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/kern_timeout.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-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 37 */ 38 39 #include <sys/cdefs.h> 40 __FBSDID("$FreeBSD$"); 41 42 #include "opt_callout_profiling.h" 43 #include "opt_ddb.h" 44 #include "opt_rss.h" 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/bus.h> 49 #include <sys/callout.h> 50 #include <sys/domainset.h> 51 #include <sys/file.h> 52 #include <sys/interrupt.h> 53 #include <sys/kernel.h> 54 #include <sys/ktr.h> 55 #include <sys/lock.h> 56 #include <sys/malloc.h> 57 #include <sys/mutex.h> 58 #include <sys/proc.h> 59 #include <sys/sdt.h> 60 #include <sys/sleepqueue.h> 61 #include <sys/sysctl.h> 62 #include <sys/smp.h> 63 64 #ifdef DDB 65 #include <ddb/ddb.h> 66 #include <ddb/db_sym.h> 67 #include <machine/_inttypes.h> 68 #endif 69 70 #ifdef SMP 71 #include <machine/cpu.h> 72 #endif 73 74 DPCPU_DECLARE(sbintime_t, hardclocktime); 75 76 SDT_PROVIDER_DEFINE(callout_execute); 77 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *"); 78 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *"); 79 80 #ifdef CALLOUT_PROFILING 81 static int avg_depth; 82 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0, 83 "Average number of items examined per softclock call. Units = 1/1000"); 84 static int avg_gcalls; 85 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0, 86 "Average number of Giant callouts made per softclock call. Units = 1/1000"); 87 static int avg_lockcalls; 88 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0, 89 "Average number of lock callouts made per softclock call. Units = 1/1000"); 90 static int avg_mpcalls; 91 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0, 92 "Average number of MP callouts made per softclock call. Units = 1/1000"); 93 static int avg_depth_dir; 94 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0, 95 "Average number of direct callouts examined per callout_process call. " 96 "Units = 1/1000"); 97 static int avg_lockcalls_dir; 98 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD, 99 &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per " 100 "callout_process call. Units = 1/1000"); 101 static int avg_mpcalls_dir; 102 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir, 103 0, "Average number of MP direct callouts made per callout_process call. " 104 "Units = 1/1000"); 105 #endif 106 107 static int ncallout; 108 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0, 109 "Number of entries in callwheel and size of timeout() preallocation"); 110 111 #ifdef RSS 112 static int pin_default_swi = 1; 113 static int pin_pcpu_swi = 1; 114 #else 115 static int pin_default_swi = 0; 116 static int pin_pcpu_swi = 0; 117 #endif 118 119 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi, 120 0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)"); 121 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi, 122 0, "Pin the per-CPU swis (except PCPU 0, which is also default)"); 123 124 /* 125 * TODO: 126 * allocate more timeout table slots when table overflows. 127 */ 128 static u_int __read_mostly callwheelsize; 129 static u_int __read_mostly callwheelmask; 130 131 /* 132 * The callout cpu exec entities represent informations necessary for 133 * describing the state of callouts currently running on the CPU and the ones 134 * necessary for migrating callouts to the new callout cpu. In particular, 135 * the first entry of the array cc_exec_entity holds informations for callout 136 * running in SWI thread context, while the second one holds informations 137 * for callout running directly from hardware interrupt context. 138 * The cached informations are very important for deferring migration when 139 * the migrating callout is already running. 140 */ 141 struct cc_exec { 142 struct callout *cc_curr; 143 callout_func_t *cc_drain; 144 void *cc_last_func; 145 void *cc_last_arg; 146 #ifdef SMP 147 callout_func_t *ce_migration_func; 148 void *ce_migration_arg; 149 sbintime_t ce_migration_time; 150 sbintime_t ce_migration_prec; 151 int ce_migration_cpu; 152 #endif 153 bool cc_cancel; 154 bool cc_waiting; 155 }; 156 157 /* 158 * There is one struct callout_cpu per cpu, holding all relevant 159 * state for the callout processing thread on the individual CPU. 160 */ 161 struct callout_cpu { 162 struct mtx_padalign cc_lock; 163 struct cc_exec cc_exec_entity[2]; 164 struct callout *cc_next; 165 struct callout_list *cc_callwheel; 166 struct callout_tailq cc_expireq; 167 sbintime_t cc_firstevent; 168 sbintime_t cc_lastscan; 169 void *cc_cookie; 170 u_int cc_bucket; 171 u_int cc_inited; 172 #ifdef KTR 173 char cc_ktr_event_name[20]; 174 #endif 175 }; 176 177 #define callout_migrating(c) ((c)->c_iflags & CALLOUT_DFRMIGRATION) 178 179 #define cc_exec_curr(cc, dir) cc->cc_exec_entity[dir].cc_curr 180 #define cc_exec_last_func(cc, dir) cc->cc_exec_entity[dir].cc_last_func 181 #define cc_exec_last_arg(cc, dir) cc->cc_exec_entity[dir].cc_last_arg 182 #define cc_exec_drain(cc, dir) cc->cc_exec_entity[dir].cc_drain 183 #define cc_exec_next(cc) cc->cc_next 184 #define cc_exec_cancel(cc, dir) cc->cc_exec_entity[dir].cc_cancel 185 #define cc_exec_waiting(cc, dir) cc->cc_exec_entity[dir].cc_waiting 186 #ifdef SMP 187 #define cc_migration_func(cc, dir) cc->cc_exec_entity[dir].ce_migration_func 188 #define cc_migration_arg(cc, dir) cc->cc_exec_entity[dir].ce_migration_arg 189 #define cc_migration_cpu(cc, dir) cc->cc_exec_entity[dir].ce_migration_cpu 190 #define cc_migration_time(cc, dir) cc->cc_exec_entity[dir].ce_migration_time 191 #define cc_migration_prec(cc, dir) cc->cc_exec_entity[dir].ce_migration_prec 192 193 static struct callout_cpu cc_cpu[MAXCPU]; 194 #define CPUBLOCK MAXCPU 195 #define CC_CPU(cpu) (&cc_cpu[(cpu)]) 196 #define CC_SELF() CC_CPU(PCPU_GET(cpuid)) 197 #else 198 static struct callout_cpu cc_cpu; 199 #define CC_CPU(cpu) (&cc_cpu) 200 #define CC_SELF() (&cc_cpu) 201 #endif 202 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock) 203 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock) 204 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED) 205 206 static int __read_mostly cc_default_cpu; 207 208 static void callout_cpu_init(struct callout_cpu *cc, int cpu); 209 static void softclock_call_cc(struct callout *c, struct callout_cpu *cc, 210 #ifdef CALLOUT_PROFILING 211 int *mpcalls, int *lockcalls, int *gcalls, 212 #endif 213 int direct); 214 215 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures"); 216 217 /** 218 * Locked by cc_lock: 219 * cc_curr - If a callout is in progress, it is cc_curr. 220 * If cc_curr is non-NULL, threads waiting in 221 * callout_drain() will be woken up as soon as the 222 * relevant callout completes. 223 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held 224 * guarantees that the current callout will not run. 225 * The softclock() function sets this to 0 before it 226 * drops callout_lock to acquire c_lock, and it calls 227 * the handler only if curr_cancelled is still 0 after 228 * cc_lock is successfully acquired. 229 * cc_waiting - If a thread is waiting in callout_drain(), then 230 * callout_wait is nonzero. Set only when 231 * cc_curr is non-NULL. 232 */ 233 234 /* 235 * Resets the execution entity tied to a specific callout cpu. 236 */ 237 static void 238 cc_cce_cleanup(struct callout_cpu *cc, int direct) 239 { 240 241 cc_exec_curr(cc, direct) = NULL; 242 cc_exec_cancel(cc, direct) = false; 243 cc_exec_waiting(cc, direct) = false; 244 #ifdef SMP 245 cc_migration_cpu(cc, direct) = CPUBLOCK; 246 cc_migration_time(cc, direct) = 0; 247 cc_migration_prec(cc, direct) = 0; 248 cc_migration_func(cc, direct) = NULL; 249 cc_migration_arg(cc, direct) = NULL; 250 #endif 251 } 252 253 /* 254 * Checks if migration is requested by a specific callout cpu. 255 */ 256 static int 257 cc_cce_migrating(struct callout_cpu *cc, int direct) 258 { 259 260 #ifdef SMP 261 return (cc_migration_cpu(cc, direct) != CPUBLOCK); 262 #else 263 return (0); 264 #endif 265 } 266 267 /* 268 * Kernel low level callwheel initialization 269 * called on the BSP during kernel startup. 270 */ 271 static void 272 callout_callwheel_init(void *dummy) 273 { 274 struct callout_cpu *cc; 275 int cpu; 276 277 /* 278 * Calculate the size of the callout wheel and the preallocated 279 * timeout() structures. 280 * XXX: Clip callout to result of previous function of maxusers 281 * maximum 384. This is still huge, but acceptable. 282 */ 283 ncallout = imin(16 + maxproc + maxfiles, 18508); 284 TUNABLE_INT_FETCH("kern.ncallout", &ncallout); 285 286 /* 287 * Calculate callout wheel size, should be next power of two higher 288 * than 'ncallout'. 289 */ 290 callwheelsize = 1 << fls(ncallout); 291 callwheelmask = callwheelsize - 1; 292 293 /* 294 * Fetch whether we're pinning the swi's or not. 295 */ 296 TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi); 297 TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi); 298 299 /* 300 * Initialize callout wheels. The software interrupt threads 301 * are created later. 302 */ 303 cc_default_cpu = PCPU_GET(cpuid); 304 CPU_FOREACH(cpu) { 305 cc = CC_CPU(cpu); 306 callout_cpu_init(cc, cpu); 307 } 308 } 309 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL); 310 311 /* 312 * Initialize the per-cpu callout structures. 313 */ 314 static void 315 callout_cpu_init(struct callout_cpu *cc, int cpu) 316 { 317 int i; 318 319 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE); 320 cc->cc_inited = 1; 321 cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) * 322 callwheelsize, M_CALLOUT, 323 DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK); 324 for (i = 0; i < callwheelsize; i++) 325 LIST_INIT(&cc->cc_callwheel[i]); 326 TAILQ_INIT(&cc->cc_expireq); 327 cc->cc_firstevent = SBT_MAX; 328 for (i = 0; i < 2; i++) 329 cc_cce_cleanup(cc, i); 330 #ifdef KTR 331 snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name), 332 "callwheel cpu %d", cpu); 333 #endif 334 } 335 336 #ifdef SMP 337 /* 338 * Switches the cpu tied to a specific callout. 339 * The function expects a locked incoming callout cpu and returns with 340 * locked outcoming callout cpu. 341 */ 342 static struct callout_cpu * 343 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu) 344 { 345 struct callout_cpu *new_cc; 346 347 MPASS(c != NULL && cc != NULL); 348 CC_LOCK_ASSERT(cc); 349 350 /* 351 * Avoid interrupts and preemption firing after the callout cpu 352 * is blocked in order to avoid deadlocks as the new thread 353 * may be willing to acquire the callout cpu lock. 354 */ 355 c->c_cpu = CPUBLOCK; 356 spinlock_enter(); 357 CC_UNLOCK(cc); 358 new_cc = CC_CPU(new_cpu); 359 CC_LOCK(new_cc); 360 spinlock_exit(); 361 c->c_cpu = new_cpu; 362 return (new_cc); 363 } 364 #endif 365 366 /* 367 * Start softclock threads. 368 */ 369 static void 370 start_softclock(void *dummy) 371 { 372 struct callout_cpu *cc; 373 char name[MAXCOMLEN]; 374 int cpu; 375 bool pin_swi; 376 struct intr_event *ie; 377 378 CPU_FOREACH(cpu) { 379 cc = CC_CPU(cpu); 380 snprintf(name, sizeof(name), "clock (%d)", cpu); 381 ie = NULL; 382 if (swi_add(&ie, name, softclock, cc, SWI_CLOCK, 383 INTR_MPSAFE, &cc->cc_cookie)) 384 panic("died while creating standard software ithreads"); 385 if (cpu == cc_default_cpu) 386 pin_swi = pin_default_swi; 387 else 388 pin_swi = pin_pcpu_swi; 389 if (pin_swi && (intr_event_bind(ie, cpu) != 0)) { 390 printf("%s: %s clock couldn't be pinned to cpu %d\n", 391 __func__, 392 cpu == cc_default_cpu ? "default" : "per-cpu", 393 cpu); 394 } 395 } 396 } 397 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL); 398 399 #define CC_HASH_SHIFT 8 400 401 static inline u_int 402 callout_hash(sbintime_t sbt) 403 { 404 405 return (sbt >> (32 - CC_HASH_SHIFT)); 406 } 407 408 static inline u_int 409 callout_get_bucket(sbintime_t sbt) 410 { 411 412 return (callout_hash(sbt) & callwheelmask); 413 } 414 415 void 416 callout_process(sbintime_t now) 417 { 418 struct callout *c, *next; 419 struct callout_cpu *cc; 420 struct callout_list *sc; 421 sbintime_t first, last, lookahead, max, tmp_max; 422 u_int firstb, lastb, nowb; 423 #ifdef CALLOUT_PROFILING 424 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0; 425 #endif 426 427 cc = CC_SELF(); 428 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET); 429 430 /* Compute the buckets of the last scan and present times. */ 431 firstb = callout_hash(cc->cc_lastscan); 432 cc->cc_lastscan = now; 433 nowb = callout_hash(now); 434 435 /* Compute the last bucket and minimum time of the bucket after it. */ 436 if (nowb == firstb) 437 lookahead = (SBT_1S / 16); 438 else if (nowb - firstb == 1) 439 lookahead = (SBT_1S / 8); 440 else 441 lookahead = SBT_1S; 442 first = last = now; 443 first += (lookahead / 2); 444 last += lookahead; 445 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT)); 446 lastb = callout_hash(last) - 1; 447 max = last; 448 449 /* 450 * Check if we wrapped around the entire wheel from the last scan. 451 * In case, we need to scan entirely the wheel for pending callouts. 452 */ 453 if (lastb - firstb >= callwheelsize) { 454 lastb = firstb + callwheelsize - 1; 455 if (nowb - firstb >= callwheelsize) 456 nowb = lastb; 457 } 458 459 /* Iterate callwheel from firstb to nowb and then up to lastb. */ 460 do { 461 sc = &cc->cc_callwheel[firstb & callwheelmask]; 462 LIST_FOREACH_SAFE(c, sc, c_links.le, next) { 463 /* Run the callout if present time within allowed. */ 464 if (c->c_time <= now) { 465 /* 466 * Consumer told us the callout may be run 467 * directly from hardware interrupt context. 468 */ 469 if (c->c_iflags & CALLOUT_DIRECT) { 470 #ifdef CALLOUT_PROFILING 471 ++depth_dir; 472 #endif 473 cc_exec_next(cc) = next; 474 cc->cc_bucket = firstb & callwheelmask; 475 LIST_REMOVE(c, c_links.le); 476 softclock_call_cc(c, cc, 477 #ifdef CALLOUT_PROFILING 478 &mpcalls_dir, &lockcalls_dir, NULL, 479 #endif 480 1); 481 next = cc_exec_next(cc); 482 cc_exec_next(cc) = NULL; 483 } else { 484 LIST_REMOVE(c, c_links.le); 485 TAILQ_INSERT_TAIL(&cc->cc_expireq, 486 c, c_links.tqe); 487 c->c_iflags |= CALLOUT_PROCESSED; 488 } 489 } else if (c->c_time >= max) { 490 /* 491 * Skip events in the distant future. 492 */ 493 ; 494 } else if (c->c_time > last) { 495 /* 496 * Event minimal time is bigger than present 497 * maximal time, so it cannot be aggregated. 498 */ 499 lastb = nowb; 500 } else { 501 /* 502 * Update first and last time, respecting this 503 * event. 504 */ 505 if (c->c_time < first) 506 first = c->c_time; 507 tmp_max = c->c_time + c->c_precision; 508 if (tmp_max < last) 509 last = tmp_max; 510 } 511 } 512 /* Proceed with the next bucket. */ 513 firstb++; 514 /* 515 * Stop if we looked after present time and found 516 * some event we can't execute at now. 517 * Stop if we looked far enough into the future. 518 */ 519 } while (((int)(firstb - lastb)) <= 0); 520 cc->cc_firstevent = last; 521 cpu_new_callout(curcpu, last, first); 522 523 #ifdef CALLOUT_PROFILING 524 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8; 525 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8; 526 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8; 527 #endif 528 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET); 529 /* 530 * swi_sched acquires the thread lock, so we don't want to call it 531 * with cc_lock held; incorrect locking order. 532 */ 533 if (!TAILQ_EMPTY(&cc->cc_expireq)) 534 swi_sched(cc->cc_cookie, 0); 535 } 536 537 static struct callout_cpu * 538 callout_lock(struct callout *c) 539 { 540 struct callout_cpu *cc; 541 int cpu; 542 543 for (;;) { 544 cpu = c->c_cpu; 545 #ifdef SMP 546 if (cpu == CPUBLOCK) { 547 while (c->c_cpu == CPUBLOCK) 548 cpu_spinwait(); 549 continue; 550 } 551 #endif 552 cc = CC_CPU(cpu); 553 CC_LOCK(cc); 554 if (cpu == c->c_cpu) 555 break; 556 CC_UNLOCK(cc); 557 } 558 return (cc); 559 } 560 561 static void 562 callout_cc_add(struct callout *c, struct callout_cpu *cc, 563 sbintime_t sbt, sbintime_t precision, void (*func)(void *), 564 void *arg, int flags) 565 { 566 int bucket; 567 568 CC_LOCK_ASSERT(cc); 569 if (sbt < cc->cc_lastscan) 570 sbt = cc->cc_lastscan; 571 c->c_arg = arg; 572 c->c_iflags |= CALLOUT_PENDING; 573 c->c_iflags &= ~CALLOUT_PROCESSED; 574 c->c_flags |= CALLOUT_ACTIVE; 575 if (flags & C_DIRECT_EXEC) 576 c->c_iflags |= CALLOUT_DIRECT; 577 c->c_func = func; 578 c->c_time = sbt; 579 c->c_precision = precision; 580 bucket = callout_get_bucket(c->c_time); 581 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x", 582 c, (int)(c->c_precision >> 32), 583 (u_int)(c->c_precision & 0xffffffff)); 584 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le); 585 if (cc->cc_bucket == bucket) 586 cc_exec_next(cc) = c; 587 588 /* 589 * Inform the eventtimers(4) subsystem there's a new callout 590 * that has been inserted, but only if really required. 591 */ 592 if (SBT_MAX - c->c_time < c->c_precision) 593 c->c_precision = SBT_MAX - c->c_time; 594 sbt = c->c_time + c->c_precision; 595 if (sbt < cc->cc_firstevent) { 596 cc->cc_firstevent = sbt; 597 cpu_new_callout(c->c_cpu, sbt, c->c_time); 598 } 599 } 600 601 static void 602 softclock_call_cc(struct callout *c, struct callout_cpu *cc, 603 #ifdef CALLOUT_PROFILING 604 int *mpcalls, int *lockcalls, int *gcalls, 605 #endif 606 int direct) 607 { 608 struct rm_priotracker tracker; 609 callout_func_t *c_func, *drain; 610 void *c_arg; 611 struct lock_class *class; 612 struct lock_object *c_lock; 613 uintptr_t lock_status; 614 int c_iflags; 615 #ifdef SMP 616 struct callout_cpu *new_cc; 617 callout_func_t *new_func; 618 void *new_arg; 619 int flags, new_cpu; 620 sbintime_t new_prec, new_time; 621 #endif 622 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 623 sbintime_t sbt1, sbt2; 624 struct timespec ts2; 625 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */ 626 static callout_func_t *lastfunc; 627 #endif 628 629 KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING, 630 ("softclock_call_cc: pend %p %x", c, c->c_iflags)); 631 KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE, 632 ("softclock_call_cc: act %p %x", c, c->c_flags)); 633 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL; 634 lock_status = 0; 635 if (c->c_iflags & CALLOUT_SHAREDLOCK) { 636 if (class == &lock_class_rm) 637 lock_status = (uintptr_t)&tracker; 638 else 639 lock_status = 1; 640 } 641 c_lock = c->c_lock; 642 c_func = c->c_func; 643 c_arg = c->c_arg; 644 c_iflags = c->c_iflags; 645 c->c_iflags &= ~CALLOUT_PENDING; 646 647 cc_exec_curr(cc, direct) = c; 648 cc_exec_last_func(cc, direct) = c_func; 649 cc_exec_last_arg(cc, direct) = c_arg; 650 cc_exec_cancel(cc, direct) = false; 651 cc_exec_drain(cc, direct) = NULL; 652 CC_UNLOCK(cc); 653 if (c_lock != NULL) { 654 class->lc_lock(c_lock, lock_status); 655 /* 656 * The callout may have been cancelled 657 * while we switched locks. 658 */ 659 if (cc_exec_cancel(cc, direct)) { 660 class->lc_unlock(c_lock); 661 goto skip; 662 } 663 /* The callout cannot be stopped now. */ 664 cc_exec_cancel(cc, direct) = true; 665 if (c_lock == &Giant.lock_object) { 666 #ifdef CALLOUT_PROFILING 667 (*gcalls)++; 668 #endif 669 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p", 670 c, c_func, c_arg); 671 } else { 672 #ifdef CALLOUT_PROFILING 673 (*lockcalls)++; 674 #endif 675 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p", 676 c, c_func, c_arg); 677 } 678 } else { 679 #ifdef CALLOUT_PROFILING 680 (*mpcalls)++; 681 #endif 682 CTR3(KTR_CALLOUT, "callout %p func %p arg %p", 683 c, c_func, c_arg); 684 } 685 KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running", 686 "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct); 687 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 688 sbt1 = sbinuptime(); 689 #endif 690 THREAD_NO_SLEEPING(); 691 SDT_PROBE1(callout_execute, , , callout__start, c); 692 c_func(c_arg); 693 SDT_PROBE1(callout_execute, , , callout__end, c); 694 THREAD_SLEEPING_OK(); 695 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 696 sbt2 = sbinuptime(); 697 sbt2 -= sbt1; 698 if (sbt2 > maxdt) { 699 if (lastfunc != c_func || sbt2 > maxdt * 2) { 700 ts2 = sbttots(sbt2); 701 printf( 702 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n", 703 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec); 704 } 705 maxdt = sbt2; 706 lastfunc = c_func; 707 } 708 #endif 709 KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle"); 710 CTR1(KTR_CALLOUT, "callout %p finished", c); 711 if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0) 712 class->lc_unlock(c_lock); 713 skip: 714 CC_LOCK(cc); 715 KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr")); 716 cc_exec_curr(cc, direct) = NULL; 717 if (cc_exec_drain(cc, direct)) { 718 drain = cc_exec_drain(cc, direct); 719 cc_exec_drain(cc, direct) = NULL; 720 CC_UNLOCK(cc); 721 drain(c_arg); 722 CC_LOCK(cc); 723 } 724 if (cc_exec_waiting(cc, direct)) { 725 /* 726 * There is someone waiting for the 727 * callout to complete. 728 * If the callout was scheduled for 729 * migration just cancel it. 730 */ 731 if (cc_cce_migrating(cc, direct)) { 732 cc_cce_cleanup(cc, direct); 733 734 /* 735 * It should be assert here that the callout is not 736 * destroyed but that is not easy. 737 */ 738 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 739 } 740 cc_exec_waiting(cc, direct) = false; 741 CC_UNLOCK(cc); 742 wakeup(&cc_exec_waiting(cc, direct)); 743 CC_LOCK(cc); 744 } else if (cc_cce_migrating(cc, direct)) { 745 #ifdef SMP 746 /* 747 * If the callout was scheduled for 748 * migration just perform it now. 749 */ 750 new_cpu = cc_migration_cpu(cc, direct); 751 new_time = cc_migration_time(cc, direct); 752 new_prec = cc_migration_prec(cc, direct); 753 new_func = cc_migration_func(cc, direct); 754 new_arg = cc_migration_arg(cc, direct); 755 cc_cce_cleanup(cc, direct); 756 757 /* 758 * It should be assert here that the callout is not destroyed 759 * but that is not easy. 760 * 761 * As first thing, handle deferred callout stops. 762 */ 763 if (!callout_migrating(c)) { 764 CTR3(KTR_CALLOUT, 765 "deferred cancelled %p func %p arg %p", 766 c, new_func, new_arg); 767 return; 768 } 769 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 770 771 new_cc = callout_cpu_switch(c, cc, new_cpu); 772 flags = (direct) ? C_DIRECT_EXEC : 0; 773 callout_cc_add(c, new_cc, new_time, new_prec, new_func, 774 new_arg, flags); 775 CC_UNLOCK(new_cc); 776 CC_LOCK(cc); 777 #else 778 panic("migration should not happen"); 779 #endif 780 } 781 } 782 783 /* 784 * The callout mechanism is based on the work of Adam M. Costello and 785 * George Varghese, published in a technical report entitled "Redesigning 786 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 787 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 788 * used in this implementation was published by G. Varghese and T. Lauck in 789 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 790 * the Efficient Implementation of a Timer Facility" in the Proceedings of 791 * the 11th ACM Annual Symposium on Operating Systems Principles, 792 * Austin, Texas Nov 1987. 793 */ 794 795 /* 796 * Software (low priority) clock interrupt. 797 * Run periodic events from timeout queue. 798 */ 799 void 800 softclock(void *arg) 801 { 802 struct callout_cpu *cc; 803 struct callout *c; 804 #ifdef CALLOUT_PROFILING 805 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0; 806 #endif 807 808 cc = (struct callout_cpu *)arg; 809 CC_LOCK(cc); 810 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) { 811 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 812 softclock_call_cc(c, cc, 813 #ifdef CALLOUT_PROFILING 814 &mpcalls, &lockcalls, &gcalls, 815 #endif 816 0); 817 #ifdef CALLOUT_PROFILING 818 ++depth; 819 #endif 820 } 821 #ifdef CALLOUT_PROFILING 822 avg_depth += (depth * 1000 - avg_depth) >> 8; 823 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8; 824 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8; 825 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8; 826 #endif 827 CC_UNLOCK(cc); 828 } 829 830 void 831 callout_when(sbintime_t sbt, sbintime_t precision, int flags, 832 sbintime_t *res, sbintime_t *prec_res) 833 { 834 sbintime_t to_sbt, to_pr; 835 836 if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) { 837 *res = sbt; 838 *prec_res = precision; 839 return; 840 } 841 if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt) 842 sbt = tick_sbt; 843 if ((flags & C_HARDCLOCK) != 0 || sbt >= sbt_tickthreshold) { 844 /* 845 * Obtain the time of the last hardclock() call on 846 * this CPU directly from the kern_clocksource.c. 847 * This value is per-CPU, but it is equal for all 848 * active ones. 849 */ 850 #ifdef __LP64__ 851 to_sbt = DPCPU_GET(hardclocktime); 852 #else 853 spinlock_enter(); 854 to_sbt = DPCPU_GET(hardclocktime); 855 spinlock_exit(); 856 #endif 857 if (cold && to_sbt == 0) 858 to_sbt = sbinuptime(); 859 if ((flags & C_HARDCLOCK) == 0) 860 to_sbt += tick_sbt; 861 } else 862 to_sbt = sbinuptime(); 863 if (SBT_MAX - to_sbt < sbt) 864 to_sbt = SBT_MAX; 865 else 866 to_sbt += sbt; 867 *res = to_sbt; 868 to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp : 869 sbt >> C_PRELGET(flags)); 870 *prec_res = to_pr > precision ? to_pr : precision; 871 } 872 873 /* 874 * New interface; clients allocate their own callout structures. 875 * 876 * callout_reset() - establish or change a timeout 877 * callout_stop() - disestablish a timeout 878 * callout_init() - initialize a callout structure so that it can 879 * safely be passed to callout_reset() and callout_stop() 880 * 881 * <sys/callout.h> defines three convenience macros: 882 * 883 * callout_active() - returns truth if callout has not been stopped, 884 * drained, or deactivated since the last time the callout was 885 * reset. 886 * callout_pending() - returns truth if callout is still waiting for timeout 887 * callout_deactivate() - marks the callout as having been serviced 888 */ 889 int 890 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec, 891 callout_func_t *ftn, void *arg, int cpu, int flags) 892 { 893 sbintime_t to_sbt, precision; 894 struct callout_cpu *cc; 895 int cancelled, direct; 896 int ignore_cpu=0; 897 898 cancelled = 0; 899 if (cpu == -1) { 900 ignore_cpu = 1; 901 } else if ((cpu >= MAXCPU) || 902 ((CC_CPU(cpu))->cc_inited == 0)) { 903 /* Invalid CPU spec */ 904 panic("Invalid CPU in callout %d", cpu); 905 } 906 callout_when(sbt, prec, flags, &to_sbt, &precision); 907 908 /* 909 * This flag used to be added by callout_cc_add, but the 910 * first time you call this we could end up with the 911 * wrong direct flag if we don't do it before we add. 912 */ 913 if (flags & C_DIRECT_EXEC) { 914 direct = 1; 915 } else { 916 direct = 0; 917 } 918 KASSERT(!direct || c->c_lock == NULL || 919 (LOCK_CLASS(c->c_lock)->lc_flags & LC_SPINLOCK), 920 ("%s: direct callout %p has non-spin lock", __func__, c)); 921 cc = callout_lock(c); 922 /* 923 * Don't allow migration if the user does not care. 924 */ 925 if (ignore_cpu) { 926 cpu = c->c_cpu; 927 } 928 929 if (cc_exec_curr(cc, direct) == c) { 930 /* 931 * We're being asked to reschedule a callout which is 932 * currently in progress. If there is a lock then we 933 * can cancel the callout if it has not really started. 934 */ 935 if (c->c_lock != NULL && !cc_exec_cancel(cc, direct)) 936 cancelled = cc_exec_cancel(cc, direct) = true; 937 if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) { 938 /* 939 * Someone has called callout_drain to kill this 940 * callout. Don't reschedule. 941 */ 942 CTR4(KTR_CALLOUT, "%s %p func %p arg %p", 943 cancelled ? "cancelled" : "failed to cancel", 944 c, c->c_func, c->c_arg); 945 CC_UNLOCK(cc); 946 return (cancelled); 947 } 948 #ifdef SMP 949 if (callout_migrating(c)) { 950 /* 951 * This only occurs when a second callout_reset_sbt_on 952 * is made after a previous one moved it into 953 * deferred migration (below). Note we do *not* change 954 * the prev_cpu even though the previous target may 955 * be different. 956 */ 957 cc_migration_cpu(cc, direct) = cpu; 958 cc_migration_time(cc, direct) = to_sbt; 959 cc_migration_prec(cc, direct) = precision; 960 cc_migration_func(cc, direct) = ftn; 961 cc_migration_arg(cc, direct) = arg; 962 cancelled = 1; 963 CC_UNLOCK(cc); 964 return (cancelled); 965 } 966 #endif 967 } 968 if (c->c_iflags & CALLOUT_PENDING) { 969 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 970 if (cc_exec_next(cc) == c) 971 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 972 LIST_REMOVE(c, c_links.le); 973 } else { 974 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 975 } 976 cancelled = 1; 977 c->c_iflags &= ~ CALLOUT_PENDING; 978 c->c_flags &= ~ CALLOUT_ACTIVE; 979 } 980 981 #ifdef SMP 982 /* 983 * If the callout must migrate try to perform it immediately. 984 * If the callout is currently running, just defer the migration 985 * to a more appropriate moment. 986 */ 987 if (c->c_cpu != cpu) { 988 if (cc_exec_curr(cc, direct) == c) { 989 /* 990 * Pending will have been removed since we are 991 * actually executing the callout on another 992 * CPU. That callout should be waiting on the 993 * lock the caller holds. If we set both 994 * active/and/pending after we return and the 995 * lock on the executing callout proceeds, it 996 * will then see pending is true and return. 997 * At the return from the actual callout execution 998 * the migration will occur in softclock_call_cc 999 * and this new callout will be placed on the 1000 * new CPU via a call to callout_cpu_switch() which 1001 * will get the lock on the right CPU followed 1002 * by a call callout_cc_add() which will add it there. 1003 * (see above in softclock_call_cc()). 1004 */ 1005 cc_migration_cpu(cc, direct) = cpu; 1006 cc_migration_time(cc, direct) = to_sbt; 1007 cc_migration_prec(cc, direct) = precision; 1008 cc_migration_func(cc, direct) = ftn; 1009 cc_migration_arg(cc, direct) = arg; 1010 c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING); 1011 c->c_flags |= CALLOUT_ACTIVE; 1012 CTR6(KTR_CALLOUT, 1013 "migration of %p func %p arg %p in %d.%08x to %u deferred", 1014 c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1015 (u_int)(to_sbt & 0xffffffff), cpu); 1016 CC_UNLOCK(cc); 1017 return (cancelled); 1018 } 1019 cc = callout_cpu_switch(c, cc, cpu); 1020 } 1021 #endif 1022 1023 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, flags); 1024 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x", 1025 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1026 (u_int)(to_sbt & 0xffffffff)); 1027 CC_UNLOCK(cc); 1028 1029 return (cancelled); 1030 } 1031 1032 /* 1033 * Common idioms that can be optimized in the future. 1034 */ 1035 int 1036 callout_schedule_on(struct callout *c, int to_ticks, int cpu) 1037 { 1038 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu); 1039 } 1040 1041 int 1042 callout_schedule(struct callout *c, int to_ticks) 1043 { 1044 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu); 1045 } 1046 1047 int 1048 _callout_stop_safe(struct callout *c, int flags, callout_func_t *drain) 1049 { 1050 struct callout_cpu *cc, *old_cc; 1051 struct lock_class *class; 1052 int direct, sq_locked, use_lock; 1053 int cancelled, not_on_a_list; 1054 1055 if ((flags & CS_DRAIN) != 0) 1056 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock, 1057 "calling %s", __func__); 1058 1059 KASSERT((flags & CS_DRAIN) == 0 || drain == NULL, 1060 ("Cannot set drain callback and CS_DRAIN flag at the same time")); 1061 1062 /* 1063 * Some old subsystems don't hold Giant while running a callout_stop(), 1064 * so just discard this check for the moment. 1065 */ 1066 if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) { 1067 if (c->c_lock == &Giant.lock_object) 1068 use_lock = mtx_owned(&Giant); 1069 else { 1070 use_lock = 1; 1071 class = LOCK_CLASS(c->c_lock); 1072 class->lc_assert(c->c_lock, LA_XLOCKED); 1073 } 1074 } else 1075 use_lock = 0; 1076 if (c->c_iflags & CALLOUT_DIRECT) { 1077 direct = 1; 1078 } else { 1079 direct = 0; 1080 } 1081 sq_locked = 0; 1082 old_cc = NULL; 1083 again: 1084 cc = callout_lock(c); 1085 1086 if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) == 1087 (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) && 1088 ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) { 1089 /* 1090 * Special case where this slipped in while we 1091 * were migrating *as* the callout is about to 1092 * execute. The caller probably holds the lock 1093 * the callout wants. 1094 * 1095 * Get rid of the migration first. Then set 1096 * the flag that tells this code *not* to 1097 * try to remove it from any lists (its not 1098 * on one yet). When the callout wheel runs, 1099 * it will ignore this callout. 1100 */ 1101 c->c_iflags &= ~CALLOUT_PENDING; 1102 c->c_flags &= ~CALLOUT_ACTIVE; 1103 not_on_a_list = 1; 1104 } else { 1105 not_on_a_list = 0; 1106 } 1107 1108 /* 1109 * If the callout was migrating while the callout cpu lock was 1110 * dropped, just drop the sleepqueue lock and check the states 1111 * again. 1112 */ 1113 if (sq_locked != 0 && cc != old_cc) { 1114 #ifdef SMP 1115 CC_UNLOCK(cc); 1116 sleepq_release(&cc_exec_waiting(old_cc, direct)); 1117 sq_locked = 0; 1118 old_cc = NULL; 1119 goto again; 1120 #else 1121 panic("migration should not happen"); 1122 #endif 1123 } 1124 1125 /* 1126 * If the callout is running, try to stop it or drain it. 1127 */ 1128 if (cc_exec_curr(cc, direct) == c) { 1129 /* 1130 * Succeed we to stop it or not, we must clear the 1131 * active flag - this is what API users expect. If we're 1132 * draining and the callout is currently executing, first wait 1133 * until it finishes. 1134 */ 1135 if ((flags & CS_DRAIN) == 0) 1136 c->c_flags &= ~CALLOUT_ACTIVE; 1137 1138 if ((flags & CS_DRAIN) != 0) { 1139 /* 1140 * The current callout is running (or just 1141 * about to run) and blocking is allowed, so 1142 * just wait for the current invocation to 1143 * finish. 1144 */ 1145 if (cc_exec_curr(cc, direct) == c) { 1146 /* 1147 * Use direct calls to sleepqueue interface 1148 * instead of cv/msleep in order to avoid 1149 * a LOR between cc_lock and sleepqueue 1150 * chain spinlocks. This piece of code 1151 * emulates a msleep_spin() call actually. 1152 * 1153 * If we already have the sleepqueue chain 1154 * locked, then we can safely block. If we 1155 * don't already have it locked, however, 1156 * we have to drop the cc_lock to lock 1157 * it. This opens several races, so we 1158 * restart at the beginning once we have 1159 * both locks. If nothing has changed, then 1160 * we will end up back here with sq_locked 1161 * set. 1162 */ 1163 if (!sq_locked) { 1164 CC_UNLOCK(cc); 1165 sleepq_lock( 1166 &cc_exec_waiting(cc, direct)); 1167 sq_locked = 1; 1168 old_cc = cc; 1169 goto again; 1170 } 1171 1172 /* 1173 * Migration could be cancelled here, but 1174 * as long as it is still not sure when it 1175 * will be packed up, just let softclock() 1176 * take care of it. 1177 */ 1178 cc_exec_waiting(cc, direct) = true; 1179 DROP_GIANT(); 1180 CC_UNLOCK(cc); 1181 sleepq_add( 1182 &cc_exec_waiting(cc, direct), 1183 &cc->cc_lock.lock_object, "codrain", 1184 SLEEPQ_SLEEP, 0); 1185 sleepq_wait( 1186 &cc_exec_waiting(cc, direct), 1187 0); 1188 sq_locked = 0; 1189 old_cc = NULL; 1190 1191 /* Reacquire locks previously released. */ 1192 PICKUP_GIANT(); 1193 goto again; 1194 } 1195 c->c_flags &= ~CALLOUT_ACTIVE; 1196 } else if (use_lock && 1197 !cc_exec_cancel(cc, direct) && (drain == NULL)) { 1198 1199 /* 1200 * The current callout is waiting for its 1201 * lock which we hold. Cancel the callout 1202 * and return. After our caller drops the 1203 * lock, the callout will be skipped in 1204 * softclock(). This *only* works with a 1205 * callout_stop() *not* callout_drain() or 1206 * callout_async_drain(). 1207 */ 1208 cc_exec_cancel(cc, direct) = true; 1209 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1210 c, c->c_func, c->c_arg); 1211 KASSERT(!cc_cce_migrating(cc, direct), 1212 ("callout wrongly scheduled for migration")); 1213 if (callout_migrating(c)) { 1214 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1215 #ifdef SMP 1216 cc_migration_cpu(cc, direct) = CPUBLOCK; 1217 cc_migration_time(cc, direct) = 0; 1218 cc_migration_prec(cc, direct) = 0; 1219 cc_migration_func(cc, direct) = NULL; 1220 cc_migration_arg(cc, direct) = NULL; 1221 #endif 1222 } 1223 CC_UNLOCK(cc); 1224 KASSERT(!sq_locked, ("sleepqueue chain locked")); 1225 return (1); 1226 } else if (callout_migrating(c)) { 1227 /* 1228 * The callout is currently being serviced 1229 * and the "next" callout is scheduled at 1230 * its completion with a migration. We remove 1231 * the migration flag so it *won't* get rescheduled, 1232 * but we can't stop the one thats running so 1233 * we return 0. 1234 */ 1235 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1236 #ifdef SMP 1237 /* 1238 * We can't call cc_cce_cleanup here since 1239 * if we do it will remove .ce_curr and 1240 * its still running. This will prevent a 1241 * reschedule of the callout when the 1242 * execution completes. 1243 */ 1244 cc_migration_cpu(cc, direct) = CPUBLOCK; 1245 cc_migration_time(cc, direct) = 0; 1246 cc_migration_prec(cc, direct) = 0; 1247 cc_migration_func(cc, direct) = NULL; 1248 cc_migration_arg(cc, direct) = NULL; 1249 #endif 1250 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p", 1251 c, c->c_func, c->c_arg); 1252 if (drain) { 1253 KASSERT(cc_exec_drain(cc, direct) == NULL, 1254 ("callout drain function already set to %p", 1255 cc_exec_drain(cc, direct))); 1256 cc_exec_drain(cc, direct) = drain; 1257 } 1258 CC_UNLOCK(cc); 1259 return (0); 1260 } else { 1261 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1262 c, c->c_func, c->c_arg); 1263 if (drain) { 1264 KASSERT(cc_exec_drain(cc, direct) == NULL, 1265 ("callout drain function already set to %p", 1266 cc_exec_drain(cc, direct))); 1267 cc_exec_drain(cc, direct) = drain; 1268 } 1269 } 1270 KASSERT(!sq_locked, ("sleepqueue chain still locked")); 1271 cancelled = 0; 1272 } else 1273 cancelled = 1; 1274 1275 if (sq_locked) 1276 sleepq_release(&cc_exec_waiting(cc, direct)); 1277 1278 if ((c->c_iflags & CALLOUT_PENDING) == 0) { 1279 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1280 c, c->c_func, c->c_arg); 1281 /* 1282 * For not scheduled and not executing callout return 1283 * negative value. 1284 */ 1285 if (cc_exec_curr(cc, direct) != c) 1286 cancelled = -1; 1287 CC_UNLOCK(cc); 1288 return (cancelled); 1289 } 1290 1291 c->c_iflags &= ~CALLOUT_PENDING; 1292 c->c_flags &= ~CALLOUT_ACTIVE; 1293 1294 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1295 c, c->c_func, c->c_arg); 1296 if (not_on_a_list == 0) { 1297 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1298 if (cc_exec_next(cc) == c) 1299 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1300 LIST_REMOVE(c, c_links.le); 1301 } else { 1302 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1303 } 1304 } 1305 CC_UNLOCK(cc); 1306 return (cancelled); 1307 } 1308 1309 void 1310 callout_init(struct callout *c, int mpsafe) 1311 { 1312 bzero(c, sizeof *c); 1313 if (mpsafe) { 1314 c->c_lock = NULL; 1315 c->c_iflags = CALLOUT_RETURNUNLOCKED; 1316 } else { 1317 c->c_lock = &Giant.lock_object; 1318 c->c_iflags = 0; 1319 } 1320 c->c_cpu = cc_default_cpu; 1321 } 1322 1323 void 1324 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags) 1325 { 1326 bzero(c, sizeof *c); 1327 c->c_lock = lock; 1328 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0, 1329 ("callout_init_lock: bad flags %d", flags)); 1330 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0, 1331 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock")); 1332 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & LC_SLEEPABLE), 1333 ("%s: callout %p has sleepable lock", __func__, c)); 1334 c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK); 1335 c->c_cpu = cc_default_cpu; 1336 } 1337 1338 static int 1339 flssbt(sbintime_t sbt) 1340 { 1341 1342 sbt += (uint64_t)sbt >> 1; 1343 if (sizeof(long) >= sizeof(sbintime_t)) 1344 return (flsl(sbt)); 1345 if (sbt >= SBT_1S) 1346 return (flsl(((uint64_t)sbt) >> 32) + 32); 1347 return (flsl(sbt)); 1348 } 1349 1350 /* 1351 * Dump immediate statistic snapshot of the scheduled callouts. 1352 */ 1353 static int 1354 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS) 1355 { 1356 struct callout *tmp; 1357 struct callout_cpu *cc; 1358 struct callout_list *sc; 1359 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t; 1360 int ct[64], cpr[64], ccpbk[32]; 1361 int error, val, i, count, tcum, pcum, maxc, c, medc; 1362 int cpu; 1363 1364 val = 0; 1365 error = sysctl_handle_int(oidp, &val, 0, req); 1366 if (error != 0 || req->newptr == NULL) 1367 return (error); 1368 count = maxc = 0; 1369 st = spr = maxt = maxpr = 0; 1370 bzero(ccpbk, sizeof(ccpbk)); 1371 bzero(ct, sizeof(ct)); 1372 bzero(cpr, sizeof(cpr)); 1373 now = sbinuptime(); 1374 CPU_FOREACH(cpu) { 1375 cc = CC_CPU(cpu); 1376 CC_LOCK(cc); 1377 for (i = 0; i < callwheelsize; i++) { 1378 sc = &cc->cc_callwheel[i]; 1379 c = 0; 1380 LIST_FOREACH(tmp, sc, c_links.le) { 1381 c++; 1382 t = tmp->c_time - now; 1383 if (t < 0) 1384 t = 0; 1385 st += t / SBT_1US; 1386 spr += tmp->c_precision / SBT_1US; 1387 if (t > maxt) 1388 maxt = t; 1389 if (tmp->c_precision > maxpr) 1390 maxpr = tmp->c_precision; 1391 ct[flssbt(t)]++; 1392 cpr[flssbt(tmp->c_precision)]++; 1393 } 1394 if (c > maxc) 1395 maxc = c; 1396 ccpbk[fls(c + c / 2)]++; 1397 count += c; 1398 } 1399 CC_UNLOCK(cc); 1400 } 1401 1402 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++) 1403 tcum += ct[i]; 1404 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1405 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++) 1406 pcum += cpr[i]; 1407 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1408 for (i = 0, c = 0; i < 32 && c < count / 2; i++) 1409 c += ccpbk[i]; 1410 medc = (i >= 2) ? (1 << (i - 2)) : 0; 1411 1412 printf("Scheduled callouts statistic snapshot:\n"); 1413 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n", 1414 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT); 1415 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n", 1416 medc, 1417 count / callwheelsize / mp_ncpus, 1418 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000, 1419 maxc); 1420 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1421 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32, 1422 (st / count) / 1000000, (st / count) % 1000000, 1423 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32); 1424 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1425 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32, 1426 (spr / count) / 1000000, (spr / count) % 1000000, 1427 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32); 1428 printf(" Distribution: \tbuckets\t time\t tcum\t" 1429 " prec\t pcum\n"); 1430 for (i = 0, tcum = pcum = 0; i < 64; i++) { 1431 if (ct[i] == 0 && cpr[i] == 0) 1432 continue; 1433 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0; 1434 tcum += ct[i]; 1435 pcum += cpr[i]; 1436 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n", 1437 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32, 1438 i - 1 - (32 - CC_HASH_SHIFT), 1439 ct[i], tcum, cpr[i], pcum); 1440 } 1441 return (error); 1442 } 1443 SYSCTL_PROC(_kern, OID_AUTO, callout_stat, 1444 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1445 0, 0, sysctl_kern_callout_stat, "I", 1446 "Dump immediate statistic snapshot of the scheduled callouts"); 1447 1448 #ifdef DDB 1449 static void 1450 _show_callout(struct callout *c) 1451 { 1452 1453 db_printf("callout %p\n", c); 1454 #define C_DB_PRINTF(f, e) db_printf(" %s = " f "\n", #e, c->e); 1455 db_printf(" &c_links = %p\n", &(c->c_links)); 1456 C_DB_PRINTF("%" PRId64, c_time); 1457 C_DB_PRINTF("%" PRId64, c_precision); 1458 C_DB_PRINTF("%p", c_arg); 1459 C_DB_PRINTF("%p", c_func); 1460 C_DB_PRINTF("%p", c_lock); 1461 C_DB_PRINTF("%#x", c_flags); 1462 C_DB_PRINTF("%#x", c_iflags); 1463 C_DB_PRINTF("%d", c_cpu); 1464 #undef C_DB_PRINTF 1465 } 1466 1467 DB_SHOW_COMMAND(callout, db_show_callout) 1468 { 1469 1470 if (!have_addr) { 1471 db_printf("usage: show callout <struct callout *>\n"); 1472 return; 1473 } 1474 1475 _show_callout((struct callout *)addr); 1476 } 1477 1478 static void 1479 _show_last_callout(int cpu, int direct, const char *dirstr) 1480 { 1481 struct callout_cpu *cc; 1482 void *func, *arg; 1483 1484 cc = CC_CPU(cpu); 1485 func = cc_exec_last_func(cc, direct); 1486 arg = cc_exec_last_arg(cc, direct); 1487 db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func); 1488 db_printsym((db_expr_t)func, DB_STGY_ANY); 1489 db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg); 1490 } 1491 1492 DB_SHOW_COMMAND(callout_last, db_show_callout_last) 1493 { 1494 int cpu, last; 1495 1496 if (have_addr) { 1497 if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) { 1498 db_printf("no such cpu: %d\n", (int)addr); 1499 return; 1500 } 1501 cpu = last = addr; 1502 } else { 1503 cpu = 0; 1504 last = mp_maxid; 1505 } 1506 1507 while (cpu <= last) { 1508 if (!CPU_ABSENT(cpu)) { 1509 _show_last_callout(cpu, 0, ""); 1510 _show_last_callout(cpu, 1, " direct"); 1511 } 1512 cpu++; 1513 } 1514 } 1515 #endif /* DDB */
Cache object: e7234aedb9854287aaa718fa4e80550b
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]