xref: /freebsd/sys/kern/subr_epoch.c (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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/param.h>
30 #include <sys/systm.h>
31 #include <sys/counter.h>
32 #include <sys/epoch.h>
33 #include <sys/gtaskqueue.h>
34 #include <sys/kernel.h>
35 #include <sys/limits.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mutex.h>
39 #include <sys/pcpu.h>
40 #include <sys/proc.h>
41 #include <sys/sched.h>
42 #include <sys/sx.h>
43 #include <sys/smp.h>
44 #include <sys/sysctl.h>
45 #include <sys/turnstile.h>
46 #ifdef EPOCH_TRACE
47 #include <machine/stdarg.h>
48 #include <sys/stack.h>
49 #include <sys/tree.h>
50 #endif
51 #include <vm/vm.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_kern.h>
54 #include <vm/uma.h>
55 
56 #include <machine/stack.h>
57 
58 #include <ck_epoch.h>
59 
60 #ifdef __amd64__
61 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
62 #else
63 #define EPOCH_ALIGN CACHE_LINE_SIZE
64 #endif
65 
66 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
67 typedef struct epoch_record {
68 	ck_epoch_record_t er_record;
69 	struct epoch_context er_drain_ctx;
70 	struct epoch *er_parent;
71 	volatile struct epoch_tdlist er_tdlist;
72 	volatile uint32_t er_gen;
73 	uint32_t er_cpuid;
74 #ifdef INVARIANTS
75 	/* Used to verify record ownership for non-preemptible epochs. */
76 	struct thread *er_td;
77 #endif
78 } __aligned(EPOCH_ALIGN)     *epoch_record_t;
79 
80 struct epoch {
81 	struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
82 	epoch_record_t e_pcpu_record;
83 	int	e_in_use;
84 	int	e_flags;
85 	struct sx e_drain_sx;
86 	struct mtx e_drain_mtx;
87 	volatile int e_drain_count;
88 	const char *e_name;
89 };
90 
91 /* arbitrary --- needs benchmarking */
92 #define MAX_ADAPTIVE_SPIN 100
93 #define MAX_EPOCHS 64
94 
95 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
96 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
97     "epoch information");
98 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
99     "epoch stats");
100 
101 /* Stats. */
102 static counter_u64_t block_count;
103 
104 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
105     &block_count, "# of times a thread was in an epoch when epoch_wait was called");
106 static counter_u64_t migrate_count;
107 
108 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
109     &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
110 static counter_u64_t turnstile_count;
111 
112 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
113     &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
114 static counter_u64_t switch_count;
115 
116 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
117     &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
118 static counter_u64_t epoch_call_count;
119 
120 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
121     &epoch_call_count, "# of times a callback was deferred");
122 static counter_u64_t epoch_call_task_count;
123 
124 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
125     &epoch_call_task_count, "# of times a callback task was run");
126 
127 TAILQ_HEAD (threadlist, thread);
128 
129 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
130     ck_epoch_entry_container)
131 
132 static struct epoch epoch_array[MAX_EPOCHS];
133 
134 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
135 DPCPU_DEFINE(int, epoch_cb_count);
136 
137 static __read_mostly int inited;
138 __read_mostly epoch_t global_epoch;
139 __read_mostly epoch_t global_epoch_preempt;
140 
141 static void epoch_call_task(void *context __unused);
142 static 	uma_zone_t pcpu_zone_record;
143 
144 static struct sx epoch_sx;
145 
146 #define	EPOCH_LOCK() sx_xlock(&epoch_sx)
147 #define	EPOCH_UNLOCK() sx_xunlock(&epoch_sx)
148 
149 static epoch_record_t
150 epoch_currecord(epoch_t epoch)
151 {
152 
153 	return (zpcpu_get(epoch->e_pcpu_record));
154 }
155 
156 #ifdef EPOCH_TRACE
157 struct stackentry {
158 	RB_ENTRY(stackentry) se_node;
159 	struct stack se_stack;
160 };
161 
162 static int
163 stackentry_compare(struct stackentry *a, struct stackentry *b)
164 {
165 
166 	if (a->se_stack.depth > b->se_stack.depth)
167 		return (1);
168 	if (a->se_stack.depth < b->se_stack.depth)
169 		return (-1);
170 	for (int i = 0; i < a->se_stack.depth; i++) {
171 		if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
172 			return (1);
173 		if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
174 			return (-1);
175 	}
176 
177 	return (0);
178 }
179 
180 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
181 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
182 
183 static struct mtx epoch_stacks_lock;
184 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
185 
186 static bool epoch_trace_stack_print = true;
187 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
188     &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
189 
190 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
191 static inline void
192 epoch_trace_report(const char *fmt, ...)
193 {
194 	va_list ap;
195 	struct stackentry se, *new;
196 
197 	stack_save(&se.se_stack);
198 
199 	/* Tree is never reduced - go lockless. */
200 	if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
201 		return;
202 
203 	new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
204 	if (new != NULL) {
205 		bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
206 
207 		mtx_lock(&epoch_stacks_lock);
208 		new = RB_INSERT(stacktree, &epoch_stacks, new);
209 		mtx_unlock(&epoch_stacks_lock);
210 		if (new != NULL)
211 			free(new, M_STACK);
212 	}
213 
214 	va_start(ap, fmt);
215 	(void)vprintf(fmt, ap);
216 	va_end(ap);
217 	if (epoch_trace_stack_print)
218 		stack_print_ddb(&se.se_stack);
219 }
220 
221 static inline void
222 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
223     const char *file, int line)
224 {
225 	epoch_tracker_t iet;
226 
227 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink) {
228 		if (iet->et_epoch != epoch)
229 			continue;
230 		epoch_trace_report("Recursively entering epoch %s "
231 		    "at %s:%d, previously entered at %s:%d\n",
232 		    epoch->e_name, file, line,
233 		    iet->et_file, iet->et_line);
234 	}
235 	et->et_epoch = epoch;
236 	et->et_file = file;
237 	et->et_line = line;
238 	et->et_flags = 0;
239 	SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
240 }
241 
242 static inline void
243 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
244     const char *file, int line)
245 {
246 
247 	if (SLIST_FIRST(&td->td_epochs) != et) {
248 		epoch_trace_report("Exiting epoch %s in a not nested order "
249 		    "at %s:%d. Most recently entered %s at %s:%d\n",
250 		    epoch->e_name,
251 		    file, line,
252 		    SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
253 		    SLIST_FIRST(&td->td_epochs)->et_file,
254 		    SLIST_FIRST(&td->td_epochs)->et_line);
255 		/* This will panic if et is not anywhere on td_epochs. */
256 		SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
257 	} else
258 		SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
259 	if (et->et_flags & ET_REPORT_EXIT)
260 		printf("Td %p exiting epoch %s at %s:%d\n", td, epoch->e_name,
261 		    file, line);
262 }
263 
264 /* Used by assertions that check thread state before going to sleep. */
265 void
266 epoch_trace_list(struct thread *td)
267 {
268 	epoch_tracker_t iet;
269 
270 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
271 		printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
272 		    iet->et_file, iet->et_line);
273 }
274 
275 void
276 epoch_where_report(epoch_t epoch)
277 {
278 	epoch_record_t er;
279 	struct epoch_tracker *tdwait;
280 
281 	MPASS(epoch != NULL);
282 	MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
283 	MPASS(!THREAD_CAN_SLEEP());
284 	critical_enter();
285 	er = epoch_currecord(epoch);
286 	TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
287 		if (tdwait->et_td == curthread)
288 			break;
289 	critical_exit();
290 	if (tdwait != NULL) {
291 		tdwait->et_flags |= ET_REPORT_EXIT;
292 		printf("Td %p entered epoch %s at %s:%d\n", curthread,
293 		    epoch->e_name, tdwait->et_file, tdwait->et_line);
294 	}
295 }
296 #endif /* EPOCH_TRACE */
297 
298 static void
299 epoch_init(void *arg __unused)
300 {
301 	int cpu;
302 
303 	block_count = counter_u64_alloc(M_WAITOK);
304 	migrate_count = counter_u64_alloc(M_WAITOK);
305 	turnstile_count = counter_u64_alloc(M_WAITOK);
306 	switch_count = counter_u64_alloc(M_WAITOK);
307 	epoch_call_count = counter_u64_alloc(M_WAITOK);
308 	epoch_call_task_count = counter_u64_alloc(M_WAITOK);
309 
310 	pcpu_zone_record = uma_zcreate("epoch_record pcpu",
311 	    sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
312 	    UMA_ALIGN_PTR, UMA_ZONE_PCPU);
313 	CPU_FOREACH(cpu) {
314 		GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
315 		    epoch_call_task, NULL);
316 		taskqgroup_attach_cpu(qgroup_softirq,
317 		    DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
318 		    "epoch call task");
319 	}
320 #ifdef EPOCH_TRACE
321 	SLIST_INIT(&thread0.td_epochs);
322 #endif
323 	sx_init(&epoch_sx, "epoch-sx");
324 	inited = 1;
325 	global_epoch = epoch_alloc("Global", 0);
326 	global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
327 }
328 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
329 
330 #if !defined(EARLY_AP_STARTUP)
331 static void
332 epoch_init_smp(void *dummy __unused)
333 {
334 	inited = 2;
335 }
336 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
337 #endif
338 
339 static void
340 epoch_ctor(epoch_t epoch)
341 {
342 	epoch_record_t er;
343 	int cpu;
344 
345 	epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
346 	CPU_FOREACH(cpu) {
347 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
348 		bzero(er, sizeof(*er));
349 		ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
350 		TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
351 		er->er_cpuid = cpu;
352 		er->er_parent = epoch;
353 	}
354 }
355 
356 static void
357 epoch_adjust_prio(struct thread *td, u_char prio)
358 {
359 
360 	thread_lock(td);
361 	sched_prio(td, prio);
362 	thread_unlock(td);
363 }
364 
365 epoch_t
366 epoch_alloc(const char *name, int flags)
367 {
368 	epoch_t epoch;
369 	int i;
370 
371 	MPASS(name != NULL);
372 
373 	if (__predict_false(!inited))
374 		panic("%s called too early in boot", __func__);
375 
376 	EPOCH_LOCK();
377 
378 	/*
379 	 * Find a free index in the epoch array. If no free index is
380 	 * found, try to use the index after the last one.
381 	 */
382 	for (i = 0;; i++) {
383 		/*
384 		 * If too many epochs are currently allocated,
385 		 * return NULL.
386 		 */
387 		if (i == MAX_EPOCHS) {
388 			epoch = NULL;
389 			goto done;
390 		}
391 		if (epoch_array[i].e_in_use == 0)
392 			break;
393 	}
394 
395 	epoch = epoch_array + i;
396 	ck_epoch_init(&epoch->e_epoch);
397 	epoch_ctor(epoch);
398 	epoch->e_flags = flags;
399 	epoch->e_name = name;
400 	sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
401 	mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
402 
403 	/*
404 	 * Set e_in_use last, because when this field is set the
405 	 * epoch_call_task() function will start scanning this epoch
406 	 * structure.
407 	 */
408 	atomic_store_rel_int(&epoch->e_in_use, 1);
409 done:
410 	EPOCH_UNLOCK();
411 	return (epoch);
412 }
413 
414 void
415 epoch_free(epoch_t epoch)
416 {
417 #ifdef INVARIANTS
418 	int cpu;
419 #endif
420 
421 	EPOCH_LOCK();
422 
423 	MPASS(epoch->e_in_use != 0);
424 
425 	epoch_drain_callbacks(epoch);
426 
427 	atomic_store_rel_int(&epoch->e_in_use, 0);
428 	/*
429 	 * Make sure the epoch_call_task() function see e_in_use equal
430 	 * to zero, by calling epoch_wait() on the global_epoch:
431 	 */
432 	epoch_wait(global_epoch);
433 #ifdef INVARIANTS
434 	CPU_FOREACH(cpu) {
435 		epoch_record_t er;
436 
437 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
438 
439 		/*
440 		 * Sanity check: none of the records should be in use anymore.
441 		 * We drained callbacks above and freeing the pcpu records is
442 		 * imminent.
443 		 */
444 		MPASS(er->er_td == NULL);
445 		MPASS(TAILQ_EMPTY(&er->er_tdlist));
446 	}
447 #endif
448 	uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
449 	mtx_destroy(&epoch->e_drain_mtx);
450 	sx_destroy(&epoch->e_drain_sx);
451 	memset(epoch, 0, sizeof(*epoch));
452 
453 	EPOCH_UNLOCK();
454 }
455 
456 #define INIT_CHECK(epoch)					\
457 	do {							\
458 		if (__predict_false((epoch) == NULL))		\
459 			return;					\
460 	} while (0)
461 
462 void
463 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
464 {
465 	struct epoch_record *er;
466 	struct thread *td;
467 
468 	MPASS(cold || epoch != NULL);
469 	td = curthread;
470 	MPASS(kstack_contains(td, (vm_offset_t)et, sizeof(*et)));
471 
472 	INIT_CHECK(epoch);
473 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
474 
475 #ifdef EPOCH_TRACE
476 	epoch_trace_enter(td, epoch, et, file, line);
477 #endif
478 	et->et_td = td;
479 	THREAD_NO_SLEEPING();
480 	critical_enter();
481 	sched_pin();
482 	et->et_old_priority = td->td_priority;
483 	er = epoch_currecord(epoch);
484 	/* Record-level tracking is reserved for non-preemptible epochs. */
485 	MPASS(er->er_td == NULL);
486 	TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
487 	ck_epoch_begin(&er->er_record, &et->et_section);
488 	critical_exit();
489 }
490 
491 void
492 epoch_enter(epoch_t epoch)
493 {
494 	epoch_record_t er;
495 
496 	MPASS(cold || epoch != NULL);
497 	INIT_CHECK(epoch);
498 	critical_enter();
499 	er = epoch_currecord(epoch);
500 #ifdef INVARIANTS
501 	if (er->er_record.active == 0) {
502 		MPASS(er->er_td == NULL);
503 		er->er_td = curthread;
504 	} else {
505 		/* We've recursed, just make sure our accounting isn't wrong. */
506 		MPASS(er->er_td == curthread);
507 	}
508 #endif
509 	ck_epoch_begin(&er->er_record, NULL);
510 }
511 
512 void
513 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
514 {
515 	struct epoch_record *er;
516 	struct thread *td;
517 
518 	INIT_CHECK(epoch);
519 	td = curthread;
520 	critical_enter();
521 	sched_unpin();
522 	THREAD_SLEEPING_OK();
523 	er = epoch_currecord(epoch);
524 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
525 	MPASS(et != NULL);
526 	MPASS(et->et_td == td);
527 #ifdef INVARIANTS
528 	et->et_td = (void*)0xDEADBEEF;
529 	/* Record-level tracking is reserved for non-preemptible epochs. */
530 	MPASS(er->er_td == NULL);
531 #endif
532 	ck_epoch_end(&er->er_record, &et->et_section);
533 	TAILQ_REMOVE(&er->er_tdlist, et, et_link);
534 	er->er_gen++;
535 	if (__predict_false(et->et_old_priority != td->td_priority))
536 		epoch_adjust_prio(td, et->et_old_priority);
537 	critical_exit();
538 #ifdef EPOCH_TRACE
539 	epoch_trace_exit(td, epoch, et, file, line);
540 #endif
541 }
542 
543 void
544 epoch_exit(epoch_t epoch)
545 {
546 	epoch_record_t er;
547 
548 	INIT_CHECK(epoch);
549 	er = epoch_currecord(epoch);
550 	ck_epoch_end(&er->er_record, NULL);
551 #ifdef INVARIANTS
552 	MPASS(er->er_td == curthread);
553 	if (er->er_record.active == 0)
554 		er->er_td = NULL;
555 #endif
556 	critical_exit();
557 }
558 
559 /*
560  * epoch_block_handler_preempt() is a callback from the CK code when another
561  * thread is currently in an epoch section.
562  */
563 static void
564 epoch_block_handler_preempt(struct ck_epoch *global __unused,
565     ck_epoch_record_t *cr, void *arg __unused)
566 {
567 	epoch_record_t record;
568 	struct thread *td, *owner, *curwaittd;
569 	struct epoch_tracker *tdwait;
570 	struct turnstile *ts;
571 	struct lock_object *lock;
572 	int spincount, gen;
573 	int locksheld __unused;
574 
575 	record = __containerof(cr, struct epoch_record, er_record);
576 	td = curthread;
577 	locksheld = td->td_locks;
578 	spincount = 0;
579 	counter_u64_add(block_count, 1);
580 	/*
581 	 * We lost a race and there's no longer any threads
582 	 * on the CPU in an epoch section.
583 	 */
584 	if (TAILQ_EMPTY(&record->er_tdlist))
585 		return;
586 
587 	if (record->er_cpuid != curcpu) {
588 		/*
589 		 * If the head of the list is running, we can wait for it
590 		 * to remove itself from the list and thus save us the
591 		 * overhead of a migration
592 		 */
593 		gen = record->er_gen;
594 		thread_unlock(td);
595 		/*
596 		 * We can't actually check if the waiting thread is running
597 		 * so we simply poll for it to exit before giving up and
598 		 * migrating.
599 		 */
600 		do {
601 			cpu_spinwait();
602 		} while (!TAILQ_EMPTY(&record->er_tdlist) &&
603 				 gen == record->er_gen &&
604 				 spincount++ < MAX_ADAPTIVE_SPIN);
605 		thread_lock(td);
606 		/*
607 		 * If the generation has changed we can poll again
608 		 * otherwise we need to migrate.
609 		 */
610 		if (gen != record->er_gen)
611 			return;
612 		/*
613 		 * Being on the same CPU as that of the record on which
614 		 * we need to wait allows us access to the thread
615 		 * list associated with that CPU. We can then examine the
616 		 * oldest thread in the queue and wait on its turnstile
617 		 * until it resumes and so on until a grace period
618 		 * elapses.
619 		 *
620 		 */
621 		counter_u64_add(migrate_count, 1);
622 		sched_bind(td, record->er_cpuid);
623 		/*
624 		 * At this point we need to return to the ck code
625 		 * to scan to see if a grace period has elapsed.
626 		 * We can't move on to check the thread list, because
627 		 * in the meantime new threads may have arrived that
628 		 * in fact belong to a different epoch.
629 		 */
630 		return;
631 	}
632 	/*
633 	 * Try to find a thread in an epoch section on this CPU
634 	 * waiting on a turnstile. Otherwise find the lowest
635 	 * priority thread (highest prio value) and drop our priority
636 	 * to match to allow it to run.
637 	 */
638 	TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
639 		/*
640 		 * Propagate our priority to any other waiters to prevent us
641 		 * from starving them. They will have their original priority
642 		 * restore on exit from epoch_wait().
643 		 */
644 		curwaittd = tdwait->et_td;
645 		if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
646 			critical_enter();
647 			thread_unlock(td);
648 			thread_lock(curwaittd);
649 			sched_prio(curwaittd, td->td_priority);
650 			thread_unlock(curwaittd);
651 			thread_lock(td);
652 			critical_exit();
653 		}
654 		if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
655 		    ((ts = curwaittd->td_blocked) != NULL)) {
656 			/*
657 			 * We unlock td to allow turnstile_wait to reacquire
658 			 * the thread lock. Before unlocking it we enter a
659 			 * critical section to prevent preemption after we
660 			 * reenable interrupts by dropping the thread lock in
661 			 * order to prevent curwaittd from getting to run.
662 			 */
663 			critical_enter();
664 			thread_unlock(td);
665 
666 			if (turnstile_lock(ts, &lock, &owner)) {
667 				if (ts == curwaittd->td_blocked) {
668 					MPASS(TD_IS_INHIBITED(curwaittd) &&
669 					    TD_ON_LOCK(curwaittd));
670 					critical_exit();
671 					turnstile_wait(ts, owner,
672 					    curwaittd->td_tsqueue);
673 					counter_u64_add(turnstile_count, 1);
674 					thread_lock(td);
675 					return;
676 				}
677 				turnstile_unlock(ts, lock);
678 			}
679 			thread_lock(td);
680 			critical_exit();
681 			KASSERT(td->td_locks == locksheld,
682 			    ("%d extra locks held", td->td_locks - locksheld));
683 		}
684 	}
685 	/*
686 	 * We didn't find any threads actually blocked on a lock
687 	 * so we have nothing to do except context switch away.
688 	 */
689 	counter_u64_add(switch_count, 1);
690 	mi_switch(SW_VOL | SWT_RELINQUISH);
691 	/*
692 	 * It is important the thread lock is dropped while yielding
693 	 * to allow other threads to acquire the lock pointed to by
694 	 * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
695 	 * thread lock before returning. Else a deadlock like
696 	 * situation might happen.
697 	 */
698 	thread_lock(td);
699 }
700 
701 void
702 epoch_wait_preempt(epoch_t epoch)
703 {
704 	struct thread *td;
705 	int was_bound;
706 	int old_cpu;
707 	int old_pinned;
708 	u_char old_prio;
709 	int locks __unused;
710 
711 	MPASS(cold || epoch != NULL);
712 	INIT_CHECK(epoch);
713 	td = curthread;
714 #ifdef INVARIANTS
715 	locks = curthread->td_locks;
716 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
717 	if ((epoch->e_flags & EPOCH_LOCKED) == 0)
718 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
719 		    "epoch_wait() can be long running");
720 	KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
721 	    "of an epoch section of the same epoch"));
722 #endif
723 	DROP_GIANT();
724 	thread_lock(td);
725 
726 	old_cpu = PCPU_GET(cpuid);
727 	old_pinned = td->td_pinned;
728 	old_prio = td->td_priority;
729 	was_bound = sched_is_bound(td);
730 	sched_unbind(td);
731 	td->td_pinned = 0;
732 	sched_bind(td, old_cpu);
733 
734 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
735 	    NULL);
736 
737 	/* restore CPU binding, if any */
738 	if (was_bound != 0) {
739 		sched_bind(td, old_cpu);
740 	} else {
741 		/* get thread back to initial CPU, if any */
742 		if (old_pinned != 0)
743 			sched_bind(td, old_cpu);
744 		sched_unbind(td);
745 	}
746 	/* restore pinned after bind */
747 	td->td_pinned = old_pinned;
748 
749 	/* restore thread priority */
750 	sched_prio(td, old_prio);
751 	thread_unlock(td);
752 	PICKUP_GIANT();
753 	KASSERT(td->td_locks == locks,
754 	    ("%d residual locks held", td->td_locks - locks));
755 }
756 
757 static void
758 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
759     void *arg __unused)
760 {
761 	cpu_spinwait();
762 }
763 
764 void
765 epoch_wait(epoch_t epoch)
766 {
767 
768 	MPASS(cold || epoch != NULL);
769 	INIT_CHECK(epoch);
770 	MPASS(epoch->e_flags == 0);
771 	critical_enter();
772 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
773 	critical_exit();
774 }
775 
776 void
777 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
778 {
779 	epoch_record_t er;
780 	ck_epoch_entry_t *cb;
781 
782 	cb = (void *)ctx;
783 
784 	MPASS(callback);
785 	/* too early in boot to have epoch set up */
786 	if (__predict_false(epoch == NULL))
787 		goto boottime;
788 #if !defined(EARLY_AP_STARTUP)
789 	if (__predict_false(inited < 2))
790 		goto boottime;
791 #endif
792 
793 	critical_enter();
794 	*DPCPU_PTR(epoch_cb_count) += 1;
795 	er = epoch_currecord(epoch);
796 	ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
797 	critical_exit();
798 	return;
799 boottime:
800 	callback(ctx);
801 }
802 
803 static void
804 epoch_call_task(void *arg __unused)
805 {
806 	ck_stack_entry_t *cursor, *head, *next;
807 	ck_epoch_record_t *record;
808 	epoch_record_t er;
809 	epoch_t epoch;
810 	ck_stack_t cb_stack;
811 	int i, npending, total;
812 
813 	ck_stack_init(&cb_stack);
814 	critical_enter();
815 	epoch_enter(global_epoch);
816 	for (total = i = 0; i != MAX_EPOCHS; i++) {
817 		epoch = epoch_array + i;
818 		if (__predict_false(
819 		    atomic_load_acq_int(&epoch->e_in_use) == 0))
820 			continue;
821 		er = epoch_currecord(epoch);
822 		record = &er->er_record;
823 		if ((npending = record->n_pending) == 0)
824 			continue;
825 		ck_epoch_poll_deferred(record, &cb_stack);
826 		total += npending - record->n_pending;
827 	}
828 	epoch_exit(global_epoch);
829 	*DPCPU_PTR(epoch_cb_count) -= total;
830 	critical_exit();
831 
832 	counter_u64_add(epoch_call_count, total);
833 	counter_u64_add(epoch_call_task_count, 1);
834 
835 	head = ck_stack_batch_pop_npsc(&cb_stack);
836 	for (cursor = head; cursor != NULL; cursor = next) {
837 		struct ck_epoch_entry *entry =
838 		    ck_epoch_entry_container(cursor);
839 
840 		next = CK_STACK_NEXT(cursor);
841 		entry->function(entry);
842 	}
843 }
844 
845 static int
846 in_epoch_verbose_preempt(epoch_t epoch, int dump_onfail)
847 {
848 	epoch_record_t er;
849 	struct epoch_tracker *tdwait;
850 	struct thread *td;
851 
852 	MPASS(epoch != NULL);
853 	MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
854 	td = curthread;
855 	if (THREAD_CAN_SLEEP())
856 		return (0);
857 	critical_enter();
858 	er = epoch_currecord(epoch);
859 	TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
860 		if (tdwait->et_td == td) {
861 			critical_exit();
862 			return (1);
863 		}
864 #ifdef INVARIANTS
865 	if (dump_onfail) {
866 		MPASS(td->td_pinned);
867 		printf("cpu: %d id: %d\n", curcpu, td->td_tid);
868 		TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
869 			printf("td_tid: %d ", tdwait->et_td->td_tid);
870 		printf("\n");
871 	}
872 #endif
873 	critical_exit();
874 	return (0);
875 }
876 
877 #ifdef INVARIANTS
878 static void
879 epoch_assert_nocpu(epoch_t epoch, struct thread *td)
880 {
881 	epoch_record_t er;
882 	int cpu;
883 	bool crit;
884 
885 	crit = td->td_critnest > 0;
886 
887 	/* Check for a critical section mishap. */
888 	CPU_FOREACH(cpu) {
889 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
890 		KASSERT(er->er_td != td,
891 		    ("%s critical section in epoch '%s', from cpu %d",
892 		    (crit ? "exited" : "re-entered"), epoch->e_name, cpu));
893 	}
894 }
895 #else
896 #define	epoch_assert_nocpu(e, td) do {} while (0)
897 #endif
898 
899 int
900 in_epoch_verbose(epoch_t epoch, int dump_onfail)
901 {
902 	epoch_record_t er;
903 	struct thread *td;
904 
905 	if (__predict_false((epoch) == NULL))
906 		return (0);
907 	if ((epoch->e_flags & EPOCH_PREEMPT) != 0)
908 		return (in_epoch_verbose_preempt(epoch, dump_onfail));
909 
910 	/*
911 	 * The thread being in a critical section is a necessary
912 	 * condition to be correctly inside a non-preemptible epoch,
913 	 * so it's definitely not in this epoch.
914 	 */
915 	td = curthread;
916 	if (td->td_critnest == 0) {
917 		epoch_assert_nocpu(epoch, td);
918 		return (0);
919 	}
920 
921 	/*
922 	 * The current cpu is in a critical section, so the epoch record will be
923 	 * stable for the rest of this function.  Knowing that the record is not
924 	 * active is sufficient for knowing whether we're in this epoch or not,
925 	 * since it's a pcpu record.
926 	 */
927 	er = epoch_currecord(epoch);
928 	if (er->er_record.active == 0) {
929 		epoch_assert_nocpu(epoch, td);
930 		return (0);
931 	}
932 
933 	MPASS(er->er_td == td);
934 	return (1);
935 }
936 
937 int
938 in_epoch(epoch_t epoch)
939 {
940 	return (in_epoch_verbose(epoch, 0));
941 }
942 
943 static void
944 epoch_drain_cb(struct epoch_context *ctx)
945 {
946 	struct epoch *epoch =
947 	    __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
948 
949 	if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
950 		mtx_lock(&epoch->e_drain_mtx);
951 		wakeup(epoch);
952 		mtx_unlock(&epoch->e_drain_mtx);
953 	}
954 }
955 
956 void
957 epoch_drain_callbacks(epoch_t epoch)
958 {
959 	epoch_record_t er;
960 	struct thread *td;
961 	int was_bound;
962 	int old_pinned;
963 	int old_cpu;
964 	int cpu;
965 
966 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
967 	    "epoch_drain_callbacks() may sleep!");
968 
969 	/* too early in boot to have epoch set up */
970 	if (__predict_false(epoch == NULL))
971 		return;
972 #if !defined(EARLY_AP_STARTUP)
973 	if (__predict_false(inited < 2))
974 		return;
975 #endif
976 	DROP_GIANT();
977 
978 	sx_xlock(&epoch->e_drain_sx);
979 	mtx_lock(&epoch->e_drain_mtx);
980 
981 	td = curthread;
982 	thread_lock(td);
983 	old_cpu = PCPU_GET(cpuid);
984 	old_pinned = td->td_pinned;
985 	was_bound = sched_is_bound(td);
986 	sched_unbind(td);
987 	td->td_pinned = 0;
988 
989 	CPU_FOREACH(cpu)
990 		epoch->e_drain_count++;
991 	CPU_FOREACH(cpu) {
992 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
993 		sched_bind(td, cpu);
994 		epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
995 	}
996 
997 	/* restore CPU binding, if any */
998 	if (was_bound != 0) {
999 		sched_bind(td, old_cpu);
1000 	} else {
1001 		/* get thread back to initial CPU, if any */
1002 		if (old_pinned != 0)
1003 			sched_bind(td, old_cpu);
1004 		sched_unbind(td);
1005 	}
1006 	/* restore pinned after bind */
1007 	td->td_pinned = old_pinned;
1008 
1009 	thread_unlock(td);
1010 
1011 	while (epoch->e_drain_count != 0)
1012 		msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
1013 
1014 	mtx_unlock(&epoch->e_drain_mtx);
1015 	sx_xunlock(&epoch->e_drain_sx);
1016 
1017 	PICKUP_GIANT();
1018 }
1019