xref: /freebsd/sys/kern/kern_switch.c (revision 1e413cf93298b5b97441a21d9a50fdcd0ee9945e)
1 /*-
2  * Copyright (c) 2001 Jake Burkholder <jake@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
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.
13  *
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
24  * SUCH DAMAGE.
25  */
26 
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_sched.h"
32 
33 #ifndef KERN_SWITCH_INCLUDE
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kdb.h>
37 #include <sys/kernel.h>
38 #include <sys/ktr.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/proc.h>
42 #include <sys/queue.h>
43 #include <sys/sched.h>
44 #else  /* KERN_SWITCH_INCLUDE */
45 #if defined(SMP) && (defined(__i386__) || defined(__amd64__))
46 #include <sys/smp.h>
47 #endif
48 #if defined(SMP) && defined(SCHED_4BSD)
49 #include <sys/sysctl.h>
50 #endif
51 
52 #include <machine/cpu.h>
53 
54 /* Uncomment this to enable logging of critical_enter/exit. */
55 #if 0
56 #define	KTR_CRITICAL	KTR_SCHED
57 #else
58 #define	KTR_CRITICAL	0
59 #endif
60 
61 #ifdef FULL_PREEMPTION
62 #ifndef PREEMPTION
63 #error "The FULL_PREEMPTION option requires the PREEMPTION option"
64 #endif
65 #endif
66 
67 CTASSERT((RQB_BPW * RQB_LEN) == RQ_NQS);
68 
69 /*
70  * kern.sched.preemption allows user space to determine if preemption support
71  * is compiled in or not.  It is not currently a boot or runtime flag that
72  * can be changed.
73  */
74 #ifdef PREEMPTION
75 static int kern_sched_preemption = 1;
76 #else
77 static int kern_sched_preemption = 0;
78 #endif
79 SYSCTL_INT(_kern_sched, OID_AUTO, preemption, CTLFLAG_RD,
80     &kern_sched_preemption, 0, "Kernel preemption enabled");
81 
82 #ifdef SCHED_STATS
83 long switch_preempt;
84 long switch_owepreempt;
85 long switch_turnstile;
86 long switch_sleepq;
87 long switch_sleepqtimo;
88 long switch_relinquish;
89 long switch_needresched;
90 static SYSCTL_NODE(_kern_sched, OID_AUTO, stats, CTLFLAG_RW, 0, "switch stats");
91 SYSCTL_INT(_kern_sched_stats, OID_AUTO, preempt, CTLFLAG_RD, &switch_preempt, 0, "");
92 SYSCTL_INT(_kern_sched_stats, OID_AUTO, owepreempt, CTLFLAG_RD, &switch_owepreempt, 0, "");
93 SYSCTL_INT(_kern_sched_stats, OID_AUTO, turnstile, CTLFLAG_RD, &switch_turnstile, 0, "");
94 SYSCTL_INT(_kern_sched_stats, OID_AUTO, sleepq, CTLFLAG_RD, &switch_sleepq, 0, "");
95 SYSCTL_INT(_kern_sched_stats, OID_AUTO, sleepqtimo, CTLFLAG_RD, &switch_sleepqtimo, 0, "");
96 SYSCTL_INT(_kern_sched_stats, OID_AUTO, relinquish, CTLFLAG_RD, &switch_relinquish, 0, "");
97 SYSCTL_INT(_kern_sched_stats, OID_AUTO, needresched, CTLFLAG_RD, &switch_needresched, 0, "");
98 static int
99 sysctl_stats_reset(SYSCTL_HANDLER_ARGS)
100 {
101         int error;
102 	int val;
103 
104         val = 0;
105         error = sysctl_handle_int(oidp, &val, 0, req);
106         if (error != 0 || req->newptr == NULL)
107                 return (error);
108         if (val == 0)
109                 return (0);
110 	switch_preempt = 0;
111 	switch_owepreempt = 0;
112 	switch_turnstile = 0;
113 	switch_sleepq = 0;
114 	switch_sleepqtimo = 0;
115 	switch_relinquish = 0;
116 	switch_needresched = 0;
117 
118 	return (0);
119 }
120 
121 SYSCTL_PROC(_kern_sched_stats, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_WR, NULL,
122     0, sysctl_stats_reset, "I", "Reset scheduler statistics");
123 #endif
124 
125 /************************************************************************
126  * Functions that manipulate runnability from a thread perspective.	*
127  ************************************************************************/
128 /*
129  * Select the thread that will be run next.
130  */
131 struct thread *
132 choosethread(void)
133 {
134 	struct thread *td;
135 
136 retry:
137 	td = sched_choose();
138 
139 	/*
140 	 * If we are in panic, only allow system threads,
141 	 * plus the one we are running in, to be run.
142 	 */
143 	if (panicstr && ((td->td_proc->p_flag & P_SYSTEM) == 0 &&
144 	    (td->td_flags & TDF_INPANIC) == 0)) {
145 		/* note that it is no longer on the run queue */
146 		TD_SET_CAN_RUN(td);
147 		goto retry;
148 	}
149 
150 	TD_SET_RUNNING(td);
151 	return (td);
152 }
153 
154 /*
155  * Kernel thread preemption implementation.  Critical sections mark
156  * regions of code in which preemptions are not allowed.
157  */
158 void
159 critical_enter(void)
160 {
161 	struct thread *td;
162 
163 	td = curthread;
164 	td->td_critnest++;
165 	CTR4(KTR_CRITICAL, "critical_enter by thread %p (%ld, %s) to %d", td,
166 	    (long)td->td_proc->p_pid, td->td_name, td->td_critnest);
167 }
168 
169 void
170 critical_exit(void)
171 {
172 	struct thread *td;
173 
174 	td = curthread;
175 	KASSERT(td->td_critnest != 0,
176 	    ("critical_exit: td_critnest == 0"));
177 
178 	if (td->td_critnest == 1) {
179 		td->td_critnest = 0;
180 		if (td->td_owepreempt) {
181 			td->td_critnest = 1;
182 			thread_lock(td);
183 			td->td_critnest--;
184 			SCHED_STAT_INC(switch_owepreempt);
185 			mi_switch(SW_INVOL|SW_PREEMPT, NULL);
186 			thread_unlock(td);
187 		}
188 	} else
189 		td->td_critnest--;
190 
191 	CTR4(KTR_CRITICAL, "critical_exit by thread %p (%ld, %s) to %d", td,
192 	    (long)td->td_proc->p_pid, td->td_name, td->td_critnest);
193 }
194 
195 /*
196  * This function is called when a thread is about to be put on run queue
197  * because it has been made runnable or its priority has been adjusted.  It
198  * determines if the new thread should be immediately preempted to.  If so,
199  * it switches to it and eventually returns true.  If not, it returns false
200  * so that the caller may place the thread on an appropriate run queue.
201  */
202 int
203 maybe_preempt(struct thread *td)
204 {
205 #ifdef PREEMPTION
206 	struct thread *ctd;
207 	int cpri, pri;
208 #endif
209 
210 #ifdef PREEMPTION
211 	/*
212 	 * The new thread should not preempt the current thread if any of the
213 	 * following conditions are true:
214 	 *
215 	 *  - The kernel is in the throes of crashing (panicstr).
216 	 *  - The current thread has a higher (numerically lower) or
217 	 *    equivalent priority.  Note that this prevents curthread from
218 	 *    trying to preempt to itself.
219 	 *  - It is too early in the boot for context switches (cold is set).
220 	 *  - The current thread has an inhibitor set or is in the process of
221 	 *    exiting.  In this case, the current thread is about to switch
222 	 *    out anyways, so there's no point in preempting.  If we did,
223 	 *    the current thread would not be properly resumed as well, so
224 	 *    just avoid that whole landmine.
225 	 *  - If the new thread's priority is not a realtime priority and
226 	 *    the current thread's priority is not an idle priority and
227 	 *    FULL_PREEMPTION is disabled.
228 	 *
229 	 * If all of these conditions are false, but the current thread is in
230 	 * a nested critical section, then we have to defer the preemption
231 	 * until we exit the critical section.  Otherwise, switch immediately
232 	 * to the new thread.
233 	 */
234 	ctd = curthread;
235 	THREAD_LOCK_ASSERT(td, MA_OWNED);
236 	KASSERT ((ctd->td_sched != NULL && ctd->td_sched->ts_thread == ctd),
237 	  ("thread has no (or wrong) sched-private part."));
238 	KASSERT((td->td_inhibitors == 0),
239 			("maybe_preempt: trying to run inhibited thread"));
240 	pri = td->td_priority;
241 	cpri = ctd->td_priority;
242 	if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
243 	    TD_IS_INHIBITED(ctd))
244 		return (0);
245 #ifndef FULL_PREEMPTION
246 	if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
247 		return (0);
248 #endif
249 
250 	if (ctd->td_critnest > 1) {
251 		CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
252 		    ctd->td_critnest);
253 		ctd->td_owepreempt = 1;
254 		return (0);
255 	}
256 	/*
257 	 * Thread is runnable but not yet put on system run queue.
258 	 */
259 	MPASS(ctd->td_lock == td->td_lock);
260 	MPASS(TD_ON_RUNQ(td));
261 	TD_SET_RUNNING(td);
262 	CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
263 	    td->td_proc->p_pid, td->td_name);
264 	SCHED_STAT_INC(switch_preempt);
265 	mi_switch(SW_INVOL|SW_PREEMPT, td);
266 	/*
267 	 * td's lock pointer may have changed.  We have to return with it
268 	 * locked.
269 	 */
270 	spinlock_enter();
271 	thread_unlock(ctd);
272 	thread_lock(td);
273 	spinlock_exit();
274 	return (1);
275 #else
276 	return (0);
277 #endif
278 }
279 
280 #if 0
281 #ifndef PREEMPTION
282 /* XXX: There should be a non-static version of this. */
283 static void
284 printf_caddr_t(void *data)
285 {
286 	printf("%s", (char *)data);
287 }
288 static char preempt_warning[] =
289     "WARNING: Kernel preemption is disabled, expect reduced performance.\n";
290 SYSINIT(preempt_warning, SI_SUB_COPYRIGHT, SI_ORDER_ANY, printf_caddr_t,
291     preempt_warning)
292 #endif
293 #endif
294 
295 /************************************************************************
296  * SYSTEM RUN QUEUE manipulations and tests				*
297  ************************************************************************/
298 /*
299  * Initialize a run structure.
300  */
301 void
302 runq_init(struct runq *rq)
303 {
304 	int i;
305 
306 	bzero(rq, sizeof *rq);
307 	for (i = 0; i < RQ_NQS; i++)
308 		TAILQ_INIT(&rq->rq_queues[i]);
309 }
310 
311 /*
312  * Clear the status bit of the queue corresponding to priority level pri,
313  * indicating that it is empty.
314  */
315 static __inline void
316 runq_clrbit(struct runq *rq, int pri)
317 {
318 	struct rqbits *rqb;
319 
320 	rqb = &rq->rq_status;
321 	CTR4(KTR_RUNQ, "runq_clrbit: bits=%#x %#x bit=%#x word=%d",
322 	    rqb->rqb_bits[RQB_WORD(pri)],
323 	    rqb->rqb_bits[RQB_WORD(pri)] & ~RQB_BIT(pri),
324 	    RQB_BIT(pri), RQB_WORD(pri));
325 	rqb->rqb_bits[RQB_WORD(pri)] &= ~RQB_BIT(pri);
326 }
327 
328 /*
329  * Find the index of the first non-empty run queue.  This is done by
330  * scanning the status bits, a set bit indicates a non-empty queue.
331  */
332 static __inline int
333 runq_findbit(struct runq *rq)
334 {
335 	struct rqbits *rqb;
336 	int pri;
337 	int i;
338 
339 	rqb = &rq->rq_status;
340 	for (i = 0; i < RQB_LEN; i++)
341 		if (rqb->rqb_bits[i]) {
342 			pri = RQB_FFS(rqb->rqb_bits[i]) + (i << RQB_L2BPW);
343 			CTR3(KTR_RUNQ, "runq_findbit: bits=%#x i=%d pri=%d",
344 			    rqb->rqb_bits[i], i, pri);
345 			return (pri);
346 		}
347 
348 	return (-1);
349 }
350 
351 static __inline int
352 runq_findbit_from(struct runq *rq, u_char pri)
353 {
354 	struct rqbits *rqb;
355 	rqb_word_t mask;
356 	int i;
357 
358 	/*
359 	 * Set the mask for the first word so we ignore priorities before 'pri'.
360 	 */
361 	mask = (rqb_word_t)-1 << (pri & (RQB_BPW - 1));
362 	rqb = &rq->rq_status;
363 again:
364 	for (i = RQB_WORD(pri); i < RQB_LEN; mask = -1, i++) {
365 		mask = rqb->rqb_bits[i] & mask;
366 		if (mask == 0)
367 			continue;
368 		pri = RQB_FFS(mask) + (i << RQB_L2BPW);
369 		CTR3(KTR_RUNQ, "runq_findbit_from: bits=%#x i=%d pri=%d",
370 		    mask, i, pri);
371 		return (pri);
372 	}
373 	if (pri == 0)
374 		return (-1);
375 	/*
376 	 * Wrap back around to the beginning of the list just once so we
377 	 * scan the whole thing.
378 	 */
379 	pri = 0;
380 	goto again;
381 }
382 
383 /*
384  * Set the status bit of the queue corresponding to priority level pri,
385  * indicating that it is non-empty.
386  */
387 static __inline void
388 runq_setbit(struct runq *rq, int pri)
389 {
390 	struct rqbits *rqb;
391 
392 	rqb = &rq->rq_status;
393 	CTR4(KTR_RUNQ, "runq_setbit: bits=%#x %#x bit=%#x word=%d",
394 	    rqb->rqb_bits[RQB_WORD(pri)],
395 	    rqb->rqb_bits[RQB_WORD(pri)] | RQB_BIT(pri),
396 	    RQB_BIT(pri), RQB_WORD(pri));
397 	rqb->rqb_bits[RQB_WORD(pri)] |= RQB_BIT(pri);
398 }
399 
400 /*
401  * Add the thread to the queue specified by its priority, and set the
402  * corresponding status bit.
403  */
404 void
405 runq_add(struct runq *rq, struct td_sched *ts, int flags)
406 {
407 	struct rqhead *rqh;
408 	int pri;
409 
410 	pri = ts->ts_thread->td_priority / RQ_PPQ;
411 	ts->ts_rqindex = pri;
412 	runq_setbit(rq, pri);
413 	rqh = &rq->rq_queues[pri];
414 	CTR5(KTR_RUNQ, "runq_add: td=%p ts=%p pri=%d %d rqh=%p",
415 	    ts->ts_thread, ts, ts->ts_thread->td_priority, pri, rqh);
416 	if (flags & SRQ_PREEMPTED) {
417 		TAILQ_INSERT_HEAD(rqh, ts, ts_procq);
418 	} else {
419 		TAILQ_INSERT_TAIL(rqh, ts, ts_procq);
420 	}
421 }
422 
423 void
424 runq_add_pri(struct runq *rq, struct td_sched *ts, u_char pri, int flags)
425 {
426 	struct rqhead *rqh;
427 
428 	KASSERT(pri < RQ_NQS, ("runq_add_pri: %d out of range", pri));
429 	ts->ts_rqindex = pri;
430 	runq_setbit(rq, pri);
431 	rqh = &rq->rq_queues[pri];
432 	CTR5(KTR_RUNQ, "runq_add_pri: td=%p ke=%p pri=%d idx=%d rqh=%p",
433 	    ts->ts_thread, ts, ts->ts_thread->td_priority, pri, rqh);
434 	if (flags & SRQ_PREEMPTED) {
435 		TAILQ_INSERT_HEAD(rqh, ts, ts_procq);
436 	} else {
437 		TAILQ_INSERT_TAIL(rqh, ts, ts_procq);
438 	}
439 }
440 /*
441  * Return true if there are runnable processes of any priority on the run
442  * queue, false otherwise.  Has no side effects, does not modify the run
443  * queue structure.
444  */
445 int
446 runq_check(struct runq *rq)
447 {
448 	struct rqbits *rqb;
449 	int i;
450 
451 	rqb = &rq->rq_status;
452 	for (i = 0; i < RQB_LEN; i++)
453 		if (rqb->rqb_bits[i]) {
454 			CTR2(KTR_RUNQ, "runq_check: bits=%#x i=%d",
455 			    rqb->rqb_bits[i], i);
456 			return (1);
457 		}
458 	CTR0(KTR_RUNQ, "runq_check: empty");
459 
460 	return (0);
461 }
462 
463 #if defined(SMP) && defined(SCHED_4BSD)
464 int runq_fuzz = 1;
465 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
466 #endif
467 
468 /*
469  * Find the highest priority process on the run queue.
470  */
471 struct td_sched *
472 runq_choose(struct runq *rq)
473 {
474 	struct rqhead *rqh;
475 	struct td_sched *ts;
476 	int pri;
477 
478 	while ((pri = runq_findbit(rq)) != -1) {
479 		rqh = &rq->rq_queues[pri];
480 #if defined(SMP) && defined(SCHED_4BSD)
481 		/* fuzz == 1 is normal.. 0 or less are ignored */
482 		if (runq_fuzz > 1) {
483 			/*
484 			 * In the first couple of entries, check if
485 			 * there is one for our CPU as a preference.
486 			 */
487 			int count = runq_fuzz;
488 			int cpu = PCPU_GET(cpuid);
489 			struct td_sched *ts2;
490 			ts2 = ts = TAILQ_FIRST(rqh);
491 
492 			while (count-- && ts2) {
493 				if (ts->ts_thread->td_lastcpu == cpu) {
494 					ts = ts2;
495 					break;
496 				}
497 				ts2 = TAILQ_NEXT(ts2, ts_procq);
498 			}
499 		} else
500 #endif
501 			ts = TAILQ_FIRST(rqh);
502 		KASSERT(ts != NULL, ("runq_choose: no proc on busy queue"));
503 		CTR3(KTR_RUNQ,
504 		    "runq_choose: pri=%d td_sched=%p rqh=%p", pri, ts, rqh);
505 		return (ts);
506 	}
507 	CTR1(KTR_RUNQ, "runq_choose: idleproc pri=%d", pri);
508 
509 	return (NULL);
510 }
511 
512 struct td_sched *
513 runq_choose_from(struct runq *rq, u_char idx)
514 {
515 	struct rqhead *rqh;
516 	struct td_sched *ts;
517 	int pri;
518 
519 	if ((pri = runq_findbit_from(rq, idx)) != -1) {
520 		rqh = &rq->rq_queues[pri];
521 		ts = TAILQ_FIRST(rqh);
522 		KASSERT(ts != NULL, ("runq_choose: no proc on busy queue"));
523 		CTR4(KTR_RUNQ,
524 		    "runq_choose_from: pri=%d kse=%p idx=%d rqh=%p",
525 		    pri, ts, ts->ts_rqindex, rqh);
526 		return (ts);
527 	}
528 	CTR1(KTR_RUNQ, "runq_choose_from: idleproc pri=%d", pri);
529 
530 	return (NULL);
531 }
532 /*
533  * Remove the thread from the queue specified by its priority, and clear the
534  * corresponding status bit if the queue becomes empty.
535  * Caller must set state afterwards.
536  */
537 void
538 runq_remove(struct runq *rq, struct td_sched *ts)
539 {
540 
541 	runq_remove_idx(rq, ts, NULL);
542 }
543 
544 void
545 runq_remove_idx(struct runq *rq, struct td_sched *ts, u_char *idx)
546 {
547 	struct rqhead *rqh;
548 	u_char pri;
549 
550 	KASSERT(ts->ts_thread->td_flags & TDF_INMEM,
551 		("runq_remove_idx: thread swapped out"));
552 	pri = ts->ts_rqindex;
553 	KASSERT(pri < RQ_NQS, ("runq_remove_idx: Invalid index %d\n", pri));
554 	rqh = &rq->rq_queues[pri];
555 	CTR5(KTR_RUNQ, "runq_remove_idx: td=%p, ts=%p pri=%d %d rqh=%p",
556 	    ts->ts_thread, ts, ts->ts_thread->td_priority, pri, rqh);
557 	{
558 		struct td_sched *nts;
559 
560 		TAILQ_FOREACH(nts, rqh, ts_procq)
561 			if (nts == ts)
562 				break;
563 		if (ts != nts)
564 			panic("runq_remove_idx: ts %p not on rqindex %d",
565 			    ts, pri);
566 	}
567 	TAILQ_REMOVE(rqh, ts, ts_procq);
568 	if (TAILQ_EMPTY(rqh)) {
569 		CTR0(KTR_RUNQ, "runq_remove_idx: empty");
570 		runq_clrbit(rq, pri);
571 		if (idx != NULL && *idx == pri)
572 			*idx = (pri + 1) % RQ_NQS;
573 	}
574 }
575 
576 /****** functions that are temporarily here ***********/
577 #include <vm/uma.h>
578 
579 /*
580  *  Allocate scheduler specific per-process resources.
581  * The thread and proc have already been linked in.
582  *
583  * Called from:
584  *  proc_init() (UMA init method)
585  */
586 void
587 sched_newproc(struct proc *p, struct thread *td)
588 {
589 }
590 
591 /*
592  * thread is being either created or recycled.
593  * Fix up the per-scheduler resources associated with it.
594  * Called from:
595  *  sched_fork_thread()
596  *  thread_dtor()  (*may go away)
597  *  thread_init()  (*may go away)
598  */
599 void
600 sched_newthread(struct thread *td)
601 {
602 	struct td_sched *ts;
603 
604 	ts = (struct td_sched *) (td + 1);
605 	bzero(ts, sizeof(*ts));
606 	td->td_sched     = ts;
607 	ts->ts_thread	= td;
608 }
609 
610 #endif /* KERN_SWITCH_INCLUDE */
611