xref: /freebsd/sys/kern/kern_switch.c (revision d91f8db5f1822c43cd256f19aae1d059e4b25a26)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2001 Jake Burkholder <jake@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_sched.h"
33 
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 #include <sys/smp.h>
45 #include <sys/sysctl.h>
46 
47 #include <machine/cpu.h>
48 
49 /* Uncomment this to enable logging of critical_enter/exit. */
50 #if 0
51 #define	KTR_CRITICAL	KTR_SCHED
52 #else
53 #define	KTR_CRITICAL	0
54 #endif
55 
56 #ifdef FULL_PREEMPTION
57 #ifndef PREEMPTION
58 #error "The FULL_PREEMPTION option requires the PREEMPTION option"
59 #endif
60 #endif
61 
62 CTASSERT((RQB_BPW * RQB_LEN) == RQ_NQS);
63 
64 /*
65  * kern.sched.preemption allows user space to determine if preemption support
66  * is compiled in or not.  It is not currently a boot or runtime flag that
67  * can be changed.
68  */
69 #ifdef PREEMPTION
70 static int kern_sched_preemption = 1;
71 #else
72 static int kern_sched_preemption = 0;
73 #endif
74 SYSCTL_INT(_kern_sched, OID_AUTO, preemption, CTLFLAG_RD,
75     &kern_sched_preemption, 0, "Kernel preemption enabled");
76 
77 /*
78  * Support for scheduler stats exported via kern.sched.stats.  All stats may
79  * be reset with kern.sched.stats.reset = 1.  Stats may be defined elsewhere
80  * with SCHED_STAT_DEFINE().
81  */
82 #ifdef SCHED_STATS
83 SYSCTL_NODE(_kern_sched, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
84     "switch stats");
85 
86 /* Switch reasons from mi_switch(9). */
87 DPCPU_DEFINE(long, sched_switch_stats[SWT_COUNT]);
88 SCHED_STAT_DEFINE_VAR(owepreempt,
89     &DPCPU_NAME(sched_switch_stats[SWT_OWEPREEMPT]), "");
90 SCHED_STAT_DEFINE_VAR(turnstile,
91     &DPCPU_NAME(sched_switch_stats[SWT_TURNSTILE]), "");
92 SCHED_STAT_DEFINE_VAR(sleepq,
93     &DPCPU_NAME(sched_switch_stats[SWT_SLEEPQ]), "");
94 SCHED_STAT_DEFINE_VAR(relinquish,
95     &DPCPU_NAME(sched_switch_stats[SWT_RELINQUISH]), "");
96 SCHED_STAT_DEFINE_VAR(needresched,
97     &DPCPU_NAME(sched_switch_stats[SWT_NEEDRESCHED]), "");
98 SCHED_STAT_DEFINE_VAR(idle,
99     &DPCPU_NAME(sched_switch_stats[SWT_IDLE]), "");
100 SCHED_STAT_DEFINE_VAR(iwait,
101     &DPCPU_NAME(sched_switch_stats[SWT_IWAIT]), "");
102 SCHED_STAT_DEFINE_VAR(suspend,
103     &DPCPU_NAME(sched_switch_stats[SWT_SUSPEND]), "");
104 SCHED_STAT_DEFINE_VAR(remotepreempt,
105     &DPCPU_NAME(sched_switch_stats[SWT_REMOTEPREEMPT]), "");
106 SCHED_STAT_DEFINE_VAR(remotewakeidle,
107     &DPCPU_NAME(sched_switch_stats[SWT_REMOTEWAKEIDLE]), "");
108 SCHED_STAT_DEFINE_VAR(bind,
109     &DPCPU_NAME(sched_switch_stats[SWT_BIND]), "");
110 
111 static int
112 sysctl_stats_reset(SYSCTL_HANDLER_ARGS)
113 {
114 	struct sysctl_oid *p;
115 	uintptr_t counter;
116         int error;
117 	int val;
118 	int i;
119 
120         val = 0;
121         error = sysctl_handle_int(oidp, &val, 0, req);
122         if (error != 0 || req->newptr == NULL)
123                 return (error);
124         if (val == 0)
125                 return (0);
126 	/*
127 	 * Traverse the list of children of _kern_sched_stats and reset each
128 	 * to 0.  Skip the reset entry.
129 	 */
130 	RB_FOREACH(p, sysctl_oid_list, oidp->oid_parent) {
131 		if (p == oidp || p->oid_arg1 == NULL)
132 			continue;
133 		counter = (uintptr_t)p->oid_arg1;
134 		CPU_FOREACH(i) {
135 			*(long *)(dpcpu_off[i] + counter) = 0;
136 		}
137 	}
138 	return (0);
139 }
140 
141 SYSCTL_PROC(_kern_sched_stats, OID_AUTO, reset,
142     CTLTYPE_INT | CTLFLAG_WR | CTLFLAG_MPSAFE, NULL, 0,
143     sysctl_stats_reset, "I",
144     "Reset scheduler statistics");
145 #endif
146 
147 /************************************************************************
148  * Functions that manipulate runnability from a thread perspective.	*
149  ************************************************************************/
150 /*
151  * Select the thread that will be run next.
152  */
153 
154 static __noinline struct thread *
155 choosethread_panic(struct thread *td)
156 {
157 
158 	/*
159 	 * If we are in panic, only allow system threads,
160 	 * plus the one we are running in, to be run.
161 	 */
162 retry:
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 */
166 		TD_SET_CAN_RUN(td);
167 		td = sched_choose();
168 		goto retry;
169 	}
170 
171 	TD_SET_RUNNING(td);
172 	return (td);
173 }
174 
175 struct thread *
176 choosethread(void)
177 {
178 	struct thread *td;
179 
180 	td = sched_choose();
181 
182 	if (KERNEL_PANICKED())
183 		return (choosethread_panic(td));
184 
185 	TD_SET_RUNNING(td);
186 	return (td);
187 }
188 
189 /*
190  * Kernel thread preemption implementation.  Critical sections mark
191  * regions of code in which preemptions are not allowed.
192  *
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.
198  */
199 void
200 critical_enter_KBI(void)
201 {
202 #ifdef KTR
203 	struct thread *td = curthread;
204 #endif
205 	critical_enter();
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);
208 }
209 
210 void __noinline
211 critical_exit_preempt(void)
212 {
213 	struct thread *td;
214 	int flags;
215 
216 	/*
217 	 * If td_critnest is 0, it is possible that we are going to get
218 	 * preempted again before reaching the code below. This happens
219 	 * rarely and is harmless. However, this means td_owepreempt may
220 	 * now be unset.
221 	 */
222 	td = curthread;
223 	if (td->td_critnest != 0)
224 		return;
225 	if (kdb_active)
226 		return;
227 
228 	/*
229 	 * Microoptimization: we committed to switch,
230 	 * disable preemption in interrupt handlers
231 	 * while spinning for the thread lock.
232 	 */
233 	td->td_critnest = 1;
234 	thread_lock(td);
235 	td->td_critnest--;
236 	flags = SW_INVOL | SW_PREEMPT;
237 	if (TD_IS_IDLETHREAD(td))
238 		flags |= SWT_IDLE;
239 	else
240 		flags |= SWT_OWEPREEMPT;
241 	mi_switch(flags);
242 }
243 
244 void
245 critical_exit_KBI(void)
246 {
247 #ifdef KTR
248 	struct thread *td = curthread;
249 #endif
250 	critical_exit();
251 	CTR4(KTR_CRITICAL, "critical_exit by thread %p (%ld, %s) to %d", td,
252 	    (long)td->td_proc->p_pid, td->td_name, td->td_critnest);
253 }
254 
255 /************************************************************************
256  * SYSTEM RUN QUEUE manipulations and tests				*
257  ************************************************************************/
258 /*
259  * Initialize a run structure.
260  */
261 void
262 runq_init(struct runq *rq)
263 {
264 	int i;
265 
266 	bzero(rq, sizeof *rq);
267 	for (i = 0; i < RQ_NQS; i++)
268 		TAILQ_INIT(&rq->rq_queues[i]);
269 }
270 
271 /*
272  * Clear the status bit of the queue corresponding to priority level pri,
273  * indicating that it is empty.
274  */
275 static __inline void
276 runq_clrbit(struct runq *rq, int pri)
277 {
278 	struct rqbits *rqb;
279 
280 	rqb = &rq->rq_status;
281 	CTR4(KTR_RUNQ, "runq_clrbit: bits=%#x %#x bit=%#x word=%d",
282 	    rqb->rqb_bits[RQB_WORD(pri)],
283 	    rqb->rqb_bits[RQB_WORD(pri)] & ~RQB_BIT(pri),
284 	    RQB_BIT(pri), RQB_WORD(pri));
285 	rqb->rqb_bits[RQB_WORD(pri)] &= ~RQB_BIT(pri);
286 }
287 
288 /*
289  * Find the index of the first non-empty run queue.  This is done by
290  * scanning the status bits, a set bit indicates a non-empty queue.
291  */
292 static __inline int
293 runq_findbit(struct runq *rq)
294 {
295 	struct rqbits *rqb;
296 	int pri;
297 	int i;
298 
299 	rqb = &rq->rq_status;
300 	for (i = 0; i < RQB_LEN; i++)
301 		if (rqb->rqb_bits[i]) {
302 			pri = RQB_FFS(rqb->rqb_bits[i]) + (i << RQB_L2BPW);
303 			CTR3(KTR_RUNQ, "runq_findbit: bits=%#x i=%d pri=%d",
304 			    rqb->rqb_bits[i], i, pri);
305 			return (pri);
306 		}
307 
308 	return (-1);
309 }
310 
311 static __inline int
312 runq_findbit_from(struct runq *rq, u_char pri)
313 {
314 	struct rqbits *rqb;
315 	rqb_word_t mask;
316 	int i;
317 
318 	/*
319 	 * Set the mask for the first word so we ignore priorities before 'pri'.
320 	 */
321 	mask = (rqb_word_t)-1 << (pri & (RQB_BPW - 1));
322 	rqb = &rq->rq_status;
323 again:
324 	for (i = RQB_WORD(pri); i < RQB_LEN; mask = -1, i++) {
325 		mask = rqb->rqb_bits[i] & mask;
326 		if (mask == 0)
327 			continue;
328 		pri = RQB_FFS(mask) + (i << RQB_L2BPW);
329 		CTR3(KTR_RUNQ, "runq_findbit_from: bits=%#x i=%d pri=%d",
330 		    mask, i, pri);
331 		return (pri);
332 	}
333 	if (pri == 0)
334 		return (-1);
335 	/*
336 	 * Wrap back around to the beginning of the list just once so we
337 	 * scan the whole thing.
338 	 */
339 	pri = 0;
340 	goto again;
341 }
342 
343 /*
344  * Set the status bit of the queue corresponding to priority level pri,
345  * indicating that it is non-empty.
346  */
347 static __inline void
348 runq_setbit(struct runq *rq, int pri)
349 {
350 	struct rqbits *rqb;
351 
352 	rqb = &rq->rq_status;
353 	CTR4(KTR_RUNQ, "runq_setbit: bits=%#x %#x bit=%#x word=%d",
354 	    rqb->rqb_bits[RQB_WORD(pri)],
355 	    rqb->rqb_bits[RQB_WORD(pri)] | RQB_BIT(pri),
356 	    RQB_BIT(pri), RQB_WORD(pri));
357 	rqb->rqb_bits[RQB_WORD(pri)] |= RQB_BIT(pri);
358 }
359 
360 /*
361  * Add the thread to the queue specified by its priority, and set the
362  * corresponding status bit.
363  */
364 void
365 runq_add(struct runq *rq, struct thread *td, int flags)
366 {
367 	struct rqhead *rqh;
368 	int pri;
369 
370 	pri = td->td_priority / RQ_PPQ;
371 	td->td_rqindex = pri;
372 	runq_setbit(rq, pri);
373 	rqh = &rq->rq_queues[pri];
374 	CTR4(KTR_RUNQ, "runq_add: td=%p pri=%d %d rqh=%p",
375 	    td, td->td_priority, pri, rqh);
376 	if (flags & SRQ_PREEMPTED) {
377 		TAILQ_INSERT_HEAD(rqh, td, td_runq);
378 	} else {
379 		TAILQ_INSERT_TAIL(rqh, td, td_runq);
380 	}
381 }
382 
383 void
384 runq_add_pri(struct runq *rq, struct thread *td, u_char pri, int flags)
385 {
386 	struct rqhead *rqh;
387 
388 	KASSERT(pri < RQ_NQS, ("runq_add_pri: %d out of range", pri));
389 	td->td_rqindex = pri;
390 	runq_setbit(rq, pri);
391 	rqh = &rq->rq_queues[pri];
392 	CTR4(KTR_RUNQ, "runq_add_pri: td=%p pri=%d idx=%d rqh=%p",
393 	    td, td->td_priority, pri, rqh);
394 	if (flags & SRQ_PREEMPTED) {
395 		TAILQ_INSERT_HEAD(rqh, td, td_runq);
396 	} else {
397 		TAILQ_INSERT_TAIL(rqh, td, td_runq);
398 	}
399 }
400 /*
401  * Return true if there are runnable processes of any priority on the run
402  * queue, false otherwise.  Has no side effects, does not modify the run
403  * queue structure.
404  */
405 int
406 runq_check(struct runq *rq)
407 {
408 	struct rqbits *rqb;
409 	int i;
410 
411 	rqb = &rq->rq_status;
412 	for (i = 0; i < RQB_LEN; i++)
413 		if (rqb->rqb_bits[i]) {
414 			CTR2(KTR_RUNQ, "runq_check: bits=%#x i=%d",
415 			    rqb->rqb_bits[i], i);
416 			return (1);
417 		}
418 	CTR0(KTR_RUNQ, "runq_check: empty");
419 
420 	return (0);
421 }
422 
423 /*
424  * Find the highest priority process on the run queue.
425  */
426 struct thread *
427 runq_choose_fuzz(struct runq *rq, int fuzz)
428 {
429 	struct rqhead *rqh;
430 	struct thread *td;
431 	int pri;
432 
433 	while ((pri = runq_findbit(rq)) != -1) {
434 		rqh = &rq->rq_queues[pri];
435 		/* fuzz == 1 is normal.. 0 or less are ignored */
436 		if (fuzz > 1) {
437 			/*
438 			 * In the first couple of entries, check if
439 			 * there is one for our CPU as a preference.
440 			 */
441 			int count = fuzz;
442 			int cpu = PCPU_GET(cpuid);
443 			struct thread *td2;
444 			td2 = td = TAILQ_FIRST(rqh);
445 
446 			while (count-- && td2) {
447 				if (td2->td_lastcpu == cpu) {
448 					td = td2;
449 					break;
450 				}
451 				td2 = TAILQ_NEXT(td2, td_runq);
452 			}
453 		} else
454 			td = TAILQ_FIRST(rqh);
455 		KASSERT(td != NULL, ("runq_choose_fuzz: no proc on busy queue"));
456 		CTR3(KTR_RUNQ,
457 		    "runq_choose_fuzz: pri=%d thread=%p rqh=%p", pri, td, rqh);
458 		return (td);
459 	}
460 	CTR1(KTR_RUNQ, "runq_choose_fuzz: idleproc pri=%d", pri);
461 
462 	return (NULL);
463 }
464 
465 /*
466  * Find the highest priority process on the run queue.
467  */
468 struct thread *
469 runq_choose(struct runq *rq)
470 {
471 	struct rqhead *rqh;
472 	struct thread *td;
473 	int pri;
474 
475 	while ((pri = runq_findbit(rq)) != -1) {
476 		rqh = &rq->rq_queues[pri];
477 		td = TAILQ_FIRST(rqh);
478 		KASSERT(td != NULL, ("runq_choose: no thread on busy queue"));
479 		CTR3(KTR_RUNQ,
480 		    "runq_choose: pri=%d thread=%p rqh=%p", pri, td, rqh);
481 		return (td);
482 	}
483 	CTR1(KTR_RUNQ, "runq_choose: idlethread pri=%d", pri);
484 
485 	return (NULL);
486 }
487 
488 struct thread *
489 runq_choose_from(struct runq *rq, u_char idx)
490 {
491 	struct rqhead *rqh;
492 	struct thread *td;
493 	int pri;
494 
495 	if ((pri = runq_findbit_from(rq, idx)) != -1) {
496 		rqh = &rq->rq_queues[pri];
497 		td = TAILQ_FIRST(rqh);
498 		KASSERT(td != NULL, ("runq_choose: no thread on busy queue"));
499 		CTR4(KTR_RUNQ,
500 		    "runq_choose_from: pri=%d thread=%p idx=%d rqh=%p",
501 		    pri, td, td->td_rqindex, rqh);
502 		return (td);
503 	}
504 	CTR1(KTR_RUNQ, "runq_choose_from: idlethread pri=%d", pri);
505 
506 	return (NULL);
507 }
508 /*
509  * Remove the thread from the queue specified by its priority, and clear the
510  * corresponding status bit if the queue becomes empty.
511  * Caller must set state afterwards.
512  */
513 void
514 runq_remove(struct runq *rq, struct thread *td)
515 {
516 
517 	runq_remove_idx(rq, td, NULL);
518 }
519 
520 void
521 runq_remove_idx(struct runq *rq, struct thread *td, u_char *idx)
522 {
523 	struct rqhead *rqh;
524 	u_char pri;
525 
526 	KASSERT(td->td_flags & TDF_INMEM,
527 		("runq_remove_idx: thread swapped out"));
528 	pri = td->td_rqindex;
529 	KASSERT(pri < RQ_NQS, ("runq_remove_idx: Invalid index %d\n", pri));
530 	rqh = &rq->rq_queues[pri];
531 	CTR4(KTR_RUNQ, "runq_remove_idx: td=%p, pri=%d %d rqh=%p",
532 	    td, td->td_priority, pri, rqh);
533 	TAILQ_REMOVE(rqh, td, td_runq);
534 	if (TAILQ_EMPTY(rqh)) {
535 		CTR0(KTR_RUNQ, "runq_remove_idx: empty");
536 		runq_clrbit(rq, pri);
537 		if (idx != NULL && *idx == pri)
538 			*idx = (pri + 1) % RQ_NQS;
539 	}
540 }
541