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