xref: /freebsd/sys/kern/subr_epoch.c (revision 5bb3134a8c21cb87b30e135ef168483f0333dabb)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2018, Matthew Macy <mmacy@freebsd.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/counter.h>
35 #include <sys/epoch.h>
36 #include <sys/gtaskqueue.h>
37 #include <sys/kernel.h>
38 #include <sys/limits.h>
39 #include <sys/lock.h>
40 #include <sys/malloc.h>
41 #include <sys/mutex.h>
42 #include <sys/pcpu.h>
43 #include <sys/proc.h>
44 #include <sys/sched.h>
45 #include <sys/sx.h>
46 #include <sys/smp.h>
47 #include <sys/sysctl.h>
48 #include <sys/turnstile.h>
49 #ifdef EPOCH_TRACE
50 #include <machine/stdarg.h>
51 #include <sys/stack.h>
52 #include <sys/tree.h>
53 #endif
54 #include <vm/vm.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_kern.h>
57 #include <vm/uma.h>
58 
59 #include <ck_epoch.h>
60 
61 #ifdef __amd64__
62 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
63 #else
64 #define EPOCH_ALIGN CACHE_LINE_SIZE
65 #endif
66 
67 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
68 typedef struct epoch_record {
69 	ck_epoch_record_t er_record;
70 	struct epoch_context er_drain_ctx;
71 	struct epoch *er_parent;
72 	volatile struct epoch_tdlist er_tdlist;
73 	volatile uint32_t er_gen;
74 	uint32_t er_cpuid;
75 #ifdef INVARIANTS
76 	/* Used to verify record ownership for non-preemptible epochs. */
77 	struct thread *er_td;
78 #endif
79 } __aligned(EPOCH_ALIGN)     *epoch_record_t;
80 
81 struct epoch {
82 	struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
83 	epoch_record_t e_pcpu_record;
84 	int	e_in_use;
85 	int	e_flags;
86 	struct sx e_drain_sx;
87 	struct mtx e_drain_mtx;
88 	volatile int e_drain_count;
89 	const char *e_name;
90 };
91 
92 /* arbitrary --- needs benchmarking */
93 #define MAX_ADAPTIVE_SPIN 100
94 #define MAX_EPOCHS 64
95 
96 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
97 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
98     "epoch information");
99 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
100     "epoch stats");
101 
102 /* Stats. */
103 static counter_u64_t block_count;
104 
105 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
106     &block_count, "# of times a thread was in an epoch when epoch_wait was called");
107 static counter_u64_t migrate_count;
108 
109 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
110     &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
111 static counter_u64_t turnstile_count;
112 
113 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
114     &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
115 static counter_u64_t switch_count;
116 
117 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
118     &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
119 static counter_u64_t epoch_call_count;
120 
121 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
122     &epoch_call_count, "# of times a callback was deferred");
123 static counter_u64_t epoch_call_task_count;
124 
125 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
126     &epoch_call_task_count, "# of times a callback task was run");
127 
128 TAILQ_HEAD (threadlist, thread);
129 
130 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
131     ck_epoch_entry_container)
132 
133 static struct epoch epoch_array[MAX_EPOCHS];
134 
135 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
136 DPCPU_DEFINE(int, epoch_cb_count);
137 
138 static __read_mostly int inited;
139 __read_mostly epoch_t global_epoch;
140 __read_mostly epoch_t global_epoch_preempt;
141 
142 static void epoch_call_task(void *context __unused);
143 static 	uma_zone_t pcpu_zone_record;
144 
145 static struct sx epoch_sx;
146 
147 #define	EPOCH_LOCK() sx_xlock(&epoch_sx)
148 #define	EPOCH_UNLOCK() sx_xunlock(&epoch_sx)
149 
150 #ifdef EPOCH_TRACE
151 struct stackentry {
152 	RB_ENTRY(stackentry) se_node;
153 	struct stack se_stack;
154 };
155 
156 static int
157 stackentry_compare(struct stackentry *a, struct stackentry *b)
158 {
159 
160 	if (a->se_stack.depth > b->se_stack.depth)
161 		return (1);
162 	if (a->se_stack.depth < b->se_stack.depth)
163 		return (-1);
164 	for (int i = 0; i < a->se_stack.depth; i++) {
165 		if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
166 			return (1);
167 		if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
168 			return (-1);
169 	}
170 
171 	return (0);
172 }
173 
174 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
175 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
176 
177 static struct mtx epoch_stacks_lock;
178 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
179 
180 static bool epoch_trace_stack_print = true;
181 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
182     &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
183 
184 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
185 static inline void
186 epoch_trace_report(const char *fmt, ...)
187 {
188 	va_list ap;
189 	struct stackentry se, *new;
190 
191 	stack_zero(&se.se_stack);	/* XXX: is it really needed? */
192 	stack_save(&se.se_stack);
193 
194 	/* Tree is never reduced - go lockless. */
195 	if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
196 		return;
197 
198 	new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
199 	if (new != NULL) {
200 		bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
201 
202 		mtx_lock(&epoch_stacks_lock);
203 		new = RB_INSERT(stacktree, &epoch_stacks, new);
204 		mtx_unlock(&epoch_stacks_lock);
205 		if (new != NULL)
206 			free(new, M_STACK);
207 	}
208 
209 	va_start(ap, fmt);
210 	(void)vprintf(fmt, ap);
211 	va_end(ap);
212 	if (epoch_trace_stack_print)
213 		stack_print_ddb(&se.se_stack);
214 }
215 
216 static inline void
217 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
218     const char *file, int line)
219 {
220 	epoch_tracker_t iet;
221 
222 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink) {
223 		if (iet->et_epoch != epoch)
224 			continue;
225 		epoch_trace_report("Recursively entering epoch %s "
226 		    "at %s:%d, previously entered at %s:%d\n",
227 		    epoch->e_name, file, line,
228 		    iet->et_file, iet->et_line);
229 	}
230 	et->et_epoch = epoch;
231 	et->et_file = file;
232 	et->et_line = line;
233 	SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
234 }
235 
236 static inline void
237 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
238     const char *file, int line)
239 {
240 
241 	if (SLIST_FIRST(&td->td_epochs) != et) {
242 		epoch_trace_report("Exiting epoch %s in a not nested order "
243 		    "at %s:%d. Most recently entered %s at %s:%d\n",
244 		    epoch->e_name,
245 		    file, line,
246 		    SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
247 		    SLIST_FIRST(&td->td_epochs)->et_file,
248 		    SLIST_FIRST(&td->td_epochs)->et_line);
249 		/* This will panic if et is not anywhere on td_epochs. */
250 		SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
251 	} else
252 		SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
253 }
254 
255 /* Used by assertions that check thread state before going to sleep. */
256 void
257 epoch_trace_list(struct thread *td)
258 {
259 	epoch_tracker_t iet;
260 
261 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
262 		printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
263 		    iet->et_file, iet->et_line);
264 }
265 #endif /* EPOCH_TRACE */
266 
267 static void
268 epoch_init(void *arg __unused)
269 {
270 	int cpu;
271 
272 	block_count = counter_u64_alloc(M_WAITOK);
273 	migrate_count = counter_u64_alloc(M_WAITOK);
274 	turnstile_count = counter_u64_alloc(M_WAITOK);
275 	switch_count = counter_u64_alloc(M_WAITOK);
276 	epoch_call_count = counter_u64_alloc(M_WAITOK);
277 	epoch_call_task_count = counter_u64_alloc(M_WAITOK);
278 
279 	pcpu_zone_record = uma_zcreate("epoch_record pcpu",
280 	    sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
281 	    UMA_ALIGN_PTR, UMA_ZONE_PCPU);
282 	CPU_FOREACH(cpu) {
283 		GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
284 		    epoch_call_task, NULL);
285 		taskqgroup_attach_cpu(qgroup_softirq,
286 		    DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
287 		    "epoch call task");
288 	}
289 #ifdef EPOCH_TRACE
290 	SLIST_INIT(&thread0.td_epochs);
291 #endif
292 	sx_init(&epoch_sx, "epoch-sx");
293 	inited = 1;
294 	global_epoch = epoch_alloc("Global", 0);
295 	global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
296 }
297 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
298 
299 #if !defined(EARLY_AP_STARTUP)
300 static void
301 epoch_init_smp(void *dummy __unused)
302 {
303 	inited = 2;
304 }
305 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
306 #endif
307 
308 static void
309 epoch_ctor(epoch_t epoch)
310 {
311 	epoch_record_t er;
312 	int cpu;
313 
314 	epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
315 	CPU_FOREACH(cpu) {
316 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
317 		bzero(er, sizeof(*er));
318 		ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
319 		TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
320 		er->er_cpuid = cpu;
321 		er->er_parent = epoch;
322 	}
323 }
324 
325 static void
326 epoch_adjust_prio(struct thread *td, u_char prio)
327 {
328 
329 	thread_lock(td);
330 	sched_prio(td, prio);
331 	thread_unlock(td);
332 }
333 
334 epoch_t
335 epoch_alloc(const char *name, int flags)
336 {
337 	epoch_t epoch;
338 	int i;
339 
340 	MPASS(name != NULL);
341 
342 	if (__predict_false(!inited))
343 		panic("%s called too early in boot", __func__);
344 
345 	EPOCH_LOCK();
346 
347 	/*
348 	 * Find a free index in the epoch array. If no free index is
349 	 * found, try to use the index after the last one.
350 	 */
351 	for (i = 0;; i++) {
352 		/*
353 		 * If too many epochs are currently allocated,
354 		 * return NULL.
355 		 */
356 		if (i == MAX_EPOCHS) {
357 			epoch = NULL;
358 			goto done;
359 		}
360 		if (epoch_array[i].e_in_use == 0)
361 			break;
362 	}
363 
364 	epoch = epoch_array + i;
365 	ck_epoch_init(&epoch->e_epoch);
366 	epoch_ctor(epoch);
367 	epoch->e_flags = flags;
368 	epoch->e_name = name;
369 	sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
370 	mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
371 
372 	/*
373 	 * Set e_in_use last, because when this field is set the
374 	 * epoch_call_task() function will start scanning this epoch
375 	 * structure.
376 	 */
377 	atomic_store_rel_int(&epoch->e_in_use, 1);
378 done:
379 	EPOCH_UNLOCK();
380 	return (epoch);
381 }
382 
383 void
384 epoch_free(epoch_t epoch)
385 {
386 #ifdef INVARIANTS
387 	int cpu;
388 #endif
389 
390 	EPOCH_LOCK();
391 
392 	MPASS(epoch->e_in_use != 0);
393 
394 	epoch_drain_callbacks(epoch);
395 
396 	atomic_store_rel_int(&epoch->e_in_use, 0);
397 	/*
398 	 * Make sure the epoch_call_task() function see e_in_use equal
399 	 * to zero, by calling epoch_wait() on the global_epoch:
400 	 */
401 	epoch_wait(global_epoch);
402 #ifdef INVARIANTS
403 	CPU_FOREACH(cpu) {
404 		epoch_record_t er;
405 
406 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
407 
408 		/*
409 		 * Sanity check: none of the records should be in use anymore.
410 		 * We drained callbacks above and freeing the pcpu records is
411 		 * imminent.
412 		 */
413 		MPASS(er->er_td == NULL);
414 		MPASS(TAILQ_EMPTY(&er->er_tdlist));
415 	}
416 #endif
417 	uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
418 	mtx_destroy(&epoch->e_drain_mtx);
419 	sx_destroy(&epoch->e_drain_sx);
420 	memset(epoch, 0, sizeof(*epoch));
421 
422 	EPOCH_UNLOCK();
423 }
424 
425 static epoch_record_t
426 epoch_currecord(epoch_t epoch)
427 {
428 
429 	return (zpcpu_get(epoch->e_pcpu_record));
430 }
431 
432 #define INIT_CHECK(epoch)					\
433 	do {							\
434 		if (__predict_false((epoch) == NULL))		\
435 			return;					\
436 	} while (0)
437 
438 void
439 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
440 {
441 	struct epoch_record *er;
442 	struct thread *td;
443 
444 	MPASS(cold || epoch != NULL);
445 	td = curthread;
446 	MPASS((vm_offset_t)et >= td->td_kstack &&
447 	    (vm_offset_t)et + sizeof(struct epoch_tracker) <=
448 	    td->td_kstack + td->td_kstack_pages * PAGE_SIZE);
449 
450 	INIT_CHECK(epoch);
451 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
452 
453 #ifdef EPOCH_TRACE
454 	epoch_trace_enter(td, epoch, et, file, line);
455 #endif
456 	et->et_td = td;
457 	THREAD_NO_SLEEPING();
458 	critical_enter();
459 	sched_pin();
460 	et->et_old_priority = td->td_priority;
461 	er = epoch_currecord(epoch);
462 	/* Record-level tracking is reserved for non-preemptible epochs. */
463 	MPASS(er->er_td == NULL);
464 	TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
465 	ck_epoch_begin(&er->er_record, &et->et_section);
466 	critical_exit();
467 }
468 
469 void
470 epoch_enter(epoch_t epoch)
471 {
472 	epoch_record_t er;
473 
474 	MPASS(cold || epoch != NULL);
475 	INIT_CHECK(epoch);
476 	critical_enter();
477 	er = epoch_currecord(epoch);
478 #ifdef INVARIANTS
479 	if (er->er_record.active == 0) {
480 		MPASS(er->er_td == NULL);
481 		er->er_td = curthread;
482 	} else {
483 		/* We've recursed, just make sure our accounting isn't wrong. */
484 		MPASS(er->er_td == curthread);
485 	}
486 #endif
487 	ck_epoch_begin(&er->er_record, NULL);
488 }
489 
490 void
491 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
492 {
493 	struct epoch_record *er;
494 	struct thread *td;
495 
496 	INIT_CHECK(epoch);
497 	td = curthread;
498 	critical_enter();
499 	sched_unpin();
500 	THREAD_SLEEPING_OK();
501 	er = epoch_currecord(epoch);
502 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
503 	MPASS(et != NULL);
504 	MPASS(et->et_td == td);
505 #ifdef INVARIANTS
506 	et->et_td = (void*)0xDEADBEEF;
507 	/* Record-level tracking is reserved for non-preemptible epochs. */
508 	MPASS(er->er_td == NULL);
509 #endif
510 	ck_epoch_end(&er->er_record, &et->et_section);
511 	TAILQ_REMOVE(&er->er_tdlist, et, et_link);
512 	er->er_gen++;
513 	if (__predict_false(et->et_old_priority != td->td_priority))
514 		epoch_adjust_prio(td, et->et_old_priority);
515 	critical_exit();
516 #ifdef EPOCH_TRACE
517 	epoch_trace_exit(td, epoch, et, file, line);
518 #endif
519 }
520 
521 void
522 epoch_exit(epoch_t epoch)
523 {
524 	epoch_record_t er;
525 
526 	INIT_CHECK(epoch);
527 	er = epoch_currecord(epoch);
528 	ck_epoch_end(&er->er_record, NULL);
529 #ifdef INVARIANTS
530 	MPASS(er->er_td == curthread);
531 	if (er->er_record.active == 0)
532 		er->er_td = NULL;
533 #endif
534 	critical_exit();
535 }
536 
537 /*
538  * epoch_block_handler_preempt() is a callback from the CK code when another
539  * thread is currently in an epoch section.
540  */
541 static void
542 epoch_block_handler_preempt(struct ck_epoch *global __unused,
543     ck_epoch_record_t *cr, void *arg __unused)
544 {
545 	epoch_record_t record;
546 	struct thread *td, *owner, *curwaittd;
547 	struct epoch_tracker *tdwait;
548 	struct turnstile *ts;
549 	struct lock_object *lock;
550 	int spincount, gen;
551 	int locksheld __unused;
552 
553 	record = __containerof(cr, struct epoch_record, er_record);
554 	td = curthread;
555 	locksheld = td->td_locks;
556 	spincount = 0;
557 	counter_u64_add(block_count, 1);
558 	/*
559 	 * We lost a race and there's no longer any threads
560 	 * on the CPU in an epoch section.
561 	 */
562 	if (TAILQ_EMPTY(&record->er_tdlist))
563 		return;
564 
565 	if (record->er_cpuid != curcpu) {
566 		/*
567 		 * If the head of the list is running, we can wait for it
568 		 * to remove itself from the list and thus save us the
569 		 * overhead of a migration
570 		 */
571 		gen = record->er_gen;
572 		thread_unlock(td);
573 		/*
574 		 * We can't actually check if the waiting thread is running
575 		 * so we simply poll for it to exit before giving up and
576 		 * migrating.
577 		 */
578 		do {
579 			cpu_spinwait();
580 		} while (!TAILQ_EMPTY(&record->er_tdlist) &&
581 				 gen == record->er_gen &&
582 				 spincount++ < MAX_ADAPTIVE_SPIN);
583 		thread_lock(td);
584 		/*
585 		 * If the generation has changed we can poll again
586 		 * otherwise we need to migrate.
587 		 */
588 		if (gen != record->er_gen)
589 			return;
590 		/*
591 		 * Being on the same CPU as that of the record on which
592 		 * we need to wait allows us access to the thread
593 		 * list associated with that CPU. We can then examine the
594 		 * oldest thread in the queue and wait on its turnstile
595 		 * until it resumes and so on until a grace period
596 		 * elapses.
597 		 *
598 		 */
599 		counter_u64_add(migrate_count, 1);
600 		sched_bind(td, record->er_cpuid);
601 		/*
602 		 * At this point we need to return to the ck code
603 		 * to scan to see if a grace period has elapsed.
604 		 * We can't move on to check the thread list, because
605 		 * in the meantime new threads may have arrived that
606 		 * in fact belong to a different epoch.
607 		 */
608 		return;
609 	}
610 	/*
611 	 * Try to find a thread in an epoch section on this CPU
612 	 * waiting on a turnstile. Otherwise find the lowest
613 	 * priority thread (highest prio value) and drop our priority
614 	 * to match to allow it to run.
615 	 */
616 	TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
617 		/*
618 		 * Propagate our priority to any other waiters to prevent us
619 		 * from starving them. They will have their original priority
620 		 * restore on exit from epoch_wait().
621 		 */
622 		curwaittd = tdwait->et_td;
623 		if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
624 			critical_enter();
625 			thread_unlock(td);
626 			thread_lock(curwaittd);
627 			sched_prio(curwaittd, td->td_priority);
628 			thread_unlock(curwaittd);
629 			thread_lock(td);
630 			critical_exit();
631 		}
632 		if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
633 		    ((ts = curwaittd->td_blocked) != NULL)) {
634 			/*
635 			 * We unlock td to allow turnstile_wait to reacquire
636 			 * the thread lock. Before unlocking it we enter a
637 			 * critical section to prevent preemption after we
638 			 * reenable interrupts by dropping the thread lock in
639 			 * order to prevent curwaittd from getting to run.
640 			 */
641 			critical_enter();
642 			thread_unlock(td);
643 
644 			if (turnstile_lock(ts, &lock, &owner)) {
645 				if (ts == curwaittd->td_blocked) {
646 					MPASS(TD_IS_INHIBITED(curwaittd) &&
647 					    TD_ON_LOCK(curwaittd));
648 					critical_exit();
649 					turnstile_wait(ts, owner,
650 					    curwaittd->td_tsqueue);
651 					counter_u64_add(turnstile_count, 1);
652 					thread_lock(td);
653 					return;
654 				}
655 				turnstile_unlock(ts, lock);
656 			}
657 			thread_lock(td);
658 			critical_exit();
659 			KASSERT(td->td_locks == locksheld,
660 			    ("%d extra locks held", td->td_locks - locksheld));
661 		}
662 	}
663 	/*
664 	 * We didn't find any threads actually blocked on a lock
665 	 * so we have nothing to do except context switch away.
666 	 */
667 	counter_u64_add(switch_count, 1);
668 	mi_switch(SW_VOL | SWT_RELINQUISH);
669 	/*
670 	 * It is important the thread lock is dropped while yielding
671 	 * to allow other threads to acquire the lock pointed to by
672 	 * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
673 	 * thread lock before returning. Else a deadlock like
674 	 * situation might happen.
675 	 */
676 	thread_lock(td);
677 }
678 
679 void
680 epoch_wait_preempt(epoch_t epoch)
681 {
682 	struct thread *td;
683 	int was_bound;
684 	int old_cpu;
685 	int old_pinned;
686 	u_char old_prio;
687 	int locks __unused;
688 
689 	MPASS(cold || epoch != NULL);
690 	INIT_CHECK(epoch);
691 	td = curthread;
692 #ifdef INVARIANTS
693 	locks = curthread->td_locks;
694 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
695 	if ((epoch->e_flags & EPOCH_LOCKED) == 0)
696 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
697 		    "epoch_wait() can be long running");
698 	KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
699 	    "of an epoch section of the same epoch"));
700 #endif
701 	DROP_GIANT();
702 	thread_lock(td);
703 
704 	old_cpu = PCPU_GET(cpuid);
705 	old_pinned = td->td_pinned;
706 	old_prio = td->td_priority;
707 	was_bound = sched_is_bound(td);
708 	sched_unbind(td);
709 	td->td_pinned = 0;
710 	sched_bind(td, old_cpu);
711 
712 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
713 	    NULL);
714 
715 	/* restore CPU binding, if any */
716 	if (was_bound != 0) {
717 		sched_bind(td, old_cpu);
718 	} else {
719 		/* get thread back to initial CPU, if any */
720 		if (old_pinned != 0)
721 			sched_bind(td, old_cpu);
722 		sched_unbind(td);
723 	}
724 	/* restore pinned after bind */
725 	td->td_pinned = old_pinned;
726 
727 	/* restore thread priority */
728 	sched_prio(td, old_prio);
729 	thread_unlock(td);
730 	PICKUP_GIANT();
731 	KASSERT(td->td_locks == locks,
732 	    ("%d residual locks held", td->td_locks - locks));
733 }
734 
735 static void
736 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
737     void *arg __unused)
738 {
739 	cpu_spinwait();
740 }
741 
742 void
743 epoch_wait(epoch_t epoch)
744 {
745 
746 	MPASS(cold || epoch != NULL);
747 	INIT_CHECK(epoch);
748 	MPASS(epoch->e_flags == 0);
749 	critical_enter();
750 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
751 	critical_exit();
752 }
753 
754 void
755 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
756 {
757 	epoch_record_t er;
758 	ck_epoch_entry_t *cb;
759 
760 	cb = (void *)ctx;
761 
762 	MPASS(callback);
763 	/* too early in boot to have epoch set up */
764 	if (__predict_false(epoch == NULL))
765 		goto boottime;
766 #if !defined(EARLY_AP_STARTUP)
767 	if (__predict_false(inited < 2))
768 		goto boottime;
769 #endif
770 
771 	critical_enter();
772 	*DPCPU_PTR(epoch_cb_count) += 1;
773 	er = epoch_currecord(epoch);
774 	ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
775 	critical_exit();
776 	return;
777 boottime:
778 	callback(ctx);
779 }
780 
781 static void
782 epoch_call_task(void *arg __unused)
783 {
784 	ck_stack_entry_t *cursor, *head, *next;
785 	ck_epoch_record_t *record;
786 	epoch_record_t er;
787 	epoch_t epoch;
788 	ck_stack_t cb_stack;
789 	int i, npending, total;
790 
791 	ck_stack_init(&cb_stack);
792 	critical_enter();
793 	epoch_enter(global_epoch);
794 	for (total = i = 0; i != MAX_EPOCHS; i++) {
795 		epoch = epoch_array + i;
796 		if (__predict_false(
797 		    atomic_load_acq_int(&epoch->e_in_use) == 0))
798 			continue;
799 		er = epoch_currecord(epoch);
800 		record = &er->er_record;
801 		if ((npending = record->n_pending) == 0)
802 			continue;
803 		ck_epoch_poll_deferred(record, &cb_stack);
804 		total += npending - record->n_pending;
805 	}
806 	epoch_exit(global_epoch);
807 	*DPCPU_PTR(epoch_cb_count) -= total;
808 	critical_exit();
809 
810 	counter_u64_add(epoch_call_count, total);
811 	counter_u64_add(epoch_call_task_count, 1);
812 
813 	head = ck_stack_batch_pop_npsc(&cb_stack);
814 	for (cursor = head; cursor != NULL; cursor = next) {
815 		struct ck_epoch_entry *entry =
816 		    ck_epoch_entry_container(cursor);
817 
818 		next = CK_STACK_NEXT(cursor);
819 		entry->function(entry);
820 	}
821 }
822 
823 static int
824 in_epoch_verbose_preempt(epoch_t epoch, int dump_onfail)
825 {
826 	epoch_record_t er;
827 	struct epoch_tracker *tdwait;
828 	struct thread *td;
829 
830 	MPASS(epoch != NULL);
831 	MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
832 	td = curthread;
833 	if (THREAD_CAN_SLEEP())
834 		return (0);
835 	critical_enter();
836 	er = epoch_currecord(epoch);
837 	TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
838 		if (tdwait->et_td == td) {
839 			critical_exit();
840 			return (1);
841 		}
842 #ifdef INVARIANTS
843 	if (dump_onfail) {
844 		MPASS(td->td_pinned);
845 		printf("cpu: %d id: %d\n", curcpu, td->td_tid);
846 		TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
847 			printf("td_tid: %d ", tdwait->et_td->td_tid);
848 		printf("\n");
849 	}
850 #endif
851 	critical_exit();
852 	return (0);
853 }
854 
855 #ifdef INVARIANTS
856 static void
857 epoch_assert_nocpu(epoch_t epoch, struct thread *td)
858 {
859 	epoch_record_t er;
860 	int cpu;
861 	bool crit;
862 
863 	crit = td->td_critnest > 0;
864 
865 	/* Check for a critical section mishap. */
866 	CPU_FOREACH(cpu) {
867 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
868 		KASSERT(er->er_td != td,
869 		    ("%s critical section in epoch '%s', from cpu %d",
870 		    (crit ? "exited" : "re-entered"), epoch->e_name, cpu));
871 	}
872 }
873 #else
874 #define	epoch_assert_nocpu(e, td) do {} while (0)
875 #endif
876 
877 int
878 in_epoch_verbose(epoch_t epoch, int dump_onfail)
879 {
880 	epoch_record_t er;
881 	struct thread *td;
882 
883 	if (__predict_false((epoch) == NULL))
884 		return (0);
885 	if ((epoch->e_flags & EPOCH_PREEMPT) != 0)
886 		return (in_epoch_verbose_preempt(epoch, dump_onfail));
887 
888 	/*
889 	 * The thread being in a critical section is a necessary
890 	 * condition to be correctly inside a non-preemptible epoch,
891 	 * so it's definitely not in this epoch.
892 	 */
893 	td = curthread;
894 	if (td->td_critnest == 0) {
895 		epoch_assert_nocpu(epoch, td);
896 		return (0);
897 	}
898 
899 	/*
900 	 * The current cpu is in a critical section, so the epoch record will be
901 	 * stable for the rest of this function.  Knowing that the record is not
902 	 * active is sufficient for knowing whether we're in this epoch or not,
903 	 * since it's a pcpu record.
904 	 */
905 	er = epoch_currecord(epoch);
906 	if (er->er_record.active == 0) {
907 		epoch_assert_nocpu(epoch, td);
908 		return (0);
909 	}
910 
911 	MPASS(er->er_td == td);
912 	return (1);
913 }
914 
915 int
916 in_epoch(epoch_t epoch)
917 {
918 	return (in_epoch_verbose(epoch, 0));
919 }
920 
921 static void
922 epoch_drain_cb(struct epoch_context *ctx)
923 {
924 	struct epoch *epoch =
925 	    __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
926 
927 	if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
928 		mtx_lock(&epoch->e_drain_mtx);
929 		wakeup(epoch);
930 		mtx_unlock(&epoch->e_drain_mtx);
931 	}
932 }
933 
934 void
935 epoch_drain_callbacks(epoch_t epoch)
936 {
937 	epoch_record_t er;
938 	struct thread *td;
939 	int was_bound;
940 	int old_pinned;
941 	int old_cpu;
942 	int cpu;
943 
944 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
945 	    "epoch_drain_callbacks() may sleep!");
946 
947 	/* too early in boot to have epoch set up */
948 	if (__predict_false(epoch == NULL))
949 		return;
950 #if !defined(EARLY_AP_STARTUP)
951 	if (__predict_false(inited < 2))
952 		return;
953 #endif
954 	DROP_GIANT();
955 
956 	sx_xlock(&epoch->e_drain_sx);
957 	mtx_lock(&epoch->e_drain_mtx);
958 
959 	td = curthread;
960 	thread_lock(td);
961 	old_cpu = PCPU_GET(cpuid);
962 	old_pinned = td->td_pinned;
963 	was_bound = sched_is_bound(td);
964 	sched_unbind(td);
965 	td->td_pinned = 0;
966 
967 	CPU_FOREACH(cpu)
968 		epoch->e_drain_count++;
969 	CPU_FOREACH(cpu) {
970 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
971 		sched_bind(td, cpu);
972 		epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
973 	}
974 
975 	/* restore CPU binding, if any */
976 	if (was_bound != 0) {
977 		sched_bind(td, old_cpu);
978 	} else {
979 		/* get thread back to initial CPU, if any */
980 		if (old_pinned != 0)
981 			sched_bind(td, old_cpu);
982 		sched_unbind(td);
983 	}
984 	/* restore pinned after bind */
985 	td->td_pinned = old_pinned;
986 
987 	thread_unlock(td);
988 
989 	while (epoch->e_drain_count != 0)
990 		msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
991 
992 	mtx_unlock(&epoch->e_drain_mtx);
993 	sx_xunlock(&epoch->e_drain_sx);
994 
995 	PICKUP_GIANT();
996 }
997