sys/kern/subr_epoch.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   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 #include <sys/param.h>
  31 #include <sys/systm.h>
  32 #include <sys/counter.h>
  33 #include <sys/epoch.h>
  34 #include <sys/gtaskqueue.h>
  35 #include <sys/kernel.h>
  36 #include <sys/limits.h>
  37 #include <sys/lock.h>
  38 #include <sys/malloc.h>
  39 #include <sys/mutex.h>
  40 #include <sys/pcpu.h>
  41 #include <sys/proc.h>
  42 #include <sys/sched.h>
  43 #include <sys/sx.h>
  44 #include <sys/smp.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/turnstile.h>
  47 #ifdef EPOCH_TRACE
  48 #include <machine/stdarg.h>
  49 #include <sys/stack.h>
  50 #include <sys/tree.h>
  51 #endif
  52 #include <vm/vm.h>
  53 #include <vm/vm_extern.h>
  54 #include <vm/vm_kern.h>
  55 #include <vm/uma.h>
  56
  57 #include <machine/stack.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 static epoch_record_t
 151 epoch_currecord(epoch_t epoch)
 152 {
 153
 154         return (zpcpu_get(epoch->e_pcpu_record));
 155 }
 156
 157 #ifdef EPOCH_TRACE
 158 struct stackentry {
 159         RB_ENTRY(stackentry) se_node;
 160         struct stack se_stack;
 161 };
 162
 163 static int
 164 stackentry_compare(struct stackentry *a, struct stackentry *b)
 165 {
 166
 167         if (a->se_stack.depth > b->se_stack.depth)
 168                 return (1);
 169         if (a->se_stack.depth < b->se_stack.depth)
 170                 return (-1);
 171         for (int i = 0; i < a->se_stack.depth; i++) {
 172                 if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
 173                         return (1);
 174                 if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
 175                         return (-1);
 176         }
 177
 178         return (0);
 179 }
 180
 181 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
 182 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
 183
 184 static struct mtx epoch_stacks_lock;
 185 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
 186
 187 static bool epoch_trace_stack_print = true;
 188 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
 189     &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
 190
 191 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
 192 static inline void
 193 epoch_trace_report(const char *fmt, ...)
 194 {
 195         va_list ap;
 196         struct stackentry se, *new;
 197
 198         stack_save(&se.se_stack);
 199
 200         /* Tree is never reduced - go lockless. */
 201         if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
 202                 return;
 203
 204         new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
 205         if (new != NULL) {
 206                 bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
 207
 208                 mtx_lock(&epoch_stacks_lock);
 209                 new = RB_INSERT(stacktree, &epoch_stacks, new);
 210                 mtx_unlock(&epoch_stacks_lock);
 211                 if (new != NULL)
 212                         free(new, M_STACK);
 213         }
 214
 215         va_start(ap, fmt);
 216         (void)vprintf(fmt, ap);
 217         va_end(ap);
 218         if (epoch_trace_stack_print)
 219                 stack_print_ddb(&se.se_stack);
 220 }
 221
 222 static inline void
 223 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
 224     const char *file, int line)
 225 {
 226         epoch_tracker_t iet;
 227
 228         SLIST_FOREACH(iet, &td->td_epochs, et_tlink) {
 229                 if (iet->et_epoch != epoch)
 230                         continue;
 231                 epoch_trace_report("Recursively entering epoch %s "
 232                     "at %s:%d, previously entered at %s:%d\n",
 233                     epoch->e_name, file, line,
 234                     iet->et_file, iet->et_line);
 235         }
 236         et->et_epoch = epoch;
 237         et->et_file = file;
 238         et->et_line = line;
 239         et->et_flags = 0;
 240         SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
 241 }
 242
 243 static inline void
 244 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
 245     const char *file, int line)
 246 {
 247
 248         if (SLIST_FIRST(&td->td_epochs) != et) {
 249                 epoch_trace_report("Exiting epoch %s in a not nested order "
 250                     "at %s:%d. Most recently entered %s at %s:%d\n",
 251                     epoch->e_name,
 252                     file, line,
 253                     SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
 254                     SLIST_FIRST(&td->td_epochs)->et_file,
 255                     SLIST_FIRST(&td->td_epochs)->et_line);
 256                 /* This will panic if et is not anywhere on td_epochs. */
 257                 SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
 258         } else
 259                 SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
 260         if (et->et_flags & ET_REPORT_EXIT)
 261                 printf("Td %p exiting epoch %s at %s:%d\n", td, epoch->e_name,
 262                     file, line);
 263 }
 264
 265 /* Used by assertions that check thread state before going to sleep. */
 266 void
 267 epoch_trace_list(struct thread *td)
 268 {
 269         epoch_tracker_t iet;
 270
 271         SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
 272                 printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
 273                     iet->et_file, iet->et_line);
 274 }
 275
 276 void
 277 epoch_where_report(epoch_t epoch)
 278 {
 279         epoch_record_t er;
 280         struct epoch_tracker *tdwait;
 281
 282         MPASS(epoch != NULL);
 283         MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
 284         MPASS(!THREAD_CAN_SLEEP());
 285         critical_enter();
 286         er = epoch_currecord(epoch);
 287         TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 288                 if (tdwait->et_td == curthread)
 289                         break;
 290         critical_exit();
 291         if (tdwait != NULL) {
 292                 tdwait->et_flags |= ET_REPORT_EXIT;
 293                 printf("Td %p entered epoch %s at %s:%d\n", curthread,
 294                     epoch->e_name, tdwait->et_file, tdwait->et_line);
 295         }
 296 }
 297 #endif /* EPOCH_TRACE */
 298
 299 static void
 300 epoch_init(void *arg __unused)
 301 {
 302         int cpu;
 303
 304         block_count = counter_u64_alloc(M_WAITOK);
 305         migrate_count = counter_u64_alloc(M_WAITOK);
 306         turnstile_count = counter_u64_alloc(M_WAITOK);
 307         switch_count = counter_u64_alloc(M_WAITOK);
 308         epoch_call_count = counter_u64_alloc(M_WAITOK);
 309         epoch_call_task_count = counter_u64_alloc(M_WAITOK);
 310
 311         pcpu_zone_record = uma_zcreate("epoch_record pcpu",
 312             sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
 313             UMA_ALIGN_PTR, UMA_ZONE_PCPU);
 314         CPU_FOREACH(cpu) {
 315                 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
 316                     epoch_call_task, NULL);
 317                 taskqgroup_attach_cpu(qgroup_softirq,
 318                     DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
 319                     "epoch call task");
 320         }
 321 #ifdef EPOCH_TRACE
 322         SLIST_INIT(&thread0.td_epochs);
 323 #endif
 324         sx_init(&epoch_sx, "epoch-sx");
 325         inited = 1;
 326         global_epoch = epoch_alloc("Global", 0);
 327         global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
 328 }
 329 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
 330
 331 #if !defined(EARLY_AP_STARTUP)
 332 static void
 333 epoch_init_smp(void *dummy __unused)
 334 {
 335         inited = 2;
 336 }
 337 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
 338 #endif
 339
 340 static void
 341 epoch_ctor(epoch_t epoch)
 342 {
 343         epoch_record_t er;
 344         int cpu;
 345
 346         epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
 347         CPU_FOREACH(cpu) {
 348                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 349                 bzero(er, sizeof(*er));
 350                 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 351                 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 352                 er->er_cpuid = cpu;
 353                 er->er_parent = epoch;
 354         }
 355 }
 356
 357 static void
 358 epoch_adjust_prio(struct thread *td, u_char prio)
 359 {
 360
 361         thread_lock(td);
 362         sched_prio(td, prio);
 363         thread_unlock(td);
 364 }
 365
 366 epoch_t
 367 epoch_alloc(const char *name, int flags)
 368 {
 369         epoch_t epoch;
 370         int i;
 371
 372         MPASS(name != NULL);
 373
 374         if (__predict_false(!inited))
 375                 panic("%s called too early in boot", __func__);
 376
 377         EPOCH_LOCK();
 378
 379         /*
 380          * Find a free index in the epoch array. If no free index is
 381          * found, try to use the index after the last one.
 382          */
 383         for (i = 0;; i++) {
 384                 /*
 385                  * If too many epochs are currently allocated,
 386                  * return NULL.
 387                  */
 388                 if (i == MAX_EPOCHS) {
 389                         epoch = NULL;
 390                         goto done;
 391                 }
 392                 if (epoch_array[i].e_in_use == 0)
 393                         break;
 394         }
 395
 396         epoch = epoch_array + i;
 397         ck_epoch_init(&epoch->e_epoch);
 398         epoch_ctor(epoch);
 399         epoch->e_flags = flags;
 400         epoch->e_name = name;
 401         sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
 402         mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
 403
 404         /*
 405          * Set e_in_use last, because when this field is set the
 406          * epoch_call_task() function will start scanning this epoch
 407          * structure.
 408          */
 409         atomic_store_rel_int(&epoch->e_in_use, 1);
 410 done:
 411         EPOCH_UNLOCK();
 412         return (epoch);
 413 }
 414
 415 void
 416 epoch_free(epoch_t epoch)
 417 {
 418 #ifdef INVARIANTS
 419         int cpu;
 420 #endif
 421
 422         EPOCH_LOCK();
 423
 424         MPASS(epoch->e_in_use != 0);
 425
 426         epoch_drain_callbacks(epoch);
 427
 428         atomic_store_rel_int(&epoch->e_in_use, 0);
 429         /*
 430          * Make sure the epoch_call_task() function see e_in_use equal
 431          * to zero, by calling epoch_wait() on the global_epoch:
 432          */
 433         epoch_wait(global_epoch);
 434 #ifdef INVARIANTS
 435         CPU_FOREACH(cpu) {
 436                 epoch_record_t er;
 437
 438                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 439
 440                 /*
 441                  * Sanity check: none of the records should be in use anymore.
 442                  * We drained callbacks above and freeing the pcpu records is
 443                  * imminent.
 444                  */
 445                 MPASS(er->er_td == NULL);
 446                 MPASS(TAILQ_EMPTY(&er->er_tdlist));
 447         }
 448 #endif
 449         uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
 450         mtx_destroy(&epoch->e_drain_mtx);
 451         sx_destroy(&epoch->e_drain_sx);
 452         memset(epoch, 0, sizeof(*epoch));
 453
 454         EPOCH_UNLOCK();
 455 }
 456
 457 #define INIT_CHECK(epoch)                                       \
 458         do {                                                    \
 459                 if (__predict_false((epoch) == NULL))           \
 460                         return;                                 \
 461         } while (0)
 462
 463 void
 464 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
 465 {
 466         struct epoch_record *er;
 467         struct thread *td;
 468
 469         MPASS(cold || epoch != NULL);
 470         td = curthread;
 471         MPASS(kstack_contains(td, (vm_offset_t)et, sizeof(*et)));
 472
 473         INIT_CHECK(epoch);
 474         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 475
 476 #ifdef EPOCH_TRACE
 477         epoch_trace_enter(td, epoch, et, file, line);
 478 #endif
 479         et->et_td = td;
 480         THREAD_NO_SLEEPING();
 481         critical_enter();
 482         sched_pin();
 483         et->et_old_priority = td->td_priority;
 484         er = epoch_currecord(epoch);
 485         /* Record-level tracking is reserved for non-preemptible epochs. */
 486         MPASS(er->er_td == NULL);
 487         TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
 488         ck_epoch_begin(&er->er_record, &et->et_section);
 489         critical_exit();
 490 }
 491
 492 void
 493 epoch_enter(epoch_t epoch)
 494 {
 495         epoch_record_t er;
 496
 497         MPASS(cold || epoch != NULL);
 498         INIT_CHECK(epoch);
 499         critical_enter();
 500         er = epoch_currecord(epoch);
 501 #ifdef INVARIANTS
 502         if (er->er_record.active == 0) {
 503                 MPASS(er->er_td == NULL);
 504                 er->er_td = curthread;
 505         } else {
 506                 /* We've recursed, just make sure our accounting isn't wrong. */
 507                 MPASS(er->er_td == curthread);
 508         }
 509 #endif
 510         ck_epoch_begin(&er->er_record, NULL);
 511 }
 512
 513 void
 514 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
 515 {
 516         struct epoch_record *er;
 517         struct thread *td;
 518
 519         INIT_CHECK(epoch);
 520         td = curthread;
 521         critical_enter();
 522         sched_unpin();
 523         THREAD_SLEEPING_OK();
 524         er = epoch_currecord(epoch);
 525         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 526         MPASS(et != NULL);
 527         MPASS(et->et_td == td);
 528 #ifdef INVARIANTS
 529         et->et_td = (void*)0xDEADBEEF;
 530         /* Record-level tracking is reserved for non-preemptible epochs. */
 531         MPASS(er->er_td == NULL);
 532 #endif
 533         ck_epoch_end(&er->er_record, &et->et_section);
 534         TAILQ_REMOVE(&er->er_tdlist, et, et_link);
 535         er->er_gen++;
 536         if (__predict_false(et->et_old_priority != td->td_priority))
 537                 epoch_adjust_prio(td, et->et_old_priority);
 538         critical_exit();
 539 #ifdef EPOCH_TRACE
 540         epoch_trace_exit(td, epoch, et, file, line);
 541 #endif
 542 }
 543
 544 void
 545 epoch_exit(epoch_t epoch)
 546 {
 547         epoch_record_t er;
 548
 549         INIT_CHECK(epoch);
 550         er = epoch_currecord(epoch);
 551         ck_epoch_end(&er->er_record, NULL);
 552 #ifdef INVARIANTS
 553         MPASS(er->er_td == curthread);
 554         if (er->er_record.active == 0)
 555                 er->er_td = NULL;
 556 #endif
 557         critical_exit();
 558 }
 559
 560 /*
 561  * epoch_block_handler_preempt() is a callback from the CK code when another
 562  * thread is currently in an epoch section.
 563  */
 564 static void
 565 epoch_block_handler_preempt(struct ck_epoch *global __unused,
 566     ck_epoch_record_t *cr, void *arg __unused)
 567 {
 568         epoch_record_t record;
 569         struct thread *td, *owner, *curwaittd;
 570         struct epoch_tracker *tdwait;
 571         struct turnstile *ts;
 572         struct lock_object *lock;
 573         int spincount, gen;
 574         int locksheld __unused;
 575
 576         record = __containerof(cr, struct epoch_record, er_record);
 577         td = curthread;
 578         locksheld = td->td_locks;
 579         spincount = 0;
 580         counter_u64_add(block_count, 1);
 581         /*
 582          * We lost a race and there's no longer any threads
 583          * on the CPU in an epoch section.
 584          */
 585         if (TAILQ_EMPTY(&record->er_tdlist))
 586                 return;
 587
 588         if (record->er_cpuid != curcpu) {
 589                 /*
 590                  * If the head of the list is running, we can wait for it
 591                  * to remove itself from the list and thus save us the
 592                  * overhead of a migration
 593                  */
 594                 gen = record->er_gen;
 595                 thread_unlock(td);
 596                 /*
 597                  * We can't actually check if the waiting thread is running
 598                  * so we simply poll for it to exit before giving up and
 599                  * migrating.
 600                  */
 601                 do {
 602                         cpu_spinwait();
 603                 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
 604                                  gen == record->er_gen &&
 605                                  spincount++ < MAX_ADAPTIVE_SPIN);
 606                 thread_lock(td);
 607                 /*
 608                  * If the generation has changed we can poll again
 609                  * otherwise we need to migrate.
 610                  */
 611                 if (gen != record->er_gen)
 612                         return;
 613                 /*
 614                  * Being on the same CPU as that of the record on which
 615                  * we need to wait allows us access to the thread
 616                  * list associated with that CPU. We can then examine the
 617                  * oldest thread in the queue and wait on its turnstile
 618                  * until it resumes and so on until a grace period
 619                  * elapses.
 620                  *
 621                  */
 622                 counter_u64_add(migrate_count, 1);
 623                 sched_bind(td, record->er_cpuid);
 624                 /*
 625                  * At this point we need to return to the ck code
 626                  * to scan to see if a grace period has elapsed.
 627                  * We can't move on to check the thread list, because
 628                  * in the meantime new threads may have arrived that
 629                  * in fact belong to a different epoch.
 630                  */
 631                 return;
 632         }
 633         /*
 634          * Try to find a thread in an epoch section on this CPU
 635          * waiting on a turnstile. Otherwise find the lowest
 636          * priority thread (highest prio value) and drop our priority
 637          * to match to allow it to run.
 638          */
 639         TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
 640                 /*
 641                  * Propagate our priority to any other waiters to prevent us
 642                  * from starving them. They will have their original priority
 643                  * restore on exit from epoch_wait().
 644                  */
 645                 curwaittd = tdwait->et_td;
 646                 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
 647                         critical_enter();
 648                         thread_unlock(td);
 649                         thread_lock(curwaittd);
 650                         sched_prio(curwaittd, td->td_priority);
 651                         thread_unlock(curwaittd);
 652                         thread_lock(td);
 653                         critical_exit();
 654                 }
 655                 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
 656                     ((ts = curwaittd->td_blocked) != NULL)) {
 657                         /*
 658                          * We unlock td to allow turnstile_wait to reacquire
 659                          * the thread lock. Before unlocking it we enter a
 660                          * critical section to prevent preemption after we
 661                          * reenable interrupts by dropping the thread lock in
 662                          * order to prevent curwaittd from getting to run.
 663                          */
 664                         critical_enter();
 665                         thread_unlock(td);
 666
 667                         if (turnstile_lock(ts, &lock, &owner)) {
 668                                 if (ts == curwaittd->td_blocked) {
 669                                         MPASS(TD_IS_INHIBITED(curwaittd) &&
 670                                             TD_ON_LOCK(curwaittd));
 671                                         critical_exit();
 672                                         turnstile_wait(ts, owner,
 673                                             curwaittd->td_tsqueue);
 674                                         counter_u64_add(turnstile_count, 1);
 675                                         thread_lock(td);
 676                                         return;
 677                                 }
 678                                 turnstile_unlock(ts, lock);
 679                         }
 680                         thread_lock(td);
 681                         critical_exit();
 682                         KASSERT(td->td_locks == locksheld,
 683                             ("%d extra locks held", td->td_locks - locksheld));
 684                 }
 685         }
 686         /*
 687          * We didn't find any threads actually blocked on a lock
 688          * so we have nothing to do except context switch away.
 689          */
 690         counter_u64_add(switch_count, 1);
 691         mi_switch(SW_VOL | SWT_RELINQUISH);
 692         /*
 693          * It is important the thread lock is dropped while yielding
 694          * to allow other threads to acquire the lock pointed to by
 695          * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
 696          * thread lock before returning. Else a deadlock like
 697          * situation might happen.
 698          */
 699         thread_lock(td);
 700 }
 701
 702 void
 703 epoch_wait_preempt(epoch_t epoch)
 704 {
 705         struct thread *td;
 706         int was_bound;
 707         int old_cpu;
 708         int old_pinned;
 709         u_char old_prio;
 710         int locks __unused;
 711
 712         MPASS(cold || epoch != NULL);
 713         INIT_CHECK(epoch);
 714         td = curthread;
 715 #ifdef INVARIANTS
 716         locks = curthread->td_locks;
 717         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 718         if ((epoch->e_flags & EPOCH_LOCKED) == 0)
 719                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 720                     "epoch_wait() can be long running");
 721         KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
 722             "of an epoch section of the same epoch"));
 723 #endif
 724         DROP_GIANT();
 725         thread_lock(td);
 726
 727         old_cpu = PCPU_GET(cpuid);
 728         old_pinned = td->td_pinned;
 729         old_prio = td->td_priority;
 730         was_bound = sched_is_bound(td);
 731         sched_unbind(td);
 732         td->td_pinned = 0;
 733         sched_bind(td, old_cpu);
 734
 735         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
 736             NULL);
 737
 738         /* restore CPU binding, if any */
 739         if (was_bound != 0) {
 740                 sched_bind(td, old_cpu);
 741         } else {
 742                 /* get thread back to initial CPU, if any */
 743                 if (old_pinned != 0)
 744                         sched_bind(td, old_cpu);
 745                 sched_unbind(td);
 746         }
 747         /* restore pinned after bind */
 748         td->td_pinned = old_pinned;
 749
 750         /* restore thread priority */
 751         sched_prio(td, old_prio);
 752         thread_unlock(td);
 753         PICKUP_GIANT();
 754         KASSERT(td->td_locks == locks,
 755             ("%d residual locks held", td->td_locks - locks));
 756 }
 757
 758 static void
 759 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
 760     void *arg __unused)
 761 {
 762         cpu_spinwait();
 763 }
 764
 765 void
 766 epoch_wait(epoch_t epoch)
 767 {
 768
 769         MPASS(cold || epoch != NULL);
 770         INIT_CHECK(epoch);
 771         MPASS(epoch->e_flags == 0);
 772         critical_enter();
 773         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
 774         critical_exit();
 775 }
 776
 777 void
 778 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
 779 {
 780         epoch_record_t er;
 781         ck_epoch_entry_t *cb;
 782
 783         cb = (void *)ctx;
 784
 785         MPASS(callback);
 786         /* too early in boot to have epoch set up */
 787         if (__predict_false(epoch == NULL))
 788                 goto boottime;
 789 #if !defined(EARLY_AP_STARTUP)
 790         if (__predict_false(inited < 2))
 791                 goto boottime;
 792 #endif
 793
 794         critical_enter();
 795         *DPCPU_PTR(epoch_cb_count) += 1;
 796         er = epoch_currecord(epoch);
 797         ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
 798         critical_exit();
 799         return;
 800 boottime:
 801         callback(ctx);
 802 }
 803
 804 static void
 805 epoch_call_task(void *arg __unused)
 806 {
 807         ck_stack_entry_t *cursor, *head, *next;
 808         ck_epoch_record_t *record;
 809         epoch_record_t er;
 810         epoch_t epoch;
 811         ck_stack_t cb_stack;
 812         int i, npending, total;
 813
 814         ck_stack_init(&cb_stack);
 815         critical_enter();
 816         epoch_enter(global_epoch);
 817         for (total = i = 0; i != MAX_EPOCHS; i++) {
 818                 epoch = epoch_array + i;
 819                 if (__predict_false(
 820                     atomic_load_acq_int(&epoch->e_in_use) == 0))
 821                         continue;
 822                 er = epoch_currecord(epoch);
 823                 record = &er->er_record;
 824                 if ((npending = record->n_pending) == 0)
 825                         continue;
 826                 ck_epoch_poll_deferred(record, &cb_stack);
 827                 total += npending - record->n_pending;
 828         }
 829         epoch_exit(global_epoch);
 830         *DPCPU_PTR(epoch_cb_count) -= total;
 831         critical_exit();
 832
 833         counter_u64_add(epoch_call_count, total);
 834         counter_u64_add(epoch_call_task_count, 1);
 835
 836         head = ck_stack_batch_pop_npsc(&cb_stack);
 837         for (cursor = head; cursor != NULL; cursor = next) {
 838                 struct ck_epoch_entry *entry =
 839                     ck_epoch_entry_container(cursor);
 840
 841                 next = CK_STACK_NEXT(cursor);
 842                 entry->function(entry);
 843         }
 844 }
 845
 846 static int
 847 in_epoch_verbose_preempt(epoch_t epoch, int dump_onfail)
 848 {
 849         epoch_record_t er;
 850         struct epoch_tracker *tdwait;
 851         struct thread *td;
 852
 853         MPASS(epoch != NULL);
 854         MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
 855         td = curthread;
 856         if (THREAD_CAN_SLEEP())
 857                 return (0);
 858         critical_enter();
 859         er = epoch_currecord(epoch);
 860         TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 861                 if (tdwait->et_td == td) {
 862                         critical_exit();
 863                         return (1);
 864                 }
 865 #ifdef INVARIANTS
 866         if (dump_onfail) {
 867                 MPASS(td->td_pinned);
 868                 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
 869                 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 870                         printf("td_tid: %d ", tdwait->et_td->td_tid);
 871                 printf("\n");
 872         }
 873 #endif
 874         critical_exit();
 875         return (0);
 876 }
 877
 878 #ifdef INVARIANTS
 879 static void
 880 epoch_assert_nocpu(epoch_t epoch, struct thread *td)
 881 {
 882         epoch_record_t er;
 883         int cpu;
 884         bool crit;
 885
 886         crit = td->td_critnest > 0;
 887
 888         /* Check for a critical section mishap. */
 889         CPU_FOREACH(cpu) {
 890                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 891                 KASSERT(er->er_td != td,
 892                     ("%s critical section in epoch '%s', from cpu %d",
 893                     (crit ? "exited" : "re-entered"), epoch->e_name, cpu));
 894         }
 895 }
 896 #else
 897 #define epoch_assert_nocpu(e, td) do {} while (0)
 898 #endif
 899
 900 int
 901 in_epoch_verbose(epoch_t epoch, int dump_onfail)
 902 {
 903         epoch_record_t er;
 904         struct thread *td;
 905
 906         if (__predict_false((epoch) == NULL))
 907                 return (0);
 908         if ((epoch->e_flags & EPOCH_PREEMPT) != 0)
 909                 return (in_epoch_verbose_preempt(epoch, dump_onfail));
 910
 911         /*
 912          * The thread being in a critical section is a necessary
 913          * condition to be correctly inside a non-preemptible epoch,
 914          * so it's definitely not in this epoch.
 915          */
 916         td = curthread;
 917         if (td->td_critnest == 0) {
 918                 epoch_assert_nocpu(epoch, td);
 919                 return (0);
 920         }
 921
 922         /*
 923          * The current cpu is in a critical section, so the epoch record will be
 924          * stable for the rest of this function.  Knowing that the record is not
 925          * active is sufficient for knowing whether we're in this epoch or not,
 926          * since it's a pcpu record.
 927          */
 928         er = epoch_currecord(epoch);
 929         if (er->er_record.active == 0) {
 930                 epoch_assert_nocpu(epoch, td);
 931                 return (0);
 932         }
 933
 934         MPASS(er->er_td == td);
 935         return (1);
 936 }
 937
 938 int
 939 in_epoch(epoch_t epoch)
 940 {
 941         return (in_epoch_verbose(epoch, 0));
 942 }
 943
 944 static void
 945 epoch_drain_cb(struct epoch_context *ctx)
 946 {
 947         struct epoch *epoch =
 948             __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
 949
 950         if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
 951                 mtx_lock(&epoch->e_drain_mtx);
 952                 wakeup(epoch);
 953                 mtx_unlock(&epoch->e_drain_mtx);
 954         }
 955 }
 956
 957 void
 958 epoch_drain_callbacks(epoch_t epoch)
 959 {
 960         epoch_record_t er;
 961         struct thread *td;
 962         int was_bound;
 963         int old_pinned;
 964         int old_cpu;
 965         int cpu;
 966
 967         WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 968             "epoch_drain_callbacks() may sleep!");
 969
 970         /* too early in boot to have epoch set up */
 971         if (__predict_false(epoch == NULL))
 972                 return;
 973 #if !defined(EARLY_AP_STARTUP)
 974         if (__predict_false(inited < 2))
 975                 return;
 976 #endif
 977         DROP_GIANT();
 978
 979         sx_xlock(&epoch->e_drain_sx);
 980         mtx_lock(&epoch->e_drain_mtx);
 981
 982         td = curthread;
 983         thread_lock(td);
 984         old_cpu = PCPU_GET(cpuid);
 985         old_pinned = td->td_pinned;
 986         was_bound = sched_is_bound(td);
 987         sched_unbind(td);
 988         td->td_pinned = 0;
 989
 990         CPU_FOREACH(cpu)
 991                 epoch->e_drain_count++;
 992         CPU_FOREACH(cpu) {
 993                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 994                 sched_bind(td, cpu);
 995                 epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
 996         }
 997
 998         /* restore CPU binding, if any */
 999         if (was_bound != 0) {
1000                 sched_bind(td, old_cpu);
1001         } else {
1002                 /* get thread back to initial CPU, if any */
1003                 if (old_pinned != 0)
1004                         sched_bind(td, old_cpu);
1005                 sched_unbind(td);
1006         }
1007         /* restore pinned after bind */
1008         td->td_pinned = old_pinned;
1009
1010         thread_unlock(td);
1011
1012         while (epoch->e_drain_count != 0)
1013                 msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
1014
1015         mtx_unlock(&epoch->e_drain_mtx);
1016         sx_xunlock(&epoch->e_drain_sx);
1017
1018         PICKUP_GIANT();
1019 }