xref: /freebsd/sys/kern/subr_epoch.c (revision a134ebd6e63f658f2d3d04ac0c60d23bcaa86dd7)
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 static MALLOC_DEFINE(M_EPOCH, "epoch", "epoch based reclamation");
62 
63 #ifdef __amd64__
64 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
65 #else
66 #define EPOCH_ALIGN CACHE_LINE_SIZE
67 #endif
68 
69 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
70 typedef struct epoch_record {
71 	ck_epoch_record_t er_record;
72 	struct epoch_context er_drain_ctx;
73 	struct epoch *er_parent;
74 	volatile struct epoch_tdlist er_tdlist;
75 	volatile uint32_t er_gen;
76 	uint32_t er_cpuid;
77 } __aligned(EPOCH_ALIGN)     *epoch_record_t;
78 
79 struct epoch {
80 	struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
81 	epoch_record_t e_pcpu_record;
82 	int	e_idx;
83 	int	e_flags;
84 	struct sx e_drain_sx;
85 	struct mtx e_drain_mtx;
86 	volatile int e_drain_count;
87 	const char *e_name;
88 };
89 
90 /* arbitrary --- needs benchmarking */
91 #define MAX_ADAPTIVE_SPIN 100
92 #define MAX_EPOCHS 64
93 
94 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
95 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
96     "epoch information");
97 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
98     "epoch stats");
99 
100 /* Stats. */
101 static counter_u64_t block_count;
102 
103 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
104     &block_count, "# of times a thread was in an epoch when epoch_wait was called");
105 static counter_u64_t migrate_count;
106 
107 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
108     &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
109 static counter_u64_t turnstile_count;
110 
111 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
112     &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
113 static counter_u64_t switch_count;
114 
115 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
116     &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
117 static counter_u64_t epoch_call_count;
118 
119 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
120     &epoch_call_count, "# of times a callback was deferred");
121 static counter_u64_t epoch_call_task_count;
122 
123 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
124     &epoch_call_task_count, "# of times a callback task was run");
125 
126 TAILQ_HEAD (threadlist, thread);
127 
128 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
129     ck_epoch_entry_container)
130 
131 epoch_t	allepochs[MAX_EPOCHS];
132 
133 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
134 DPCPU_DEFINE(int, epoch_cb_count);
135 
136 static __read_mostly int inited;
137 static __read_mostly int epoch_count;
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 #ifdef EPOCH_TRACE
145 struct stackentry {
146 	RB_ENTRY(stackentry) se_node;
147 	struct stack se_stack;
148 };
149 
150 static int
151 stackentry_compare(struct stackentry *a, struct stackentry *b)
152 {
153 
154 	if (a->se_stack.depth > b->se_stack.depth)
155 		return (1);
156 	if (a->se_stack.depth < b->se_stack.depth)
157 		return (-1);
158 	for (int i = 0; i < a->se_stack.depth; i++) {
159 		if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
160 			return (1);
161 		if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
162 			return (-1);
163 	}
164 
165 	return (0);
166 }
167 
168 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
169 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
170 
171 static struct mtx epoch_stacks_lock;
172 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
173 
174 static bool epoch_trace_stack_print = true;
175 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
176     &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
177 
178 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
179 static inline void
180 epoch_trace_report(const char *fmt, ...)
181 {
182 	va_list ap;
183 	struct stackentry se, *new;
184 
185 	stack_zero(&se.se_stack);	/* XXX: is it really needed? */
186 	stack_save(&se.se_stack);
187 
188 	/* Tree is never reduced - go lockless. */
189 	if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
190 		return;
191 
192 	new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
193 	if (new != NULL) {
194 		bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
195 
196 		mtx_lock(&epoch_stacks_lock);
197 		new = RB_INSERT(stacktree, &epoch_stacks, new);
198 		mtx_unlock(&epoch_stacks_lock);
199 		if (new != NULL)
200 			free(new, M_STACK);
201 	}
202 
203 	va_start(ap, fmt);
204 	(void)vprintf(fmt, ap);
205 	va_end(ap);
206 	if (epoch_trace_stack_print)
207 		stack_print_ddb(&se.se_stack);
208 }
209 
210 static inline void
211 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
212     const char *file, int line)
213 {
214 	epoch_tracker_t iet;
215 
216 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
217 		if (iet->et_epoch == epoch)
218 			epoch_trace_report("Recursively entering epoch %s "
219 			    "at %s:%d, previously entered at %s:%d\n",
220 			    epoch->e_name, file, line,
221 			    iet->et_file, iet->et_line);
222 	et->et_epoch = epoch;
223 	et->et_file = file;
224 	et->et_line = line;
225 	SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
226 }
227 
228 static inline void
229 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
230     const char *file, int line)
231 {
232 
233 	if (SLIST_FIRST(&td->td_epochs) != et) {
234 		epoch_trace_report("Exiting epoch %s in a not nested order "
235 		    "at %s:%d. Most recently entered %s at %s:%d\n",
236 		    epoch->e_name,
237 		    file, line,
238 		    SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
239 		    SLIST_FIRST(&td->td_epochs)->et_file,
240 		    SLIST_FIRST(&td->td_epochs)->et_line);
241 		/* This will panic if et is not anywhere on td_epochs. */
242 		SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
243 	} else
244 		SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
245 }
246 
247 /* Used by assertions that check thread state before going to sleep. */
248 void
249 epoch_trace_list(struct thread *td)
250 {
251 	epoch_tracker_t iet;
252 
253 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
254 		printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
255 		    iet->et_file, iet->et_line);
256 }
257 #endif /* EPOCH_TRACE */
258 
259 static void
260 epoch_init(void *arg __unused)
261 {
262 	int cpu;
263 
264 	block_count = counter_u64_alloc(M_WAITOK);
265 	migrate_count = counter_u64_alloc(M_WAITOK);
266 	turnstile_count = counter_u64_alloc(M_WAITOK);
267 	switch_count = counter_u64_alloc(M_WAITOK);
268 	epoch_call_count = counter_u64_alloc(M_WAITOK);
269 	epoch_call_task_count = counter_u64_alloc(M_WAITOK);
270 
271 	pcpu_zone_record = uma_zcreate("epoch_record pcpu",
272 	    sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
273 	    UMA_ALIGN_PTR, UMA_ZONE_PCPU);
274 	CPU_FOREACH(cpu) {
275 		GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
276 		    epoch_call_task, NULL);
277 		taskqgroup_attach_cpu(qgroup_softirq,
278 		    DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
279 		    "epoch call task");
280 	}
281 #ifdef EPOCH_TRACE
282 	SLIST_INIT(&thread0.td_epochs);
283 #endif
284 	inited = 1;
285 	global_epoch = epoch_alloc("Global", 0);
286 	global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
287 }
288 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
289 
290 #if !defined(EARLY_AP_STARTUP)
291 static void
292 epoch_init_smp(void *dummy __unused)
293 {
294 	inited = 2;
295 }
296 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
297 #endif
298 
299 static void
300 epoch_ctor(epoch_t epoch)
301 {
302 	epoch_record_t er;
303 	int cpu;
304 
305 	epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
306 	CPU_FOREACH(cpu) {
307 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
308 		bzero(er, sizeof(*er));
309 		ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
310 		TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
311 		er->er_cpuid = cpu;
312 		er->er_parent = epoch;
313 	}
314 }
315 
316 static void
317 epoch_adjust_prio(struct thread *td, u_char prio)
318 {
319 
320 	thread_lock(td);
321 	sched_prio(td, prio);
322 	thread_unlock(td);
323 }
324 
325 epoch_t
326 epoch_alloc(const char *name, int flags)
327 {
328 	epoch_t epoch;
329 
330 	if (__predict_false(!inited))
331 		panic("%s called too early in boot", __func__);
332 	epoch = malloc(sizeof(struct epoch), M_EPOCH, M_ZERO | M_WAITOK);
333 	ck_epoch_init(&epoch->e_epoch);
334 	epoch_ctor(epoch);
335 	MPASS(epoch_count < MAX_EPOCHS - 2);
336 	epoch->e_flags = flags;
337 	epoch->e_idx = epoch_count;
338 	epoch->e_name = name;
339 	sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
340 	mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
341 	allepochs[epoch_count++] = epoch;
342 	return (epoch);
343 }
344 
345 void
346 epoch_free(epoch_t epoch)
347 {
348 
349 	epoch_drain_callbacks(epoch);
350 	allepochs[epoch->e_idx] = NULL;
351 	epoch_wait(global_epoch);
352 	uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
353 	mtx_destroy(&epoch->e_drain_mtx);
354 	sx_destroy(&epoch->e_drain_sx);
355 	free(epoch, M_EPOCH);
356 }
357 
358 static epoch_record_t
359 epoch_currecord(epoch_t epoch)
360 {
361 
362 	return (zpcpu_get(epoch->e_pcpu_record));
363 }
364 
365 #define INIT_CHECK(epoch)					\
366 	do {							\
367 		if (__predict_false((epoch) == NULL))		\
368 			return;					\
369 	} while (0)
370 
371 void
372 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
373 {
374 	struct epoch_record *er;
375 	struct thread *td;
376 
377 	MPASS(cold || epoch != NULL);
378 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
379 	td = curthread;
380 	MPASS((vm_offset_t)et >= td->td_kstack &&
381 	    (vm_offset_t)et + sizeof(struct epoch_tracker) <=
382 	    td->td_kstack + td->td_kstack_pages * PAGE_SIZE);
383 
384 	INIT_CHECK(epoch);
385 #ifdef EPOCH_TRACE
386 	epoch_trace_enter(td, epoch, et, file, line);
387 #endif
388 	et->et_td = td;
389 	THREAD_NO_SLEEPING();
390 	critical_enter();
391 	sched_pin();
392 	td->td_pre_epoch_prio = td->td_priority;
393 	er = epoch_currecord(epoch);
394 	TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
395 	ck_epoch_begin(&er->er_record, &et->et_section);
396 	critical_exit();
397 }
398 
399 void
400 epoch_enter(epoch_t epoch)
401 {
402 	epoch_record_t er;
403 
404 	MPASS(cold || epoch != NULL);
405 	INIT_CHECK(epoch);
406 	critical_enter();
407 	er = epoch_currecord(epoch);
408 	ck_epoch_begin(&er->er_record, NULL);
409 }
410 
411 void
412 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
413 {
414 	struct epoch_record *er;
415 	struct thread *td;
416 
417 	INIT_CHECK(epoch);
418 	td = curthread;
419 	critical_enter();
420 	sched_unpin();
421 	THREAD_SLEEPING_OK();
422 	er = epoch_currecord(epoch);
423 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
424 	MPASS(et != NULL);
425 	MPASS(et->et_td == td);
426 #ifdef INVARIANTS
427 	et->et_td = (void*)0xDEADBEEF;
428 #endif
429 	ck_epoch_end(&er->er_record, &et->et_section);
430 	TAILQ_REMOVE(&er->er_tdlist, et, et_link);
431 	er->er_gen++;
432 	if (__predict_false(td->td_pre_epoch_prio != td->td_priority))
433 		epoch_adjust_prio(td, td->td_pre_epoch_prio);
434 	critical_exit();
435 #ifdef EPOCH_TRACE
436 	epoch_trace_exit(td, epoch, et, file, line);
437 #endif
438 }
439 
440 void
441 epoch_exit(epoch_t epoch)
442 {
443 	epoch_record_t er;
444 
445 	INIT_CHECK(epoch);
446 	er = epoch_currecord(epoch);
447 	ck_epoch_end(&er->er_record, NULL);
448 	critical_exit();
449 }
450 
451 /*
452  * epoch_block_handler_preempt() is a callback from the CK code when another
453  * thread is currently in an epoch section.
454  */
455 static void
456 epoch_block_handler_preempt(struct ck_epoch *global __unused,
457     ck_epoch_record_t *cr, void *arg __unused)
458 {
459 	epoch_record_t record;
460 	struct thread *td, *owner, *curwaittd;
461 	struct epoch_tracker *tdwait;
462 	struct turnstile *ts;
463 	struct lock_object *lock;
464 	int spincount, gen;
465 	int locksheld __unused;
466 
467 	record = __containerof(cr, struct epoch_record, er_record);
468 	td = curthread;
469 	locksheld = td->td_locks;
470 	spincount = 0;
471 	counter_u64_add(block_count, 1);
472 	/*
473 	 * We lost a race and there's no longer any threads
474 	 * on the CPU in an epoch section.
475 	 */
476 	if (TAILQ_EMPTY(&record->er_tdlist))
477 		return;
478 
479 	if (record->er_cpuid != curcpu) {
480 		/*
481 		 * If the head of the list is running, we can wait for it
482 		 * to remove itself from the list and thus save us the
483 		 * overhead of a migration
484 		 */
485 		gen = record->er_gen;
486 		thread_unlock(td);
487 		/*
488 		 * We can't actually check if the waiting thread is running
489 		 * so we simply poll for it to exit before giving up and
490 		 * migrating.
491 		 */
492 		do {
493 			cpu_spinwait();
494 		} while (!TAILQ_EMPTY(&record->er_tdlist) &&
495 				 gen == record->er_gen &&
496 				 spincount++ < MAX_ADAPTIVE_SPIN);
497 		thread_lock(td);
498 		/*
499 		 * If the generation has changed we can poll again
500 		 * otherwise we need to migrate.
501 		 */
502 		if (gen != record->er_gen)
503 			return;
504 		/*
505 		 * Being on the same CPU as that of the record on which
506 		 * we need to wait allows us access to the thread
507 		 * list associated with that CPU. We can then examine the
508 		 * oldest thread in the queue and wait on its turnstile
509 		 * until it resumes and so on until a grace period
510 		 * elapses.
511 		 *
512 		 */
513 		counter_u64_add(migrate_count, 1);
514 		sched_bind(td, record->er_cpuid);
515 		/*
516 		 * At this point we need to return to the ck code
517 		 * to scan to see if a grace period has elapsed.
518 		 * We can't move on to check the thread list, because
519 		 * in the meantime new threads may have arrived that
520 		 * in fact belong to a different epoch.
521 		 */
522 		return;
523 	}
524 	/*
525 	 * Try to find a thread in an epoch section on this CPU
526 	 * waiting on a turnstile. Otherwise find the lowest
527 	 * priority thread (highest prio value) and drop our priority
528 	 * to match to allow it to run.
529 	 */
530 	TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
531 		/*
532 		 * Propagate our priority to any other waiters to prevent us
533 		 * from starving them. They will have their original priority
534 		 * restore on exit from epoch_wait().
535 		 */
536 		curwaittd = tdwait->et_td;
537 		if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
538 			critical_enter();
539 			thread_unlock(td);
540 			thread_lock(curwaittd);
541 			sched_prio(curwaittd, td->td_priority);
542 			thread_unlock(curwaittd);
543 			thread_lock(td);
544 			critical_exit();
545 		}
546 		if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
547 		    ((ts = curwaittd->td_blocked) != NULL)) {
548 			/*
549 			 * We unlock td to allow turnstile_wait to reacquire
550 			 * the thread lock. Before unlocking it we enter a
551 			 * critical section to prevent preemption after we
552 			 * reenable interrupts by dropping the thread lock in
553 			 * order to prevent curwaittd from getting to run.
554 			 */
555 			critical_enter();
556 			thread_unlock(td);
557 
558 			if (turnstile_lock(ts, &lock, &owner)) {
559 				if (ts == curwaittd->td_blocked) {
560 					MPASS(TD_IS_INHIBITED(curwaittd) &&
561 					    TD_ON_LOCK(curwaittd));
562 					critical_exit();
563 					turnstile_wait(ts, owner,
564 					    curwaittd->td_tsqueue);
565 					counter_u64_add(turnstile_count, 1);
566 					thread_lock(td);
567 					return;
568 				}
569 				turnstile_unlock(ts, lock);
570 			}
571 			thread_lock(td);
572 			critical_exit();
573 			KASSERT(td->td_locks == locksheld,
574 			    ("%d extra locks held", td->td_locks - locksheld));
575 		}
576 	}
577 	/*
578 	 * We didn't find any threads actually blocked on a lock
579 	 * so we have nothing to do except context switch away.
580 	 */
581 	counter_u64_add(switch_count, 1);
582 	mi_switch(SW_VOL | SWT_RELINQUISH);
583 	/*
584 	 * It is important the thread lock is dropped while yielding
585 	 * to allow other threads to acquire the lock pointed to by
586 	 * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
587 	 * thread lock before returning. Else a deadlock like
588 	 * situation might happen.
589 	 */
590 	thread_lock(td);
591 }
592 
593 void
594 epoch_wait_preempt(epoch_t epoch)
595 {
596 	struct thread *td;
597 	int was_bound;
598 	int old_cpu;
599 	int old_pinned;
600 	u_char old_prio;
601 	int locks __unused;
602 
603 	MPASS(cold || epoch != NULL);
604 	INIT_CHECK(epoch);
605 	td = curthread;
606 #ifdef INVARIANTS
607 	locks = curthread->td_locks;
608 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
609 	if ((epoch->e_flags & EPOCH_LOCKED) == 0)
610 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
611 		    "epoch_wait() can be long running");
612 	KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
613 	    "of an epoch section of the same epoch"));
614 #endif
615 	DROP_GIANT();
616 	thread_lock(td);
617 
618 	old_cpu = PCPU_GET(cpuid);
619 	old_pinned = td->td_pinned;
620 	old_prio = td->td_priority;
621 	was_bound = sched_is_bound(td);
622 	sched_unbind(td);
623 	td->td_pinned = 0;
624 	sched_bind(td, old_cpu);
625 
626 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
627 	    NULL);
628 
629 	/* restore CPU binding, if any */
630 	if (was_bound != 0) {
631 		sched_bind(td, old_cpu);
632 	} else {
633 		/* get thread back to initial CPU, if any */
634 		if (old_pinned != 0)
635 			sched_bind(td, old_cpu);
636 		sched_unbind(td);
637 	}
638 	/* restore pinned after bind */
639 	td->td_pinned = old_pinned;
640 
641 	/* restore thread priority */
642 	sched_prio(td, old_prio);
643 	thread_unlock(td);
644 	PICKUP_GIANT();
645 	KASSERT(td->td_locks == locks,
646 	    ("%d residual locks held", td->td_locks - locks));
647 }
648 
649 static void
650 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
651     void *arg __unused)
652 {
653 	cpu_spinwait();
654 }
655 
656 void
657 epoch_wait(epoch_t epoch)
658 {
659 
660 	MPASS(cold || epoch != NULL);
661 	INIT_CHECK(epoch);
662 	MPASS(epoch->e_flags == 0);
663 	critical_enter();
664 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
665 	critical_exit();
666 }
667 
668 void
669 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
670 {
671 	epoch_record_t er;
672 	ck_epoch_entry_t *cb;
673 
674 	cb = (void *)ctx;
675 
676 	MPASS(callback);
677 	/* too early in boot to have epoch set up */
678 	if (__predict_false(epoch == NULL))
679 		goto boottime;
680 #if !defined(EARLY_AP_STARTUP)
681 	if (__predict_false(inited < 2))
682 		goto boottime;
683 #endif
684 
685 	critical_enter();
686 	*DPCPU_PTR(epoch_cb_count) += 1;
687 	er = epoch_currecord(epoch);
688 	ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
689 	critical_exit();
690 	return;
691 boottime:
692 	callback(ctx);
693 }
694 
695 static void
696 epoch_call_task(void *arg __unused)
697 {
698 	ck_stack_entry_t *cursor, *head, *next;
699 	ck_epoch_record_t *record;
700 	epoch_record_t er;
701 	epoch_t epoch;
702 	ck_stack_t cb_stack;
703 	int i, npending, total;
704 
705 	ck_stack_init(&cb_stack);
706 	critical_enter();
707 	epoch_enter(global_epoch);
708 	for (total = i = 0; i < epoch_count; i++) {
709 		if (__predict_false((epoch = allepochs[i]) == NULL))
710 			continue;
711 		er = epoch_currecord(epoch);
712 		record = &er->er_record;
713 		if ((npending = record->n_pending) == 0)
714 			continue;
715 		ck_epoch_poll_deferred(record, &cb_stack);
716 		total += npending - record->n_pending;
717 	}
718 	epoch_exit(global_epoch);
719 	*DPCPU_PTR(epoch_cb_count) -= total;
720 	critical_exit();
721 
722 	counter_u64_add(epoch_call_count, total);
723 	counter_u64_add(epoch_call_task_count, 1);
724 
725 	head = ck_stack_batch_pop_npsc(&cb_stack);
726 	for (cursor = head; cursor != NULL; cursor = next) {
727 		struct ck_epoch_entry *entry =
728 		    ck_epoch_entry_container(cursor);
729 
730 		next = CK_STACK_NEXT(cursor);
731 		entry->function(entry);
732 	}
733 }
734 
735 int
736 in_epoch_verbose(epoch_t epoch, int dump_onfail)
737 {
738 	struct epoch_tracker *tdwait;
739 	struct thread *td;
740 	epoch_record_t er;
741 
742 	td = curthread;
743 	if (THREAD_CAN_SLEEP())
744 		return (0);
745 	if (__predict_false((epoch) == NULL))
746 		return (0);
747 	critical_enter();
748 	er = epoch_currecord(epoch);
749 	TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
750 		if (tdwait->et_td == td) {
751 			critical_exit();
752 			return (1);
753 		}
754 #ifdef INVARIANTS
755 	if (dump_onfail) {
756 		MPASS(td->td_pinned);
757 		printf("cpu: %d id: %d\n", curcpu, td->td_tid);
758 		TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
759 			printf("td_tid: %d ", tdwait->et_td->td_tid);
760 		printf("\n");
761 	}
762 #endif
763 	critical_exit();
764 	return (0);
765 }
766 
767 int
768 in_epoch(epoch_t epoch)
769 {
770 	return (in_epoch_verbose(epoch, 0));
771 }
772 
773 static void
774 epoch_drain_cb(struct epoch_context *ctx)
775 {
776 	struct epoch *epoch =
777 	    __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
778 
779 	if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
780 		mtx_lock(&epoch->e_drain_mtx);
781 		wakeup(epoch);
782 		mtx_unlock(&epoch->e_drain_mtx);
783 	}
784 }
785 
786 void
787 epoch_drain_callbacks(epoch_t epoch)
788 {
789 	epoch_record_t er;
790 	struct thread *td;
791 	int was_bound;
792 	int old_pinned;
793 	int old_cpu;
794 	int cpu;
795 
796 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
797 	    "epoch_drain_callbacks() may sleep!");
798 
799 	/* too early in boot to have epoch set up */
800 	if (__predict_false(epoch == NULL))
801 		return;
802 #if !defined(EARLY_AP_STARTUP)
803 	if (__predict_false(inited < 2))
804 		return;
805 #endif
806 	DROP_GIANT();
807 
808 	sx_xlock(&epoch->e_drain_sx);
809 	mtx_lock(&epoch->e_drain_mtx);
810 
811 	td = curthread;
812 	thread_lock(td);
813 	old_cpu = PCPU_GET(cpuid);
814 	old_pinned = td->td_pinned;
815 	was_bound = sched_is_bound(td);
816 	sched_unbind(td);
817 	td->td_pinned = 0;
818 
819 	CPU_FOREACH(cpu)
820 		epoch->e_drain_count++;
821 	CPU_FOREACH(cpu) {
822 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
823 		sched_bind(td, cpu);
824 		epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
825 	}
826 
827 	/* restore CPU binding, if any */
828 	if (was_bound != 0) {
829 		sched_bind(td, old_cpu);
830 	} else {
831 		/* get thread back to initial CPU, if any */
832 		if (old_pinned != 0)
833 			sched_bind(td, old_cpu);
834 		sched_unbind(td);
835 	}
836 	/* restore pinned after bind */
837 	td->td_pinned = old_pinned;
838 
839 	thread_unlock(td);
840 
841 	while (epoch->e_drain_count != 0)
842 		msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
843 
844 	mtx_unlock(&epoch->e_drain_mtx);
845 	sx_xunlock(&epoch->e_drain_sx);
846 
847 	PICKUP_GIANT();
848 }
849