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/sx.h>
  47 #include <sys/smp.h>
  48 #include <sys/sysctl.h>
  49 #include <sys/turnstile.h>
  50 #include <vm/vm.h>
  51 #include <vm/vm_extern.h>
  52 #include <vm/vm_kern.h>
  53 #include <vm/uma.h>
  54
  55 #include <ck_epoch.h>
  56
  57 #ifdef __amd64__
  58 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
  59 #else
  60 #define EPOCH_ALIGN CACHE_LINE_SIZE
  61 #endif
  62
  63 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
  64 typedef struct epoch_record {
  65         ck_epoch_record_t er_record;
  66         volatile struct epoch_tdlist er_tdlist;
  67         volatile uint32_t er_gen;
  68         uint32_t er_cpuid;
  69         /* fields above are part of KBI and cannot be modified */
  70         struct epoch_context er_drain_ctx;
  71         struct epoch *er_parent;
  72 } __aligned(EPOCH_ALIGN)     *epoch_record_t;
  73
  74 struct epoch {
  75         struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
  76         epoch_record_t e_pcpu_record;
  77         int     e_in_use;
  78         int     e_flags;
  79         /* fields above are part of KBI and cannot be modified */
  80         struct sx e_drain_sx;
  81         struct mtx e_drain_mtx;
  82         volatile int e_drain_count;
  83 };
  84
  85 /* arbitrary --- needs benchmarking */
  86 #define MAX_ADAPTIVE_SPIN 100
  87 #define MAX_EPOCHS 64
  88
  89 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
  90 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW, 0, "epoch information");
  91 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW, 0, "epoch stats");
  92
  93 /* Stats. */
  94 static counter_u64_t block_count;
  95
  96 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
  97     &block_count, "# of times a thread was in an epoch when epoch_wait was called");
  98 static counter_u64_t migrate_count;
  99
 100 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
 101     &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
 102 static counter_u64_t turnstile_count;
 103
 104 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
 105     &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
 106 static counter_u64_t switch_count;
 107
 108 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
 109     &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
 110 static counter_u64_t epoch_call_count;
 111
 112 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
 113     &epoch_call_count, "# of times a callback was deferred");
 114 static counter_u64_t epoch_call_task_count;
 115
 116 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
 117     &epoch_call_task_count, "# of times a callback task was run");
 118
 119 TAILQ_HEAD (threadlist, thread);
 120
 121 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
 122     ck_epoch_entry_container)
 123
 124 static struct epoch epoch_array[MAX_EPOCHS];
 125
 126 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
 127 DPCPU_DEFINE(int, epoch_cb_count);
 128
 129 static __read_mostly int inited;
 130 __read_mostly epoch_t global_epoch;
 131 __read_mostly epoch_t global_epoch_preempt;
 132
 133 static void epoch_call_task(void *context __unused);
 134 static  uma_zone_t pcpu_zone_record;
 135
 136 static struct sx epoch_sx;
 137
 138 #define EPOCH_LOCK() sx_xlock(&epoch_sx)
 139 #define EPOCH_UNLOCK() sx_xunlock(&epoch_sx)
 140
 141 static void
 142 epoch_init(void *arg __unused)
 143 {
 144         int cpu;
 145
 146         block_count = counter_u64_alloc(M_WAITOK);
 147         migrate_count = counter_u64_alloc(M_WAITOK);
 148         turnstile_count = counter_u64_alloc(M_WAITOK);
 149         switch_count = counter_u64_alloc(M_WAITOK);
 150         epoch_call_count = counter_u64_alloc(M_WAITOK);
 151         epoch_call_task_count = counter_u64_alloc(M_WAITOK);
 152
 153         pcpu_zone_record = uma_zcreate("epoch_record pcpu",
 154             sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
 155             UMA_ALIGN_PTR, UMA_ZONE_PCPU);
 156         CPU_FOREACH(cpu) {
 157                 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
 158                     epoch_call_task, NULL);
 159                 taskqgroup_attach_cpu(qgroup_softirq,
 160                     DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, -1,
 161                     "epoch call task");
 162         }
 163         sx_init(&epoch_sx, "epoch-sx");
 164         inited = 1;
 165         global_epoch = epoch_alloc(0);
 166         global_epoch_preempt = epoch_alloc(EPOCH_PREEMPT);
 167 }
 168 SYSINIT(epoch, SI_SUB_TASKQ + 1, SI_ORDER_FIRST, epoch_init, NULL);
 169
 170 #if !defined(EARLY_AP_STARTUP)
 171 static void
 172 epoch_init_smp(void *dummy __unused)
 173 {
 174         inited = 2;
 175 }
 176 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
 177 #endif
 178
 179 static void
 180 epoch_ctor(epoch_t epoch)
 181 {
 182         epoch_record_t er;
 183         int cpu;
 184
 185         epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
 186         CPU_FOREACH(cpu) {
 187                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 188                 bzero(er, sizeof(*er));
 189                 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 190                 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 191                 er->er_cpuid = cpu;
 192                 er->er_parent = epoch;
 193         }
 194 }
 195
 196 static void
 197 epoch_adjust_prio(struct thread *td, u_char prio)
 198 {
 199
 200         thread_lock(td);
 201         sched_prio(td, prio);
 202         thread_unlock(td);
 203 }
 204
 205 epoch_t
 206 epoch_alloc(int flags)
 207 {
 208         epoch_t epoch;
 209         int i;
 210
 211         if (__predict_false(!inited))
 212                 panic("%s called too early in boot", __func__);
 213
 214         EPOCH_LOCK();
 215
 216         /*
 217          * Find a free index in the epoch array. If no free index is
 218          * found, try to use the index after the last one.
 219          */
 220         for (i = 0;; i++) {
 221                 /*
 222                  * If too many epochs are currently allocated,
 223                  * return NULL.
 224                  */
 225                 if (i == MAX_EPOCHS) {
 226                         epoch = NULL;
 227                         goto done;
 228                 }
 229                 if (epoch_array[i].e_in_use == 0)
 230                         break;
 231         }
 232
 233         epoch = epoch_array + i;
 234         ck_epoch_init(&epoch->e_epoch);
 235         epoch_ctor(epoch);
 236         epoch->e_flags = flags;
 237         sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
 238         mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
 239
 240         /*
 241          * Set e_in_use last, because when this field is set the
 242          * epoch_call_task() function will start scanning this epoch
 243          * structure.
 244          */
 245         atomic_store_rel_int(&epoch->e_in_use, 1);
 246 done:
 247         EPOCH_UNLOCK();
 248         return (epoch);
 249 }
 250
 251 void
 252 epoch_free(epoch_t epoch)
 253 {
 254
 255         EPOCH_LOCK();
 256
 257         MPASS(epoch->e_in_use != 0);
 258
 259         epoch_drain_callbacks(epoch);
 260
 261         atomic_store_rel_int(&epoch->e_in_use, 0);
 262         /*
 263          * Make sure the epoch_call_task() function see e_in_use equal
 264          * to zero, by calling epoch_wait() on the global_epoch:
 265          */
 266         epoch_wait(global_epoch);
 267         uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
 268         mtx_destroy(&epoch->e_drain_mtx);
 269         sx_destroy(&epoch->e_drain_sx);
 270         memset(epoch, 0, sizeof(*epoch));
 271
 272         EPOCH_UNLOCK();
 273 }
 274
 275 static epoch_record_t
 276 epoch_currecord(epoch_t epoch)
 277 {
 278
 279         return (zpcpu_get_cpu(epoch->e_pcpu_record, curcpu));
 280 }
 281
 282 #define INIT_CHECK(epoch)                                       \
 283         do {                                                    \
 284                 if (__predict_false((epoch) == NULL))           \
 285                         return;                                 \
 286         } while (0)
 287
 288 void
 289 epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et)
 290 {
 291         struct epoch_record *er;
 292         struct thread *td;
 293
 294         MPASS(cold || epoch != NULL);
 295         INIT_CHECK(epoch);
 296         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 297 #ifdef EPOCH_TRACKER_DEBUG
 298         et->et_magic_pre = EPOCH_MAGIC0;
 299         et->et_magic_post = EPOCH_MAGIC1;
 300 #endif
 301         td = curthread;
 302         et->et_td = td;
 303         td->td_epochnest++;
 304         critical_enter();
 305         sched_pin();
 306
 307         td->td_pre_epoch_prio = td->td_priority;
 308         er = epoch_currecord(epoch);
 309         TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
 310         ck_epoch_begin(&er->er_record, &et->et_section);
 311         critical_exit();
 312 }
 313
 314 void
 315 epoch_enter(epoch_t epoch)
 316 {
 317         struct thread *td;
 318         epoch_record_t er;
 319
 320         MPASS(cold || epoch != NULL);
 321         INIT_CHECK(epoch);
 322         td = curthread;
 323
 324         td->td_epochnest++;
 325         critical_enter();
 326         er = epoch_currecord(epoch);
 327         ck_epoch_begin(&er->er_record, NULL);
 328 }
 329
 330 void
 331 epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et)
 332 {
 333         struct epoch_record *er;
 334         struct thread *td;
 335
 336         INIT_CHECK(epoch);
 337         td = curthread;
 338         critical_enter();
 339         sched_unpin();
 340         MPASS(td->td_epochnest);
 341         td->td_epochnest--;
 342         er = epoch_currecord(epoch);
 343         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 344         MPASS(et != NULL);
 345         MPASS(et->et_td == td);
 346 #ifdef EPOCH_TRACKER_DEBUG
 347         MPASS(et->et_magic_pre == EPOCH_MAGIC0);
 348         MPASS(et->et_magic_post == EPOCH_MAGIC1);
 349         et->et_magic_pre = 0;
 350         et->et_magic_post = 0;
 351 #endif
 352 #ifdef INVARIANTS
 353         et->et_td = (void*)0xDEADBEEF;
 354 #endif
 355         ck_epoch_end(&er->er_record, &et->et_section);
 356         TAILQ_REMOVE(&er->er_tdlist, et, et_link);
 357         er->er_gen++;
 358         if (__predict_false(td->td_pre_epoch_prio != td->td_priority))
 359                 epoch_adjust_prio(td, td->td_pre_epoch_prio);
 360         critical_exit();
 361 }
 362
 363 void
 364 epoch_exit(epoch_t epoch)
 365 {
 366         struct thread *td;
 367         epoch_record_t er;
 368
 369         INIT_CHECK(epoch);
 370         td = curthread;
 371         MPASS(td->td_epochnest);
 372         td->td_epochnest--;
 373         er = epoch_currecord(epoch);
 374         ck_epoch_end(&er->er_record, NULL);
 375         critical_exit();
 376 }
 377
 378 /*
 379  * epoch_block_handler_preempt() is a callback from the CK code when another
 380  * thread is currently in an epoch section.
 381  */
 382 static void
 383 epoch_block_handler_preempt(struct ck_epoch *global __unused,
 384     ck_epoch_record_t *cr, void *arg __unused)
 385 {
 386         epoch_record_t record;
 387         struct thread *td, *owner, *curwaittd;
 388         struct epoch_tracker *tdwait;
 389         struct turnstile *ts;
 390         struct lock_object *lock;
 391         int spincount, gen;
 392         int locksheld __unused;
 393
 394         record = __containerof(cr, struct epoch_record, er_record);
 395         td = curthread;
 396         locksheld = td->td_locks;
 397         spincount = 0;
 398         counter_u64_add(block_count, 1);
 399         /*
 400          * We lost a race and there's no longer any threads
 401          * on the CPU in an epoch section.
 402          */
 403         if (TAILQ_EMPTY(&record->er_tdlist))
 404                 return;
 405
 406         if (record->er_cpuid != curcpu) {
 407                 /*
 408                  * If the head of the list is running, we can wait for it
 409                  * to remove itself from the list and thus save us the
 410                  * overhead of a migration
 411                  */
 412                 gen = record->er_gen;
 413                 thread_unlock(td);
 414                 /*
 415                  * We can't actually check if the waiting thread is running
 416                  * so we simply poll for it to exit before giving up and
 417                  * migrating.
 418                  */
 419                 do {
 420                         cpu_spinwait();
 421                 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
 422                                  gen == record->er_gen &&
 423                                  spincount++ < MAX_ADAPTIVE_SPIN);
 424                 thread_lock(td);
 425                 /*
 426                  * If the generation has changed we can poll again
 427                  * otherwise we need to migrate.
 428                  */
 429                 if (gen != record->er_gen)
 430                         return;
 431                 /*
 432                  * Being on the same CPU as that of the record on which
 433                  * we need to wait allows us access to the thread
 434                  * list associated with that CPU. We can then examine the
 435                  * oldest thread in the queue and wait on its turnstile
 436                  * until it resumes and so on until a grace period
 437                  * elapses.
 438                  *
 439                  */
 440                 counter_u64_add(migrate_count, 1);
 441                 sched_bind(td, record->er_cpuid);
 442                 /*
 443                  * At this point we need to return to the ck code
 444                  * to scan to see if a grace period has elapsed.
 445                  * We can't move on to check the thread list, because
 446                  * in the meantime new threads may have arrived that
 447                  * in fact belong to a different epoch.
 448                  */
 449                 return;
 450         }
 451         /*
 452          * Try to find a thread in an epoch section on this CPU
 453          * waiting on a turnstile. Otherwise find the lowest
 454          * priority thread (highest prio value) and drop our priority
 455          * to match to allow it to run.
 456          */
 457         TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
 458                 /*
 459                  * Propagate our priority to any other waiters to prevent us
 460                  * from starving them. They will have their original priority
 461                  * restore on exit from epoch_wait().
 462                  */
 463                 curwaittd = tdwait->et_td;
 464                 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
 465                         critical_enter();
 466                         thread_unlock(td);
 467                         thread_lock(curwaittd);
 468                         sched_prio(curwaittd, td->td_priority);
 469                         thread_unlock(curwaittd);
 470                         thread_lock(td);
 471                         critical_exit();
 472                 }
 473                 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
 474                     ((ts = curwaittd->td_blocked) != NULL)) {
 475                         /*
 476                          * We unlock td to allow turnstile_wait to reacquire
 477                          * the thread lock. Before unlocking it we enter a
 478                          * critical section to prevent preemption after we
 479                          * reenable interrupts by dropping the thread lock in
 480                          * order to prevent curwaittd from getting to run.
 481                          */
 482                         critical_enter();
 483                         thread_unlock(td);
 484
 485                         if (turnstile_lock(ts, &lock, &owner)) {
 486                                 if (ts == curwaittd->td_blocked) {
 487                                         MPASS(TD_IS_INHIBITED(curwaittd) &&
 488                                             TD_ON_LOCK(curwaittd));
 489                                         critical_exit();
 490                                         turnstile_wait(ts, owner,
 491                                             curwaittd->td_tsqueue);
 492                                         counter_u64_add(turnstile_count, 1);
 493                                         thread_lock(td);
 494                                         return;
 495                                 }
 496                                 turnstile_unlock(ts, lock);
 497                         }
 498                         thread_lock(td);
 499                         critical_exit();
 500                         KASSERT(td->td_locks == locksheld,
 501                             ("%d extra locks held", td->td_locks - locksheld));
 502                 }
 503         }
 504         /*
 505          * We didn't find any threads actually blocked on a lock
 506          * so we have nothing to do except context switch away.
 507          */
 508         counter_u64_add(switch_count, 1);
 509         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 510
 511         /*
 512          * Release the thread lock while yielding to
 513          * allow other threads to acquire the lock
 514          * pointed to by TDQ_LOCKPTR(td). Else a
 515          * deadlock like situation might happen. (HPS)
 516          */
 517         thread_unlock(td);
 518         thread_lock(td);
 519 }
 520
 521 void
 522 epoch_wait_preempt(epoch_t epoch)
 523 {
 524         struct thread *td;
 525         int was_bound;
 526         int old_cpu;
 527         int old_pinned;
 528         u_char old_prio;
 529         int locks __unused;
 530
 531         MPASS(cold || epoch != NULL);
 532         INIT_CHECK(epoch);
 533         td = curthread;
 534 #ifdef INVARIANTS
 535         locks = curthread->td_locks;
 536         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 537         if ((epoch->e_flags & EPOCH_LOCKED) == 0)
 538                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 539                     "epoch_wait() can be long running");
 540         KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
 541             "of an epoch section of the same epoch"));
 542 #endif
 543         DROP_GIANT();
 544         thread_lock(td);
 545
 546         old_cpu = PCPU_GET(cpuid);
 547         old_pinned = td->td_pinned;
 548         old_prio = td->td_priority;
 549         was_bound = sched_is_bound(td);
 550         sched_unbind(td);
 551         td->td_pinned = 0;
 552         sched_bind(td, old_cpu);
 553
 554         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
 555             NULL);
 556
 557         /* restore CPU binding, if any */
 558         if (was_bound != 0) {
 559                 sched_bind(td, old_cpu);
 560         } else {
 561                 /* get thread back to initial CPU, if any */
 562                 if (old_pinned != 0)
 563                         sched_bind(td, old_cpu);
 564                 sched_unbind(td);
 565         }
 566         /* restore pinned after bind */
 567         td->td_pinned = old_pinned;
 568
 569         /* restore thread priority */
 570         sched_prio(td, old_prio);
 571         thread_unlock(td);
 572         PICKUP_GIANT();
 573         KASSERT(td->td_locks == locks,
 574             ("%d residual locks held", td->td_locks - locks));
 575 }
 576
 577 static void
 578 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
 579     void *arg __unused)
 580 {
 581         cpu_spinwait();
 582 }
 583
 584 void
 585 epoch_wait(epoch_t epoch)
 586 {
 587
 588         MPASS(cold || epoch != NULL);
 589         INIT_CHECK(epoch);
 590         MPASS(epoch->e_flags == 0);
 591         critical_enter();
 592         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
 593         critical_exit();
 594 }
 595
 596 void
 597 epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t))
 598 {
 599         epoch_record_t er;
 600         ck_epoch_entry_t *cb;
 601
 602         cb = (void *)ctx;
 603
 604         MPASS(callback);
 605         /* too early in boot to have epoch set up */
 606         if (__predict_false(epoch == NULL))
 607                 goto boottime;
 608 #if !defined(EARLY_AP_STARTUP)
 609         if (__predict_false(inited < 2))
 610                 goto boottime;
 611 #endif
 612
 613         critical_enter();
 614         *DPCPU_PTR(epoch_cb_count) += 1;
 615         er = epoch_currecord(epoch);
 616         ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
 617         critical_exit();
 618         return;
 619 boottime:
 620         callback(ctx);
 621 }
 622
 623 static void
 624 epoch_call_task(void *arg __unused)
 625 {
 626         ck_stack_entry_t *cursor, *head, *next;
 627         ck_epoch_record_t *record;
 628         epoch_record_t er;
 629         epoch_t epoch;
 630         ck_stack_t cb_stack;
 631         int i, npending, total;
 632
 633         ck_stack_init(&cb_stack);
 634         critical_enter();
 635         epoch_enter(global_epoch);
 636         for (total = i = 0; i != MAX_EPOCHS; i++) {
 637                 epoch = epoch_array + i;
 638                 if (__predict_false(
 639                     atomic_load_acq_int(&epoch->e_in_use) == 0))
 640                         continue;
 641                 er = epoch_currecord(epoch);
 642                 record = &er->er_record;
 643                 if ((npending = record->n_pending) == 0)
 644                         continue;
 645                 ck_epoch_poll_deferred(record, &cb_stack);
 646                 total += npending - record->n_pending;
 647         }
 648         epoch_exit(global_epoch);
 649         *DPCPU_PTR(epoch_cb_count) -= total;
 650         critical_exit();
 651
 652         counter_u64_add(epoch_call_count, total);
 653         counter_u64_add(epoch_call_task_count, 1);
 654
 655         head = ck_stack_batch_pop_npsc(&cb_stack);
 656         for (cursor = head; cursor != NULL; cursor = next) {
 657                 struct ck_epoch_entry *entry =
 658                     ck_epoch_entry_container(cursor);
 659
 660                 next = CK_STACK_NEXT(cursor);
 661                 entry->function(entry);
 662         }
 663 }
 664
 665 int
 666 in_epoch_verbose(epoch_t epoch, int dump_onfail)
 667 {
 668         struct epoch_tracker *tdwait;
 669         struct thread *td;
 670         epoch_record_t er;
 671
 672         td = curthread;
 673         if (td->td_epochnest == 0)
 674                 return (0);
 675         if (__predict_false((epoch) == NULL))
 676                 return (0);
 677         critical_enter();
 678         er = epoch_currecord(epoch);
 679         TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 680                 if (tdwait->et_td == td) {
 681                         critical_exit();
 682                         return (1);
 683                 }
 684 #ifdef INVARIANTS
 685         if (dump_onfail) {
 686                 MPASS(td->td_pinned);
 687                 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
 688                 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 689                         printf("td_tid: %d ", tdwait->et_td->td_tid);
 690                 printf("\n");
 691         }
 692 #endif
 693         critical_exit();
 694         return (0);
 695 }
 696
 697 int
 698 in_epoch(epoch_t epoch)
 699 {
 700         return (in_epoch_verbose(epoch, 0));
 701 }
 702
 703 static void
 704 epoch_drain_cb(struct epoch_context *ctx)
 705 {
 706         struct epoch *epoch =
 707             __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
 708
 709         if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
 710                 mtx_lock(&epoch->e_drain_mtx);
 711                 wakeup(epoch);
 712                 mtx_unlock(&epoch->e_drain_mtx);
 713         }
 714 }
 715
 716 void
 717 epoch_drain_callbacks(epoch_t epoch)
 718 {
 719         epoch_record_t er;
 720         struct thread *td;
 721         int was_bound;
 722         int old_pinned;
 723         int old_cpu;
 724         int cpu;
 725
 726         WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 727             "epoch_drain_callbacks() may sleep!");
 728
 729         /* too early in boot to have epoch set up */
 730         if (__predict_false(epoch == NULL))
 731                 return;
 732 #if !defined(EARLY_AP_STARTUP)
 733         if (__predict_false(inited < 2))
 734                 return;
 735 #endif
 736         DROP_GIANT();
 737
 738         sx_xlock(&epoch->e_drain_sx);
 739         mtx_lock(&epoch->e_drain_mtx);
 740
 741         td = curthread;
 742         thread_lock(td);
 743         old_cpu = PCPU_GET(cpuid);
 744         old_pinned = td->td_pinned;
 745         was_bound = sched_is_bound(td);
 746         sched_unbind(td);
 747         td->td_pinned = 0;
 748
 749         CPU_FOREACH(cpu)
 750                 epoch->e_drain_count++;
 751         CPU_FOREACH(cpu) {
 752                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 753                 sched_bind(td, cpu);
 754                 epoch_call(epoch, &er->er_drain_ctx, &epoch_drain_cb);
 755         }
 756
 757         /* restore CPU binding, if any */
 758         if (was_bound != 0) {
 759                 sched_bind(td, old_cpu);
 760         } else {
 761                 /* get thread back to initial CPU, if any */
 762                 if (old_pinned != 0)
 763                         sched_bind(td, old_cpu);
 764                 sched_unbind(td);
 765         }
 766         /* restore pinned after bind */
 767         td->td_pinned = old_pinned;
 768
 769         thread_unlock(td);
 770
 771         while (epoch->e_drain_count != 0)
 772                 msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
 773
 774         mtx_unlock(&epoch->e_drain_mtx);
 775         sx_xunlock(&epoch->e_drain_sx);
 776
 777         PICKUP_GIANT();
 778 }
 779
 780 /* for binary compatibility */
 781
 782 struct epoch_tracker_KBI {
 783         void *datap[3];
 784 #ifdef EPOCH_TRACKER_DEBUG
 785         int datai[5];
 786 #else
 787         int datai[1];
 788 #endif
 789 } __aligned(sizeof(void *));
 790
 791 CTASSERT(sizeof(struct epoch_tracker_KBI) >= sizeof(struct epoch_tracker));
 792
 793 void
 794 epoch_enter_preempt_KBI(epoch_t epoch, epoch_tracker_t et)
 795 {
 796         epoch_enter_preempt(epoch, et);
 797 }
 798
 799 void
 800 epoch_exit_preempt_KBI(epoch_t epoch, epoch_tracker_t et)
 801 {
 802         epoch_exit_preempt(epoch, et);
 803 }
 804
 805 void
 806 epoch_enter_KBI(epoch_t epoch)
 807 {
 808         epoch_enter(epoch);
 809 }
 810
 811 void
 812 epoch_exit_KBI(epoch_t epoch)
 813 {
 814         epoch_exit(epoch);
 815 }