sys/kern/subr_epoch.c

   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/types.h>
  34 #include <sys/systm.h>
  35 #include <sys/counter.h>
  36 #include <sys/epoch.h>
  37 #include <sys/gtaskqueue.h>
  38 #include <sys/kernel.h>
  39 #include <sys/limits.h>
  40 #include <sys/lock.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mutex.h>
  43 #include <sys/pcpu.h>
  44 #include <sys/proc.h>
  45 #include <sys/sched.h>
  46 #include <sys/smp.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/turnstile.h>
  49 #include <vm/vm.h>
  50 #include <vm/vm_extern.h>
  51 #include <vm/vm_kern.h>
  52 #include <vm/uma.h>
  53
  54 #include <ck_epoch.h>
  55
  56 static MALLOC_DEFINE(M_EPOCH, "epoch", "epoch based reclamation");
  57
  58 #ifdef __amd64__
  59 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
  60 #else
  61 #define EPOCH_ALIGN CACHE_LINE_SIZE
  62 #endif
  63
  64 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
  65 typedef struct epoch_record {
  66         ck_epoch_record_t er_record;
  67         volatile struct epoch_tdlist er_tdlist;
  68         volatile uint32_t er_gen;
  69         uint32_t er_cpuid;
  70 } __aligned(EPOCH_ALIGN)     *epoch_record_t;
  71
  72 struct epoch {
  73         struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
  74         epoch_record_t e_pcpu_record;
  75         int     e_idx;
  76         int     e_flags;
  77 };
  78
  79 /* arbitrary --- needs benchmarking */
  80 #define MAX_ADAPTIVE_SPIN 100
  81 #define MAX_EPOCHS 64
  82
  83 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
  84 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW, 0, "epoch information");
  85 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW, 0, "epoch stats");
  86
  87 /* Stats. */
  88 static counter_u64_t block_count;
  89
  90 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
  91     &block_count, "# of times a thread was in an epoch when epoch_wait was called");
  92 static counter_u64_t migrate_count;
  93
  94 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
  95     &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
  96 static counter_u64_t turnstile_count;
  97
  98 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
  99     &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
 100 static counter_u64_t switch_count;
 101
 102 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
 103     &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
 104 static counter_u64_t epoch_call_count;
 105
 106 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
 107     &epoch_call_count, "# of times a callback was deferred");
 108 static counter_u64_t epoch_call_task_count;
 109
 110 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
 111     &epoch_call_task_count, "# of times a callback task was run");
 112
 113 TAILQ_HEAD (threadlist, thread);
 114
 115 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
 116     ck_epoch_entry_container)
 117
 118 epoch_t allepochs[MAX_EPOCHS];
 119
 120 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
 121 DPCPU_DEFINE(int, epoch_cb_count);
 122
 123 static __read_mostly int inited;
 124 static __read_mostly int epoch_count;
 125 __read_mostly epoch_t global_epoch;
 126 __read_mostly epoch_t global_epoch_preempt;
 127
 128 static void epoch_call_task(void *context __unused);
 129 static  uma_zone_t pcpu_zone_record;
 130
 131 static void
 132 epoch_init(void *arg __unused)
 133 {
 134         int cpu;
 135
 136         block_count = counter_u64_alloc(M_WAITOK);
 137         migrate_count = counter_u64_alloc(M_WAITOK);
 138         turnstile_count = counter_u64_alloc(M_WAITOK);
 139         switch_count = counter_u64_alloc(M_WAITOK);
 140         epoch_call_count = counter_u64_alloc(M_WAITOK);
 141         epoch_call_task_count = counter_u64_alloc(M_WAITOK);
 142
 143         pcpu_zone_record = uma_zcreate("epoch_record pcpu",
 144             sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
 145             UMA_ALIGN_PTR, UMA_ZONE_PCPU);
 146         CPU_FOREACH(cpu) {
 147                 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
 148                     epoch_call_task, NULL);
 149                 taskqgroup_attach_cpu(qgroup_softirq,
 150                     DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
 151                     "epoch call task");
 152         }
 153         inited = 1;
 154         global_epoch = epoch_alloc(0);
 155         global_epoch_preempt = epoch_alloc(EPOCH_PREEMPT);
 156 }
 157 SYSINIT(epoch, SI_SUB_TASKQ + 1, SI_ORDER_FIRST, epoch_init, NULL);
 158
 159 #if !defined(EARLY_AP_STARTUP)
 160 static void
 161 epoch_init_smp(void *dummy __unused)
 162 {
 163         inited = 2;
 164 }
 165 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
 166 #endif
 167
 168 static void
 169 epoch_ctor(epoch_t epoch)
 170 {
 171         epoch_record_t er;
 172         int cpu;
 173
 174         epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
 175         CPU_FOREACH(cpu) {
 176                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 177                 bzero(er, sizeof(*er));
 178                 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 179                 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 180                 er->er_cpuid = cpu;
 181         }
 182 }
 183
 184 static void
 185 epoch_adjust_prio(struct thread *td, u_char prio)
 186 {
 187
 188         thread_lock(td);
 189         sched_prio(td, prio);
 190         thread_unlock(td);
 191 }
 192
 193 epoch_t
 194 epoch_alloc(int flags)
 195 {
 196         epoch_t epoch;
 197
 198         if (__predict_false(!inited))
 199                 panic("%s called too early in boot", __func__);
 200         epoch = malloc(sizeof(struct epoch), M_EPOCH, M_ZERO | M_WAITOK);
 201         ck_epoch_init(&epoch->e_epoch);
 202         epoch_ctor(epoch);
 203         MPASS(epoch_count < MAX_EPOCHS - 2);
 204         epoch->e_flags = flags;
 205         epoch->e_idx = epoch_count;
 206         allepochs[epoch_count++] = epoch;
 207         return (epoch);
 208 }
 209
 210 void
 211 epoch_free(epoch_t epoch)
 212 {
 213 #ifdef INVARIANTS
 214         struct epoch_record *er;
 215         int cpu;
 216
 217         CPU_FOREACH(cpu) {
 218                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 219                 MPASS(TAILQ_EMPTY(&er->er_tdlist));
 220         }
 221 #endif
 222         allepochs[epoch->e_idx] = NULL;
 223         epoch_wait(global_epoch);
 224         uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
 225         free(epoch, M_EPOCH);
 226 }
 227
 228 static epoch_record_t
 229 epoch_currecord(epoch_t epoch)
 230 {
 231
 232         return (zpcpu_get_cpu(epoch->e_pcpu_record, curcpu));
 233 }
 234
 235 #define INIT_CHECK(epoch)                                       \
 236         do {                                                    \
 237                 if (__predict_false((epoch) == NULL))           \
 238                         return;                                 \
 239         } while (0)
 240
 241 void
 242 epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et)
 243 {
 244         struct epoch_record *er;
 245         struct thread *td;
 246
 247         MPASS(cold || epoch != NULL);
 248         INIT_CHECK(epoch);
 249         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 250 #ifdef EPOCH_TRACKER_DEBUG
 251         et->et_magic_pre = EPOCH_MAGIC0;
 252         et->et_magic_post = EPOCH_MAGIC1;
 253 #endif
 254         td = curthread;
 255         et->et_td = td;
 256         td->td_epochnest++;
 257         critical_enter();
 258         sched_pin();
 259
 260         td->td_pre_epoch_prio = td->td_priority;
 261         er = epoch_currecord(epoch);
 262         TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
 263         ck_epoch_begin(&er->er_record, &et->et_section);
 264         critical_exit();
 265 }
 266
 267 void
 268 epoch_enter(epoch_t epoch)
 269 {
 270         struct thread *td;
 271         epoch_record_t er;
 272
 273         MPASS(cold || epoch != NULL);
 274         INIT_CHECK(epoch);
 275         td = curthread;
 276
 277         td->td_epochnest++;
 278         critical_enter();
 279         er = epoch_currecord(epoch);
 280         ck_epoch_begin(&er->er_record, NULL);
 281 }
 282
 283 void
 284 epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et)
 285 {
 286         struct epoch_record *er;
 287         struct thread *td;
 288
 289         INIT_CHECK(epoch);
 290         td = curthread;
 291         critical_enter();
 292         sched_unpin();
 293         MPASS(td->td_epochnest);
 294         td->td_epochnest--;
 295         er = epoch_currecord(epoch);
 296         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 297         MPASS(et != NULL);
 298         MPASS(et->et_td == td);
 299 #ifdef EPOCH_TRACKER_DEBUG
 300         MPASS(et->et_magic_pre == EPOCH_MAGIC0);
 301         MPASS(et->et_magic_post == EPOCH_MAGIC1);
 302         et->et_magic_pre = 0;
 303         et->et_magic_post = 0;
 304 #endif
 305 #ifdef INVARIANTS
 306         et->et_td = (void*)0xDEADBEEF;
 307 #endif
 308         ck_epoch_end(&er->er_record, &et->et_section);
 309         TAILQ_REMOVE(&er->er_tdlist, et, et_link);
 310         er->er_gen++;
 311         if (__predict_false(td->td_pre_epoch_prio != td->td_priority))
 312                 epoch_adjust_prio(td, td->td_pre_epoch_prio);
 313         critical_exit();
 314 }
 315
 316 void
 317 epoch_exit(epoch_t epoch)
 318 {
 319         struct thread *td;
 320         epoch_record_t er;
 321
 322         INIT_CHECK(epoch);
 323         td = curthread;
 324         MPASS(td->td_epochnest);
 325         td->td_epochnest--;
 326         er = epoch_currecord(epoch);
 327         ck_epoch_end(&er->er_record, NULL);
 328         critical_exit();
 329 }
 330
 331 /*
 332  * epoch_block_handler_preempt() is a callback from the CK code when another
 333  * thread is currently in an epoch section.
 334  */
 335 static void
 336 epoch_block_handler_preempt(struct ck_epoch *global __unused,
 337     ck_epoch_record_t *cr, void *arg __unused)
 338 {
 339         epoch_record_t record;
 340         struct thread *td, *owner, *curwaittd;
 341         struct epoch_tracker *tdwait;
 342         struct turnstile *ts;
 343         struct lock_object *lock;
 344         int spincount, gen;
 345         int locksheld __unused;
 346
 347         record = __containerof(cr, struct epoch_record, er_record);
 348         td = curthread;
 349         locksheld = td->td_locks;
 350         spincount = 0;
 351         counter_u64_add(block_count, 1);
 352         /*
 353          * We lost a race and there's no longer any threads
 354          * on the CPU in an epoch section.
 355          */
 356         if (TAILQ_EMPTY(&record->er_tdlist))
 357                 return;
 358
 359         if (record->er_cpuid != curcpu) {
 360                 /*
 361                  * If the head of the list is running, we can wait for it
 362                  * to remove itself from the list and thus save us the
 363                  * overhead of a migration
 364                  */
 365                 gen = record->er_gen;
 366                 thread_unlock(td);
 367                 /*
 368                  * We can't actually check if the waiting thread is running
 369                  * so we simply poll for it to exit before giving up and
 370                  * migrating.
 371                  */
 372                 do {
 373                         cpu_spinwait();
 374                 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
 375                                  gen == record->er_gen &&
 376                                  spincount++ < MAX_ADAPTIVE_SPIN);
 377                 thread_lock(td);
 378                 /*
 379                  * If the generation has changed we can poll again
 380                  * otherwise we need to migrate.
 381                  */
 382                 if (gen != record->er_gen)
 383                         return;
 384                 /*
 385                  * Being on the same CPU as that of the record on which
 386                  * we need to wait allows us access to the thread
 387                  * list associated with that CPU. We can then examine the
 388                  * oldest thread in the queue and wait on its turnstile
 389                  * until it resumes and so on until a grace period
 390                  * elapses.
 391                  *
 392                  */
 393                 counter_u64_add(migrate_count, 1);
 394                 sched_bind(td, record->er_cpuid);
 395                 /*
 396                  * At this point we need to return to the ck code
 397                  * to scan to see if a grace period has elapsed.
 398                  * We can't move on to check the thread list, because
 399                  * in the meantime new threads may have arrived that
 400                  * in fact belong to a different epoch.
 401                  */
 402                 return;
 403         }
 404         /*
 405          * Try to find a thread in an epoch section on this CPU
 406          * waiting on a turnstile. Otherwise find the lowest
 407          * priority thread (highest prio value) and drop our priority
 408          * to match to allow it to run.
 409          */
 410         TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
 411                 /*
 412                  * Propagate our priority to any other waiters to prevent us
 413                  * from starving them. They will have their original priority
 414                  * restore on exit from epoch_wait().
 415                  */
 416                 curwaittd = tdwait->et_td;
 417                 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
 418                         critical_enter();
 419                         thread_unlock(td);
 420                         thread_lock(curwaittd);
 421                         sched_prio(curwaittd, td->td_priority);
 422                         thread_unlock(curwaittd);
 423                         thread_lock(td);
 424                         critical_exit();
 425                 }
 426                 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
 427                     ((ts = curwaittd->td_blocked) != NULL)) {
 428                         /*
 429                          * We unlock td to allow turnstile_wait to reacquire
 430                          * the thread lock. Before unlocking it we enter a
 431                          * critical section to prevent preemption after we
 432                          * reenable interrupts by dropping the thread lock in
 433                          * order to prevent curwaittd from getting to run.
 434                          */
 435                         critical_enter();
 436                         thread_unlock(td);
 437                         owner = turnstile_lock(ts, &lock);
 438                         /*
 439                          * The owner pointer indicates that the lock succeeded.
 440                          * Only in case we hold the lock and the turnstile we
 441                          * locked is still the one that curwaittd is blocked on
 442                          * can we continue. Otherwise the turnstile pointer has
 443                          * been changed out from underneath us, as in the case
 444                          * where the lock holder has signalled curwaittd,
 445                          * and we need to continue.
 446                          */
 447                         if (owner != NULL && ts == curwaittd->td_blocked) {
 448                                 MPASS(TD_IS_INHIBITED(curwaittd) &&
 449                                     TD_ON_LOCK(curwaittd));
 450                                 critical_exit();
 451                                 turnstile_wait(ts, owner, curwaittd->td_tsqueue);
 452                                 counter_u64_add(turnstile_count, 1);
 453                                 thread_lock(td);
 454                                 return;
 455                         } else if (owner != NULL)
 456                                 turnstile_unlock(ts, lock);
 457                         thread_lock(td);
 458                         critical_exit();
 459                         KASSERT(td->td_locks == locksheld,
 460                             ("%d extra locks held", td->td_locks - locksheld));
 461                 }
 462         }
 463         /*
 464          * We didn't find any threads actually blocked on a lock
 465          * so we have nothing to do except context switch away.
 466          */
 467         counter_u64_add(switch_count, 1);
 468         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 469
 470         /*
 471          * Release the thread lock while yielding to
 472          * allow other threads to acquire the lock
 473          * pointed to by TDQ_LOCKPTR(td). Else a
 474          * deadlock like situation might happen. (HPS)
 475          */
 476         thread_unlock(td);
 477         thread_lock(td);
 478 }
 479
 480 void
 481 epoch_wait_preempt(epoch_t epoch)
 482 {
 483         struct thread *td;
 484         int was_bound;
 485         int old_cpu;
 486         int old_pinned;
 487         u_char old_prio;
 488         int locks __unused;
 489
 490         MPASS(cold || epoch != NULL);
 491         INIT_CHECK(epoch);
 492         td = curthread;
 493 #ifdef INVARIANTS
 494         locks = curthread->td_locks;
 495         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 496         if ((epoch->e_flags & EPOCH_LOCKED) == 0)
 497                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 498                     "epoch_wait() can be long running");
 499         KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
 500             "of an epoch section of the same epoch"));
 501 #endif
 502         thread_lock(td);
 503         DROP_GIANT();
 504
 505         old_cpu = PCPU_GET(cpuid);
 506         old_pinned = td->td_pinned;
 507         old_prio = td->td_priority;
 508         was_bound = sched_is_bound(td);
 509         sched_unbind(td);
 510         td->td_pinned = 0;
 511         sched_bind(td, old_cpu);
 512
 513         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
 514             NULL);
 515
 516         /* restore CPU binding, if any */
 517         if (was_bound != 0) {
 518                 sched_bind(td, old_cpu);
 519         } else {
 520                 /* get thread back to initial CPU, if any */
 521                 if (old_pinned != 0)
 522                         sched_bind(td, old_cpu);
 523                 sched_unbind(td);
 524         }
 525         /* restore pinned after bind */
 526         td->td_pinned = old_pinned;
 527
 528         /* restore thread priority */
 529         sched_prio(td, old_prio);
 530         thread_unlock(td);
 531         PICKUP_GIANT();
 532         KASSERT(td->td_locks == locks,
 533             ("%d residual locks held", td->td_locks - locks));
 534 }
 535
 536 static void
 537 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
 538     void *arg __unused)
 539 {
 540         cpu_spinwait();
 541 }
 542
 543 void
 544 epoch_wait(epoch_t epoch)
 545 {
 546
 547         MPASS(cold || epoch != NULL);
 548         INIT_CHECK(epoch);
 549         MPASS(epoch->e_flags == 0);
 550         critical_enter();
 551         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
 552         critical_exit();
 553 }
 554
 555 void
 556 epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t))
 557 {
 558         epoch_record_t er;
 559         ck_epoch_entry_t *cb;
 560
 561         cb = (void *)ctx;
 562
 563         MPASS(callback);
 564         /* too early in boot to have epoch set up */
 565         if (__predict_false(epoch == NULL))
 566                 goto boottime;
 567 #if !defined(EARLY_AP_STARTUP)
 568         if (__predict_false(inited < 2))
 569                 goto boottime;
 570 #endif
 571
 572         critical_enter();
 573         *DPCPU_PTR(epoch_cb_count) += 1;
 574         er = epoch_currecord(epoch);
 575         ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
 576         critical_exit();
 577         return;
 578 boottime:
 579         callback(ctx);
 580 }
 581
 582 static void
 583 epoch_call_task(void *arg __unused)
 584 {
 585         ck_stack_entry_t *cursor, *head, *next;
 586         ck_epoch_record_t *record;
 587         epoch_record_t er;
 588         epoch_t epoch;
 589         ck_stack_t cb_stack;
 590         int i, npending, total;
 591
 592         ck_stack_init(&cb_stack);
 593         critical_enter();
 594         epoch_enter(global_epoch);
 595         for (total = i = 0; i < epoch_count; i++) {
 596                 if (__predict_false((epoch = allepochs[i]) == NULL))
 597                         continue;
 598                 er = epoch_currecord(epoch);
 599                 record = &er->er_record;
 600                 if ((npending = record->n_pending) == 0)
 601                         continue;
 602                 ck_epoch_poll_deferred(record, &cb_stack);
 603                 total += npending - record->n_pending;
 604         }
 605         epoch_exit(global_epoch);
 606         *DPCPU_PTR(epoch_cb_count) -= total;
 607         critical_exit();
 608
 609         counter_u64_add(epoch_call_count, total);
 610         counter_u64_add(epoch_call_task_count, 1);
 611
 612         head = ck_stack_batch_pop_npsc(&cb_stack);
 613         for (cursor = head; cursor != NULL; cursor = next) {
 614                 struct ck_epoch_entry *entry =
 615                     ck_epoch_entry_container(cursor);
 616
 617                 next = CK_STACK_NEXT(cursor);
 618                 entry->function(entry);
 619         }
 620 }
 621
 622 int
 623 in_epoch_verbose(epoch_t epoch, int dump_onfail)
 624 {
 625         struct epoch_tracker *tdwait;
 626         struct thread *td;
 627         epoch_record_t er;
 628
 629         td = curthread;
 630         if (td->td_epochnest == 0)
 631                 return (0);
 632         if (__predict_false((epoch) == NULL))
 633                 return (0);
 634         critical_enter();
 635         er = epoch_currecord(epoch);
 636         TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 637                 if (tdwait->et_td == td) {
 638                         critical_exit();
 639                         return (1);
 640                 }
 641 #ifdef INVARIANTS
 642         if (dump_onfail) {
 643                 MPASS(td->td_pinned);
 644                 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
 645                 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 646                         printf("td_tid: %d ", tdwait->et_td->td_tid);
 647                 printf("\n");
 648         }
 649 #endif
 650         critical_exit();
 651         return (0);
 652 }
 653
 654 int
 655 in_epoch(epoch_t epoch)
 656 {
 657         return (in_epoch_verbose(epoch, 0));
 658 }
 659
 660 void
 661 epoch_thread_init(struct thread *td)
 662 {
 663
 664         td->td_et = malloc(sizeof(struct epoch_tracker), M_EPOCH, M_WAITOK);
 665 }
 666
 667 void
 668 epoch_thread_fini(struct thread *td)
 669 {
 670
 671         free(td->td_et, M_EPOCH);
 672 }