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 #ifdef __amd64__
  62 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
  63 #else
  64 #define EPOCH_ALIGN CACHE_LINE_SIZE
  65 #endif
  66
  67 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
  68 typedef struct epoch_record {
  69         ck_epoch_record_t er_record;
  70         struct epoch_context er_drain_ctx;
  71         struct epoch *er_parent;
  72         volatile struct epoch_tdlist er_tdlist;
  73         volatile uint32_t er_gen;
  74         uint32_t er_cpuid;
  75 #ifdef INVARIANTS
  76         /* Used to verify record ownership for non-preemptible epochs. */
  77         struct thread *er_td;
  78 #endif
  79 } __aligned(EPOCH_ALIGN)     *epoch_record_t;
  80
  81 struct epoch {
  82         struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
  83         epoch_record_t e_pcpu_record;
  84         int     e_in_use;
  85         int     e_flags;
  86         struct sx e_drain_sx;
  87         struct mtx e_drain_mtx;
  88         volatile int e_drain_count;
  89         const char *e_name;
  90 };
  91
  92 /* arbitrary --- needs benchmarking */
  93 #define MAX_ADAPTIVE_SPIN 100
  94 #define MAX_EPOCHS 64
  95
  96 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
  97 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
  98     "epoch information");
  99 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 100     "epoch stats");
 101
 102 /* Stats. */
 103 static counter_u64_t block_count;
 104
 105 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
 106     &block_count, "# of times a thread was in an epoch when epoch_wait was called");
 107 static counter_u64_t migrate_count;
 108
 109 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
 110     &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
 111 static counter_u64_t turnstile_count;
 112
 113 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
 114     &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
 115 static counter_u64_t switch_count;
 116
 117 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
 118     &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
 119 static counter_u64_t epoch_call_count;
 120
 121 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
 122     &epoch_call_count, "# of times a callback was deferred");
 123 static counter_u64_t epoch_call_task_count;
 124
 125 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
 126     &epoch_call_task_count, "# of times a callback task was run");
 127
 128 TAILQ_HEAD (threadlist, thread);
 129
 130 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
 131     ck_epoch_entry_container)
 132
 133 static struct epoch epoch_array[MAX_EPOCHS];
 134
 135 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
 136 DPCPU_DEFINE(int, epoch_cb_count);
 137
 138 static __read_mostly int inited;
 139 __read_mostly epoch_t global_epoch;
 140 __read_mostly epoch_t global_epoch_preempt;
 141
 142 static void epoch_call_task(void *context __unused);
 143 static  uma_zone_t pcpu_zone_record;
 144
 145 static struct sx epoch_sx;
 146
 147 #define EPOCH_LOCK() sx_xlock(&epoch_sx)
 148 #define EPOCH_UNLOCK() sx_xunlock(&epoch_sx)
 149
 150 #ifdef EPOCH_TRACE
 151 struct stackentry {
 152         RB_ENTRY(stackentry) se_node;
 153         struct stack se_stack;
 154 };
 155
 156 static int
 157 stackentry_compare(struct stackentry *a, struct stackentry *b)
 158 {
 159
 160         if (a->se_stack.depth > b->se_stack.depth)
 161                 return (1);
 162         if (a->se_stack.depth < b->se_stack.depth)
 163                 return (-1);
 164         for (int i = 0; i < a->se_stack.depth; i++) {
 165                 if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
 166                         return (1);
 167                 if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
 168                         return (-1);
 169         }
 170
 171         return (0);
 172 }
 173
 174 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
 175 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
 176
 177 static struct mtx epoch_stacks_lock;
 178 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
 179
 180 static bool epoch_trace_stack_print = true;
 181 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
 182     &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
 183
 184 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
 185 static inline void
 186 epoch_trace_report(const char *fmt, ...)
 187 {
 188         va_list ap;
 189         struct stackentry se, *new;
 190
 191         stack_zero(&se.se_stack);       /* XXX: is it really needed? */
 192         stack_save(&se.se_stack);
 193
 194         /* Tree is never reduced - go lockless. */
 195         if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
 196                 return;
 197
 198         new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
 199         if (new != NULL) {
 200                 bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
 201
 202                 mtx_lock(&epoch_stacks_lock);
 203                 new = RB_INSERT(stacktree, &epoch_stacks, new);
 204                 mtx_unlock(&epoch_stacks_lock);
 205                 if (new != NULL)
 206                         free(new, M_STACK);
 207         }
 208
 209         va_start(ap, fmt);
 210         (void)vprintf(fmt, ap);
 211         va_end(ap);
 212         if (epoch_trace_stack_print)
 213                 stack_print_ddb(&se.se_stack);
 214 }
 215
 216 static inline void
 217 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
 218     const char *file, int line)
 219 {
 220         epoch_tracker_t iet;
 221
 222         SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
 223                 if (iet->et_epoch == epoch)
 224                         epoch_trace_report("Recursively entering epoch %s "
 225                             "at %s:%d, previously entered at %s:%d\n",
 226                             epoch->e_name, file, line,
 227                             iet->et_file, iet->et_line);
 228         et->et_epoch = epoch;
 229         et->et_file = file;
 230         et->et_line = line;
 231         SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
 232 }
 233
 234 static inline void
 235 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
 236     const char *file, int line)
 237 {
 238
 239         if (SLIST_FIRST(&td->td_epochs) != et) {
 240                 epoch_trace_report("Exiting epoch %s in a not nested order "
 241                     "at %s:%d. Most recently entered %s at %s:%d\n",
 242                     epoch->e_name,
 243                     file, line,
 244                     SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
 245                     SLIST_FIRST(&td->td_epochs)->et_file,
 246                     SLIST_FIRST(&td->td_epochs)->et_line);
 247                 /* This will panic if et is not anywhere on td_epochs. */
 248                 SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
 249         } else
 250                 SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
 251 }
 252
 253 /* Used by assertions that check thread state before going to sleep. */
 254 void
 255 epoch_trace_list(struct thread *td)
 256 {
 257         epoch_tracker_t iet;
 258
 259         SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
 260                 printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
 261                     iet->et_file, iet->et_line);
 262 }
 263 #endif /* EPOCH_TRACE */
 264
 265 static void
 266 epoch_init(void *arg __unused)
 267 {
 268         int cpu;
 269
 270         block_count = counter_u64_alloc(M_WAITOK);
 271         migrate_count = counter_u64_alloc(M_WAITOK);
 272         turnstile_count = counter_u64_alloc(M_WAITOK);
 273         switch_count = counter_u64_alloc(M_WAITOK);
 274         epoch_call_count = counter_u64_alloc(M_WAITOK);
 275         epoch_call_task_count = counter_u64_alloc(M_WAITOK);
 276
 277         pcpu_zone_record = uma_zcreate("epoch_record pcpu",
 278             sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
 279             UMA_ALIGN_PTR, UMA_ZONE_PCPU);
 280         CPU_FOREACH(cpu) {
 281                 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
 282                     epoch_call_task, NULL);
 283                 taskqgroup_attach_cpu(qgroup_softirq,
 284                     DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
 285                     "epoch call task");
 286         }
 287 #ifdef EPOCH_TRACE
 288         SLIST_INIT(&thread0.td_epochs);
 289 #endif
 290         sx_init(&epoch_sx, "epoch-sx");
 291         inited = 1;
 292         global_epoch = epoch_alloc("Global", 0);
 293         global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
 294 }
 295 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
 296
 297 #if !defined(EARLY_AP_STARTUP)
 298 static void
 299 epoch_init_smp(void *dummy __unused)
 300 {
 301         inited = 2;
 302 }
 303 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
 304 #endif
 305
 306 static void
 307 epoch_ctor(epoch_t epoch)
 308 {
 309         epoch_record_t er;
 310         int cpu;
 311
 312         epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
 313         CPU_FOREACH(cpu) {
 314                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 315                 bzero(er, sizeof(*er));
 316                 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 317                 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 318                 er->er_cpuid = cpu;
 319                 er->er_parent = epoch;
 320         }
 321 }
 322
 323 static void
 324 epoch_adjust_prio(struct thread *td, u_char prio)
 325 {
 326
 327         thread_lock(td);
 328         sched_prio(td, prio);
 329         thread_unlock(td);
 330 }
 331
 332 epoch_t
 333 epoch_alloc(const char *name, int flags)
 334 {
 335         epoch_t epoch;
 336         int i;
 337
 338         MPASS(name != NULL);
 339
 340         if (__predict_false(!inited))
 341                 panic("%s called too early in boot", __func__);
 342
 343         EPOCH_LOCK();
 344
 345         /*
 346          * Find a free index in the epoch array. If no free index is
 347          * found, try to use the index after the last one.
 348          */
 349         for (i = 0;; i++) {
 350                 /*
 351                  * If too many epochs are currently allocated,
 352                  * return NULL.
 353                  */
 354                 if (i == MAX_EPOCHS) {
 355                         epoch = NULL;
 356                         goto done;
 357                 }
 358                 if (epoch_array[i].e_in_use == 0)
 359                         break;
 360         }
 361
 362         epoch = epoch_array + i;
 363         ck_epoch_init(&epoch->e_epoch);
 364         epoch_ctor(epoch);
 365         epoch->e_flags = flags;
 366         epoch->e_name = name;
 367         sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
 368         mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
 369
 370         /*
 371          * Set e_in_use last, because when this field is set the
 372          * epoch_call_task() function will start scanning this epoch
 373          * structure.
 374          */
 375         atomic_store_rel_int(&epoch->e_in_use, 1);
 376 done:
 377         EPOCH_UNLOCK();
 378         return (epoch);
 379 }
 380
 381 void
 382 epoch_free(epoch_t epoch)
 383 {
 384 #ifdef INVARIANTS
 385         int cpu;
 386 #endif
 387
 388         EPOCH_LOCK();
 389
 390         MPASS(epoch->e_in_use != 0);
 391
 392         epoch_drain_callbacks(epoch);
 393
 394         atomic_store_rel_int(&epoch->e_in_use, 0);
 395         /*
 396          * Make sure the epoch_call_task() function see e_in_use equal
 397          * to zero, by calling epoch_wait() on the global_epoch:
 398          */
 399         epoch_wait(global_epoch);
 400 #ifdef INVARIANTS
 401         CPU_FOREACH(cpu) {
 402                 epoch_record_t er;
 403
 404                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 405
 406                 /*
 407                  * Sanity check: none of the records should be in use anymore.
 408                  * We drained callbacks above and freeing the pcpu records is
 409                  * imminent.
 410                  */
 411                 MPASS(er->er_td == NULL);
 412                 MPASS(TAILQ_EMPTY(&er->er_tdlist));
 413         }
 414 #endif
 415         uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
 416         mtx_destroy(&epoch->e_drain_mtx);
 417         sx_destroy(&epoch->e_drain_sx);
 418         memset(epoch, 0, sizeof(*epoch));
 419
 420         EPOCH_UNLOCK();
 421 }
 422
 423 static epoch_record_t
 424 epoch_currecord(epoch_t epoch)
 425 {
 426
 427         return (zpcpu_get(epoch->e_pcpu_record));
 428 }
 429
 430 #define INIT_CHECK(epoch)                                       \
 431         do {                                                    \
 432                 if (__predict_false((epoch) == NULL))           \
 433                         return;                                 \
 434         } while (0)
 435
 436 void
 437 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
 438 {
 439         struct epoch_record *er;
 440         struct thread *td;
 441
 442         MPASS(cold || epoch != NULL);
 443         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 444         td = curthread;
 445         MPASS((vm_offset_t)et >= td->td_kstack &&
 446             (vm_offset_t)et + sizeof(struct epoch_tracker) <=
 447             td->td_kstack + td->td_kstack_pages * PAGE_SIZE);
 448
 449         INIT_CHECK(epoch);
 450 #ifdef EPOCH_TRACE
 451         epoch_trace_enter(td, epoch, et, file, line);
 452 #endif
 453         et->et_td = td;
 454         THREAD_NO_SLEEPING();
 455         critical_enter();
 456         sched_pin();
 457         td->td_pre_epoch_prio = td->td_priority;
 458         er = epoch_currecord(epoch);
 459         /* Record-level tracking is reserved for non-preemptible epochs. */
 460         MPASS(er->er_td == NULL);
 461         TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
 462         ck_epoch_begin(&er->er_record, &et->et_section);
 463         critical_exit();
 464 }
 465
 466 void
 467 epoch_enter(epoch_t epoch)
 468 {
 469         epoch_record_t er;
 470
 471         MPASS(cold || epoch != NULL);
 472         INIT_CHECK(epoch);
 473         critical_enter();
 474         er = epoch_currecord(epoch);
 475 #ifdef INVARIANTS
 476         if (er->er_record.active == 0) {
 477                 MPASS(er->er_td == NULL);
 478                 er->er_td = curthread;
 479         } else {
 480                 /* We've recursed, just make sure our accounting isn't wrong. */
 481                 MPASS(er->er_td == curthread);
 482         }
 483 #endif
 484         ck_epoch_begin(&er->er_record, NULL);
 485 }
 486
 487 void
 488 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
 489 {
 490         struct epoch_record *er;
 491         struct thread *td;
 492
 493         INIT_CHECK(epoch);
 494         td = curthread;
 495         critical_enter();
 496         sched_unpin();
 497         THREAD_SLEEPING_OK();
 498         er = epoch_currecord(epoch);
 499         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 500         MPASS(et != NULL);
 501         MPASS(et->et_td == td);
 502 #ifdef INVARIANTS
 503         et->et_td = (void*)0xDEADBEEF;
 504         /* Record-level tracking is reserved for non-preemptible epochs. */
 505         MPASS(er->er_td == NULL);
 506 #endif
 507         ck_epoch_end(&er->er_record, &et->et_section);
 508         TAILQ_REMOVE(&er->er_tdlist, et, et_link);
 509         er->er_gen++;
 510         if (__predict_false(td->td_pre_epoch_prio != td->td_priority))
 511                 epoch_adjust_prio(td, td->td_pre_epoch_prio);
 512         critical_exit();
 513 #ifdef EPOCH_TRACE
 514         epoch_trace_exit(td, epoch, et, file, line);
 515 #endif
 516 }
 517
 518 void
 519 epoch_exit(epoch_t epoch)
 520 {
 521         epoch_record_t er;
 522
 523         INIT_CHECK(epoch);
 524         er = epoch_currecord(epoch);
 525         ck_epoch_end(&er->er_record, NULL);
 526 #ifdef INVARIANTS
 527         MPASS(er->er_td == curthread);
 528         if (er->er_record.active == 0)
 529                 er->er_td = NULL;
 530 #endif
 531         critical_exit();
 532 }
 533
 534 /*
 535  * epoch_block_handler_preempt() is a callback from the CK code when another
 536  * thread is currently in an epoch section.
 537  */
 538 static void
 539 epoch_block_handler_preempt(struct ck_epoch *global __unused,
 540     ck_epoch_record_t *cr, void *arg __unused)
 541 {
 542         epoch_record_t record;
 543         struct thread *td, *owner, *curwaittd;
 544         struct epoch_tracker *tdwait;
 545         struct turnstile *ts;
 546         struct lock_object *lock;
 547         int spincount, gen;
 548         int locksheld __unused;
 549
 550         record = __containerof(cr, struct epoch_record, er_record);
 551         td = curthread;
 552         locksheld = td->td_locks;
 553         spincount = 0;
 554         counter_u64_add(block_count, 1);
 555         /*
 556          * We lost a race and there's no longer any threads
 557          * on the CPU in an epoch section.
 558          */
 559         if (TAILQ_EMPTY(&record->er_tdlist))
 560                 return;
 561
 562         if (record->er_cpuid != curcpu) {
 563                 /*
 564                  * If the head of the list is running, we can wait for it
 565                  * to remove itself from the list and thus save us the
 566                  * overhead of a migration
 567                  */
 568                 gen = record->er_gen;
 569                 thread_unlock(td);
 570                 /*
 571                  * We can't actually check if the waiting thread is running
 572                  * so we simply poll for it to exit before giving up and
 573                  * migrating.
 574                  */
 575                 do {
 576                         cpu_spinwait();
 577                 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
 578                                  gen == record->er_gen &&
 579                                  spincount++ < MAX_ADAPTIVE_SPIN);
 580                 thread_lock(td);
 581                 /*
 582                  * If the generation has changed we can poll again
 583                  * otherwise we need to migrate.
 584                  */
 585                 if (gen != record->er_gen)
 586                         return;
 587                 /*
 588                  * Being on the same CPU as that of the record on which
 589                  * we need to wait allows us access to the thread
 590                  * list associated with that CPU. We can then examine the
 591                  * oldest thread in the queue and wait on its turnstile
 592                  * until it resumes and so on until a grace period
 593                  * elapses.
 594                  *
 595                  */
 596                 counter_u64_add(migrate_count, 1);
 597                 sched_bind(td, record->er_cpuid);
 598                 /*
 599                  * At this point we need to return to the ck code
 600                  * to scan to see if a grace period has elapsed.
 601                  * We can't move on to check the thread list, because
 602                  * in the meantime new threads may have arrived that
 603                  * in fact belong to a different epoch.
 604                  */
 605                 return;
 606         }
 607         /*
 608          * Try to find a thread in an epoch section on this CPU
 609          * waiting on a turnstile. Otherwise find the lowest
 610          * priority thread (highest prio value) and drop our priority
 611          * to match to allow it to run.
 612          */
 613         TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
 614                 /*
 615                  * Propagate our priority to any other waiters to prevent us
 616                  * from starving them. They will have their original priority
 617                  * restore on exit from epoch_wait().
 618                  */
 619                 curwaittd = tdwait->et_td;
 620                 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
 621                         critical_enter();
 622                         thread_unlock(td);
 623                         thread_lock(curwaittd);
 624                         sched_prio(curwaittd, td->td_priority);
 625                         thread_unlock(curwaittd);
 626                         thread_lock(td);
 627                         critical_exit();
 628                 }
 629                 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
 630                     ((ts = curwaittd->td_blocked) != NULL)) {
 631                         /*
 632                          * We unlock td to allow turnstile_wait to reacquire
 633                          * the thread lock. Before unlocking it we enter a
 634                          * critical section to prevent preemption after we
 635                          * reenable interrupts by dropping the thread lock in
 636                          * order to prevent curwaittd from getting to run.
 637                          */
 638                         critical_enter();
 639                         thread_unlock(td);
 640
 641                         if (turnstile_lock(ts, &lock, &owner)) {
 642                                 if (ts == curwaittd->td_blocked) {
 643                                         MPASS(TD_IS_INHIBITED(curwaittd) &&
 644                                             TD_ON_LOCK(curwaittd));
 645                                         critical_exit();
 646                                         turnstile_wait(ts, owner,
 647                                             curwaittd->td_tsqueue);
 648                                         counter_u64_add(turnstile_count, 1);
 649                                         thread_lock(td);
 650                                         return;
 651                                 }
 652                                 turnstile_unlock(ts, lock);
 653                         }
 654                         thread_lock(td);
 655                         critical_exit();
 656                         KASSERT(td->td_locks == locksheld,
 657                             ("%d extra locks held", td->td_locks - locksheld));
 658                 }
 659         }
 660         /*
 661          * We didn't find any threads actually blocked on a lock
 662          * so we have nothing to do except context switch away.
 663          */
 664         counter_u64_add(switch_count, 1);
 665         mi_switch(SW_VOL | SWT_RELINQUISH);
 666         /*
 667          * It is important the thread lock is dropped while yielding
 668          * to allow other threads to acquire the lock pointed to by
 669          * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
 670          * thread lock before returning. Else a deadlock like
 671          * situation might happen.
 672          */
 673         thread_lock(td);
 674 }
 675
 676 void
 677 epoch_wait_preempt(epoch_t epoch)
 678 {
 679         struct thread *td;
 680         int was_bound;
 681         int old_cpu;
 682         int old_pinned;
 683         u_char old_prio;
 684         int locks __unused;
 685
 686         MPASS(cold || epoch != NULL);
 687         INIT_CHECK(epoch);
 688         td = curthread;
 689 #ifdef INVARIANTS
 690         locks = curthread->td_locks;
 691         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 692         if ((epoch->e_flags & EPOCH_LOCKED) == 0)
 693                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 694                     "epoch_wait() can be long running");
 695         KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
 696             "of an epoch section of the same epoch"));
 697 #endif
 698         DROP_GIANT();
 699         thread_lock(td);
 700
 701         old_cpu = PCPU_GET(cpuid);
 702         old_pinned = td->td_pinned;
 703         old_prio = td->td_priority;
 704         was_bound = sched_is_bound(td);
 705         sched_unbind(td);
 706         td->td_pinned = 0;
 707         sched_bind(td, old_cpu);
 708
 709         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
 710             NULL);
 711
 712         /* restore CPU binding, if any */
 713         if (was_bound != 0) {
 714                 sched_bind(td, old_cpu);
 715         } else {
 716                 /* get thread back to initial CPU, if any */
 717                 if (old_pinned != 0)
 718                         sched_bind(td, old_cpu);
 719                 sched_unbind(td);
 720         }
 721         /* restore pinned after bind */
 722         td->td_pinned = old_pinned;
 723
 724         /* restore thread priority */
 725         sched_prio(td, old_prio);
 726         thread_unlock(td);
 727         PICKUP_GIANT();
 728         KASSERT(td->td_locks == locks,
 729             ("%d residual locks held", td->td_locks - locks));
 730 }
 731
 732 static void
 733 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
 734     void *arg __unused)
 735 {
 736         cpu_spinwait();
 737 }
 738
 739 void
 740 epoch_wait(epoch_t epoch)
 741 {
 742
 743         MPASS(cold || epoch != NULL);
 744         INIT_CHECK(epoch);
 745         MPASS(epoch->e_flags == 0);
 746         critical_enter();
 747         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
 748         critical_exit();
 749 }
 750
 751 void
 752 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
 753 {
 754         epoch_record_t er;
 755         ck_epoch_entry_t *cb;
 756
 757         cb = (void *)ctx;
 758
 759         MPASS(callback);
 760         /* too early in boot to have epoch set up */
 761         if (__predict_false(epoch == NULL))
 762                 goto boottime;
 763 #if !defined(EARLY_AP_STARTUP)
 764         if (__predict_false(inited < 2))
 765                 goto boottime;
 766 #endif
 767
 768         critical_enter();
 769         *DPCPU_PTR(epoch_cb_count) += 1;
 770         er = epoch_currecord(epoch);
 771         ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
 772         critical_exit();
 773         return;
 774 boottime:
 775         callback(ctx);
 776 }
 777
 778 static void
 779 epoch_call_task(void *arg __unused)
 780 {
 781         ck_stack_entry_t *cursor, *head, *next;
 782         ck_epoch_record_t *record;
 783         epoch_record_t er;
 784         epoch_t epoch;
 785         ck_stack_t cb_stack;
 786         int i, npending, total;
 787
 788         ck_stack_init(&cb_stack);
 789         critical_enter();
 790         epoch_enter(global_epoch);
 791         for (total = i = 0; i != MAX_EPOCHS; i++) {
 792                 epoch = epoch_array + i;
 793                 if (__predict_false(
 794                     atomic_load_acq_int(&epoch->e_in_use) == 0))
 795                         continue;
 796                 er = epoch_currecord(epoch);
 797                 record = &er->er_record;
 798                 if ((npending = record->n_pending) == 0)
 799                         continue;
 800                 ck_epoch_poll_deferred(record, &cb_stack);
 801                 total += npending - record->n_pending;
 802         }
 803         epoch_exit(global_epoch);
 804         *DPCPU_PTR(epoch_cb_count) -= total;
 805         critical_exit();
 806
 807         counter_u64_add(epoch_call_count, total);
 808         counter_u64_add(epoch_call_task_count, 1);
 809
 810         head = ck_stack_batch_pop_npsc(&cb_stack);
 811         for (cursor = head; cursor != NULL; cursor = next) {
 812                 struct ck_epoch_entry *entry =
 813                     ck_epoch_entry_container(cursor);
 814
 815                 next = CK_STACK_NEXT(cursor);
 816                 entry->function(entry);
 817         }
 818 }
 819
 820 static int
 821 in_epoch_verbose_preempt(epoch_t epoch, int dump_onfail)
 822 {
 823         epoch_record_t er;
 824         struct epoch_tracker *tdwait;
 825         struct thread *td;
 826
 827         MPASS(epoch != NULL);
 828         MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
 829         td = curthread;
 830         if (THREAD_CAN_SLEEP())
 831                 return (0);
 832         critical_enter();
 833         er = epoch_currecord(epoch);
 834         TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 835                 if (tdwait->et_td == td) {
 836                         critical_exit();
 837                         return (1);
 838                 }
 839 #ifdef INVARIANTS
 840         if (dump_onfail) {
 841                 MPASS(td->td_pinned);
 842                 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
 843                 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 844                         printf("td_tid: %d ", tdwait->et_td->td_tid);
 845                 printf("\n");
 846         }
 847 #endif
 848         critical_exit();
 849         return (0);
 850 }
 851
 852 #ifdef INVARIANTS
 853 static void
 854 epoch_assert_nocpu(epoch_t epoch, struct thread *td)
 855 {
 856         epoch_record_t er;
 857         int cpu;
 858         bool crit;
 859
 860         crit = td->td_critnest > 0;
 861
 862         /* Check for a critical section mishap. */
 863         CPU_FOREACH(cpu) {
 864                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 865                 KASSERT(er->er_td != td,
 866                     ("%s critical section in epoch '%s', from cpu %d",
 867                     (crit ? "exited" : "re-entered"), epoch->e_name, cpu));
 868         }
 869 }
 870 #else
 871 #define epoch_assert_nocpu(e, td)
 872 #endif
 873
 874 int
 875 in_epoch_verbose(epoch_t epoch, int dump_onfail)
 876 {
 877         epoch_record_t er;
 878         struct thread *td;
 879
 880         if (__predict_false((epoch) == NULL))
 881                 return (0);
 882         if ((epoch->e_flags & EPOCH_PREEMPT) != 0)
 883                 return (in_epoch_verbose_preempt(epoch, dump_onfail));
 884
 885         /*
 886          * The thread being in a critical section is a necessary
 887          * condition to be correctly inside a non-preemptible epoch,
 888          * so it's definitely not in this epoch.
 889          */
 890         td = curthread;
 891         if (td->td_critnest == 0) {
 892                 epoch_assert_nocpu(epoch, td);
 893                 return (0);
 894         }
 895
 896         /*
 897          * The current cpu is in a critical section, so the epoch record will be
 898          * stable for the rest of this function.  Knowing that the record is not
 899          * active is sufficient for knowing whether we're in this epoch or not,
 900          * since it's a pcpu record.
 901          */
 902         er = epoch_currecord(epoch);
 903         if (er->er_record.active == 0) {
 904                 epoch_assert_nocpu(epoch, td);
 905                 return (0);
 906         }
 907
 908         MPASS(er->er_td == td);
 909         return (1);
 910 }
 911
 912 int
 913 in_epoch(epoch_t epoch)
 914 {
 915         return (in_epoch_verbose(epoch, 0));
 916 }
 917
 918 static void
 919 epoch_drain_cb(struct epoch_context *ctx)
 920 {
 921         struct epoch *epoch =
 922             __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
 923
 924         if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
 925                 mtx_lock(&epoch->e_drain_mtx);
 926                 wakeup(epoch);
 927                 mtx_unlock(&epoch->e_drain_mtx);
 928         }
 929 }
 930
 931 void
 932 epoch_drain_callbacks(epoch_t epoch)
 933 {
 934         epoch_record_t er;
 935         struct thread *td;
 936         int was_bound;
 937         int old_pinned;
 938         int old_cpu;
 939         int cpu;
 940
 941         WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 942             "epoch_drain_callbacks() may sleep!");
 943
 944         /* too early in boot to have epoch set up */
 945         if (__predict_false(epoch == NULL))
 946                 return;
 947 #if !defined(EARLY_AP_STARTUP)
 948         if (__predict_false(inited < 2))
 949                 return;
 950 #endif
 951         DROP_GIANT();
 952
 953         sx_xlock(&epoch->e_drain_sx);
 954         mtx_lock(&epoch->e_drain_mtx);
 955
 956         td = curthread;
 957         thread_lock(td);
 958         old_cpu = PCPU_GET(cpuid);
 959         old_pinned = td->td_pinned;
 960         was_bound = sched_is_bound(td);
 961         sched_unbind(td);
 962         td->td_pinned = 0;
 963
 964         CPU_FOREACH(cpu)
 965                 epoch->e_drain_count++;
 966         CPU_FOREACH(cpu) {
 967                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 968                 sched_bind(td, cpu);
 969                 epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
 970         }
 971
 972         /* restore CPU binding, if any */
 973         if (was_bound != 0) {
 974                 sched_bind(td, old_cpu);
 975         } else {
 976                 /* get thread back to initial CPU, if any */
 977                 if (old_pinned != 0)
 978                         sched_bind(td, old_cpu);
 979                 sched_unbind(td);
 980         }
 981         /* restore pinned after bind */
 982         td->td_pinned = old_pinned;
 983
 984         thread_unlock(td);
 985
 986         while (epoch->e_drain_count != 0)
 987                 msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
 988
 989         mtx_unlock(&epoch->e_drain_mtx);
 990         sx_xunlock(&epoch->e_drain_sx);
 991
 992         PICKUP_GIANT();
 993 }