sys/kern/subr_epoch.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2018, Matthew Macy <mmacy@freebsd.org>
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  *
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/types.h>
  34 #include <sys/systm.h>
  35 #include <sys/counter.h>
  36 #include <sys/epoch.h>
  37 #include <sys/gtaskqueue.h>
  38 #include <sys/kernel.h>
  39 #include <sys/limits.h>
  40 #include <sys/lock.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mutex.h>
  43 #include <sys/pcpu.h>
  44 #include <sys/proc.h>
  45 #include <sys/sched.h>
  46 #include <sys/smp.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/turnstile.h>
  49 #include <vm/vm.h>
  50 #include <vm/vm_extern.h>
  51 #include <vm/vm_kern.h>
  52 #include <vm/uma.h>
  53
  54 #include <ck_epoch.h>
  55
  56 static MALLOC_DEFINE(M_EPOCH, "epoch", "epoch based reclamation");
  57
  58 /* arbitrary --- needs benchmarking */
  59 #define MAX_ADAPTIVE_SPIN 1000
  60 #define MAX_EPOCHS 64
  61
  62 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
  63 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW, 0, "epoch information");
  64 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW, 0, "epoch stats");
  65
  66 /* Stats. */
  67 static counter_u64_t block_count;
  68
  69 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
  70     &block_count, "# of times a thread was in an epoch when epoch_wait was called");
  71 static counter_u64_t migrate_count;
  72
  73 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
  74     &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
  75 static counter_u64_t turnstile_count;
  76
  77 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
  78     &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
  79 static counter_u64_t switch_count;
  80
  81 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
  82     &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
  83 static counter_u64_t epoch_call_count;
  84
  85 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
  86     &epoch_call_count, "# of times a callback was deferred");
  87 static counter_u64_t epoch_call_task_count;
  88
  89 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
  90     &epoch_call_task_count, "# of times a callback task was run");
  91
  92 TAILQ_HEAD (threadlist, thread);
  93
  94 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
  95     ck_epoch_entry_container)
  96
  97 epoch_t allepochs[MAX_EPOCHS];
  98
  99 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
 100 DPCPU_DEFINE(int, epoch_cb_count);
 101
 102 static __read_mostly int inited;
 103 static __read_mostly int epoch_count;
 104 __read_mostly epoch_t global_epoch;
 105 __read_mostly epoch_t global_epoch_preempt;
 106
 107 static void epoch_call_task(void *context __unused);
 108 static  uma_zone_t pcpu_zone_record;
 109
 110 static void
 111 epoch_init(void *arg __unused)
 112 {
 113         int cpu;
 114
 115         block_count = counter_u64_alloc(M_WAITOK);
 116         migrate_count = counter_u64_alloc(M_WAITOK);
 117         turnstile_count = counter_u64_alloc(M_WAITOK);
 118         switch_count = counter_u64_alloc(M_WAITOK);
 119         epoch_call_count = counter_u64_alloc(M_WAITOK);
 120         epoch_call_task_count = counter_u64_alloc(M_WAITOK);
 121
 122         pcpu_zone_record = uma_zcreate("epoch_record pcpu", sizeof(struct epoch_record),
 123             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
 124         CPU_FOREACH(cpu) {
 125                 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0, epoch_call_task, NULL);
 126                 taskqgroup_attach_cpu(qgroup_softirq, DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, -1, "epoch call task");
 127         }
 128         inited = 1;
 129         global_epoch = epoch_alloc(0);
 130         global_epoch_preempt = epoch_alloc(EPOCH_PREEMPT);
 131 }
 132 SYSINIT(epoch, SI_SUB_TASKQ + 1, SI_ORDER_FIRST, epoch_init, NULL);
 133
 134 #if !defined(EARLY_AP_STARTUP)
 135 static void
 136 epoch_init_smp(void *dummy __unused)
 137 {
 138         inited = 2;
 139 }
 140 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
 141 #endif
 142
 143 static void
 144 epoch_ctor(epoch_t epoch)
 145 {
 146         epoch_record_t er;
 147         int cpu;
 148
 149         epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
 150         CPU_FOREACH(cpu) {
 151                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 152                 bzero(er, sizeof(*er));
 153                 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
 154                 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
 155                 er->er_cpuid = cpu;
 156         }
 157 }
 158
 159 epoch_t
 160 epoch_alloc(int flags)
 161 {
 162         epoch_t epoch;
 163
 164         if (__predict_false(!inited))
 165                 panic("%s called too early in boot", __func__);
 166         epoch = malloc(sizeof(struct epoch), M_EPOCH, M_ZERO | M_WAITOK);
 167         ck_epoch_init(&epoch->e_epoch);
 168         epoch_ctor(epoch);
 169         MPASS(epoch_count < MAX_EPOCHS - 2);
 170         epoch->e_flags = flags;
 171         epoch->e_idx = epoch_count;
 172         allepochs[epoch_count++] = epoch;
 173         return (epoch);
 174 }
 175
 176 void
 177 epoch_free(epoch_t epoch)
 178 {
 179 #ifdef INVARIANTS
 180         struct epoch_record *er;
 181         int cpu;
 182
 183         CPU_FOREACH(cpu) {
 184                 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
 185                 MPASS(TAILQ_EMPTY(&er->er_tdlist));
 186         }
 187 #endif
 188         allepochs[epoch->e_idx] = NULL;
 189         epoch_wait(global_epoch);
 190         uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
 191         free(epoch, M_EPOCH);
 192 }
 193
 194 void
 195 epoch_enter_preempt_KBI(epoch_t epoch, epoch_tracker_t et)
 196 {
 197
 198         epoch_enter_preempt(epoch, et);
 199 }
 200
 201 void
 202 epoch_exit_preempt_KBI(epoch_t epoch, epoch_tracker_t et)
 203 {
 204
 205         epoch_exit_preempt(epoch, et);
 206 }
 207
 208 void
 209 epoch_enter_KBI(epoch_t epoch)
 210 {
 211
 212         epoch_enter(epoch);
 213 }
 214
 215 void
 216 epoch_exit_KBI(epoch_t epoch)
 217 {
 218
 219         epoch_exit(epoch);
 220 }
 221
 222 /*
 223  * epoch_block_handler_preempt is a callback from the ck code when another thread is
 224  * currently in an epoch section.
 225  */
 226 static void
 227 epoch_block_handler_preempt(struct ck_epoch *global __unused, ck_epoch_record_t *cr,
 228     void *arg __unused)
 229 {
 230         epoch_record_t record;
 231         struct thread *td, *owner, *curwaittd;
 232         struct epoch_thread *tdwait;
 233         struct turnstile *ts;
 234         struct lock_object *lock;
 235         int spincount, gen;
 236         int locksheld __unused;
 237
 238         record = __containerof(cr, struct epoch_record, er_record);
 239         td = curthread;
 240         locksheld = td->td_locks;
 241         spincount = 0;
 242         counter_u64_add(block_count, 1);
 243         if (record->er_cpuid != curcpu) {
 244                 /*
 245                  * If the head of the list is running, we can wait for it
 246                  * to remove itself from the list and thus save us the
 247                  * overhead of a migration
 248                  */
 249                 if ((tdwait = TAILQ_FIRST(&record->er_tdlist)) != NULL &&
 250                     TD_IS_RUNNING(tdwait->et_td)) {
 251                         gen = record->er_gen;
 252                         thread_unlock(td);
 253                         do {
 254                                 cpu_spinwait();
 255                         } while (tdwait == TAILQ_FIRST(&record->er_tdlist) &&
 256                             gen == record->er_gen && TD_IS_RUNNING(tdwait->et_td) &&
 257                             spincount++ < MAX_ADAPTIVE_SPIN);
 258                         thread_lock(td);
 259                         return;
 260                 }
 261                 /*
 262                  * Being on the same CPU as that of the record on which
 263                  * we need to wait allows us access to the thread
 264                  * list associated with that CPU. We can then examine the
 265                  * oldest thread in the queue and wait on its turnstile
 266                  * until it resumes and so on until a grace period
 267                  * elapses.
 268                  *
 269                  */
 270                 counter_u64_add(migrate_count, 1);
 271                 sched_bind(td, record->er_cpuid);
 272                 /*
 273                  * At this point we need to return to the ck code
 274                  * to scan to see if a grace period has elapsed.
 275                  * We can't move on to check the thread list, because
 276                  * in the meantime new threads may have arrived that
 277                  * in fact belong to a different epoch.
 278                  */
 279                 return;
 280         }
 281         /*
 282          * Try to find a thread in an epoch section on this CPU
 283          * waiting on a turnstile. Otherwise find the lowest
 284          * priority thread (highest prio value) and drop our priority
 285          * to match to allow it to run.
 286          */
 287         TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
 288                 /*
 289                  * Propagate our priority to any other waiters to prevent us
 290                  * from starving them. They will have their original priority
 291                  * restore on exit from epoch_wait().
 292                  */
 293                 curwaittd = tdwait->et_td;
 294                 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
 295                         critical_enter();
 296                         thread_unlock(td);
 297                         thread_lock(curwaittd);
 298                         sched_prio(curwaittd, td->td_priority);
 299                         thread_unlock(curwaittd);
 300                         thread_lock(td);
 301                         critical_exit();
 302                 }
 303                 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
 304                     ((ts = curwaittd->td_blocked) != NULL)) {
 305                         /*
 306                          * We unlock td to allow turnstile_wait to reacquire the
 307                          * the thread lock. Before unlocking it we enter a critical
 308                          * section to prevent preemption after we reenable interrupts
 309                          * by dropping the thread lock in order to prevent curwaittd
 310                          * from getting to run.
 311                          */
 312                         critical_enter();
 313                         thread_unlock(td);
 314                         owner = turnstile_lock(ts, &lock);
 315                         /*
 316                          * The owner pointer indicates that the lock succeeded. Only
 317                          * in case we hold the lock and the turnstile we locked is still
 318                          * the one that curwaittd is blocked on can we continue. Otherwise
 319                          * The turnstile pointer has been changed out from underneath
 320                          * us, as in the case where the lock holder has signalled curwaittd,
 321                          * and we need to continue.
 322                          */
 323                         if (owner != NULL && ts == curwaittd->td_blocked) {
 324                                 MPASS(TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd));
 325                                 critical_exit();
 326                                 turnstile_wait(ts, owner, curwaittd->td_tsqueue);
 327                                 counter_u64_add(turnstile_count, 1);
 328                                 thread_lock(td);
 329                                 return;
 330                         } else if (owner != NULL)
 331                                 turnstile_unlock(ts, lock);
 332                         thread_lock(td);
 333                         critical_exit();
 334                         KASSERT(td->td_locks == locksheld,
 335                             ("%d extra locks held", td->td_locks - locksheld));
 336                 }
 337         }
 338         /*
 339          * We didn't find any threads actually blocked on a lock
 340          * so we have nothing to do except context switch away.
 341          */
 342         counter_u64_add(switch_count, 1);
 343         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 344
 345         /*
 346          * Release the thread lock while yielding to
 347          * allow other threads to acquire the lock
 348          * pointed to by TDQ_LOCKPTR(td). Else a
 349          * deadlock like situation might happen. (HPS)
 350          */
 351         thread_unlock(td);
 352         thread_lock(td);
 353 }
 354
 355 void
 356 epoch_wait_preempt(epoch_t epoch)
 357 {
 358         struct thread *td;
 359         int was_bound;
 360         int old_cpu;
 361         int old_pinned;
 362         u_char old_prio;
 363         int locks __unused;
 364
 365         MPASS(cold || epoch != NULL);
 366         INIT_CHECK(epoch);
 367         td = curthread;
 368 #ifdef INVARIANTS
 369         locks = curthread->td_locks;
 370         MPASS(epoch->e_flags & EPOCH_PREEMPT);
 371         if ((epoch->e_flags & EPOCH_LOCKED) == 0)
 372                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
 373                     "epoch_wait() can be long running");
 374         KASSERT(!in_epoch(epoch),
 375                         ("epoch_wait_preempt() called in the middle "
 376                          "of an epoch section of the same epoch"));
 377 #endif
 378         thread_lock(td);
 379         DROP_GIANT();
 380
 381         old_cpu = PCPU_GET(cpuid);
 382         old_pinned = td->td_pinned;
 383         old_prio = td->td_priority;
 384         was_bound = sched_is_bound(td);
 385         sched_unbind(td);
 386         td->td_pinned = 0;
 387         sched_bind(td, old_cpu);
 388
 389         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt, NULL);
 390
 391         /* restore CPU binding, if any */
 392         if (was_bound != 0) {
 393                 sched_bind(td, old_cpu);
 394         } else {
 395                 /* get thread back to initial CPU, if any */
 396                 if (old_pinned != 0)
 397                         sched_bind(td, old_cpu);
 398                 sched_unbind(td);
 399         }
 400         /* restore pinned after bind */
 401         td->td_pinned = old_pinned;
 402
 403         /* restore thread priority */
 404         sched_prio(td, old_prio);
 405         thread_unlock(td);
 406         PICKUP_GIANT();
 407         KASSERT(td->td_locks == locks,
 408             ("%d residual locks held", td->td_locks - locks));
 409 }
 410
 411 static void
 412 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
 413     void *arg __unused)
 414 {
 415         cpu_spinwait();
 416 }
 417
 418 void
 419 epoch_wait(epoch_t epoch)
 420 {
 421
 422         MPASS(cold || epoch != NULL);
 423         INIT_CHECK(epoch);
 424         MPASS(epoch->e_flags == 0);
 425         critical_enter();
 426         ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
 427         critical_exit();
 428 }
 429
 430 void
 431 epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t))
 432 {
 433         epoch_record_t er;
 434         ck_epoch_entry_t *cb;
 435
 436         cb = (void *)ctx;
 437
 438         MPASS(callback);
 439         /* too early in boot to have epoch set up */
 440         if (__predict_false(epoch == NULL))
 441                 goto boottime;
 442 #if !defined(EARLY_AP_STARTUP)
 443         if (__predict_false(inited < 2))
 444                 goto boottime;
 445 #endif
 446
 447         critical_enter();
 448         *DPCPU_PTR(epoch_cb_count) += 1;
 449         er = epoch_currecord(epoch);
 450         ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
 451         critical_exit();
 452         return;
 453 boottime:
 454         callback(ctx);
 455 }
 456
 457 static void
 458 epoch_call_task(void *arg __unused)
 459 {
 460         ck_stack_entry_t *cursor, *head, *next;
 461         ck_epoch_record_t *record;
 462         epoch_record_t er;
 463         epoch_t epoch;
 464         ck_stack_t cb_stack;
 465         int i, npending, total;
 466
 467         ck_stack_init(&cb_stack);
 468         critical_enter();
 469         epoch_enter(global_epoch);
 470         for (total = i = 0; i < epoch_count; i++) {
 471                 if (__predict_false((epoch = allepochs[i]) == NULL))
 472                         continue;
 473                 er = epoch_currecord(epoch);
 474                 record = &er->er_record;
 475                 if ((npending = record->n_pending) == 0)
 476                         continue;
 477                 ck_epoch_poll_deferred(record, &cb_stack);
 478                 total += npending - record->n_pending;
 479         }
 480         epoch_exit(global_epoch);
 481         *DPCPU_PTR(epoch_cb_count) -= total;
 482         critical_exit();
 483
 484         counter_u64_add(epoch_call_count, total);
 485         counter_u64_add(epoch_call_task_count, 1);
 486
 487         head = ck_stack_batch_pop_npsc(&cb_stack);
 488         for (cursor = head; cursor != NULL; cursor = next) {
 489                 struct ck_epoch_entry *entry =
 490                 ck_epoch_entry_container(cursor);
 491
 492                 next = CK_STACK_NEXT(cursor);
 493                 entry->function(entry);
 494         }
 495 }
 496
 497 int
 498 in_epoch_verbose(epoch_t epoch, int dump_onfail)
 499 {
 500         struct epoch_thread *tdwait;
 501         struct thread *td;
 502         epoch_record_t er;
 503
 504         td = curthread;
 505         if (td->td_epochnest == 0)
 506                 return (0);
 507         if (__predict_false((epoch) == NULL))
 508                 return (0);
 509         critical_enter();
 510         er = epoch_currecord(epoch);
 511         TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 512                 if (tdwait->et_td == td) {
 513                         critical_exit();
 514                         return (1);
 515                 }
 516 #ifdef INVARIANTS
 517         if (dump_onfail) {
 518                 MPASS(td->td_pinned);
 519                 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
 520                 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
 521                         printf("td_tid: %d ", tdwait->et_td->td_tid);
 522                 printf("\n");
 523         }
 524 #endif
 525         critical_exit();
 526         return (0);
 527 }
 528
 529 int
 530 in_epoch(epoch_t epoch)
 531 {
 532         return (in_epoch_verbose(epoch, 0));
 533 }
 534
 535 void
 536 epoch_adjust_prio(struct thread *td, u_char prio)
 537 {
 538         thread_lock(td);
 539         sched_prio(td, prio);
 540         thread_unlock(td);
 541 }