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
  53 #include <ck_epoch.h>
  54
  55 static MALLOC_DEFINE(M_EPOCH, "epoch", "epoch based reclamation");
  56
  57 /* arbitrary --- needs benchmarking */
  58 #define MAX_ADAPTIVE_SPIN 1000
  59 #define MAX_EPOCHS 64
  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 CTASSERT(sizeof(epoch_section_t) == sizeof(ck_epoch_section_t));
  68 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
  69 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW, 0, "epoch information");
  70 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW, 0, "epoch stats");
  71
  72
  73 /* Stats. */
  74 static counter_u64_t block_count;
  75 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
  76                                    &block_count, "# of times a thread was in an epoch when epoch_wait was called");
  77 static counter_u64_t migrate_count;
  78 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
  79                                    &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
  80 static counter_u64_t turnstile_count;
  81 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
  82                                    &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
  83 static counter_u64_t switch_count;
  84 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
  85                                    &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
  86 static counter_u64_t epoch_call_count;
  87 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
  88                                    &epoch_call_count, "# of times a callback was deferred");
  89 static counter_u64_t epoch_call_task_count;
  90 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
  91                                    &epoch_call_task_count, "# of times a callback task was run");
  92
  93 TAILQ_HEAD(threadlist, thread);
  94
  95 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
  96     ck_epoch_entry_container)
  97
  98 typedef struct epoch_record {
  99         ck_epoch_record_t er_record;
 100         volatile struct threadlist er_tdlist;
 101         volatile uint32_t er_gen;
 102         uint32_t er_cpuid;
 103 } *epoch_record_t;
 104
 105 struct epoch_pcpu_state {
 106         struct epoch_record eps_record;
 107 } __aligned(EPOCH_ALIGN);
 108
 109 struct epoch {
 110         struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
 111         struct epoch_pcpu_state *e_pcpu_dom[MAXMEMDOM] __aligned(EPOCH_ALIGN);
 112         int e_idx;
 113         int e_flags;
 114         struct epoch_pcpu_state *e_pcpu[0];
 115 };
 116
 117 epoch_t allepochs[MAX_EPOCHS];
 118
 119 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
 120 DPCPU_DEFINE(int, epoch_cb_count);
 121
 122 static __read_mostly int domcount[MAXMEMDOM];
 123 static __read_mostly int domoffsets[MAXMEMDOM];
 124 static __read_mostly int inited;
 125 static __read_mostly int epoch_count;
 126 __read_mostly epoch_t global_epoch;
 127 __read_mostly epoch_t global_epoch_preempt;
 128
 129 static void epoch_call_task(void *context __unused);
 130
 131 #if defined(__powerpc64__) || defined(__powerpc__) || !defined(NUMA)
 132 static bool usedomains = false;
 133 #else
 134 static bool usedomains = true;
 135 #endif
 136 static void
 137 epoch_init(void *arg __unused)
 138 {
 139         int domain, cpu;
 140
 141         block_count = counter_u64_alloc(M_WAITOK);
 142         migrate_count = counter_u64_alloc(M_WAITOK);
 143         turnstile_count = counter_u64_alloc(M_WAITOK);
 144         switch_count = counter_u64_alloc(M_WAITOK);
 145         epoch_call_count = counter_u64_alloc(M_WAITOK);
 146         epoch_call_task_count = counter_u64_alloc(M_WAITOK);
 147         if (usedomains == false)
 148                 goto done;
 149         domain = 0;
 150         domoffsets[0] = 0;
 151         for (domain = 0; domain < vm_ndomains; domain++) {
 152                 domcount[domain] = CPU_COUNT(&cpuset_domain[domain]);
 153                 if (bootverbose)
 154                         printf("domcount[%d] %d\n", domain, domcount[domain]);
 155         }
 156         for (domain = 1; domain < vm_ndomains; domain++)
 157                 domoffsets[domain] = domoffsets[domain-1] + domcount[domain-1];
 158
 159         for (domain = 0; domain < vm_ndomains; domain++) {
 160                 if (domcount[domain] == 0) {
 161                         usedomains = false;
 162                         break;
 163                 }
 164         }
 165  done:
 166         CPU_FOREACH(cpu) {
 167                 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0, epoch_call_task, NULL);
 168                 taskqgroup_attach_cpu(qgroup_softirq, DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, -1, "epoch call task");
 169         }
 170         inited = 1;
 171         global_epoch = epoch_alloc(0);
 172         global_epoch_preempt = epoch_alloc(EPOCH_PREEMPT);
 173 }
 174 SYSINIT(epoch, SI_SUB_TASKQ + 1, SI_ORDER_FIRST, epoch_init, NULL);
 175
 176 static void
 177 epoch_init_numa(epoch_t epoch)
 178 {
 179         int domain, cpu_offset;
 180         struct epoch_pcpu_state *eps;
 181         epoch_record_t er;
 182
 183         for (domain = 0; domain < vm_ndomains; domain++) {
 184                 eps = malloc_domain(sizeof(*eps)*domcount[domain], M_EPOCH,
 185                                                         domain, M_ZERO|M_WAITOK);
 186                 epoch->e_pcpu_dom[domain] = eps;
 187                 cpu_offset = domoffsets[domain];
 188                 for (int i = 0; i < domcount[domain]; i++, eps++) {
 189                         epoch->e_pcpu[cpu_offset + i] = eps;
 190                         er = &eps->eps_record;
 191                         ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 192                         TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 193                         er->er_cpuid = cpu_offset + i;
 194                 }
 195         }
 196 }
 197
 198 static void
 199 epoch_init_legacy(epoch_t epoch)
 200 {
 201         struct epoch_pcpu_state *eps;
 202         epoch_record_t er;
 203
 204         eps = malloc(sizeof(*eps)*mp_ncpus, M_EPOCH, M_ZERO|M_WAITOK);
 205         epoch->e_pcpu_dom[0] = eps;
 206         for (int i = 0; i < mp_ncpus; i++, eps++) {
 207                 epoch->e_pcpu[i] = eps;
 208                 er = &eps->eps_record;
 209                 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 210                 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 211                 er->er_cpuid = i;
 212         }
 213 }
 214
 215 epoch_t
 216 epoch_alloc(int flags)
 217 {
 218         epoch_t epoch;
 219
 220         if (__predict_false(!inited))
 221                 panic("%s called too early in boot", __func__);
 222         epoch = malloc(sizeof(struct epoch) + mp_ncpus*sizeof(void*),
 223                                    M_EPOCH, M_ZERO|M_WAITOK);
 224         ck_epoch_init(&epoch->e_epoch);
 225         if (usedomains)
 226                 epoch_init_numa(epoch);
 227         else
 228                 epoch_init_legacy(epoch);
 229         MPASS(epoch_count < MAX_EPOCHS-2);
 230         epoch->e_flags = flags;
 231         epoch->e_idx = epoch_count;
 232         allepochs[epoch_count++] = epoch;
 233         return (epoch);
 234 }
 235
 236 void
 237 epoch_free(epoch_t epoch)
 238 {
 239         int domain;
 240 #ifdef INVARIANTS
 241         struct epoch_pcpu_state *eps;
 242         int cpu;
 243
 244         CPU_FOREACH(cpu) {
 245                 eps = epoch->e_pcpu[cpu];
 246                 MPASS(TAILQ_EMPTY(&eps->eps_record.er_tdlist));
 247         }
 248 #endif
 249         allepochs[epoch->e_idx] = NULL;
 250         epoch_wait(global_epoch);
 251         if (usedomains)
 252                 for (domain = 0; domain < vm_ndomains; domain++)
 253                         free_domain(epoch->e_pcpu_dom[domain], M_EPOCH);
 254         else
 255                 free(epoch->e_pcpu_dom[0], M_EPOCH);
 256         free(epoch, M_EPOCH);
 257 }
 258
 259 #define INIT_CHECK(epoch)                                                               \
 260         do {                                                                                    \
 261                 if (__predict_false((epoch) == NULL))           \
 262                         return;                                                                 \
 263         } while (0)
 264
 265 void
 266 epoch_enter_preempt_internal(epoch_t epoch, struct thread *td)
 267 {
 268         struct epoch_pcpu_state *eps;
 269
 270         MPASS(cold || epoch != NULL);
 271         INIT_CHECK(epoch);
 272         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 273         critical_enter();
 274         td->td_pre_epoch_prio = td->td_priority;
 275         eps = epoch->e_pcpu[curcpu];
 276 #ifdef INVARIANTS
 277         MPASS(td->td_epochnest < UCHAR_MAX - 2);
 278         if (td->td_epochnest > 1) {
 279                 struct thread *curtd;
 280                 int found = 0;
 281
 282                 TAILQ_FOREACH(curtd, &eps->eps_record.er_tdlist, td_epochq)
 283                         if (curtd == td)
 284                                 found = 1;
 285                 KASSERT(found, ("recursing on a second epoch"));
 286                 critical_exit();
 287                 return;
 288         }
 289 #endif
 290         TAILQ_INSERT_TAIL(&eps->eps_record.er_tdlist, td, td_epochq);
 291         sched_pin();
 292         ck_epoch_begin(&eps->eps_record.er_record, (ck_epoch_section_t*)&td->td_epoch_section);
 293         critical_exit();
 294 }
 295
 296
 297 void
 298 epoch_enter(epoch_t epoch)
 299 {
 300         ck_epoch_record_t *record;
 301         struct thread *td;
 302
 303         MPASS(cold || epoch != NULL);
 304         td = curthread;
 305
 306         critical_enter();
 307         td->td_epochnest++;
 308         record = &epoch->e_pcpu[curcpu]->eps_record.er_record;
 309         ck_epoch_begin(record, NULL);
 310 }
 311
 312 void
 313 epoch_exit_preempt_internal(epoch_t epoch, struct thread *td)
 314 {
 315         struct epoch_pcpu_state *eps;
 316
 317         MPASS(td->td_epochnest == 0);
 318         INIT_CHECK(epoch);
 319         critical_enter();
 320         eps = epoch->e_pcpu[curcpu];
 321
 322         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 323         ck_epoch_end(&eps->eps_record.er_record, (ck_epoch_section_t*)&td->td_epoch_section);
 324         TAILQ_REMOVE(&eps->eps_record.er_tdlist, td, td_epochq);
 325         eps->eps_record.er_gen++;
 326         sched_unpin();
 327         if (__predict_false(td->td_pre_epoch_prio != td->td_priority)) {
 328                 thread_lock(td);
 329                 sched_prio(td, td->td_pre_epoch_prio);
 330                 thread_unlock(td);
 331         }
 332         critical_exit();
 333 }
 334
 335 void
 336 epoch_exit(epoch_t epoch)
 337 {
 338         ck_epoch_record_t *record;
 339         struct thread *td;
 340
 341         td = curthread;
 342         td->td_epochnest--;
 343         record = &epoch->e_pcpu[curcpu]->eps_record.er_record;
 344         ck_epoch_end(record, NULL);
 345         critical_exit();
 346 }
 347
 348 /*
 349  * epoch_block_handler_preempt is a callback from the ck code when another thread is
 350  * currently in an epoch section.
 351  */
 352 static void
 353 epoch_block_handler_preempt(struct ck_epoch *global __unused, ck_epoch_record_t *cr,
 354                                         void *arg __unused)
 355 {
 356         epoch_record_t record;
 357         struct thread *td, *tdwait, *owner;
 358         struct turnstile *ts;
 359         struct lock_object *lock;
 360         int spincount, gen;
 361
 362         record = __containerof(cr, struct epoch_record, er_record);
 363         td = curthread;
 364         spincount = 0;
 365         counter_u64_add(block_count, 1);
 366         if (record->er_cpuid != curcpu) {
 367                 /*
 368                  * If the head of the list is running, we can wait for it
 369                  * to remove itself from the list and thus save us the
 370                  * overhead of a migration
 371                  */
 372                 if ((tdwait = TAILQ_FIRST(&record->er_tdlist)) != NULL &&
 373                         TD_IS_RUNNING(tdwait)) {
 374                         gen = record->er_gen;
 375                         thread_unlock(td);
 376                         do {
 377                                 cpu_spinwait();
 378                         } while (tdwait == TAILQ_FIRST(&record->er_tdlist) &&
 379                                          gen == record->er_gen && TD_IS_RUNNING(tdwait) &&
 380                                          spincount++ < MAX_ADAPTIVE_SPIN);
 381                         thread_lock(td);
 382                         return;
 383                 }
 384
 385                 /*
 386                  * Being on the same CPU as that of the record on which
 387                  * we need to wait allows us access to the thread
 388                  * list associated with that CPU. We can then examine the
 389                  * oldest thread in the queue and wait on its turnstile
 390                  * until it resumes and so on until a grace period
 391                  * elapses.
 392                  *
 393                  */
 394                 counter_u64_add(migrate_count, 1);
 395                 sched_bind(td, record->er_cpuid);
 396                 /*
 397                  * At this point we need to return to the ck code
 398                  * to scan to see if a grace period has elapsed.
 399                  * We can't move on to check the thread list, because
 400                  * in the meantime new threads may have arrived that
 401                  * in fact belong to a different epoch.
 402                  */
 403                 return;
 404         }
 405         /*
 406          * Try to find a thread in an epoch section on this CPU
 407          * waiting on a turnstile. Otherwise find the lowest
 408          * priority thread (highest prio value) and drop our priority
 409          * to match to allow it to run.
 410          */
 411         TAILQ_FOREACH(tdwait, &record->er_tdlist, td_epochq) {
 412                 /*
 413                  * Propagate our priority to any other waiters to prevent us
 414                  * from starving them. They will have their original priority
 415                  * restore on exit from epoch_wait().
 416                  */
 417                 if (!TD_IS_INHIBITED(tdwait) && tdwait->td_priority > td->td_priority) {
 418                         critical_enter();
 419                         thread_unlock(td);
 420                         thread_lock(tdwait);
 421                         sched_prio(tdwait, td->td_priority);
 422                         thread_unlock(tdwait);
 423                         thread_lock(td);
 424                         critical_exit();
 425                 }
 426                 if (TD_IS_INHIBITED(tdwait) && TD_ON_LOCK(tdwait) &&
 427                         ((ts = tdwait->td_blocked) != NULL)) {
 428                         /*
 429                          * We unlock td to allow turnstile_wait to reacquire the
 430                          * the thread lock. Before unlocking it we enter a critical
 431                          * section to prevent preemption after we reenable interrupts
 432                          * by dropping the thread lock in order to prevent tdwait
 433                          * 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. Only
 440                          * in case we hold the lock and the turnstile we locked is still
 441                          * the one that tdwait is blocked on can we continue. Otherwise
 442                          * The turnstile pointer has been changed out from underneath
 443                          * us, as in the case where the lock holder has signalled tdwait,
 444                          * and we need to continue.
 445                          */
 446                         if (owner != NULL && ts == tdwait->td_blocked) {
 447                                 MPASS(TD_IS_INHIBITED(tdwait) && TD_ON_LOCK(tdwait));
 448                                 critical_exit();
 449                                 turnstile_wait(ts, owner, tdwait->td_tsqueue);
 450                                 counter_u64_add(turnstile_count, 1);
 451                                 thread_lock(td);
 452                                 return;
 453                         } else if (owner != NULL)
 454                                 turnstile_unlock(ts, lock);
 455                         thread_lock(td);
 456                         critical_exit();
 457                         KASSERT(td->td_locks == 0,
 458                                         ("%d locks held", td->td_locks));
 459                 }
 460         }
 461         /*
 462          * We didn't find any threads actually blocked on a lock
 463          * so we have nothing to do except context switch away.
 464          */
 465         counter_u64_add(switch_count, 1);
 466         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 467
 468         /*
 469          * Release the thread lock while yielding to
 470          * allow other threads to acquire the lock
 471          * pointed to by TDQ_LOCKPTR(td). Else a
 472          * deadlock like situation might happen. (HPS)
 473          */
 474         thread_unlock(td);
 475         thread_lock(td);
 476 }
 477
 478 void
 479 epoch_wait_preempt(epoch_t epoch)
 480 {
 481         struct thread *td;
 482         int was_bound;
 483         int old_cpu;
 484         int old_pinned;
 485         u_char old_prio;
 486 #ifdef INVARIANTS
 487         int locks;
 488
 489         locks = curthread->td_locks;
 490 #endif
 491
 492         MPASS(cold || epoch != NULL);
 493         INIT_CHECK(epoch);
 494
 495         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 496         WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 497             "epoch_wait() can sleep");
 498
 499         td = curthread;
 500         KASSERT(td->td_epochnest == 0, ("epoch_wait() in the middle of an epoch section"));
 501         thread_lock(td);
 502
 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, NULL);
 514
 515         /* restore CPU binding, if any */
 516         if (was_bound != 0) {
 517                 sched_bind(td, old_cpu);
 518         } else {
 519                 /* get thread back to initial CPU, if any */
 520                 if (old_pinned != 0)
 521                         sched_bind(td, old_cpu);
 522                 sched_unbind(td);
 523         }
 524         /* restore pinned after bind */
 525         td->td_pinned = old_pinned;
 526
 527         /* restore thread priority */
 528         sched_prio(td, old_prio);
 529         thread_unlock(td);
 530         PICKUP_GIANT();
 531         KASSERT(td->td_locks == locks,
 532                         ("%d residual locks held", td->td_locks - locks));
 533 }
 534
 535 static void
 536 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
 537                                         void *arg __unused)
 538 {
 539         cpu_spinwait();
 540 }
 541
 542 void
 543 epoch_wait(epoch_t epoch)
 544 {
 545
 546         MPASS(cold || epoch != NULL);
 547         INIT_CHECK(epoch);
 548         MPASS(epoch->e_flags == 0);
 549         critical_enter();
 550         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
 551         critical_exit();
 552 }
 553
 554 void
 555 epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t))
 556 {
 557         struct epoch_pcpu_state *eps;
 558         ck_epoch_entry_t *cb;
 559
 560         cb = (void *)ctx;
 561
 562         MPASS(callback);
 563         /* too early in boot to have epoch set up */
 564         if (__predict_false(epoch == NULL))
 565                 goto boottime;
 566
 567         critical_enter();
 568         *DPCPU_PTR(epoch_cb_count) += 1;
 569         eps = epoch->e_pcpu[curcpu];
 570         ck_epoch_call(&eps->eps_record.er_record, cb, (ck_epoch_cb_t*)callback);
 571         critical_exit();
 572         return;
 573  boottime:
 574         callback(ctx);
 575 }
 576
 577 static void
 578 epoch_call_task(void *arg __unused)
 579 {
 580         ck_stack_entry_t *cursor, *head, *next;
 581         ck_epoch_record_t *record;
 582         epoch_t epoch;
 583         ck_stack_t cb_stack;
 584         int i, npending, total;
 585
 586         ck_stack_init(&cb_stack);
 587         critical_enter();
 588         epoch_enter(global_epoch);
 589         for (total = i = 0; i < epoch_count; i++) {
 590                 if (__predict_false((epoch = allepochs[i]) == NULL))
 591                         continue;
 592                 record = &epoch->e_pcpu[curcpu]->eps_record.er_record;
 593                 if ((npending = record->n_pending) == 0)
 594                         continue;
 595                 ck_epoch_poll_deferred(record, &cb_stack);
 596                 total += npending - record->n_pending;
 597         }
 598         epoch_exit(global_epoch);
 599         *DPCPU_PTR(epoch_cb_count) -= total;
 600         critical_exit();
 601
 602         counter_u64_add(epoch_call_count, total);
 603         counter_u64_add(epoch_call_task_count, 1);
 604
 605         head = ck_stack_batch_pop_npsc(&cb_stack);
 606         for (cursor = head; cursor != NULL; cursor = next) {
 607                 struct ck_epoch_entry *entry =
 608                     ck_epoch_entry_container(cursor);
 609                 next = CK_STACK_NEXT(cursor);
 610                 entry->function(entry);
 611         }
 612 }
 613
 614 int
 615 in_epoch(void)
 616 {
 617         return (curthread->td_epochnest != 0);
 618 }