2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (c) 2018, Matthew Macy <mmacy@freebsd.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
29 #include <sys/cdefs.h>
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/counter.h>
33 #include <sys/epoch.h>
34 #include <sys/gtaskqueue.h>
35 #include <sys/kernel.h>
36 #include <sys/limits.h>
38 #include <sys/malloc.h>
39 #include <sys/mutex.h>
42 #include <sys/sched.h>
45 #include <sys/sysctl.h>
46 #include <sys/turnstile.h>
48 #include <machine/stdarg.h>
49 #include <sys/stack.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_kern.h>
60 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
62 #define EPOCH_ALIGN CACHE_LINE_SIZE
65 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
66 typedef struct epoch_record {
67 ck_epoch_record_t er_record;
68 struct epoch_context er_drain_ctx;
69 struct epoch *er_parent;
70 volatile struct epoch_tdlist er_tdlist;
71 volatile uint32_t er_gen;
74 /* Used to verify record ownership for non-preemptible epochs. */
77 } __aligned(EPOCH_ALIGN) *epoch_record_t;
80 struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
81 epoch_record_t e_pcpu_record;
85 struct mtx e_drain_mtx;
86 volatile int e_drain_count;
90 /* arbitrary --- needs benchmarking */
91 #define MAX_ADAPTIVE_SPIN 100
94 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
95 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
97 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101 static counter_u64_t block_count;
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;
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;
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;
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;
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;
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");
126 TAILQ_HEAD (threadlist, thread);
128 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
129 ck_epoch_entry_container)
131 static struct epoch epoch_array[MAX_EPOCHS];
133 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
134 DPCPU_DEFINE(int, epoch_cb_count);
136 static __read_mostly int inited;
137 __read_mostly epoch_t global_epoch;
138 __read_mostly epoch_t global_epoch_preempt;
140 static void epoch_call_task(void *context __unused);
141 static uma_zone_t pcpu_zone_record;
143 static struct sx epoch_sx;
145 #define EPOCH_LOCK() sx_xlock(&epoch_sx)
146 #define EPOCH_UNLOCK() sx_xunlock(&epoch_sx)
150 RB_ENTRY(stackentry) se_node;
151 struct stack se_stack;
155 stackentry_compare(struct stackentry *a, struct stackentry *b)
158 if (a->se_stack.depth > b->se_stack.depth)
160 if (a->se_stack.depth < b->se_stack.depth)
162 for (int i = 0; i < a->se_stack.depth; i++) {
163 if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
165 if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
172 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
173 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
175 static struct mtx epoch_stacks_lock;
176 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
178 static bool epoch_trace_stack_print = true;
179 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
180 &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
182 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
184 epoch_trace_report(const char *fmt, ...)
187 struct stackentry se, *new;
189 stack_zero(&se.se_stack); /* XXX: is it really needed? */
190 stack_save(&se.se_stack);
192 /* Tree is never reduced - go lockless. */
193 if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
196 new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
198 bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
200 mtx_lock(&epoch_stacks_lock);
201 new = RB_INSERT(stacktree, &epoch_stacks, new);
202 mtx_unlock(&epoch_stacks_lock);
208 (void)vprintf(fmt, ap);
210 if (epoch_trace_stack_print)
211 stack_print_ddb(&se.se_stack);
215 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
216 const char *file, int line)
220 SLIST_FOREACH(iet, &td->td_epochs, et_tlink) {
221 if (iet->et_epoch != epoch)
223 epoch_trace_report("Recursively entering epoch %s "
224 "at %s:%d, previously entered at %s:%d\n",
225 epoch->e_name, file, line,
226 iet->et_file, iet->et_line);
228 et->et_epoch = epoch;
231 SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
235 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
236 const char *file, int line)
239 if (SLIST_FIRST(&td->td_epochs) != et) {
240 epoch_trace_report("Exiting epoch %s in a not nested order "
241 "at %s:%d. Most recently entered %s at %s:%d\n",
244 SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
245 SLIST_FIRST(&td->td_epochs)->et_file,
246 SLIST_FIRST(&td->td_epochs)->et_line);
247 /* This will panic if et is not anywhere on td_epochs. */
248 SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
250 SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
253 /* Used by assertions that check thread state before going to sleep. */
255 epoch_trace_list(struct thread *td)
259 SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
260 printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
261 iet->et_file, iet->et_line);
263 #endif /* EPOCH_TRACE */
266 epoch_init(void *arg __unused)
270 block_count = counter_u64_alloc(M_WAITOK);
271 migrate_count = counter_u64_alloc(M_WAITOK);
272 turnstile_count = counter_u64_alloc(M_WAITOK);
273 switch_count = counter_u64_alloc(M_WAITOK);
274 epoch_call_count = counter_u64_alloc(M_WAITOK);
275 epoch_call_task_count = counter_u64_alloc(M_WAITOK);
277 pcpu_zone_record = uma_zcreate("epoch_record pcpu",
278 sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
279 UMA_ALIGN_PTR, UMA_ZONE_PCPU);
281 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
282 epoch_call_task, NULL);
283 taskqgroup_attach_cpu(qgroup_softirq,
284 DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
288 SLIST_INIT(&thread0.td_epochs);
290 sx_init(&epoch_sx, "epoch-sx");
292 global_epoch = epoch_alloc("Global", 0);
293 global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
295 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
297 #if !defined(EARLY_AP_STARTUP)
299 epoch_init_smp(void *dummy __unused)
303 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
307 epoch_ctor(epoch_t epoch)
312 epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
314 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
315 bzero(er, sizeof(*er));
316 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
317 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
319 er->er_parent = epoch;
324 epoch_adjust_prio(struct thread *td, u_char prio)
328 sched_prio(td, prio);
333 epoch_alloc(const char *name, int flags)
340 if (__predict_false(!inited))
341 panic("%s called too early in boot", __func__);
346 * Find a free index in the epoch array. If no free index is
347 * found, try to use the index after the last one.
351 * If too many epochs are currently allocated,
354 if (i == MAX_EPOCHS) {
358 if (epoch_array[i].e_in_use == 0)
362 epoch = epoch_array + i;
363 ck_epoch_init(&epoch->e_epoch);
365 epoch->e_flags = flags;
366 epoch->e_name = name;
367 sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
368 mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
371 * Set e_in_use last, because when this field is set the
372 * epoch_call_task() function will start scanning this epoch
375 atomic_store_rel_int(&epoch->e_in_use, 1);
382 epoch_free(epoch_t epoch)
390 MPASS(epoch->e_in_use != 0);
392 epoch_drain_callbacks(epoch);
394 atomic_store_rel_int(&epoch->e_in_use, 0);
396 * Make sure the epoch_call_task() function see e_in_use equal
397 * to zero, by calling epoch_wait() on the global_epoch:
399 epoch_wait(global_epoch);
404 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
407 * Sanity check: none of the records should be in use anymore.
408 * We drained callbacks above and freeing the pcpu records is
411 MPASS(er->er_td == NULL);
412 MPASS(TAILQ_EMPTY(&er->er_tdlist));
415 uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
416 mtx_destroy(&epoch->e_drain_mtx);
417 sx_destroy(&epoch->e_drain_sx);
418 memset(epoch, 0, sizeof(*epoch));
423 static epoch_record_t
424 epoch_currecord(epoch_t epoch)
427 return (zpcpu_get(epoch->e_pcpu_record));
430 #define INIT_CHECK(epoch) \
432 if (__predict_false((epoch) == NULL)) \
437 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
439 struct epoch_record *er;
442 MPASS(cold || epoch != NULL);
444 MPASS(kstack_contains(td, (vm_offset_t)et, sizeof(*et)));
447 MPASS(epoch->e_flags & EPOCH_PREEMPT);
450 epoch_trace_enter(td, epoch, et, file, line);
453 THREAD_NO_SLEEPING();
456 et->et_old_priority = td->td_priority;
457 er = epoch_currecord(epoch);
458 /* Record-level tracking is reserved for non-preemptible epochs. */
459 MPASS(er->er_td == NULL);
460 TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
461 ck_epoch_begin(&er->er_record, &et->et_section);
466 epoch_enter(epoch_t epoch)
470 MPASS(cold || epoch != NULL);
473 er = epoch_currecord(epoch);
475 if (er->er_record.active == 0) {
476 MPASS(er->er_td == NULL);
477 er->er_td = curthread;
479 /* We've recursed, just make sure our accounting isn't wrong. */
480 MPASS(er->er_td == curthread);
483 ck_epoch_begin(&er->er_record, NULL);
487 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
489 struct epoch_record *er;
496 THREAD_SLEEPING_OK();
497 er = epoch_currecord(epoch);
498 MPASS(epoch->e_flags & EPOCH_PREEMPT);
500 MPASS(et->et_td == td);
502 et->et_td = (void*)0xDEADBEEF;
503 /* Record-level tracking is reserved for non-preemptible epochs. */
504 MPASS(er->er_td == NULL);
506 ck_epoch_end(&er->er_record, &et->et_section);
507 TAILQ_REMOVE(&er->er_tdlist, et, et_link);
509 if (__predict_false(et->et_old_priority != td->td_priority))
510 epoch_adjust_prio(td, et->et_old_priority);
513 epoch_trace_exit(td, epoch, et, file, line);
518 epoch_exit(epoch_t epoch)
523 er = epoch_currecord(epoch);
524 ck_epoch_end(&er->er_record, NULL);
526 MPASS(er->er_td == curthread);
527 if (er->er_record.active == 0)
534 * epoch_block_handler_preempt() is a callback from the CK code when another
535 * thread is currently in an epoch section.
538 epoch_block_handler_preempt(struct ck_epoch *global __unused,
539 ck_epoch_record_t *cr, void *arg __unused)
541 epoch_record_t record;
542 struct thread *td, *owner, *curwaittd;
543 struct epoch_tracker *tdwait;
544 struct turnstile *ts;
545 struct lock_object *lock;
547 int locksheld __unused;
549 record = __containerof(cr, struct epoch_record, er_record);
551 locksheld = td->td_locks;
553 counter_u64_add(block_count, 1);
555 * We lost a race and there's no longer any threads
556 * on the CPU in an epoch section.
558 if (TAILQ_EMPTY(&record->er_tdlist))
561 if (record->er_cpuid != curcpu) {
563 * If the head of the list is running, we can wait for it
564 * to remove itself from the list and thus save us the
565 * overhead of a migration
567 gen = record->er_gen;
570 * We can't actually check if the waiting thread is running
571 * so we simply poll for it to exit before giving up and
576 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
577 gen == record->er_gen &&
578 spincount++ < MAX_ADAPTIVE_SPIN);
581 * If the generation has changed we can poll again
582 * otherwise we need to migrate.
584 if (gen != record->er_gen)
587 * Being on the same CPU as that of the record on which
588 * we need to wait allows us access to the thread
589 * list associated with that CPU. We can then examine the
590 * oldest thread in the queue and wait on its turnstile
591 * until it resumes and so on until a grace period
595 counter_u64_add(migrate_count, 1);
596 sched_bind(td, record->er_cpuid);
598 * At this point we need to return to the ck code
599 * to scan to see if a grace period has elapsed.
600 * We can't move on to check the thread list, because
601 * in the meantime new threads may have arrived that
602 * in fact belong to a different epoch.
607 * Try to find a thread in an epoch section on this CPU
608 * waiting on a turnstile. Otherwise find the lowest
609 * priority thread (highest prio value) and drop our priority
610 * to match to allow it to run.
612 TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
614 * Propagate our priority to any other waiters to prevent us
615 * from starving them. They will have their original priority
616 * restore on exit from epoch_wait().
618 curwaittd = tdwait->et_td;
619 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
622 thread_lock(curwaittd);
623 sched_prio(curwaittd, td->td_priority);
624 thread_unlock(curwaittd);
628 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
629 ((ts = curwaittd->td_blocked) != NULL)) {
631 * We unlock td to allow turnstile_wait to reacquire
632 * the thread lock. Before unlocking it we enter a
633 * critical section to prevent preemption after we
634 * reenable interrupts by dropping the thread lock in
635 * order to prevent curwaittd from getting to run.
640 if (turnstile_lock(ts, &lock, &owner)) {
641 if (ts == curwaittd->td_blocked) {
642 MPASS(TD_IS_INHIBITED(curwaittd) &&
643 TD_ON_LOCK(curwaittd));
645 turnstile_wait(ts, owner,
646 curwaittd->td_tsqueue);
647 counter_u64_add(turnstile_count, 1);
651 turnstile_unlock(ts, lock);
655 KASSERT(td->td_locks == locksheld,
656 ("%d extra locks held", td->td_locks - locksheld));
660 * We didn't find any threads actually blocked on a lock
661 * so we have nothing to do except context switch away.
663 counter_u64_add(switch_count, 1);
664 mi_switch(SW_VOL | SWT_RELINQUISH);
666 * It is important the thread lock is dropped while yielding
667 * to allow other threads to acquire the lock pointed to by
668 * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
669 * thread lock before returning. Else a deadlock like
670 * situation might happen.
676 epoch_wait_preempt(epoch_t epoch)
685 MPASS(cold || epoch != NULL);
689 locks = curthread->td_locks;
690 MPASS(epoch->e_flags & EPOCH_PREEMPT);
691 if ((epoch->e_flags & EPOCH_LOCKED) == 0)
692 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
693 "epoch_wait() can be long running");
694 KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
695 "of an epoch section of the same epoch"));
700 old_cpu = PCPU_GET(cpuid);
701 old_pinned = td->td_pinned;
702 old_prio = td->td_priority;
703 was_bound = sched_is_bound(td);
706 sched_bind(td, old_cpu);
708 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
711 /* restore CPU binding, if any */
712 if (was_bound != 0) {
713 sched_bind(td, old_cpu);
715 /* get thread back to initial CPU, if any */
717 sched_bind(td, old_cpu);
720 /* restore pinned after bind */
721 td->td_pinned = old_pinned;
723 /* restore thread priority */
724 sched_prio(td, old_prio);
727 KASSERT(td->td_locks == locks,
728 ("%d residual locks held", td->td_locks - locks));
732 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
739 epoch_wait(epoch_t epoch)
742 MPASS(cold || epoch != NULL);
744 MPASS(epoch->e_flags == 0);
746 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
751 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
754 ck_epoch_entry_t *cb;
759 /* too early in boot to have epoch set up */
760 if (__predict_false(epoch == NULL))
762 #if !defined(EARLY_AP_STARTUP)
763 if (__predict_false(inited < 2))
768 *DPCPU_PTR(epoch_cb_count) += 1;
769 er = epoch_currecord(epoch);
770 ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
778 epoch_call_task(void *arg __unused)
780 ck_stack_entry_t *cursor, *head, *next;
781 ck_epoch_record_t *record;
785 int i, npending, total;
787 ck_stack_init(&cb_stack);
789 epoch_enter(global_epoch);
790 for (total = i = 0; i != MAX_EPOCHS; i++) {
791 epoch = epoch_array + i;
793 atomic_load_acq_int(&epoch->e_in_use) == 0))
795 er = epoch_currecord(epoch);
796 record = &er->er_record;
797 if ((npending = record->n_pending) == 0)
799 ck_epoch_poll_deferred(record, &cb_stack);
800 total += npending - record->n_pending;
802 epoch_exit(global_epoch);
803 *DPCPU_PTR(epoch_cb_count) -= total;
806 counter_u64_add(epoch_call_count, total);
807 counter_u64_add(epoch_call_task_count, 1);
809 head = ck_stack_batch_pop_npsc(&cb_stack);
810 for (cursor = head; cursor != NULL; cursor = next) {
811 struct ck_epoch_entry *entry =
812 ck_epoch_entry_container(cursor);
814 next = CK_STACK_NEXT(cursor);
815 entry->function(entry);
820 in_epoch_verbose_preempt(epoch_t epoch, int dump_onfail)
823 struct epoch_tracker *tdwait;
826 MPASS(epoch != NULL);
827 MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
829 if (THREAD_CAN_SLEEP())
832 er = epoch_currecord(epoch);
833 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
834 if (tdwait->et_td == td) {
840 MPASS(td->td_pinned);
841 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
842 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
843 printf("td_tid: %d ", tdwait->et_td->td_tid);
853 epoch_assert_nocpu(epoch_t epoch, struct thread *td)
859 crit = td->td_critnest > 0;
861 /* Check for a critical section mishap. */
863 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
864 KASSERT(er->er_td != td,
865 ("%s critical section in epoch '%s', from cpu %d",
866 (crit ? "exited" : "re-entered"), epoch->e_name, cpu));
870 #define epoch_assert_nocpu(e, td) do {} while (0)
874 in_epoch_verbose(epoch_t epoch, int dump_onfail)
879 if (__predict_false((epoch) == NULL))
881 if ((epoch->e_flags & EPOCH_PREEMPT) != 0)
882 return (in_epoch_verbose_preempt(epoch, dump_onfail));
885 * The thread being in a critical section is a necessary
886 * condition to be correctly inside a non-preemptible epoch,
887 * so it's definitely not in this epoch.
890 if (td->td_critnest == 0) {
891 epoch_assert_nocpu(epoch, td);
896 * The current cpu is in a critical section, so the epoch record will be
897 * stable for the rest of this function. Knowing that the record is not
898 * active is sufficient for knowing whether we're in this epoch or not,
899 * since it's a pcpu record.
901 er = epoch_currecord(epoch);
902 if (er->er_record.active == 0) {
903 epoch_assert_nocpu(epoch, td);
907 MPASS(er->er_td == td);
912 in_epoch(epoch_t epoch)
914 return (in_epoch_verbose(epoch, 0));
918 epoch_drain_cb(struct epoch_context *ctx)
920 struct epoch *epoch =
921 __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
923 if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
924 mtx_lock(&epoch->e_drain_mtx);
926 mtx_unlock(&epoch->e_drain_mtx);
931 epoch_drain_callbacks(epoch_t epoch)
940 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
941 "epoch_drain_callbacks() may sleep!");
943 /* too early in boot to have epoch set up */
944 if (__predict_false(epoch == NULL))
946 #if !defined(EARLY_AP_STARTUP)
947 if (__predict_false(inited < 2))
952 sx_xlock(&epoch->e_drain_sx);
953 mtx_lock(&epoch->e_drain_mtx);
957 old_cpu = PCPU_GET(cpuid);
958 old_pinned = td->td_pinned;
959 was_bound = sched_is_bound(td);
964 epoch->e_drain_count++;
966 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
968 epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
971 /* restore CPU binding, if any */
972 if (was_bound != 0) {
973 sched_bind(td, old_cpu);
975 /* get thread back to initial CPU, if any */
977 sched_bind(td, old_cpu);
980 /* restore pinned after bind */
981 td->td_pinned = old_pinned;
985 while (epoch->e_drain_count != 0)
986 msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
988 mtx_unlock(&epoch->e_drain_mtx);
989 sx_xunlock(&epoch->e_drain_sx);