xref: /freebsd/sys/kern/subr_epoch.c (revision 4731124cace5e7a0224e29784617d2856e5c59ab)
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 static epoch_record_t
151 epoch_currecord(epoch_t epoch)
152 {
153 
154 	return (zpcpu_get(epoch->e_pcpu_record));
155 }
156 
157 #ifdef EPOCH_TRACE
158 struct stackentry {
159 	RB_ENTRY(stackentry) se_node;
160 	struct stack se_stack;
161 };
162 
163 static int
164 stackentry_compare(struct stackentry *a, struct stackentry *b)
165 {
166 
167 	if (a->se_stack.depth > b->se_stack.depth)
168 		return (1);
169 	if (a->se_stack.depth < b->se_stack.depth)
170 		return (-1);
171 	for (int i = 0; i < a->se_stack.depth; i++) {
172 		if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
173 			return (1);
174 		if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
175 			return (-1);
176 	}
177 
178 	return (0);
179 }
180 
181 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
182 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
183 
184 static struct mtx epoch_stacks_lock;
185 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
186 
187 static bool epoch_trace_stack_print = true;
188 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
189     &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
190 
191 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
192 static inline void
193 epoch_trace_report(const char *fmt, ...)
194 {
195 	va_list ap;
196 	struct stackentry se, *new;
197 
198 	stack_zero(&se.se_stack);	/* XXX: is it really needed? */
199 	stack_save(&se.se_stack);
200 
201 	/* Tree is never reduced - go lockless. */
202 	if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
203 		return;
204 
205 	new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
206 	if (new != NULL) {
207 		bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
208 
209 		mtx_lock(&epoch_stacks_lock);
210 		new = RB_INSERT(stacktree, &epoch_stacks, new);
211 		mtx_unlock(&epoch_stacks_lock);
212 		if (new != NULL)
213 			free(new, M_STACK);
214 	}
215 
216 	va_start(ap, fmt);
217 	(void)vprintf(fmt, ap);
218 	va_end(ap);
219 	if (epoch_trace_stack_print)
220 		stack_print_ddb(&se.se_stack);
221 }
222 
223 static inline void
224 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
225     const char *file, int line)
226 {
227 	epoch_tracker_t iet;
228 
229 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink) {
230 		if (iet->et_epoch != epoch)
231 			continue;
232 		epoch_trace_report("Recursively entering epoch %s "
233 		    "at %s:%d, previously entered at %s:%d\n",
234 		    epoch->e_name, file, line,
235 		    iet->et_file, iet->et_line);
236 	}
237 	et->et_epoch = epoch;
238 	et->et_file = file;
239 	et->et_line = line;
240 	et->et_flags = 0;
241 	SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
242 }
243 
244 static inline void
245 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
246     const char *file, int line)
247 {
248 
249 	if (SLIST_FIRST(&td->td_epochs) != et) {
250 		epoch_trace_report("Exiting epoch %s in a not nested order "
251 		    "at %s:%d. Most recently entered %s at %s:%d\n",
252 		    epoch->e_name,
253 		    file, line,
254 		    SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
255 		    SLIST_FIRST(&td->td_epochs)->et_file,
256 		    SLIST_FIRST(&td->td_epochs)->et_line);
257 		/* This will panic if et is not anywhere on td_epochs. */
258 		SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
259 	} else
260 		SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
261 	if (et->et_flags & ET_REPORT_EXIT)
262 		printf("Td %p exiting epoch %s at %s:%d\n", td, epoch->e_name,
263 		    file, line);
264 }
265 
266 /* Used by assertions that check thread state before going to sleep. */
267 void
268 epoch_trace_list(struct thread *td)
269 {
270 	epoch_tracker_t iet;
271 
272 	SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
273 		printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
274 		    iet->et_file, iet->et_line);
275 }
276 
277 void
278 epoch_where_report(epoch_t epoch)
279 {
280 	epoch_record_t er;
281 	struct epoch_tracker *tdwait;
282 
283 	MPASS(epoch != NULL);
284 	MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
285 	MPASS(!THREAD_CAN_SLEEP());
286 	critical_enter();
287 	er = epoch_currecord(epoch);
288 	TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
289 		if (tdwait->et_td == curthread)
290 			break;
291 	critical_exit();
292 	if (tdwait != NULL) {
293 		tdwait->et_flags |= ET_REPORT_EXIT;
294 		printf("Td %p entered epoch %s at %s:%d\n", curthread,
295 		    epoch->e_name, tdwait->et_file, tdwait->et_line);
296 	}
297 }
298 #endif /* EPOCH_TRACE */
299 
300 static void
301 epoch_init(void *arg __unused)
302 {
303 	int cpu;
304 
305 	block_count = counter_u64_alloc(M_WAITOK);
306 	migrate_count = counter_u64_alloc(M_WAITOK);
307 	turnstile_count = counter_u64_alloc(M_WAITOK);
308 	switch_count = counter_u64_alloc(M_WAITOK);
309 	epoch_call_count = counter_u64_alloc(M_WAITOK);
310 	epoch_call_task_count = counter_u64_alloc(M_WAITOK);
311 
312 	pcpu_zone_record = uma_zcreate("epoch_record pcpu",
313 	    sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
314 	    UMA_ALIGN_PTR, UMA_ZONE_PCPU);
315 	CPU_FOREACH(cpu) {
316 		GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
317 		    epoch_call_task, NULL);
318 		taskqgroup_attach_cpu(qgroup_softirq,
319 		    DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
320 		    "epoch call task");
321 	}
322 #ifdef EPOCH_TRACE
323 	SLIST_INIT(&thread0.td_epochs);
324 #endif
325 	sx_init(&epoch_sx, "epoch-sx");
326 	inited = 1;
327 	global_epoch = epoch_alloc("Global", 0);
328 	global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
329 }
330 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
331 
332 #if !defined(EARLY_AP_STARTUP)
333 static void
334 epoch_init_smp(void *dummy __unused)
335 {
336 	inited = 2;
337 }
338 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
339 #endif
340 
341 static void
342 epoch_ctor(epoch_t epoch)
343 {
344 	epoch_record_t er;
345 	int cpu;
346 
347 	epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
348 	CPU_FOREACH(cpu) {
349 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
350 		bzero(er, sizeof(*er));
351 		ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
352 		TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
353 		er->er_cpuid = cpu;
354 		er->er_parent = epoch;
355 	}
356 }
357 
358 static void
359 epoch_adjust_prio(struct thread *td, u_char prio)
360 {
361 
362 	thread_lock(td);
363 	sched_prio(td, prio);
364 	thread_unlock(td);
365 }
366 
367 epoch_t
368 epoch_alloc(const char *name, int flags)
369 {
370 	epoch_t epoch;
371 	int i;
372 
373 	MPASS(name != NULL);
374 
375 	if (__predict_false(!inited))
376 		panic("%s called too early in boot", __func__);
377 
378 	EPOCH_LOCK();
379 
380 	/*
381 	 * Find a free index in the epoch array. If no free index is
382 	 * found, try to use the index after the last one.
383 	 */
384 	for (i = 0;; i++) {
385 		/*
386 		 * If too many epochs are currently allocated,
387 		 * return NULL.
388 		 */
389 		if (i == MAX_EPOCHS) {
390 			epoch = NULL;
391 			goto done;
392 		}
393 		if (epoch_array[i].e_in_use == 0)
394 			break;
395 	}
396 
397 	epoch = epoch_array + i;
398 	ck_epoch_init(&epoch->e_epoch);
399 	epoch_ctor(epoch);
400 	epoch->e_flags = flags;
401 	epoch->e_name = name;
402 	sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
403 	mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
404 
405 	/*
406 	 * Set e_in_use last, because when this field is set the
407 	 * epoch_call_task() function will start scanning this epoch
408 	 * structure.
409 	 */
410 	atomic_store_rel_int(&epoch->e_in_use, 1);
411 done:
412 	EPOCH_UNLOCK();
413 	return (epoch);
414 }
415 
416 void
417 epoch_free(epoch_t epoch)
418 {
419 #ifdef INVARIANTS
420 	int cpu;
421 #endif
422 
423 	EPOCH_LOCK();
424 
425 	MPASS(epoch->e_in_use != 0);
426 
427 	epoch_drain_callbacks(epoch);
428 
429 	atomic_store_rel_int(&epoch->e_in_use, 0);
430 	/*
431 	 * Make sure the epoch_call_task() function see e_in_use equal
432 	 * to zero, by calling epoch_wait() on the global_epoch:
433 	 */
434 	epoch_wait(global_epoch);
435 #ifdef INVARIANTS
436 	CPU_FOREACH(cpu) {
437 		epoch_record_t er;
438 
439 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
440 
441 		/*
442 		 * Sanity check: none of the records should be in use anymore.
443 		 * We drained callbacks above and freeing the pcpu records is
444 		 * imminent.
445 		 */
446 		MPASS(er->er_td == NULL);
447 		MPASS(TAILQ_EMPTY(&er->er_tdlist));
448 	}
449 #endif
450 	uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
451 	mtx_destroy(&epoch->e_drain_mtx);
452 	sx_destroy(&epoch->e_drain_sx);
453 	memset(epoch, 0, sizeof(*epoch));
454 
455 	EPOCH_UNLOCK();
456 }
457 
458 #define INIT_CHECK(epoch)					\
459 	do {							\
460 		if (__predict_false((epoch) == NULL))		\
461 			return;					\
462 	} while (0)
463 
464 void
465 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
466 {
467 	struct epoch_record *er;
468 	struct thread *td;
469 
470 	MPASS(cold || epoch != NULL);
471 	td = curthread;
472 	MPASS((vm_offset_t)et >= td->td_kstack &&
473 	    (vm_offset_t)et + sizeof(struct epoch_tracker) <=
474 	    td->td_kstack + td->td_kstack_pages * PAGE_SIZE);
475 
476 	INIT_CHECK(epoch);
477 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
478 
479 #ifdef EPOCH_TRACE
480 	epoch_trace_enter(td, epoch, et, file, line);
481 #endif
482 	et->et_td = td;
483 	THREAD_NO_SLEEPING();
484 	critical_enter();
485 	sched_pin();
486 	et->et_old_priority = td->td_priority;
487 	er = epoch_currecord(epoch);
488 	/* Record-level tracking is reserved for non-preemptible epochs. */
489 	MPASS(er->er_td == NULL);
490 	TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
491 	ck_epoch_begin(&er->er_record, &et->et_section);
492 	critical_exit();
493 }
494 
495 void
496 epoch_enter(epoch_t epoch)
497 {
498 	epoch_record_t er;
499 
500 	MPASS(cold || epoch != NULL);
501 	INIT_CHECK(epoch);
502 	critical_enter();
503 	er = epoch_currecord(epoch);
504 #ifdef INVARIANTS
505 	if (er->er_record.active == 0) {
506 		MPASS(er->er_td == NULL);
507 		er->er_td = curthread;
508 	} else {
509 		/* We've recursed, just make sure our accounting isn't wrong. */
510 		MPASS(er->er_td == curthread);
511 	}
512 #endif
513 	ck_epoch_begin(&er->er_record, NULL);
514 }
515 
516 void
517 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
518 {
519 	struct epoch_record *er;
520 	struct thread *td;
521 
522 	INIT_CHECK(epoch);
523 	td = curthread;
524 	critical_enter();
525 	sched_unpin();
526 	THREAD_SLEEPING_OK();
527 	er = epoch_currecord(epoch);
528 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
529 	MPASS(et != NULL);
530 	MPASS(et->et_td == td);
531 #ifdef INVARIANTS
532 	et->et_td = (void*)0xDEADBEEF;
533 	/* Record-level tracking is reserved for non-preemptible epochs. */
534 	MPASS(er->er_td == NULL);
535 #endif
536 	ck_epoch_end(&er->er_record, &et->et_section);
537 	TAILQ_REMOVE(&er->er_tdlist, et, et_link);
538 	er->er_gen++;
539 	if (__predict_false(et->et_old_priority != td->td_priority))
540 		epoch_adjust_prio(td, et->et_old_priority);
541 	critical_exit();
542 #ifdef EPOCH_TRACE
543 	epoch_trace_exit(td, epoch, et, file, line);
544 #endif
545 }
546 
547 void
548 epoch_exit(epoch_t epoch)
549 {
550 	epoch_record_t er;
551 
552 	INIT_CHECK(epoch);
553 	er = epoch_currecord(epoch);
554 	ck_epoch_end(&er->er_record, NULL);
555 #ifdef INVARIANTS
556 	MPASS(er->er_td == curthread);
557 	if (er->er_record.active == 0)
558 		er->er_td = NULL;
559 #endif
560 	critical_exit();
561 }
562 
563 /*
564  * epoch_block_handler_preempt() is a callback from the CK code when another
565  * thread is currently in an epoch section.
566  */
567 static void
568 epoch_block_handler_preempt(struct ck_epoch *global __unused,
569     ck_epoch_record_t *cr, void *arg __unused)
570 {
571 	epoch_record_t record;
572 	struct thread *td, *owner, *curwaittd;
573 	struct epoch_tracker *tdwait;
574 	struct turnstile *ts;
575 	struct lock_object *lock;
576 	int spincount, gen;
577 	int locksheld __unused;
578 
579 	record = __containerof(cr, struct epoch_record, er_record);
580 	td = curthread;
581 	locksheld = td->td_locks;
582 	spincount = 0;
583 	counter_u64_add(block_count, 1);
584 	/*
585 	 * We lost a race and there's no longer any threads
586 	 * on the CPU in an epoch section.
587 	 */
588 	if (TAILQ_EMPTY(&record->er_tdlist))
589 		return;
590 
591 	if (record->er_cpuid != curcpu) {
592 		/*
593 		 * If the head of the list is running, we can wait for it
594 		 * to remove itself from the list and thus save us the
595 		 * overhead of a migration
596 		 */
597 		gen = record->er_gen;
598 		thread_unlock(td);
599 		/*
600 		 * We can't actually check if the waiting thread is running
601 		 * so we simply poll for it to exit before giving up and
602 		 * migrating.
603 		 */
604 		do {
605 			cpu_spinwait();
606 		} while (!TAILQ_EMPTY(&record->er_tdlist) &&
607 				 gen == record->er_gen &&
608 				 spincount++ < MAX_ADAPTIVE_SPIN);
609 		thread_lock(td);
610 		/*
611 		 * If the generation has changed we can poll again
612 		 * otherwise we need to migrate.
613 		 */
614 		if (gen != record->er_gen)
615 			return;
616 		/*
617 		 * Being on the same CPU as that of the record on which
618 		 * we need to wait allows us access to the thread
619 		 * list associated with that CPU. We can then examine the
620 		 * oldest thread in the queue and wait on its turnstile
621 		 * until it resumes and so on until a grace period
622 		 * elapses.
623 		 *
624 		 */
625 		counter_u64_add(migrate_count, 1);
626 		sched_bind(td, record->er_cpuid);
627 		/*
628 		 * At this point we need to return to the ck code
629 		 * to scan to see if a grace period has elapsed.
630 		 * We can't move on to check the thread list, because
631 		 * in the meantime new threads may have arrived that
632 		 * in fact belong to a different epoch.
633 		 */
634 		return;
635 	}
636 	/*
637 	 * Try to find a thread in an epoch section on this CPU
638 	 * waiting on a turnstile. Otherwise find the lowest
639 	 * priority thread (highest prio value) and drop our priority
640 	 * to match to allow it to run.
641 	 */
642 	TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
643 		/*
644 		 * Propagate our priority to any other waiters to prevent us
645 		 * from starving them. They will have their original priority
646 		 * restore on exit from epoch_wait().
647 		 */
648 		curwaittd = tdwait->et_td;
649 		if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
650 			critical_enter();
651 			thread_unlock(td);
652 			thread_lock(curwaittd);
653 			sched_prio(curwaittd, td->td_priority);
654 			thread_unlock(curwaittd);
655 			thread_lock(td);
656 			critical_exit();
657 		}
658 		if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
659 		    ((ts = curwaittd->td_blocked) != NULL)) {
660 			/*
661 			 * We unlock td to allow turnstile_wait to reacquire
662 			 * the thread lock. Before unlocking it we enter a
663 			 * critical section to prevent preemption after we
664 			 * reenable interrupts by dropping the thread lock in
665 			 * order to prevent curwaittd from getting to run.
666 			 */
667 			critical_enter();
668 			thread_unlock(td);
669 
670 			if (turnstile_lock(ts, &lock, &owner)) {
671 				if (ts == curwaittd->td_blocked) {
672 					MPASS(TD_IS_INHIBITED(curwaittd) &&
673 					    TD_ON_LOCK(curwaittd));
674 					critical_exit();
675 					turnstile_wait(ts, owner,
676 					    curwaittd->td_tsqueue);
677 					counter_u64_add(turnstile_count, 1);
678 					thread_lock(td);
679 					return;
680 				}
681 				turnstile_unlock(ts, lock);
682 			}
683 			thread_lock(td);
684 			critical_exit();
685 			KASSERT(td->td_locks == locksheld,
686 			    ("%d extra locks held", td->td_locks - locksheld));
687 		}
688 	}
689 	/*
690 	 * We didn't find any threads actually blocked on a lock
691 	 * so we have nothing to do except context switch away.
692 	 */
693 	counter_u64_add(switch_count, 1);
694 	mi_switch(SW_VOL | SWT_RELINQUISH);
695 	/*
696 	 * It is important the thread lock is dropped while yielding
697 	 * to allow other threads to acquire the lock pointed to by
698 	 * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
699 	 * thread lock before returning. Else a deadlock like
700 	 * situation might happen.
701 	 */
702 	thread_lock(td);
703 }
704 
705 void
706 epoch_wait_preempt(epoch_t epoch)
707 {
708 	struct thread *td;
709 	int was_bound;
710 	int old_cpu;
711 	int old_pinned;
712 	u_char old_prio;
713 	int locks __unused;
714 
715 	MPASS(cold || epoch != NULL);
716 	INIT_CHECK(epoch);
717 	td = curthread;
718 #ifdef INVARIANTS
719 	locks = curthread->td_locks;
720 	MPASS(epoch->e_flags & EPOCH_PREEMPT);
721 	if ((epoch->e_flags & EPOCH_LOCKED) == 0)
722 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
723 		    "epoch_wait() can be long running");
724 	KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
725 	    "of an epoch section of the same epoch"));
726 #endif
727 	DROP_GIANT();
728 	thread_lock(td);
729 
730 	old_cpu = PCPU_GET(cpuid);
731 	old_pinned = td->td_pinned;
732 	old_prio = td->td_priority;
733 	was_bound = sched_is_bound(td);
734 	sched_unbind(td);
735 	td->td_pinned = 0;
736 	sched_bind(td, old_cpu);
737 
738 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
739 	    NULL);
740 
741 	/* restore CPU binding, if any */
742 	if (was_bound != 0) {
743 		sched_bind(td, old_cpu);
744 	} else {
745 		/* get thread back to initial CPU, if any */
746 		if (old_pinned != 0)
747 			sched_bind(td, old_cpu);
748 		sched_unbind(td);
749 	}
750 	/* restore pinned after bind */
751 	td->td_pinned = old_pinned;
752 
753 	/* restore thread priority */
754 	sched_prio(td, old_prio);
755 	thread_unlock(td);
756 	PICKUP_GIANT();
757 	KASSERT(td->td_locks == locks,
758 	    ("%d residual locks held", td->td_locks - locks));
759 }
760 
761 static void
762 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
763     void *arg __unused)
764 {
765 	cpu_spinwait();
766 }
767 
768 void
769 epoch_wait(epoch_t epoch)
770 {
771 
772 	MPASS(cold || epoch != NULL);
773 	INIT_CHECK(epoch);
774 	MPASS(epoch->e_flags == 0);
775 	critical_enter();
776 	ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
777 	critical_exit();
778 }
779 
780 void
781 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
782 {
783 	epoch_record_t er;
784 	ck_epoch_entry_t *cb;
785 
786 	cb = (void *)ctx;
787 
788 	MPASS(callback);
789 	/* too early in boot to have epoch set up */
790 	if (__predict_false(epoch == NULL))
791 		goto boottime;
792 #if !defined(EARLY_AP_STARTUP)
793 	if (__predict_false(inited < 2))
794 		goto boottime;
795 #endif
796 
797 	critical_enter();
798 	*DPCPU_PTR(epoch_cb_count) += 1;
799 	er = epoch_currecord(epoch);
800 	ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
801 	critical_exit();
802 	return;
803 boottime:
804 	callback(ctx);
805 }
806 
807 static void
808 epoch_call_task(void *arg __unused)
809 {
810 	ck_stack_entry_t *cursor, *head, *next;
811 	ck_epoch_record_t *record;
812 	epoch_record_t er;
813 	epoch_t epoch;
814 	ck_stack_t cb_stack;
815 	int i, npending, total;
816 
817 	ck_stack_init(&cb_stack);
818 	critical_enter();
819 	epoch_enter(global_epoch);
820 	for (total = i = 0; i != MAX_EPOCHS; i++) {
821 		epoch = epoch_array + i;
822 		if (__predict_false(
823 		    atomic_load_acq_int(&epoch->e_in_use) == 0))
824 			continue;
825 		er = epoch_currecord(epoch);
826 		record = &er->er_record;
827 		if ((npending = record->n_pending) == 0)
828 			continue;
829 		ck_epoch_poll_deferred(record, &cb_stack);
830 		total += npending - record->n_pending;
831 	}
832 	epoch_exit(global_epoch);
833 	*DPCPU_PTR(epoch_cb_count) -= total;
834 	critical_exit();
835 
836 	counter_u64_add(epoch_call_count, total);
837 	counter_u64_add(epoch_call_task_count, 1);
838 
839 	head = ck_stack_batch_pop_npsc(&cb_stack);
840 	for (cursor = head; cursor != NULL; cursor = next) {
841 		struct ck_epoch_entry *entry =
842 		    ck_epoch_entry_container(cursor);
843 
844 		next = CK_STACK_NEXT(cursor);
845 		entry->function(entry);
846 	}
847 }
848 
849 static int
850 in_epoch_verbose_preempt(epoch_t epoch, int dump_onfail)
851 {
852 	epoch_record_t er;
853 	struct epoch_tracker *tdwait;
854 	struct thread *td;
855 
856 	MPASS(epoch != NULL);
857 	MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
858 	td = curthread;
859 	if (THREAD_CAN_SLEEP())
860 		return (0);
861 	critical_enter();
862 	er = epoch_currecord(epoch);
863 	TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
864 		if (tdwait->et_td == td) {
865 			critical_exit();
866 			return (1);
867 		}
868 #ifdef INVARIANTS
869 	if (dump_onfail) {
870 		MPASS(td->td_pinned);
871 		printf("cpu: %d id: %d\n", curcpu, td->td_tid);
872 		TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
873 			printf("td_tid: %d ", tdwait->et_td->td_tid);
874 		printf("\n");
875 	}
876 #endif
877 	critical_exit();
878 	return (0);
879 }
880 
881 #ifdef INVARIANTS
882 static void
883 epoch_assert_nocpu(epoch_t epoch, struct thread *td)
884 {
885 	epoch_record_t er;
886 	int cpu;
887 	bool crit;
888 
889 	crit = td->td_critnest > 0;
890 
891 	/* Check for a critical section mishap. */
892 	CPU_FOREACH(cpu) {
893 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
894 		KASSERT(er->er_td != td,
895 		    ("%s critical section in epoch '%s', from cpu %d",
896 		    (crit ? "exited" : "re-entered"), epoch->e_name, cpu));
897 	}
898 }
899 #else
900 #define	epoch_assert_nocpu(e, td) do {} while (0)
901 #endif
902 
903 int
904 in_epoch_verbose(epoch_t epoch, int dump_onfail)
905 {
906 	epoch_record_t er;
907 	struct thread *td;
908 
909 	if (__predict_false((epoch) == NULL))
910 		return (0);
911 	if ((epoch->e_flags & EPOCH_PREEMPT) != 0)
912 		return (in_epoch_verbose_preempt(epoch, dump_onfail));
913 
914 	/*
915 	 * The thread being in a critical section is a necessary
916 	 * condition to be correctly inside a non-preemptible epoch,
917 	 * so it's definitely not in this epoch.
918 	 */
919 	td = curthread;
920 	if (td->td_critnest == 0) {
921 		epoch_assert_nocpu(epoch, td);
922 		return (0);
923 	}
924 
925 	/*
926 	 * The current cpu is in a critical section, so the epoch record will be
927 	 * stable for the rest of this function.  Knowing that the record is not
928 	 * active is sufficient for knowing whether we're in this epoch or not,
929 	 * since it's a pcpu record.
930 	 */
931 	er = epoch_currecord(epoch);
932 	if (er->er_record.active == 0) {
933 		epoch_assert_nocpu(epoch, td);
934 		return (0);
935 	}
936 
937 	MPASS(er->er_td == td);
938 	return (1);
939 }
940 
941 int
942 in_epoch(epoch_t epoch)
943 {
944 	return (in_epoch_verbose(epoch, 0));
945 }
946 
947 static void
948 epoch_drain_cb(struct epoch_context *ctx)
949 {
950 	struct epoch *epoch =
951 	    __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
952 
953 	if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
954 		mtx_lock(&epoch->e_drain_mtx);
955 		wakeup(epoch);
956 		mtx_unlock(&epoch->e_drain_mtx);
957 	}
958 }
959 
960 void
961 epoch_drain_callbacks(epoch_t epoch)
962 {
963 	epoch_record_t er;
964 	struct thread *td;
965 	int was_bound;
966 	int old_pinned;
967 	int old_cpu;
968 	int cpu;
969 
970 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
971 	    "epoch_drain_callbacks() may sleep!");
972 
973 	/* too early in boot to have epoch set up */
974 	if (__predict_false(epoch == NULL))
975 		return;
976 #if !defined(EARLY_AP_STARTUP)
977 	if (__predict_false(inited < 2))
978 		return;
979 #endif
980 	DROP_GIANT();
981 
982 	sx_xlock(&epoch->e_drain_sx);
983 	mtx_lock(&epoch->e_drain_mtx);
984 
985 	td = curthread;
986 	thread_lock(td);
987 	old_cpu = PCPU_GET(cpuid);
988 	old_pinned = td->td_pinned;
989 	was_bound = sched_is_bound(td);
990 	sched_unbind(td);
991 	td->td_pinned = 0;
992 
993 	CPU_FOREACH(cpu)
994 		epoch->e_drain_count++;
995 	CPU_FOREACH(cpu) {
996 		er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
997 		sched_bind(td, cpu);
998 		epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
999 	}
1000 
1001 	/* restore CPU binding, if any */
1002 	if (was_bound != 0) {
1003 		sched_bind(td, old_cpu);
1004 	} else {
1005 		/* get thread back to initial CPU, if any */
1006 		if (old_pinned != 0)
1007 			sched_bind(td, old_cpu);
1008 		sched_unbind(td);
1009 	}
1010 	/* restore pinned after bind */
1011 	td->td_pinned = old_pinned;
1012 
1013 	thread_unlock(td);
1014 
1015 	while (epoch->e_drain_count != 0)
1016 		msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
1017 
1018 	mtx_unlock(&epoch->e_drain_mtx);
1019 	sx_xunlock(&epoch->e_drain_sx);
1020 
1021 	PICKUP_GIANT();
1022 }
1023