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/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>
  39 #include <sys/lock.h>
  40 #include <sys/malloc.h>
  41 #include <sys/mutex.h>
  42 #include <sys/pcpu.h>
  43 #include <sys/proc.h>
  44 #include <sys/sched.h>
  45 #include <sys/sx.h>
  46 #include <sys/smp.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/turnstile.h>
  49 #ifdef EPOCH_TRACE
  50 #include <machine/stdarg.h>
  51 #include <sys/stack.h>
  52 #include <sys/tree.h>
  53 #endif
  54 #include <vm/vm.h>
  55 #include <vm/vm_extern.h>
  56 #include <vm/vm_kern.h>
  57 #include <vm/uma.h>
  58
  59 #include <ck_epoch.h>
  60
  61 static MALLOC_DEFINE(M_EPOCH, "epoch", "epoch based reclamation");
  62
  63 #ifdef __amd64__
  64 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
  65 #else
  66 #define EPOCH_ALIGN CACHE_LINE_SIZE
  67 #endif
  68
  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;
  76         uint32_t er_cpuid;
  77 } __aligned(EPOCH_ALIGN)     *epoch_record_t;
  78
  79 struct epoch {
  80         struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
  81         epoch_record_t e_pcpu_record;
  82         int     e_idx;
  83         int     e_flags;
  84         struct sx e_drain_sx;
  85         struct mtx e_drain_mtx;
  86         volatile int e_drain_count;
  87         const char *e_name;
  88 };
  89
  90 /* arbitrary --- needs benchmarking */
  91 #define MAX_ADAPTIVE_SPIN 100
  92 #define MAX_EPOCHS 64
  93
  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");
  97
  98 /* Stats. */
  99 static counter_u64_t block_count;
 100
 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;
 104
 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;
 108
 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;
 112
 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;
 116
 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;
 120
 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");
 123
 124 TAILQ_HEAD (threadlist, thread);
 125
 126 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
 127     ck_epoch_entry_container)
 128
 129 epoch_t allepochs[MAX_EPOCHS];
 130
 131 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
 132 DPCPU_DEFINE(int, epoch_cb_count);
 133
 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;
 138
 139 static void epoch_call_task(void *context __unused);
 140 static  uma_zone_t pcpu_zone_record;
 141
 142 #ifdef EPOCH_TRACE
 143 struct stackentry {
 144         RB_ENTRY(stackentry) se_node;
 145         struct stack se_stack;
 146 };
 147
 148 static int
 149 stackentry_compare(struct stackentry *a, struct stackentry *b)
 150 {
 151
 152         if (a->se_stack.depth > b->se_stack.depth)
 153                 return (1);
 154         if (a->se_stack.depth < b->se_stack.depth)
 155                 return (-1);
 156         for (int i = 0; i < a->se_stack.depth; i++) {
 157                 if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
 158                         return (1);
 159                 if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
 160                         return (-1);
 161         }
 162
 163         return (0);
 164 }
 165
 166 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
 167 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
 168
 169 static struct mtx epoch_stacks_lock;
 170 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
 171
 172 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
 173 static inline void
 174 epoch_trace_report(const char *fmt, ...)
 175 {
 176         va_list ap;
 177         struct stackentry se, *new;
 178
 179         stack_zero(&se.se_stack);       /* XXX: is it really needed? */
 180         stack_save(&se.se_stack);
 181
 182         /* Tree is never reduced - go lockless. */
 183         if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
 184                 return;
 185
 186         new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
 187         if (new != NULL) {
 188                 bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
 189
 190                 mtx_lock(&epoch_stacks_lock);
 191                 new = RB_INSERT(stacktree, &epoch_stacks, new);
 192                 mtx_unlock(&epoch_stacks_lock);
 193                 if (new != NULL)
 194                         free(new, M_STACK);
 195         }
 196
 197         va_start(ap, fmt);
 198         (void)vprintf(fmt, ap);
 199         va_end(ap);
 200         stack_print_ddb(&se.se_stack);
 201 }
 202
 203 static inline void
 204 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
 205     const char *file, int line)
 206 {
 207         epoch_tracker_t iet;
 208
 209         SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
 210                 if (iet->et_epoch == epoch)
 211                         epoch_trace_report("Recursively entering epoch %s "
 212                             "previously entered at %s:%d\n",
 213                             epoch->e_name, iet->et_file, iet->et_line);
 214         et->et_epoch = epoch;
 215         et->et_file = file;
 216         et->et_line = line;
 217         SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
 218 }
 219
 220 static inline void
 221 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
 222     const char *file, int line)
 223 {
 224
 225         if (SLIST_FIRST(&td->td_epochs) != et) {
 226                 epoch_trace_report("Exiting epoch %s in a not nested order. "
 227                     "Most recently entered %s at %s:%d\n",
 228                     epoch->e_name,
 229                     SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
 230                     SLIST_FIRST(&td->td_epochs)->et_file,
 231                     SLIST_FIRST(&td->td_epochs)->et_line);
 232                 /* This will panic if et is not anywhere on td_epochs. */
 233                 SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
 234         } else
 235                 SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
 236 }
 237 #endif /* EPOCH_TRACE */
 238
 239 static void
 240 epoch_init(void *arg __unused)
 241 {
 242         int cpu;
 243
 244         block_count = counter_u64_alloc(M_WAITOK);
 245         migrate_count = counter_u64_alloc(M_WAITOK);
 246         turnstile_count = counter_u64_alloc(M_WAITOK);
 247         switch_count = counter_u64_alloc(M_WAITOK);
 248         epoch_call_count = counter_u64_alloc(M_WAITOK);
 249         epoch_call_task_count = counter_u64_alloc(M_WAITOK);
 250
 251         pcpu_zone_record = uma_zcreate("epoch_record pcpu",
 252             sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
 253             UMA_ALIGN_PTR, UMA_ZONE_PCPU);
 254         CPU_FOREACH(cpu) {
 255                 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
 256                     epoch_call_task, NULL);
 257                 taskqgroup_attach_cpu(qgroup_softirq,
 258                     DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
 259                     "epoch call task");
 260         }
 261         SLIST_INIT(&thread0.td_epochs);
 262         inited = 1;
 263         global_epoch = epoch_alloc("Global", 0);
 264         global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
 265 }
 266 SYSINIT(epoch, SI_SUB_TASKQ + 1, SI_ORDER_FIRST, epoch_init, NULL);
 267
 268 #if !defined(EARLY_AP_STARTUP)
 269 static void
 270 epoch_init_smp(void *dummy __unused)
 271 {
 272         inited = 2;
 273 }
 274 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
 275 #endif
 276
 277 static void
 278 epoch_ctor(epoch_t epoch)
 279 {
 280         epoch_record_t er;
 281         int cpu;
 282
 283         epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
 284         CPU_FOREACH(cpu) {
 285                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 286                 bzero(er, sizeof(*er));
 287                 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 288                 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 289                 er->er_cpuid = cpu;
 290                 er->er_parent = epoch;
 291         }
 292 }
 293
 294 static void
 295 epoch_adjust_prio(struct thread *td, u_char prio)
 296 {
 297
 298         thread_lock(td);
 299         sched_prio(td, prio);
 300         thread_unlock(td);
 301 }
 302
 303 epoch_t
 304 epoch_alloc(const char *name, int flags)
 305 {
 306         epoch_t epoch;
 307
 308         if (__predict_false(!inited))
 309                 panic("%s called too early in boot", __func__);
 310         epoch = malloc(sizeof(struct epoch), M_EPOCH, M_ZERO | M_WAITOK);
 311         ck_epoch_init(&epoch->e_epoch);
 312         epoch_ctor(epoch);
 313         MPASS(epoch_count < MAX_EPOCHS - 2);
 314         epoch->e_flags = flags;
 315         epoch->e_idx = epoch_count;
 316         epoch->e_name = name;
 317         sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
 318         mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
 319         allepochs[epoch_count++] = epoch;
 320         return (epoch);
 321 }
 322
 323 void
 324 epoch_free(epoch_t epoch)
 325 {
 326
 327         epoch_drain_callbacks(epoch);
 328         allepochs[epoch->e_idx] = NULL;
 329         epoch_wait(global_epoch);
 330         uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
 331         mtx_destroy(&epoch->e_drain_mtx);
 332         sx_destroy(&epoch->e_drain_sx);
 333         free(epoch, M_EPOCH);
 334 }
 335
 336 static epoch_record_t
 337 epoch_currecord(epoch_t epoch)
 338 {
 339
 340         return (zpcpu_get_cpu(epoch->e_pcpu_record, curcpu));
 341 }
 342
 343 #define INIT_CHECK(epoch)                                       \
 344         do {                                                    \
 345                 if (__predict_false((epoch) == NULL))           \
 346                         return;                                 \
 347         } while (0)
 348
 349 void
 350 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
 351 {
 352         struct epoch_record *er;
 353         struct thread *td;
 354
 355         MPASS(cold || epoch != NULL);
 356         INIT_CHECK(epoch);
 357         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 358         td = curthread;
 359 #ifdef EPOCH_TRACE
 360         epoch_trace_enter(td, epoch, et, file, line);
 361 #endif
 362         et->et_td = td;
 363         td->td_epochnest++;
 364         critical_enter();
 365         sched_pin();
 366         td->td_pre_epoch_prio = td->td_priority;
 367         er = epoch_currecord(epoch);
 368         TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
 369         ck_epoch_begin(&er->er_record, &et->et_section);
 370         critical_exit();
 371 }
 372
 373 void
 374 epoch_enter(epoch_t epoch)
 375 {
 376         struct thread *td;
 377         epoch_record_t er;
 378
 379         MPASS(cold || epoch != NULL);
 380         INIT_CHECK(epoch);
 381         td = curthread;
 382         td->td_epochnest++;
 383         critical_enter();
 384         er = epoch_currecord(epoch);
 385         ck_epoch_begin(&er->er_record, NULL);
 386 }
 387
 388 void
 389 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
 390 {
 391         struct epoch_record *er;
 392         struct thread *td;
 393
 394         INIT_CHECK(epoch);
 395         td = curthread;
 396         critical_enter();
 397         sched_unpin();
 398         MPASS(td->td_epochnest);
 399         td->td_epochnest--;
 400         er = epoch_currecord(epoch);
 401         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 402         MPASS(et != NULL);
 403         MPASS(et->et_td == td);
 404 #ifdef INVARIANTS
 405         et->et_td = (void*)0xDEADBEEF;
 406 #endif
 407         ck_epoch_end(&er->er_record, &et->et_section);
 408         TAILQ_REMOVE(&er->er_tdlist, et, et_link);
 409         er->er_gen++;
 410         if (__predict_false(td->td_pre_epoch_prio != td->td_priority))
 411                 epoch_adjust_prio(td, td->td_pre_epoch_prio);
 412         critical_exit();
 413 #ifdef EPOCH_TRACE
 414         epoch_trace_exit(td, epoch, et, file, line);
 415 #endif
 416 }
 417
 418 void
 419 epoch_exit(epoch_t epoch)
 420 {
 421         struct thread *td;
 422         epoch_record_t er;
 423
 424         INIT_CHECK(epoch);
 425         td = curthread;
 426         MPASS(td->td_epochnest);
 427         td->td_epochnest--;
 428         er = epoch_currecord(epoch);
 429         ck_epoch_end(&er->er_record, NULL);
 430         critical_exit();
 431 }
 432
 433 /*
 434  * epoch_block_handler_preempt() is a callback from the CK code when another
 435  * thread is currently in an epoch section.
 436  */
 437 static void
 438 epoch_block_handler_preempt(struct ck_epoch *global __unused,
 439     ck_epoch_record_t *cr, void *arg __unused)
 440 {
 441         epoch_record_t record;
 442         struct thread *td, *owner, *curwaittd;
 443         struct epoch_tracker *tdwait;
 444         struct turnstile *ts;
 445         struct lock_object *lock;
 446         int spincount, gen;
 447         int locksheld __unused;
 448
 449         record = __containerof(cr, struct epoch_record, er_record);
 450         td = curthread;
 451         locksheld = td->td_locks;
 452         spincount = 0;
 453         counter_u64_add(block_count, 1);
 454         /*
 455          * We lost a race and there's no longer any threads
 456          * on the CPU in an epoch section.
 457          */
 458         if (TAILQ_EMPTY(&record->er_tdlist))
 459                 return;
 460
 461         if (record->er_cpuid != curcpu) {
 462                 /*
 463                  * If the head of the list is running, we can wait for it
 464                  * to remove itself from the list and thus save us the
 465                  * overhead of a migration
 466                  */
 467                 gen = record->er_gen;
 468                 thread_unlock(td);
 469                 /*
 470                  * We can't actually check if the waiting thread is running
 471                  * so we simply poll for it to exit before giving up and
 472                  * migrating.
 473                  */
 474                 do {
 475                         cpu_spinwait();
 476                 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
 477                                  gen == record->er_gen &&
 478                                  spincount++ < MAX_ADAPTIVE_SPIN);
 479                 thread_lock(td);
 480                 /*
 481                  * If the generation has changed we can poll again
 482                  * otherwise we need to migrate.
 483                  */
 484                 if (gen != record->er_gen)
 485                         return;
 486                 /*
 487                  * Being on the same CPU as that of the record on which
 488                  * we need to wait allows us access to the thread
 489                  * list associated with that CPU. We can then examine the
 490                  * oldest thread in the queue and wait on its turnstile
 491                  * until it resumes and so on until a grace period
 492                  * elapses.
 493                  *
 494                  */
 495                 counter_u64_add(migrate_count, 1);
 496                 sched_bind(td, record->er_cpuid);
 497                 /*
 498                  * At this point we need to return to the ck code
 499                  * to scan to see if a grace period has elapsed.
 500                  * We can't move on to check the thread list, because
 501                  * in the meantime new threads may have arrived that
 502                  * in fact belong to a different epoch.
 503                  */
 504                 return;
 505         }
 506         /*
 507          * Try to find a thread in an epoch section on this CPU
 508          * waiting on a turnstile. Otherwise find the lowest
 509          * priority thread (highest prio value) and drop our priority
 510          * to match to allow it to run.
 511          */
 512         TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
 513                 /*
 514                  * Propagate our priority to any other waiters to prevent us
 515                  * from starving them. They will have their original priority
 516                  * restore on exit from epoch_wait().
 517                  */
 518                 curwaittd = tdwait->et_td;
 519                 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
 520                         critical_enter();
 521                         thread_unlock(td);
 522                         thread_lock(curwaittd);
 523                         sched_prio(curwaittd, td->td_priority);
 524                         thread_unlock(curwaittd);
 525                         thread_lock(td);
 526                         critical_exit();
 527                 }
 528                 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
 529                     ((ts = curwaittd->td_blocked) != NULL)) {
 530                         /*
 531                          * We unlock td to allow turnstile_wait to reacquire
 532                          * the thread lock. Before unlocking it we enter a
 533                          * critical section to prevent preemption after we
 534                          * reenable interrupts by dropping the thread lock in
 535                          * order to prevent curwaittd from getting to run.
 536                          */
 537                         critical_enter();
 538                         thread_unlock(td);
 539
 540                         if (turnstile_lock(ts, &lock, &owner)) {
 541                                 if (ts == curwaittd->td_blocked) {
 542                                         MPASS(TD_IS_INHIBITED(curwaittd) &&
 543                                             TD_ON_LOCK(curwaittd));
 544                                         critical_exit();
 545                                         turnstile_wait(ts, owner,
 546                                             curwaittd->td_tsqueue);
 547                                         counter_u64_add(turnstile_count, 1);
 548                                         thread_lock(td);
 549                                         return;
 550                                 }
 551                                 turnstile_unlock(ts, lock);
 552                         }
 553                         thread_lock(td);
 554                         critical_exit();
 555                         KASSERT(td->td_locks == locksheld,
 556                             ("%d extra locks held", td->td_locks - locksheld));
 557                 }
 558         }
 559         /*
 560          * We didn't find any threads actually blocked on a lock
 561          * so we have nothing to do except context switch away.
 562          */
 563         counter_u64_add(switch_count, 1);
 564         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 565
 566         /*
 567          * Release the thread lock while yielding to
 568          * allow other threads to acquire the lock
 569          * pointed to by TDQ_LOCKPTR(td). Else a
 570          * deadlock like situation might happen. (HPS)
 571          */
 572         thread_unlock(td);
 573         thread_lock(td);
 574 }
 575
 576 void
 577 epoch_wait_preempt(epoch_t epoch)
 578 {
 579         struct thread *td;
 580         int was_bound;
 581         int old_cpu;
 582         int old_pinned;
 583         u_char old_prio;
 584         int locks __unused;
 585
 586         MPASS(cold || epoch != NULL);
 587         INIT_CHECK(epoch);
 588         td = curthread;
 589 #ifdef INVARIANTS
 590         locks = curthread->td_locks;
 591         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 592         if ((epoch->e_flags & EPOCH_LOCKED) == 0)
 593                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 594                     "epoch_wait() can be long running");
 595         KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
 596             "of an epoch section of the same epoch"));
 597 #endif
 598         thread_lock(td);
 599         DROP_GIANT();
 600
 601         old_cpu = PCPU_GET(cpuid);
 602         old_pinned = td->td_pinned;
 603         old_prio = td->td_priority;
 604         was_bound = sched_is_bound(td);
 605         sched_unbind(td);
 606         td->td_pinned = 0;
 607         sched_bind(td, old_cpu);
 608
 609         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
 610             NULL);
 611
 612         /* restore CPU binding, if any */
 613         if (was_bound != 0) {
 614                 sched_bind(td, old_cpu);
 615         } else {
 616                 /* get thread back to initial CPU, if any */
 617                 if (old_pinned != 0)
 618                         sched_bind(td, old_cpu);
 619                 sched_unbind(td);
 620         }
 621         /* restore pinned after bind */
 622         td->td_pinned = old_pinned;
 623
 624         /* restore thread priority */
 625         sched_prio(td, old_prio);
 626         thread_unlock(td);
 627         PICKUP_GIANT();
 628         KASSERT(td->td_locks == locks,
 629             ("%d residual locks held", td->td_locks - locks));
 630 }
 631
 632 static void
 633 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
 634     void *arg __unused)
 635 {
 636         cpu_spinwait();
 637 }
 638
 639 void
 640 epoch_wait(epoch_t epoch)
 641 {
 642
 643         MPASS(cold || epoch != NULL);
 644         INIT_CHECK(epoch);
 645         MPASS(epoch->e_flags == 0);
 646         critical_enter();
 647         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
 648         critical_exit();
 649 }
 650
 651 void
 652 epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t))
 653 {
 654         epoch_record_t er;
 655         ck_epoch_entry_t *cb;
 656
 657         cb = (void *)ctx;
 658
 659         MPASS(callback);
 660         /* too early in boot to have epoch set up */
 661         if (__predict_false(epoch == NULL))
 662                 goto boottime;
 663 #if !defined(EARLY_AP_STARTUP)
 664         if (__predict_false(inited < 2))
 665                 goto boottime;
 666 #endif
 667
 668         critical_enter();
 669         *DPCPU_PTR(epoch_cb_count) += 1;
 670         er = epoch_currecord(epoch);
 671         ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
 672         critical_exit();
 673         return;
 674 boottime:
 675         callback(ctx);
 676 }
 677
 678 static void
 679 epoch_call_task(void *arg __unused)
 680 {
 681         ck_stack_entry_t *cursor, *head, *next;
 682         ck_epoch_record_t *record;
 683         epoch_record_t er;
 684         epoch_t epoch;
 685         ck_stack_t cb_stack;
 686         int i, npending, total;
 687
 688         ck_stack_init(&cb_stack);
 689         critical_enter();
 690         epoch_enter(global_epoch);
 691         for (total = i = 0; i < epoch_count; i++) {
 692                 if (__predict_false((epoch = allepochs[i]) == NULL))
 693                         continue;
 694                 er = epoch_currecord(epoch);
 695                 record = &er->er_record;
 696                 if ((npending = record->n_pending) == 0)
 697                         continue;
 698                 ck_epoch_poll_deferred(record, &cb_stack);
 699                 total += npending - record->n_pending;
 700         }
 701         epoch_exit(global_epoch);
 702         *DPCPU_PTR(epoch_cb_count) -= total;
 703         critical_exit();
 704
 705         counter_u64_add(epoch_call_count, total);
 706         counter_u64_add(epoch_call_task_count, 1);
 707
 708         head = ck_stack_batch_pop_npsc(&cb_stack);
 709         for (cursor = head; cursor != NULL; cursor = next) {
 710                 struct ck_epoch_entry *entry =
 711                     ck_epoch_entry_container(cursor);
 712
 713                 next = CK_STACK_NEXT(cursor);
 714                 entry->function(entry);
 715         }
 716 }
 717
 718 int
 719 in_epoch_verbose(epoch_t epoch, int dump_onfail)
 720 {
 721         struct epoch_tracker *tdwait;
 722         struct thread *td;
 723         epoch_record_t er;
 724
 725         td = curthread;
 726         if (td->td_epochnest == 0)
 727                 return (0);
 728         if (__predict_false((epoch) == NULL))
 729                 return (0);
 730         critical_enter();
 731         er = epoch_currecord(epoch);
 732         TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 733                 if (tdwait->et_td == td) {
 734                         critical_exit();
 735                         return (1);
 736                 }
 737 #ifdef INVARIANTS
 738         if (dump_onfail) {
 739                 MPASS(td->td_pinned);
 740                 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
 741                 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 742                         printf("td_tid: %d ", tdwait->et_td->td_tid);
 743                 printf("\n");
 744         }
 745 #endif
 746         critical_exit();
 747         return (0);
 748 }
 749
 750 int
 751 in_epoch(epoch_t epoch)
 752 {
 753         return (in_epoch_verbose(epoch, 0));
 754 }
 755
 756 static void
 757 epoch_drain_cb(struct epoch_context *ctx)
 758 {
 759         struct epoch *epoch =
 760             __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
 761
 762         if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
 763                 mtx_lock(&epoch->e_drain_mtx);
 764                 wakeup(epoch);
 765                 mtx_unlock(&epoch->e_drain_mtx);
 766         }
 767 }
 768
 769 void
 770 epoch_drain_callbacks(epoch_t epoch)
 771 {
 772         epoch_record_t er;
 773         struct thread *td;
 774         int was_bound;
 775         int old_pinned;
 776         int old_cpu;
 777         int cpu;
 778
 779         WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 780             "epoch_drain_callbacks() may sleep!");
 781
 782         /* too early in boot to have epoch set up */
 783         if (__predict_false(epoch == NULL))
 784                 return;
 785 #if !defined(EARLY_AP_STARTUP)
 786         if (__predict_false(inited < 2))
 787                 return;
 788 #endif
 789         DROP_GIANT();
 790
 791         sx_xlock(&epoch->e_drain_sx);
 792         mtx_lock(&epoch->e_drain_mtx);
 793
 794         td = curthread;
 795         thread_lock(td);
 796         old_cpu = PCPU_GET(cpuid);
 797         old_pinned = td->td_pinned;
 798         was_bound = sched_is_bound(td);
 799         sched_unbind(td);
 800         td->td_pinned = 0;
 801
 802         CPU_FOREACH(cpu)
 803                 epoch->e_drain_count++;
 804         CPU_FOREACH(cpu) {
 805                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 806                 sched_bind(td, cpu);
 807                 epoch_call(epoch, &er->er_drain_ctx, &epoch_drain_cb);
 808         }
 809
 810         /* restore CPU binding, if any */
 811         if (was_bound != 0) {
 812                 sched_bind(td, old_cpu);
 813         } else {
 814                 /* get thread back to initial CPU, if any */
 815                 if (old_pinned != 0)
 816                         sched_bind(td, old_cpu);
 817                 sched_unbind(td);
 818         }
 819         /* restore pinned after bind */
 820         td->td_pinned = old_pinned;
 821
 822         thread_unlock(td);
 823
 824         while (epoch->e_drain_count != 0)
 825                 msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
 826
 827         mtx_unlock(&epoch->e_drain_mtx);
 828         sx_xunlock(&epoch->e_drain_sx);
 829
 830         PICKUP_GIANT();
 831 }
 832
 833 void
 834 epoch_thread_init(struct thread *td)
 835 {
 836
 837         td->td_et = malloc(sizeof(struct epoch_tracker), M_EPOCH, M_WAITOK);
 838 }
 839
 840 void
 841 epoch_thread_fini(struct thread *td)
 842 {
 843
 844         free(td->td_et, M_EPOCH);
 845 }