updatepri(struct thread *td)
579 ts = td_get_sched(td);
580 loadfac = loadfactor(averunnable.ldavg[0]);
581 if (ts->ts_slptime > 5 * loadfac)
584 newcpu = ts->ts_estcpu;
585 ts->ts_slptime--; /* was incremented in schedcpu() */
586 while (newcpu && --ts->ts_slptime)
587 newcpu = decay_cpu(loadfac, newcpu);
588 ts->ts_estcpu = newcpu;
593 * Compute the priority of a process when running in user mode.
594 * Arrange to reschedule if the resulting priority is better
595 * than that of the current process.
598 resetpriority(struct thread *td)
602 if (td->td_pri_class != PRI_TIMESHARE)
604 newpriority = PUSER +
605 td_get_sched(td)->ts_estcpu / INVERSE_ESTCPU_WEIGHT +
606 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
607 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
609 sched_user_prio(td, newpriority);
613 * Update the thread's priority when the associated process's user
617 resetpriority_thread(struct thread *td)
620 /* Only change threads with a time sharing user priority. */
621 if (td->td_priority < PRI_MIN_TIMESHARE ||
622 td->td_priority > PRI_MAX_TIMESHARE)
625 /* XXX the whole needresched thing is broken, but not silly. */
628 sched_prio(td, td->td_user_pri);
633 sched_setup(void *dummy)
638 /* Account for thread0. */
643 * This routine determines time constants after stathz and hz are setup.
646 sched_initticks(void *dummy)
649 realstathz = stathz ? stathz : hz;
650 sched_slice = realstathz / 10; /* ~100ms */
651 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
655 /* External interfaces start here */
658 * Very early in the boot some setup of scheduler-specific
659 * parts of proc0 and of some scheduler resources needs to be done.
668 * Set up the scheduler specific parts of thread0.
670 thread0.td_lock = &sched_lock;
671 td_get_sched(&thread0)->ts_slice = sched_slice;
672 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
679 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
681 return runq_check(&runq);
686 sched_rr_interval(void)
689 /* Convert sched_slice from stathz to hz. */
690 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
694 * We adjust the priority of the current process. The priority of a
695 * process gets worse as it accumulates CPU time. The cpu usage
696 * estimator (ts_estcpu) is increased here. resetpriority() will
697 * compute a different priority each time ts_estcpu increases by
698 * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached). The
699 * cpu usage estimator ramps up quite quickly when the process is
700 * running (linearly), and decays away exponentially, at a rate which
701 * is proportionally slower when the system is busy. The basic
702 * principle is that the system will 90% forget that the process used
703 * a lot of CPU time in 5 * loadav seconds. This causes the system to
704 * favor processes which haven't run much recently, and to round-robin
705 * among other processes.
708 sched_clock(struct thread *td)
710 struct pcpuidlestat *stat;
713 THREAD_LOCK_ASSERT(td, MA_OWNED);
714 ts = td_get_sched(td);
717 ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1);
718 if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
720 resetpriority_thread(td);
724 * Force a context switch if the current thread has used up a full
725 * time slice (default is 100ms).
727 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
728 ts->ts_slice = sched_slice;
729 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
732 stat = DPCPU_PTR(idlestat);
733 stat->oldidlecalls = stat->idlecalls;
738 * Charge child's scheduling CPU usage to parent.
741 sched_exit(struct proc *p, struct thread *td)
744 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
745 "prio:%d", td->td_priority);
747 PROC_LOCK_ASSERT(p, MA_OWNED);
748 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
752 sched_exit_thread(struct thread *td, struct thread *child)
755 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
756 "prio:%d", child->td_priority);
758 td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu +
759 td_get_sched(child)->ts_estcpu);
762 if ((child->td_flags & TDF_NOLOAD) == 0)
764 thread_unlock(child);
768 sched_fork(struct thread *td, struct thread *childtd)
770 sched_fork_thread(td, childtd);
774 sched_fork_thread(struct thread *td, struct thread *childtd)
776 struct td_sched *ts, *tsc;
778 childtd->td_oncpu = NOCPU;
779 childtd->td_lastcpu = NOCPU;
780 childtd->td_lock = &sched_lock;
781 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
782 childtd->td_priority = childtd->td_base_pri;
783 ts = td_get_sched(childtd);
784 bzero(ts, sizeof(*ts));
785 tsc = td_get_sched(td);
786 ts->ts_estcpu = tsc->ts_estcpu;
787 ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY);
792 sched_nice(struct proc *p, int nice)
796 PROC_LOCK_ASSERT(p, MA_OWNED);
798 FOREACH_THREAD_IN_PROC(p, td) {
801 resetpriority_thread(td);
807 sched_class(struct thread *td, int class)
809 THREAD_LOCK_ASSERT(td, MA_OWNED);
810 td->td_pri_class = class;
814 * Adjust the priority of a thread.
817 sched_priority(struct thread *td, u_char prio)
821 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
822 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
823 sched_tdname(curthread));
824 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
825 if (td != curthread && prio > td->td_priority) {
826 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
827 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
828 prio, KTR_ATTR_LINKED, sched_tdname(td));
829 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
832 THREAD_LOCK_ASSERT(td, MA_OWNED);
833 if (td->td_priority == prio)
835 td->td_priority = prio;
836 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
838 sched_add(td, SRQ_BORING);
843 * Update a thread's priority when it is lent another thread's
847 sched_lend_prio(struct thread *td, u_char prio)
850 td->td_flags |= TDF_BORROWING;
851 sched_priority(td, prio);
855 * Restore a thread's priority when priority propagation is
856 * over. The prio argument is the minimum priority the thread
857 * needs to have to satisfy other possible priority lending
858 * requests. If the thread's regulary priority is less
859 * important than prio the thread will keep a priority boost
863 sched_unlend_prio(struct thread *td, u_char prio)
867 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
868 td->td_base_pri <= PRI_MAX_TIMESHARE)
869 base_pri = td->td_user_pri;
871 base_pri = td->td_base_pri;
872 if (prio >= base_pri) {
873 td->td_flags &= ~TDF_BORROWING;
874 sched_prio(td, base_pri);
876 sched_lend_prio(td, prio);
880 sched_prio(struct thread *td, u_char prio)
884 /* First, update the base priority. */
885 td->td_base_pri = prio;
888 * If the thread is borrowing another thread's priority, don't ever
889 * lower the priority.
891 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
894 /* Change the real priority. */
895 oldprio = td->td_priority;
896 sched_priority(td, prio);
899 * If the thread is on a turnstile, then let the turnstile update
902 if (TD_ON_LOCK(td) && oldprio != prio)
903 turnstile_adjust(td, oldprio);
907 sched_user_prio(struct thread *td, u_char prio)
910 THREAD_LOCK_ASSERT(td, MA_OWNED);
911 td->td_base_user_pri = prio;
912 if (td->td_lend_user_pri <= prio)
914 td->td_user_pri = prio;
918 sched_lend_user_prio(struct thread *td, u_char prio)
921 THREAD_LOCK_ASSERT(td, MA_OWNED);
922 td->td_lend_user_pri = prio;
923 td->td_user_pri = min(prio, td->td_base_user_pri);
924 if (td->td_priority > td->td_user_pri)
925 sched_prio(td, td->td_user_pri);
926 else if (td->td_priority != td->td_user_pri)
927 td->td_flags |= TDF_NEEDRESCHED;
931 sched_sleep(struct thread *td, int pri)
934 THREAD_LOCK_ASSERT(td, MA_OWNED);
935 td->td_slptick = ticks;
936 td_get_sched(td)->ts_slptime = 0;
937 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
939 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
940 td->td_flags |= TDF_CANSWAP;
944 sched_switch(struct thread *td, struct thread *newtd, int flags)
952 ts = td_get_sched(td);
955 THREAD_LOCK_ASSERT(td, MA_OWNED);
958 * Switch to the sched lock to fix things up and pick
960 * Block the td_lock in order to avoid breaking the critical path.
962 if (td->td_lock != &sched_lock) {
963 mtx_lock_spin(&sched_lock);
964 tmtx = thread_lock_block(td);
967 if ((td->td_flags & TDF_NOLOAD) == 0)
970 td->td_lastcpu = td->td_oncpu;
971 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
972 (flags & SW_PREEMPT) != 0;
973 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
974 td->td_owepreempt = 0;
975 td->td_oncpu = NOCPU;
978 * At the last moment, if this thread is still marked RUNNING,
979 * then put it back on the run queue as it has not been suspended
980 * or stopped or any thing else similar. We never put the idle
981 * threads on the run queue, however.
983 if (td->td_flags & TDF_IDLETD) {
986 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
989 if (TD_IS_RUNNING(td)) {
990 /* Put us back on the run queue. */
991 sched_add(td, preempted ?
992 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
993 SRQ_OURSELF|SRQ_YIELDING);
998 * The thread we are about to run needs to be counted
999 * as if it had been added to the run queue and selected.
1005 KASSERT((newtd->td_inhibitors == 0),
1006 ("trying to run inhibited thread"));
1007 newtd->td_flags |= TDF_DIDRUN;
1008 TD_SET_RUNNING(newtd);
1009 if ((newtd->td_flags & TDF_NOLOAD) == 0)
1012 newtd = choosethread();
1013 MPASS(newtd->td_lock == &sched_lock);
1016 #if (KTR_COMPILE & KTR_SCHED) != 0
1017 if (TD_IS_IDLETHREAD(td))
1018 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1019 "prio:%d", td->td_priority);
1021 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1022 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1023 "lockname:\"%s\"", td->td_lockname);
1028 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1029 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1032 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1035 lock_profile_release_lock(&sched_lock.lock_object);
1036 #ifdef KDTRACE_HOOKS
1038 * If DTrace has set the active vtime enum to anything
1039 * other than INACTIVE (0), then it should have set the
1042 if (dtrace_vtime_active)
1043 (*dtrace_vtime_switch_func)(newtd);
1046 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1047 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1048 0, 0, __FILE__, __LINE__);
1050 * Where am I? What year is it?
1051 * We are in the same thread that went to sleep above,
1052 * but any amount of time may have passed. All our context
1053 * will still be available as will local variables.
1054 * PCPU values however may have changed as we may have
1055 * changed CPU so don't trust cached values of them.
1056 * New threads will go to fork_exit() instead of here
1057 * so if you change things here you may need to change
1060 * If the thread above was exiting it will never wake
1061 * up again here, so either it has saved everything it
1062 * needed to, or the thread_wait() or wait() will
1066 SDT_PROBE0(sched, , , on__cpu);
1068 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1069 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1072 SDT_PROBE0(sched, , , remain__cpu);
1074 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1075 "prio:%d", td->td_priority);
1078 if (td->td_flags & TDF_IDLETD)
1079 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1081 sched_lock.mtx_lock = (uintptr_t)td;
1082 td->td_oncpu = PCPU_GET(cpuid);
1083 MPASS(td->td_lock == &sched_lock);
1087 sched_wakeup(struct thread *td)
1089 struct td_sched *ts;
1091 THREAD_LOCK_ASSERT(td, MA_OWNED);
1092 ts = td_get_sched(td);
1093 td->td_flags &= ~TDF_CANSWAP;
1094 if (ts->ts_slptime > 1) {
1100 ts->ts_slice = sched_slice;
1101 sched_add(td, SRQ_BORING);
1106 forward_wakeup(int cpunum)
1109 cpuset_t dontuse, map, map2;
1113 mtx_assert(&sched_lock, MA_OWNED);
1115 CTR0(KTR_RUNQ, "forward_wakeup()");
1117 if ((!forward_wakeup_enabled) ||
1118 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1120 if (!smp_started || panicstr)
1123 forward_wakeups_requested++;
1126 * Check the idle mask we received against what we calculated
1127 * before in the old version.
1129 me = PCPU_GET(cpuid);
1131 /* Don't bother if we should be doing it ourself. */
1132 if (CPU_ISSET(me, &idle_cpus_mask) &&
1133 (cpunum == NOCPU || me == cpunum))
1136 CPU_SETOF(me, &dontuse);
1137 CPU_OR(&dontuse, &stopped_cpus);
1138 CPU_OR(&dontuse, &hlt_cpus_mask);
1140 if (forward_wakeup_use_loop) {
1141 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1143 if (!CPU_ISSET(id, &dontuse) &&
1144 pc->pc_curthread == pc->pc_idlethread) {
1150 if (forward_wakeup_use_mask) {
1151 map = idle_cpus_mask;
1152 CPU_NAND(&map, &dontuse);
1154 /* If they are both on, compare and use loop if different. */
1155 if (forward_wakeup_use_loop) {
1156 if (CPU_CMP(&map, &map2)) {
1157 printf("map != map2, loop method preferred\n");
1165 /* If we only allow a specific CPU, then mask off all the others. */
1166 if (cpunum != NOCPU) {
1167 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1168 iscpuset = CPU_ISSET(cpunum, &map);
1172 CPU_SETOF(cpunum, &map);
1174 if (!CPU_EMPTY(&map)) {
1175 forward_wakeups_delivered++;
1176 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1178 if (!CPU_ISSET(id, &map))
1180 if (cpu_idle_wakeup(pc->pc_cpuid))
1183 if (!CPU_EMPTY(&map))
1184 ipi_selected(map, IPI_AST);
1187 if (cpunum == NOCPU)
1188 printf("forward_wakeup: Idle processor not found\n");
1193 kick_other_cpu(int pri, int cpuid)
1198 pcpu = pcpu_find(cpuid);
1199 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1200 forward_wakeups_delivered++;
1201 if (!cpu_idle_wakeup(cpuid))
1202 ipi_cpu(cpuid, IPI_AST);
1206 cpri = pcpu->pc_curthread->td_priority;
1210 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1211 #if !defined(FULL_PREEMPTION)
1212 if (pri <= PRI_MAX_ITHD)
1213 #endif /* ! FULL_PREEMPTION */
1215 ipi_cpu(cpuid, IPI_PREEMPT);
1218 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1220 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1221 ipi_cpu(cpuid, IPI_AST);
1228 sched_pickcpu(struct thread *td)
1232 mtx_assert(&sched_lock, MA_OWNED);
1234 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu))
1235 best = td->td_lastcpu;
1239 if (!THREAD_CAN_SCHED(td, cpu))
1244 else if (runq_length[cpu] < runq_length[best])
1247 KASSERT(best != NOCPU, ("no valid CPUs"));
1254 sched_add(struct thread *td, int flags)
1258 struct td_sched *ts;
1263 ts = td_get_sched(td);
1264 THREAD_LOCK_ASSERT(td, MA_OWNED);
1265 KASSERT((td->td_inhibitors == 0),
1266 ("sched_add: trying to run inhibited thread"));
1267 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1268 ("sched_add: bad thread state"));
1269 KASSERT(td->td_flags & TDF_INMEM,
1270 ("sched_add: thread swapped out"));
1272 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1273 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1274 sched_tdname(curthread));
1275 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1276 KTR_ATTR_LINKED, sched_tdname(td));
1277 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1278 flags & SRQ_PREEMPTED);
1282 * Now that the thread is moving to the run-queue, set the lock
1283 * to the scheduler's lock.
1285 if (td->td_lock != &sched_lock) {
1286 mtx_lock_spin(&sched_lock);
1287 thread_lock_set(td, &sched_lock);
1292 * If SMP is started and the thread is pinned or otherwise limited to
1293 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1294 * Otherwise, queue the thread to the global run queue.
1296 * If SMP has not yet been started we must use the global run queue
1297 * as per-CPU state may not be initialized yet and we may crash if we
1298 * try to access the per-CPU run queues.
1300 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1301 ts->ts_flags & TSF_AFFINITY)) {
1302 if (td->td_pinned != 0)
1303 cpu = td->td_lastcpu;
1304 else if (td->td_flags & TDF_BOUND) {
1305 /* Find CPU from bound runq. */
1306 KASSERT(SKE_RUNQ_PCPU(ts),
1307 ("sched_add: bound td_sched not on cpu runq"));
1308 cpu = ts->ts_runq - &runq_pcpu[0];
1310 /* Find a valid CPU for our cpuset */
1311 cpu = sched_pickcpu(td);
1312 ts->ts_runq = &runq_pcpu[cpu];
1315 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1319 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1322 ts->ts_runq = &runq;
1325 if ((td->td_flags & TDF_NOLOAD) == 0)
1327 runq_add(ts->ts_runq, td, flags);
1331 cpuid = PCPU_GET(cpuid);
1332 if (single_cpu && cpu != cpuid) {
1333 kick_other_cpu(td->td_priority, cpu);
1336 tidlemsk = idle_cpus_mask;
1337 CPU_NAND(&tidlemsk, &hlt_cpus_mask);
1338 CPU_CLR(cpuid, &tidlemsk);
1340 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1341 ((flags & SRQ_INTR) == 0) &&
1342 !CPU_EMPTY(&tidlemsk))
1343 forwarded = forward_wakeup(cpu);
1347 if (!maybe_preempt(td))
1354 struct td_sched *ts;
1356 ts = td_get_sched(td);
1357 THREAD_LOCK_ASSERT(td, MA_OWNED);
1358 KASSERT((td->td_inhibitors == 0),
1359 ("sched_add: trying to run inhibited thread"));
1360 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1361 ("sched_add: bad thread state"));
1362 KASSERT(td->td_flags & TDF_INMEM,
1363 ("sched_add: thread swapped out"));
1364 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1365 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1366 sched_tdname(curthread));
1367 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1368 KTR_ATTR_LINKED, sched_tdname(td));
1369 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1370 flags & SRQ_PREEMPTED);
1373 * Now that the thread is moving to the run-queue, set the lock
1374 * to the scheduler's lock.
1376 if (td->td_lock != &sched_lock) {
1377 mtx_lock_spin(&sched_lock);
1378 thread_lock_set(td, &sched_lock);
1381 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1382 ts->ts_runq = &runq;
1384 if ((td->td_flags & TDF_NOLOAD) == 0)
1386 runq_add(ts->ts_runq, td, flags);
1387 if (!maybe_preempt(td))
1393 sched_rem(struct thread *td)
1395 struct td_sched *ts;
1397 ts = td_get_sched(td);
1398 KASSERT(td->td_flags & TDF_INMEM,
1399 ("sched_rem: thread swapped out"));
1400 KASSERT(TD_ON_RUNQ(td),
1401 ("sched_rem: thread not on run queue"));
1402 mtx_assert(&sched_lock, MA_OWNED);
1403 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1404 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1405 sched_tdname(curthread));
1406 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1408 if ((td->td_flags & TDF_NOLOAD) == 0)
1411 if (ts->ts_runq != &runq)
1412 runq_length[ts->ts_runq - runq_pcpu]--;
1414 runq_remove(ts->ts_runq, td);
1419 * Select threads to run. Note that running threads still consume a
1428 mtx_assert(&sched_lock, MA_OWNED);
1430 struct thread *tdcpu;
1433 td = runq_choose_fuzz(&runq, runq_fuzz);
1434 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1438 tdcpu->td_priority < td->td_priority)) {
1439 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1442 rq = &runq_pcpu[PCPU_GET(cpuid)];
1444 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1449 td = runq_choose(&runq);
1455 runq_length[PCPU_GET(cpuid)]--;
1457 runq_remove(rq, td);
1458 td->td_flags |= TDF_DIDRUN;
1460 KASSERT(td->td_flags & TDF_INMEM,
1461 ("sched_choose: thread swapped out"));
1464 return (PCPU_GET(idlethread));
1468 sched_preempt(struct thread *td)
1471 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1473 if (td->td_critnest > 1)
1474 td->td_owepreempt = 1;
1476 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1481 sched_userret(struct thread *td)
1484 * XXX we cheat slightly on the locking here to avoid locking in
1485 * the usual case. Setting td_priority here is essentially an
1486 * incomplete workaround for not setting it properly elsewhere.
1487 * Now that some interrupt handlers are threads, not setting it
1488 * properly elsewhere can clobber it in the window between setting
1489 * it here and returning to user mode, so don't waste time setting
1490 * it perfectly here.
1492 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1493 ("thread with borrowed priority returning to userland"));
1494 if (td->td_priority != td->td_user_pri) {
1496 td->td_priority = td->td_user_pri;
1497 td->td_base_pri = td->td_user_pri;
1503 sched_bind(struct thread *td, int cpu)
1505 struct td_sched *ts;
1507 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1508 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1510 ts = td_get_sched(td);
1512 td->td_flags |= TDF_BOUND;
1514 ts->ts_runq = &runq_pcpu[cpu];
1515 if (PCPU_GET(cpuid) == cpu)
1518 mi_switch(SW_VOL, NULL);
1523 sched_unbind(struct thread* td)
1525 THREAD_LOCK_ASSERT(td, MA_OWNED);
1526 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1527 td->td_flags &= ~TDF_BOUND;
1531 sched_is_bound(struct thread *td)
1533 THREAD_LOCK_ASSERT(td, MA_OWNED);
1534 return (td->td_flags & TDF_BOUND);
1538 sched_relinquish(struct thread *td)
1541 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1548 return (sched_tdcnt);
1552 sched_sizeof_proc(void)
1554 return (sizeof(struct proc));
1558 sched_sizeof_thread(void)
1560 return (sizeof(struct thread) + sizeof(struct td_sched));
1564 sched_pctcpu(struct thread *td)
1566 struct td_sched *ts;
1568 THREAD_LOCK_ASSERT(td, MA_OWNED);
1569 ts = td_get_sched(td);
1570 return (ts->ts_pctcpu);
1575 * Calculates the contribution to the thread cpu usage for the latest
1576 * (unfinished) second.
1579 sched_pctcpu_delta(struct thread *td)
1581 struct td_sched *ts;
1585 THREAD_LOCK_ASSERT(td, MA_OWNED);
1586 ts = td_get_sched(td);
1588 realstathz = stathz ? stathz : hz;
1589 if (ts->ts_cpticks != 0) {
1590 #if (FSHIFT >= CCPU_SHIFT)
1591 delta = (realstathz == 100)
1592 ? ((fixpt_t) ts->ts_cpticks) <<
1593 (FSHIFT - CCPU_SHIFT) :
1594 100 * (((fixpt_t) ts->ts_cpticks)
1595 << (FSHIFT - CCPU_SHIFT)) / realstathz;
1597 delta = ((FSCALE - ccpu) *
1599 FSCALE / realstathz)) >> FSHIFT;
1608 sched_estcpu(struct thread *td)
1611 return (td_get_sched(td)->ts_estcpu);
1615 * The actual idle process.
1618 sched_idletd(void *dummy)
1620 struct pcpuidlestat *stat;
1622 THREAD_NO_SLEEPING();
1623 stat = DPCPU_PTR(idlestat);
1625 mtx_assert(&Giant, MA_NOTOWNED);
1627 while (sched_runnable() == 0) {
1628 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1632 mtx_lock_spin(&sched_lock);
1633 mi_switch(SW_VOL | SWT_IDLE, NULL);
1634 mtx_unlock_spin(&sched_lock);
1639 * A CPU is entering for the first time or a thread is exiting.
1642 sched_throw(struct thread *td)
1645 * Correct spinlock nesting. The idle thread context that we are
1646 * borrowing was created so that it would start out with a single
1647 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1648 * explicitly acquired locks in this function, the nesting count
1649 * is now 2 rather than 1. Since we are nested, calling
1650 * spinlock_exit() will simply adjust the counts without allowing
1651 * spin lock using code to interrupt us.
1654 mtx_lock_spin(&sched_lock);
1656 PCPU_SET(switchtime, cpu_ticks());
1657 PCPU_SET(switchticks, ticks);
1659 lock_profile_release_lock(&sched_lock.lock_object);
1660 MPASS(td->td_lock == &sched_lock);
1661 td->td_lastcpu = td->td_oncpu;
1662 td->td_oncpu = NOCPU;
1664 mtx_assert(&sched_lock, MA_OWNED);
1665 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1666 cpu_throw(td, choosethread()); /* doesn't return */
1670 sched_fork_exit(struct thread *td)
1674 * Finish setting up thread glue so that it begins execution in a
1675 * non-nested critical section with sched_lock held but not recursed.
1677 td->td_oncpu = PCPU_GET(cpuid);
1678 sched_lock.mtx_lock = (uintptr_t)td;
1679 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1680 0, 0, __FILE__, __LINE__);
1681 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1683 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1684 "prio:%d", td->td_priority);
1685 SDT_PROBE0(sched, , , on__cpu);
1689 sched_tdname(struct thread *td)
1692 struct td_sched *ts;
1694 ts = td_get_sched(td);
1695 if (ts->ts_name[0] == '\0')
1696 snprintf(ts->ts_name, sizeof(ts->ts_name),
1697 "%s tid %d", td->td_name, td->td_tid);
1698 return (ts->ts_name);
1700 return (td->td_name);
1706 sched_clear_tdname(struct thread *td)
1708 struct td_sched *ts;
1710 ts = td_get_sched(td);
1711 ts->ts_name[0] = '\0';
1716 sched_affinity(struct thread *td)
1719 struct td_sched *ts;
1722 THREAD_LOCK_ASSERT(td, MA_OWNED);
1725 * Set the TSF_AFFINITY flag if there is at least one CPU this
1726 * thread can't run on.
1728 ts = td_get_sched(td);
1729 ts->ts_flags &= ~TSF_AFFINITY;
1731 if (!THREAD_CAN_SCHED(td, cpu)) {
1732 ts->ts_flags |= TSF_AFFINITY;
1738 * If this thread can run on all CPUs, nothing else to do.
1740 if (!(ts->ts_flags & TSF_AFFINITY))
1743 /* Pinned threads and bound threads should be left alone. */
1744 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1747 switch (td->td_state) {
1750 * If we are on a per-CPU runqueue that is in the set,
1751 * then nothing needs to be done.
1753 if (ts->ts_runq != &runq &&
1754 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1757 /* Put this thread on a valid per-CPU runqueue. */
1759 sched_add(td, SRQ_BORING);
1763 * See if our current CPU is in the set. If not, force a
1766 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1769 td->td_flags |= TDF_NEEDRESCHED;
1770 if (td != curthread)
1771 ipi_cpu(cpu, IPI_AST);