2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
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>
40 #include <sys/malloc.h>
41 #include <sys/mutex.h>
44 #include <sys/sched.h>
47 #include <sys/sysctl.h>
48 #include <sys/turnstile.h>
50 #include <machine/stdarg.h>
51 #include <sys/stack.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_kern.h>
61 static MALLOC_DEFINE(M_EPOCH, "epoch", "epoch based reclamation");
64 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
66 #define EPOCH_ALIGN CACHE_LINE_SIZE
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;
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, 0, "epoch information");
96 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW, 0, "epoch stats");
99 static counter_u64_t block_count;
101 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
102 &block_count, "# of times a thread was in an epoch when epoch_wait was called");
103 static counter_u64_t migrate_count;
105 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
106 &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
107 static counter_u64_t turnstile_count;
109 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
110 &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
111 static counter_u64_t switch_count;
113 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
114 &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
115 static counter_u64_t epoch_call_count;
117 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
118 &epoch_call_count, "# of times a callback was deferred");
119 static counter_u64_t epoch_call_task_count;
121 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
122 &epoch_call_task_count, "# of times a callback task was run");
124 TAILQ_HEAD (threadlist, thread);
126 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
127 ck_epoch_entry_container)
129 epoch_t allepochs[MAX_EPOCHS];
131 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
132 DPCPU_DEFINE(int, epoch_cb_count);
134 static __read_mostly int inited;
135 static __read_mostly int epoch_count;
136 __read_mostly epoch_t global_epoch;
137 __read_mostly epoch_t global_epoch_preempt;
139 static void epoch_call_task(void *context __unused);
140 static uma_zone_t pcpu_zone_record;
144 RB_ENTRY(stackentry) se_node;
145 struct stack se_stack;
149 stackentry_compare(struct stackentry *a, struct stackentry *b)
152 if (a->se_stack.depth > b->se_stack.depth)
154 if (a->se_stack.depth < b->se_stack.depth)
156 for (int i = 0; i < a->se_stack.depth; i++) {
157 if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
159 if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
166 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
167 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
169 static struct mtx epoch_stacks_lock;
170 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
172 static bool epoch_trace_stack_print = true;
173 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
174 &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
176 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
178 epoch_trace_report(const char *fmt, ...)
181 struct stackentry se, *new;
183 stack_zero(&se.se_stack); /* XXX: is it really needed? */
184 stack_save(&se.se_stack);
186 /* Tree is never reduced - go lockless. */
187 if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
190 new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
192 bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
194 mtx_lock(&epoch_stacks_lock);
195 new = RB_INSERT(stacktree, &epoch_stacks, new);
196 mtx_unlock(&epoch_stacks_lock);
202 (void)vprintf(fmt, ap);
204 if (epoch_trace_stack_print)
205 stack_print_ddb(&se.se_stack);
209 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
210 const char *file, int line)
214 SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
215 if (iet->et_epoch == epoch)
216 epoch_trace_report("Recursively entering epoch %s "
217 "at %s:%d, previously entered at %s:%d\n",
218 epoch->e_name, file, line,
219 iet->et_file, iet->et_line);
220 et->et_epoch = epoch;
223 SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
227 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
228 const char *file, int line)
231 if (SLIST_FIRST(&td->td_epochs) != et) {
232 epoch_trace_report("Exiting epoch %s in a not nested order "
233 "at %s:%d. Most recently entered %s at %s:%d\n",
236 SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
237 SLIST_FIRST(&td->td_epochs)->et_file,
238 SLIST_FIRST(&td->td_epochs)->et_line);
239 /* This will panic if et is not anywhere on td_epochs. */
240 SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
242 SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
245 /* Used by assertions that check thread state before going to sleep. */
247 epoch_trace_list(struct thread *td)
251 SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
252 printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
253 iet->et_file, iet->et_line);
255 #endif /* EPOCH_TRACE */
258 epoch_init(void *arg __unused)
262 block_count = counter_u64_alloc(M_WAITOK);
263 migrate_count = counter_u64_alloc(M_WAITOK);
264 turnstile_count = counter_u64_alloc(M_WAITOK);
265 switch_count = counter_u64_alloc(M_WAITOK);
266 epoch_call_count = counter_u64_alloc(M_WAITOK);
267 epoch_call_task_count = counter_u64_alloc(M_WAITOK);
269 pcpu_zone_record = uma_zcreate("epoch_record pcpu",
270 sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
271 UMA_ALIGN_PTR, UMA_ZONE_PCPU);
273 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
274 epoch_call_task, NULL);
275 taskqgroup_attach_cpu(qgroup_softirq,
276 DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
280 SLIST_INIT(&thread0.td_epochs);
283 global_epoch = epoch_alloc("Global", 0);
284 global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
286 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
288 #if !defined(EARLY_AP_STARTUP)
290 epoch_init_smp(void *dummy __unused)
294 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
298 epoch_ctor(epoch_t epoch)
303 epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
305 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
306 bzero(er, sizeof(*er));
307 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
308 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
310 er->er_parent = epoch;
315 epoch_adjust_prio(struct thread *td, u_char prio)
319 sched_prio(td, prio);
324 epoch_alloc(const char *name, int flags)
328 if (__predict_false(!inited))
329 panic("%s called too early in boot", __func__);
330 epoch = malloc(sizeof(struct epoch), M_EPOCH, M_ZERO | M_WAITOK);
331 ck_epoch_init(&epoch->e_epoch);
333 MPASS(epoch_count < MAX_EPOCHS - 2);
334 epoch->e_flags = flags;
335 epoch->e_idx = epoch_count;
336 epoch->e_name = name;
337 sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
338 mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
339 allepochs[epoch_count++] = epoch;
344 epoch_free(epoch_t epoch)
347 epoch_drain_callbacks(epoch);
348 allepochs[epoch->e_idx] = NULL;
349 epoch_wait(global_epoch);
350 uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
351 mtx_destroy(&epoch->e_drain_mtx);
352 sx_destroy(&epoch->e_drain_sx);
353 free(epoch, M_EPOCH);
356 static epoch_record_t
357 epoch_currecord(epoch_t epoch)
360 return (zpcpu_get_cpu(epoch->e_pcpu_record, curcpu));
363 #define INIT_CHECK(epoch) \
365 if (__predict_false((epoch) == NULL)) \
370 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
372 struct epoch_record *er;
375 MPASS(cold || epoch != NULL);
376 MPASS(epoch->e_flags & EPOCH_PREEMPT);
378 MPASS((vm_offset_t)et >= td->td_kstack &&
379 (vm_offset_t)et + sizeof(struct epoch_tracker) <=
380 td->td_kstack + td->td_kstack_pages * PAGE_SIZE);
384 epoch_trace_enter(td, epoch, et, file, line);
387 THREAD_NO_SLEEPING();
390 td->td_pre_epoch_prio = td->td_priority;
391 er = epoch_currecord(epoch);
392 TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
393 ck_epoch_begin(&er->er_record, &et->et_section);
398 epoch_enter(epoch_t epoch)
402 MPASS(cold || epoch != NULL);
405 er = epoch_currecord(epoch);
406 ck_epoch_begin(&er->er_record, NULL);
410 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
412 struct epoch_record *er;
419 THREAD_SLEEPING_OK();
420 er = epoch_currecord(epoch);
421 MPASS(epoch->e_flags & EPOCH_PREEMPT);
423 MPASS(et->et_td == td);
425 et->et_td = (void*)0xDEADBEEF;
427 ck_epoch_end(&er->er_record, &et->et_section);
428 TAILQ_REMOVE(&er->er_tdlist, et, et_link);
430 if (__predict_false(td->td_pre_epoch_prio != td->td_priority))
431 epoch_adjust_prio(td, td->td_pre_epoch_prio);
434 epoch_trace_exit(td, epoch, et, file, line);
439 epoch_exit(epoch_t epoch)
444 er = epoch_currecord(epoch);
445 ck_epoch_end(&er->er_record, NULL);
450 * epoch_block_handler_preempt() is a callback from the CK code when another
451 * thread is currently in an epoch section.
454 epoch_block_handler_preempt(struct ck_epoch *global __unused,
455 ck_epoch_record_t *cr, void *arg __unused)
457 epoch_record_t record;
458 struct thread *td, *owner, *curwaittd;
459 struct epoch_tracker *tdwait;
460 struct turnstile *ts;
461 struct lock_object *lock;
463 int locksheld __unused;
465 record = __containerof(cr, struct epoch_record, er_record);
467 locksheld = td->td_locks;
469 counter_u64_add(block_count, 1);
471 * We lost a race and there's no longer any threads
472 * on the CPU in an epoch section.
474 if (TAILQ_EMPTY(&record->er_tdlist))
477 if (record->er_cpuid != curcpu) {
479 * If the head of the list is running, we can wait for it
480 * to remove itself from the list and thus save us the
481 * overhead of a migration
483 gen = record->er_gen;
486 * We can't actually check if the waiting thread is running
487 * so we simply poll for it to exit before giving up and
492 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
493 gen == record->er_gen &&
494 spincount++ < MAX_ADAPTIVE_SPIN);
497 * If the generation has changed we can poll again
498 * otherwise we need to migrate.
500 if (gen != record->er_gen)
503 * Being on the same CPU as that of the record on which
504 * we need to wait allows us access to the thread
505 * list associated with that CPU. We can then examine the
506 * oldest thread in the queue and wait on its turnstile
507 * until it resumes and so on until a grace period
511 counter_u64_add(migrate_count, 1);
512 sched_bind(td, record->er_cpuid);
514 * At this point we need to return to the ck code
515 * to scan to see if a grace period has elapsed.
516 * We can't move on to check the thread list, because
517 * in the meantime new threads may have arrived that
518 * in fact belong to a different epoch.
523 * Try to find a thread in an epoch section on this CPU
524 * waiting on a turnstile. Otherwise find the lowest
525 * priority thread (highest prio value) and drop our priority
526 * to match to allow it to run.
528 TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
530 * Propagate our priority to any other waiters to prevent us
531 * from starving them. They will have their original priority
532 * restore on exit from epoch_wait().
534 curwaittd = tdwait->et_td;
535 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
538 thread_lock(curwaittd);
539 sched_prio(curwaittd, td->td_priority);
540 thread_unlock(curwaittd);
544 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
545 ((ts = curwaittd->td_blocked) != NULL)) {
547 * We unlock td to allow turnstile_wait to reacquire
548 * the thread lock. Before unlocking it we enter a
549 * critical section to prevent preemption after we
550 * reenable interrupts by dropping the thread lock in
551 * order to prevent curwaittd from getting to run.
556 if (turnstile_lock(ts, &lock, &owner)) {
557 if (ts == curwaittd->td_blocked) {
558 MPASS(TD_IS_INHIBITED(curwaittd) &&
559 TD_ON_LOCK(curwaittd));
561 turnstile_wait(ts, owner,
562 curwaittd->td_tsqueue);
563 counter_u64_add(turnstile_count, 1);
567 turnstile_unlock(ts, lock);
571 KASSERT(td->td_locks == locksheld,
572 ("%d extra locks held", td->td_locks - locksheld));
576 * We didn't find any threads actually blocked on a lock
577 * so we have nothing to do except context switch away.
579 counter_u64_add(switch_count, 1);
580 mi_switch(SW_VOL | SWT_RELINQUISH);
582 * It is important the thread lock is dropped while yielding
583 * to allow other threads to acquire the lock pointed to by
584 * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
585 * thread lock before returning. Else a deadlock like
586 * situation might happen.
592 epoch_wait_preempt(epoch_t epoch)
601 MPASS(cold || epoch != NULL);
605 locks = curthread->td_locks;
606 MPASS(epoch->e_flags & EPOCH_PREEMPT);
607 if ((epoch->e_flags & EPOCH_LOCKED) == 0)
608 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
609 "epoch_wait() can be long running");
610 KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
611 "of an epoch section of the same epoch"));
616 old_cpu = PCPU_GET(cpuid);
617 old_pinned = td->td_pinned;
618 old_prio = td->td_priority;
619 was_bound = sched_is_bound(td);
622 sched_bind(td, old_cpu);
624 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
627 /* restore CPU binding, if any */
628 if (was_bound != 0) {
629 sched_bind(td, old_cpu);
631 /* get thread back to initial CPU, if any */
633 sched_bind(td, old_cpu);
636 /* restore pinned after bind */
637 td->td_pinned = old_pinned;
639 /* restore thread priority */
640 sched_prio(td, old_prio);
643 KASSERT(td->td_locks == locks,
644 ("%d residual locks held", td->td_locks - locks));
648 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
655 epoch_wait(epoch_t epoch)
658 MPASS(cold || epoch != NULL);
660 MPASS(epoch->e_flags == 0);
662 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
667 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
670 ck_epoch_entry_t *cb;
675 /* too early in boot to have epoch set up */
676 if (__predict_false(epoch == NULL))
678 #if !defined(EARLY_AP_STARTUP)
679 if (__predict_false(inited < 2))
684 *DPCPU_PTR(epoch_cb_count) += 1;
685 er = epoch_currecord(epoch);
686 ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
694 epoch_call_task(void *arg __unused)
696 ck_stack_entry_t *cursor, *head, *next;
697 ck_epoch_record_t *record;
701 int i, npending, total;
703 ck_stack_init(&cb_stack);
705 epoch_enter(global_epoch);
706 for (total = i = 0; i < epoch_count; i++) {
707 if (__predict_false((epoch = allepochs[i]) == NULL))
709 er = epoch_currecord(epoch);
710 record = &er->er_record;
711 if ((npending = record->n_pending) == 0)
713 ck_epoch_poll_deferred(record, &cb_stack);
714 total += npending - record->n_pending;
716 epoch_exit(global_epoch);
717 *DPCPU_PTR(epoch_cb_count) -= total;
720 counter_u64_add(epoch_call_count, total);
721 counter_u64_add(epoch_call_task_count, 1);
723 head = ck_stack_batch_pop_npsc(&cb_stack);
724 for (cursor = head; cursor != NULL; cursor = next) {
725 struct ck_epoch_entry *entry =
726 ck_epoch_entry_container(cursor);
728 next = CK_STACK_NEXT(cursor);
729 entry->function(entry);
734 in_epoch_verbose(epoch_t epoch, int dump_onfail)
736 struct epoch_tracker *tdwait;
741 if (THREAD_CAN_SLEEP())
743 if (__predict_false((epoch) == NULL))
746 er = epoch_currecord(epoch);
747 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
748 if (tdwait->et_td == td) {
754 MPASS(td->td_pinned);
755 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
756 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
757 printf("td_tid: %d ", tdwait->et_td->td_tid);
766 in_epoch(epoch_t epoch)
768 return (in_epoch_verbose(epoch, 0));
772 epoch_drain_cb(struct epoch_context *ctx)
774 struct epoch *epoch =
775 __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
777 if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
778 mtx_lock(&epoch->e_drain_mtx);
780 mtx_unlock(&epoch->e_drain_mtx);
785 epoch_drain_callbacks(epoch_t epoch)
794 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
795 "epoch_drain_callbacks() may sleep!");
797 /* too early in boot to have epoch set up */
798 if (__predict_false(epoch == NULL))
800 #if !defined(EARLY_AP_STARTUP)
801 if (__predict_false(inited < 2))
806 sx_xlock(&epoch->e_drain_sx);
807 mtx_lock(&epoch->e_drain_mtx);
811 old_cpu = PCPU_GET(cpuid);
812 old_pinned = td->td_pinned;
813 was_bound = sched_is_bound(td);
818 epoch->e_drain_count++;
820 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
822 epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
825 /* restore CPU binding, if any */
826 if (was_bound != 0) {
827 sched_bind(td, old_cpu);
829 /* get thread back to initial CPU, if any */
831 sched_bind(td, old_cpu);
834 /* restore pinned after bind */
835 td->td_pinned = old_pinned;
839 while (epoch->e_drain_count != 0)
840 msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
842 mtx_unlock(&epoch->e_drain_mtx);
843 sx_xunlock(&epoch->e_drain_sx);