2 * Copyright (c) 2001 Jake Burkholder <jake@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
31 #include "opt_sched.h"
33 #include <sys/param.h>
34 #include <sys/systm.h>
36 #include <sys/kernel.h>
39 #include <sys/mutex.h>
41 #include <sys/queue.h>
42 #include <sys/sched.h>
44 #include <sys/sysctl.h>
46 #include <machine/cpu.h>
48 /* Uncomment this to enable logging of critical_enter/exit. */
50 #define KTR_CRITICAL KTR_SCHED
52 #define KTR_CRITICAL 0
55 #ifdef FULL_PREEMPTION
57 #error "The FULL_PREEMPTION option requires the PREEMPTION option"
61 CTASSERT((RQB_BPW * RQB_LEN) == RQ_NQS);
64 * kern.sched.preemption allows user space to determine if preemption support
65 * is compiled in or not. It is not currently a boot or runtime flag that
69 static int kern_sched_preemption = 1;
71 static int kern_sched_preemption = 0;
73 SYSCTL_INT(_kern_sched, OID_AUTO, preemption, CTLFLAG_RD,
74 &kern_sched_preemption, 0, "Kernel preemption enabled");
77 * Support for scheduler stats exported via kern.sched.stats. All stats may
78 * be reset with kern.sched.stats.reset = 1. Stats may be defined elsewhere
79 * with SCHED_STAT_DEFINE().
82 SYSCTL_NODE(_kern_sched, OID_AUTO, stats, CTLFLAG_RW, 0, "switch stats");
84 /* Switch reasons from mi_switch(). */
85 DPCPU_DEFINE(long, sched_switch_stats[SWT_COUNT]);
86 SCHED_STAT_DEFINE_VAR(uncategorized,
87 &DPCPU_NAME(sched_switch_stats[SWT_NONE]), "");
88 SCHED_STAT_DEFINE_VAR(preempt,
89 &DPCPU_NAME(sched_switch_stats[SWT_PREEMPT]), "");
90 SCHED_STAT_DEFINE_VAR(owepreempt,
91 &DPCPU_NAME(sched_switch_stats[SWT_OWEPREEMPT]), "");
92 SCHED_STAT_DEFINE_VAR(turnstile,
93 &DPCPU_NAME(sched_switch_stats[SWT_TURNSTILE]), "");
94 SCHED_STAT_DEFINE_VAR(sleepq,
95 &DPCPU_NAME(sched_switch_stats[SWT_SLEEPQ]), "");
96 SCHED_STAT_DEFINE_VAR(sleepqtimo,
97 &DPCPU_NAME(sched_switch_stats[SWT_SLEEPQTIMO]), "");
98 SCHED_STAT_DEFINE_VAR(relinquish,
99 &DPCPU_NAME(sched_switch_stats[SWT_RELINQUISH]), "");
100 SCHED_STAT_DEFINE_VAR(needresched,
101 &DPCPU_NAME(sched_switch_stats[SWT_NEEDRESCHED]), "");
102 SCHED_STAT_DEFINE_VAR(idle,
103 &DPCPU_NAME(sched_switch_stats[SWT_IDLE]), "");
104 SCHED_STAT_DEFINE_VAR(iwait,
105 &DPCPU_NAME(sched_switch_stats[SWT_IWAIT]), "");
106 SCHED_STAT_DEFINE_VAR(suspend,
107 &DPCPU_NAME(sched_switch_stats[SWT_SUSPEND]), "");
108 SCHED_STAT_DEFINE_VAR(remotepreempt,
109 &DPCPU_NAME(sched_switch_stats[SWT_REMOTEPREEMPT]), "");
110 SCHED_STAT_DEFINE_VAR(remotewakeidle,
111 &DPCPU_NAME(sched_switch_stats[SWT_REMOTEWAKEIDLE]), "");
114 sysctl_stats_reset(SYSCTL_HANDLER_ARGS)
116 struct sysctl_oid *p;
123 error = sysctl_handle_int(oidp, &val, 0, req);
124 if (error != 0 || req->newptr == NULL)
129 * Traverse the list of children of _kern_sched_stats and reset each
130 * to 0. Skip the reset entry.
132 SLIST_FOREACH(p, oidp->oid_parent, oid_link) {
133 if (p == oidp || p->oid_arg1 == NULL)
135 counter = (uintptr_t)p->oid_arg1;
137 *(long *)(dpcpu_off[i] + counter) = 0;
143 SYSCTL_PROC(_kern_sched_stats, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_WR, NULL,
144 0, sysctl_stats_reset, "I", "Reset scheduler statistics");
147 /************************************************************************
148 * Functions that manipulate runnability from a thread perspective. *
149 ************************************************************************/
151 * Select the thread that will be run next.
154 static __noinline struct thread *
155 choosethread_panic(struct thread *td)
159 * If we are in panic, only allow system threads,
160 * plus the one we are running in, to be run.
163 if (((td->td_proc->p_flag & P_SYSTEM) == 0 &&
164 (td->td_flags & TDF_INPANIC) == 0)) {
165 /* note that it is no longer on the run queue */
182 if (__predict_false(panicstr != NULL))
183 return (choosethread_panic(td));
190 * Kernel thread preemption implementation. Critical sections mark
191 * regions of code in which preemptions are not allowed.
193 * It might seem a good idea to inline critical_enter() but, in order
194 * to prevent instructions reordering by the compiler, a __compiler_membar()
195 * would have to be used here (the same as sched_pin()). The performance
196 * penalty imposed by the membar could, then, produce slower code than
197 * the function call itself, for most cases.
206 CTR4(KTR_CRITICAL, "critical_enter by thread %p (%ld, %s) to %d", td,
207 (long)td->td_proc->p_pid, td->td_name, td->td_critnest);
217 KASSERT(td->td_critnest != 0,
218 ("critical_exit: td_critnest == 0"));
220 if (td->td_critnest == 1) {
224 * Interrupt handlers execute critical_exit() on
225 * leave, and td_owepreempt may be left set by an
226 * interrupt handler only when td_critnest > 0. If we
227 * are decrementing td_critnest from 1 to 0, read
228 * td_owepreempt after decrementing, to not miss the
229 * preempt. Disallow compiler to reorder operations.
232 if (td->td_owepreempt && !kdb_active) {
234 * Microoptimization: we committed to switch,
235 * disable preemption in interrupt handlers
236 * while spinning for the thread lock.
241 flags = SW_INVOL | SW_PREEMPT;
242 if (TD_IS_IDLETHREAD(td))
245 flags |= SWT_OWEPREEMPT;
246 mi_switch(flags, NULL);
252 CTR4(KTR_CRITICAL, "critical_exit by thread %p (%ld, %s) to %d", td,
253 (long)td->td_proc->p_pid, td->td_name, td->td_critnest);
256 /************************************************************************
257 * SYSTEM RUN QUEUE manipulations and tests *
258 ************************************************************************/
260 * Initialize a run structure.
263 runq_init(struct runq *rq)
267 bzero(rq, sizeof *rq);
268 for (i = 0; i < RQ_NQS; i++)
269 TAILQ_INIT(&rq->rq_queues[i]);
273 * Clear the status bit of the queue corresponding to priority level pri,
274 * indicating that it is empty.
277 runq_clrbit(struct runq *rq, int pri)
281 rqb = &rq->rq_status;
282 CTR4(KTR_RUNQ, "runq_clrbit: bits=%#x %#x bit=%#x word=%d",
283 rqb->rqb_bits[RQB_WORD(pri)],
284 rqb->rqb_bits[RQB_WORD(pri)] & ~RQB_BIT(pri),
285 RQB_BIT(pri), RQB_WORD(pri));
286 rqb->rqb_bits[RQB_WORD(pri)] &= ~RQB_BIT(pri);
290 * Find the index of the first non-empty run queue. This is done by
291 * scanning the status bits, a set bit indicates a non-empty queue.
294 runq_findbit(struct runq *rq)
300 rqb = &rq->rq_status;
301 for (i = 0; i < RQB_LEN; i++)
302 if (rqb->rqb_bits[i]) {
303 pri = RQB_FFS(rqb->rqb_bits[i]) + (i << RQB_L2BPW);
304 CTR3(KTR_RUNQ, "runq_findbit: bits=%#x i=%d pri=%d",
305 rqb->rqb_bits[i], i, pri);
313 runq_findbit_from(struct runq *rq, u_char pri)
320 * Set the mask for the first word so we ignore priorities before 'pri'.
322 mask = (rqb_word_t)-1 << (pri & (RQB_BPW - 1));
323 rqb = &rq->rq_status;
325 for (i = RQB_WORD(pri); i < RQB_LEN; mask = -1, i++) {
326 mask = rqb->rqb_bits[i] & mask;
329 pri = RQB_FFS(mask) + (i << RQB_L2BPW);
330 CTR3(KTR_RUNQ, "runq_findbit_from: bits=%#x i=%d pri=%d",
337 * Wrap back around to the beginning of the list just once so we
338 * scan the whole thing.
345 * Set the status bit of the queue corresponding to priority level pri,
346 * indicating that it is non-empty.
349 runq_setbit(struct runq *rq, int pri)
353 rqb = &rq->rq_status;
354 CTR4(KTR_RUNQ, "runq_setbit: bits=%#x %#x bit=%#x word=%d",
355 rqb->rqb_bits[RQB_WORD(pri)],
356 rqb->rqb_bits[RQB_WORD(pri)] | RQB_BIT(pri),
357 RQB_BIT(pri), RQB_WORD(pri));
358 rqb->rqb_bits[RQB_WORD(pri)] |= RQB_BIT(pri);
362 * Add the thread to the queue specified by its priority, and set the
363 * corresponding status bit.
366 runq_add(struct runq *rq, struct thread *td, int flags)
371 pri = td->td_priority / RQ_PPQ;
372 td->td_rqindex = pri;
373 runq_setbit(rq, pri);
374 rqh = &rq->rq_queues[pri];
375 CTR4(KTR_RUNQ, "runq_add: td=%p pri=%d %d rqh=%p",
376 td, td->td_priority, pri, rqh);
377 if (flags & SRQ_PREEMPTED) {
378 TAILQ_INSERT_HEAD(rqh, td, td_runq);
380 TAILQ_INSERT_TAIL(rqh, td, td_runq);
385 runq_add_pri(struct runq *rq, struct thread *td, u_char pri, int flags)
389 KASSERT(pri < RQ_NQS, ("runq_add_pri: %d out of range", pri));
390 td->td_rqindex = pri;
391 runq_setbit(rq, pri);
392 rqh = &rq->rq_queues[pri];
393 CTR4(KTR_RUNQ, "runq_add_pri: td=%p pri=%d idx=%d rqh=%p",
394 td, td->td_priority, pri, rqh);
395 if (flags & SRQ_PREEMPTED) {
396 TAILQ_INSERT_HEAD(rqh, td, td_runq);
398 TAILQ_INSERT_TAIL(rqh, td, td_runq);
402 * Return true if there are runnable processes of any priority on the run
403 * queue, false otherwise. Has no side effects, does not modify the run
407 runq_check(struct runq *rq)
412 rqb = &rq->rq_status;
413 for (i = 0; i < RQB_LEN; i++)
414 if (rqb->rqb_bits[i]) {
415 CTR2(KTR_RUNQ, "runq_check: bits=%#x i=%d",
416 rqb->rqb_bits[i], i);
419 CTR0(KTR_RUNQ, "runq_check: empty");
425 * Find the highest priority process on the run queue.
428 runq_choose_fuzz(struct runq *rq, int fuzz)
434 while ((pri = runq_findbit(rq)) != -1) {
435 rqh = &rq->rq_queues[pri];
436 /* fuzz == 1 is normal.. 0 or less are ignored */
439 * In the first couple of entries, check if
440 * there is one for our CPU as a preference.
443 int cpu = PCPU_GET(cpuid);
445 td2 = td = TAILQ_FIRST(rqh);
447 while (count-- && td2) {
448 if (td2->td_lastcpu == cpu) {
452 td2 = TAILQ_NEXT(td2, td_runq);
455 td = TAILQ_FIRST(rqh);
456 KASSERT(td != NULL, ("runq_choose_fuzz: no proc on busy queue"));
458 "runq_choose_fuzz: pri=%d thread=%p rqh=%p", pri, td, rqh);
461 CTR1(KTR_RUNQ, "runq_choose_fuzz: idleproc pri=%d", pri);
467 * Find the highest priority process on the run queue.
470 runq_choose(struct runq *rq)
476 while ((pri = runq_findbit(rq)) != -1) {
477 rqh = &rq->rq_queues[pri];
478 td = TAILQ_FIRST(rqh);
479 KASSERT(td != NULL, ("runq_choose: no thread on busy queue"));
481 "runq_choose: pri=%d thread=%p rqh=%p", pri, td, rqh);
484 CTR1(KTR_RUNQ, "runq_choose: idlethread pri=%d", pri);
490 runq_choose_from(struct runq *rq, u_char idx)
496 if ((pri = runq_findbit_from(rq, idx)) != -1) {
497 rqh = &rq->rq_queues[pri];
498 td = TAILQ_FIRST(rqh);
499 KASSERT(td != NULL, ("runq_choose: no thread on busy queue"));
501 "runq_choose_from: pri=%d thread=%p idx=%d rqh=%p",
502 pri, td, td->td_rqindex, rqh);
505 CTR1(KTR_RUNQ, "runq_choose_from: idlethread pri=%d", pri);
510 * Remove the thread from the queue specified by its priority, and clear the
511 * corresponding status bit if the queue becomes empty.
512 * Caller must set state afterwards.
515 runq_remove(struct runq *rq, struct thread *td)
518 runq_remove_idx(rq, td, NULL);
522 runq_remove_idx(struct runq *rq, struct thread *td, u_char *idx)
527 KASSERT(td->td_flags & TDF_INMEM,
528 ("runq_remove_idx: thread swapped out"));
529 pri = td->td_rqindex;
530 KASSERT(pri < RQ_NQS, ("runq_remove_idx: Invalid index %d\n", pri));
531 rqh = &rq->rq_queues[pri];
532 CTR4(KTR_RUNQ, "runq_remove_idx: td=%p, pri=%d %d rqh=%p",
533 td, td->td_priority, pri, rqh);
534 TAILQ_REMOVE(rqh, td, td_runq);
535 if (TAILQ_EMPTY(rqh)) {
536 CTR0(KTR_RUNQ, "runq_remove_idx: empty");
537 runq_clrbit(rq, pri);
538 if (idx != NULL && *idx == pri)
539 *idx = (pri + 1) % RQ_NQS;