Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/kern/subr_epoch.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
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: releng/12.0/sys/kern/subr_epoch.c 337525 2018-08-09 05:18:27Z mmacy $"); 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 100 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 /* 244 * We lost a race and there's no longer any threads 245 * on the CPU in an epoch section. 246 */ 247 if (TAILQ_EMPTY(&record->er_tdlist)) 248 return; 249 250 if (record->er_cpuid != curcpu) { 251 /* 252 * If the head of the list is running, we can wait for it 253 * to remove itself from the list and thus save us the 254 * overhead of a migration 255 */ 256 gen = record->er_gen; 257 thread_unlock(td); 258 /* 259 * We can't actually check if the waiting thread is running 260 * so we simply poll for it to exit before giving up and 261 * migrating. 262 */ 263 do { 264 cpu_spinwait(); 265 } while (!TAILQ_EMPTY(&record->er_tdlist) && 266 gen == record->er_gen && 267 spincount++ < MAX_ADAPTIVE_SPIN); 268 thread_lock(td); 269 /* 270 * If the generation has changed we can poll again 271 * otherwise we need to migrate. 272 */ 273 if (gen != record->er_gen) 274 return; 275 /* 276 * Being on the same CPU as that of the record on which 277 * we need to wait allows us access to the thread 278 * list associated with that CPU. We can then examine the 279 * oldest thread in the queue and wait on its turnstile 280 * until it resumes and so on until a grace period 281 * elapses. 282 * 283 */ 284 counter_u64_add(migrate_count, 1); 285 sched_bind(td, record->er_cpuid); 286 /* 287 * At this point we need to return to the ck code 288 * to scan to see if a grace period has elapsed. 289 * We can't move on to check the thread list, because 290 * in the meantime new threads may have arrived that 291 * in fact belong to a different epoch. 292 */ 293 return; 294 } 295 /* 296 * Try to find a thread in an epoch section on this CPU 297 * waiting on a turnstile. Otherwise find the lowest 298 * priority thread (highest prio value) and drop our priority 299 * to match to allow it to run. 300 */ 301 TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) { 302 /* 303 * Propagate our priority to any other waiters to prevent us 304 * from starving them. They will have their original priority 305 * restore on exit from epoch_wait(). 306 */ 307 curwaittd = tdwait->et_td; 308 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) { 309 critical_enter(); 310 thread_unlock(td); 311 thread_lock(curwaittd); 312 sched_prio(curwaittd, td->td_priority); 313 thread_unlock(curwaittd); 314 thread_lock(td); 315 critical_exit(); 316 } 317 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) && 318 ((ts = curwaittd->td_blocked) != NULL)) { 319 /* 320 * We unlock td to allow turnstile_wait to reacquire the 321 * the thread lock. Before unlocking it we enter a critical 322 * section to prevent preemption after we reenable interrupts 323 * by dropping the thread lock in order to prevent curwaittd 324 * from getting to run. 325 */ 326 critical_enter(); 327 thread_unlock(td); 328 owner = turnstile_lock(ts, &lock); 329 /* 330 * The owner pointer indicates that the lock succeeded. Only 331 * in case we hold the lock and the turnstile we locked is still 332 * the one that curwaittd is blocked on can we continue. Otherwise 333 * The turnstile pointer has been changed out from underneath 334 * us, as in the case where the lock holder has signalled curwaittd, 335 * and we need to continue. 336 */ 337 if (owner != NULL && ts == curwaittd->td_blocked) { 338 MPASS(TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd)); 339 critical_exit(); 340 turnstile_wait(ts, owner, curwaittd->td_tsqueue); 341 counter_u64_add(turnstile_count, 1); 342 thread_lock(td); 343 return; 344 } else if (owner != NULL) 345 turnstile_unlock(ts, lock); 346 thread_lock(td); 347 critical_exit(); 348 KASSERT(td->td_locks == locksheld, 349 ("%d extra locks held", td->td_locks - locksheld)); 350 } 351 } 352 /* 353 * We didn't find any threads actually blocked on a lock 354 * so we have nothing to do except context switch away. 355 */ 356 counter_u64_add(switch_count, 1); 357 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 358 359 /* 360 * Release the thread lock while yielding to 361 * allow other threads to acquire the lock 362 * pointed to by TDQ_LOCKPTR(td). Else a 363 * deadlock like situation might happen. (HPS) 364 */ 365 thread_unlock(td); 366 thread_lock(td); 367 } 368 369 void 370 epoch_wait_preempt(epoch_t epoch) 371 { 372 struct thread *td; 373 int was_bound; 374 int old_cpu; 375 int old_pinned; 376 u_char old_prio; 377 int locks __unused; 378 379 MPASS(cold || epoch != NULL); 380 INIT_CHECK(epoch); 381 td = curthread; 382 #ifdef INVARIANTS 383 locks = curthread->td_locks; 384 MPASS(epoch->e_flags & EPOCH_PREEMPT); 385 if ((epoch->e_flags & EPOCH_LOCKED) == 0) 386 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 387 "epoch_wait() can be long running"); 388 KASSERT(!in_epoch(epoch), 389 ("epoch_wait_preempt() called in the middle " 390 "of an epoch section of the same epoch")); 391 #endif 392 thread_lock(td); 393 DROP_GIANT(); 394 395 old_cpu = PCPU_GET(cpuid); 396 old_pinned = td->td_pinned; 397 old_prio = td->td_priority; 398 was_bound = sched_is_bound(td); 399 sched_unbind(td); 400 td->td_pinned = 0; 401 sched_bind(td, old_cpu); 402 403 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt, NULL); 404 405 /* restore CPU binding, if any */ 406 if (was_bound != 0) { 407 sched_bind(td, old_cpu); 408 } else { 409 /* get thread back to initial CPU, if any */ 410 if (old_pinned != 0) 411 sched_bind(td, old_cpu); 412 sched_unbind(td); 413 } 414 /* restore pinned after bind */ 415 td->td_pinned = old_pinned; 416 417 /* restore thread priority */ 418 sched_prio(td, old_prio); 419 thread_unlock(td); 420 PICKUP_GIANT(); 421 KASSERT(td->td_locks == locks, 422 ("%d residual locks held", td->td_locks - locks)); 423 } 424 425 static void 426 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused, 427 void *arg __unused) 428 { 429 cpu_spinwait(); 430 } 431 432 void 433 epoch_wait(epoch_t epoch) 434 { 435 436 MPASS(cold || epoch != NULL); 437 INIT_CHECK(epoch); 438 MPASS(epoch->e_flags == 0); 439 critical_enter(); 440 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL); 441 critical_exit(); 442 } 443 444 void 445 epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t)) 446 { 447 epoch_record_t er; 448 ck_epoch_entry_t *cb; 449 450 cb = (void *)ctx; 451 452 MPASS(callback); 453 /* too early in boot to have epoch set up */ 454 if (__predict_false(epoch == NULL)) 455 goto boottime; 456 #if !defined(EARLY_AP_STARTUP) 457 if (__predict_false(inited < 2)) 458 goto boottime; 459 #endif 460 461 critical_enter(); 462 *DPCPU_PTR(epoch_cb_count) += 1; 463 er = epoch_currecord(epoch); 464 ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback); 465 critical_exit(); 466 return; 467 boottime: 468 callback(ctx); 469 } 470 471 static void 472 epoch_call_task(void *arg __unused) 473 { 474 ck_stack_entry_t *cursor, *head, *next; 475 ck_epoch_record_t *record; 476 epoch_record_t er; 477 epoch_t epoch; 478 ck_stack_t cb_stack; 479 int i, npending, total; 480 481 ck_stack_init(&cb_stack); 482 critical_enter(); 483 epoch_enter(global_epoch); 484 for (total = i = 0; i < epoch_count; i++) { 485 if (__predict_false((epoch = allepochs[i]) == NULL)) 486 continue; 487 er = epoch_currecord(epoch); 488 record = &er->er_record; 489 if ((npending = record->n_pending) == 0) 490 continue; 491 ck_epoch_poll_deferred(record, &cb_stack); 492 total += npending - record->n_pending; 493 } 494 epoch_exit(global_epoch); 495 *DPCPU_PTR(epoch_cb_count) -= total; 496 critical_exit(); 497 498 counter_u64_add(epoch_call_count, total); 499 counter_u64_add(epoch_call_task_count, 1); 500 501 head = ck_stack_batch_pop_npsc(&cb_stack); 502 for (cursor = head; cursor != NULL; cursor = next) { 503 struct ck_epoch_entry *entry = 504 ck_epoch_entry_container(cursor); 505 506 next = CK_STACK_NEXT(cursor); 507 entry->function(entry); 508 } 509 } 510 511 int 512 in_epoch_verbose(epoch_t epoch, int dump_onfail) 513 { 514 struct epoch_thread *tdwait; 515 struct thread *td; 516 epoch_record_t er; 517 518 td = curthread; 519 if (td->td_epochnest == 0) 520 return (0); 521 if (__predict_false((epoch) == NULL)) 522 return (0); 523 critical_enter(); 524 er = epoch_currecord(epoch); 525 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link) 526 if (tdwait->et_td == td) { 527 critical_exit(); 528 return (1); 529 } 530 #ifdef INVARIANTS 531 if (dump_onfail) { 532 MPASS(td->td_pinned); 533 printf("cpu: %d id: %d\n", curcpu, td->td_tid); 534 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link) 535 printf("td_tid: %d ", tdwait->et_td->td_tid); 536 printf("\n"); 537 } 538 #endif 539 critical_exit(); 540 return (0); 541 } 542 543 int 544 in_epoch(epoch_t epoch) 545 { 546 return (in_epoch_verbose(epoch, 0)); 547 } 548 549 void 550 epoch_adjust_prio(struct thread *td, u_char prio) 551 { 552 thread_lock(td); 553 sched_prio(td, prio); 554 thread_unlock(td); 555 }
Cache object: 4bb01369e85f35a755c0bf2de1adf8a0
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]