1/*- 2 * Copyright (c) 1982, 1986, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 35 */ 36 37#include <sys/cdefs.h> 38__FBSDID("$FreeBSD: stable/11/sys/kern/kern_timeout.c 360633 2020-05-04 16:30:36Z jhb $"); 39 40#include "opt_callout_profiling.h" 41#include "opt_ddb.h" 42#if defined(__arm__) 43#include "opt_timer.h" 44#endif 45#include "opt_rss.h" 46 47#include <sys/param.h> 48#include <sys/systm.h> 49#include <sys/bus.h> 50#include <sys/callout.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 <machine/_inttypes.h> 67#endif 68 69#ifdef SMP 70#include <machine/cpu.h> 71#endif 72 73#ifndef NO_EVENTTIMERS 74DPCPU_DECLARE(sbintime_t, hardclocktime); 75#endif 76 77SDT_PROVIDER_DEFINE(callout_execute); 78SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *"); 79SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *"); 80 81#ifdef CALLOUT_PROFILING 82static int avg_depth; 83SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0, 84 "Average number of items examined per softclock call. Units = 1/1000"); 85static int avg_gcalls; 86SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0, 87 "Average number of Giant callouts made per softclock call. Units = 1/1000"); 88static int avg_lockcalls; 89SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0, 90 "Average number of lock callouts made per softclock call. Units = 1/1000"); 91static int avg_mpcalls; 92SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0, 93 "Average number of MP callouts made per softclock call. Units = 1/1000"); 94static int avg_depth_dir; 95SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0, 96 "Average number of direct callouts examined per callout_process call. " 97 "Units = 1/1000"); 98static int avg_lockcalls_dir; 99SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD, 100 &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per " 101 "callout_process call. Units = 1/1000"); 102static int avg_mpcalls_dir; 103SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir, 104 0, "Average number of MP direct callouts made per callout_process call. " 105 "Units = 1/1000"); 106#endif 107 108static int ncallout; 109SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0, 110 "Number of entries in callwheel and size of timeout() preallocation"); 111 112#ifdef RSS 113static int pin_default_swi = 1; 114static int pin_pcpu_swi = 1; 115#else 116static int pin_default_swi = 0; 117static int pin_pcpu_swi = 0; 118#endif 119 120SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi, 121 0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)"); 122SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi, 123 0, "Pin the per-CPU swis (except PCPU 0, which is also default"); 124 125/* 126 * TODO: 127 * allocate more timeout table slots when table overflows. 128 */ 129u_int callwheelsize, 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 */ 141struct cc_exec { 142 struct callout *cc_curr; 143 callout_func_t *cc_drain; 144#ifdef SMP 145 callout_func_t *ce_migration_func; 146 void *ce_migration_arg; 147 int ce_migration_cpu; 148 sbintime_t ce_migration_time; 149 sbintime_t ce_migration_prec; 150#endif 151 bool cc_cancel; 152 bool cc_waiting; 153}; 154 155/* 156 * There is one struct callout_cpu per cpu, holding all relevant 157 * state for the callout processing thread on the individual CPU. 158 */ 159struct callout_cpu { 160 struct mtx_padalign cc_lock; 161 struct cc_exec cc_exec_entity[2]; 162 struct callout *cc_next; 163 struct callout *cc_callout; 164 struct callout_list *cc_callwheel; 165 struct callout_tailq cc_expireq; 166 struct callout_slist cc_callfree; 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 char cc_ktr_event_name[20]; 173}; 174 175#define callout_migrating(c) ((c)->c_iflags & CALLOUT_DFRMIGRATION) 176 177#define cc_exec_curr(cc, dir) cc->cc_exec_entity[dir].cc_curr 178#define cc_exec_drain(cc, dir) cc->cc_exec_entity[dir].cc_drain 179#define cc_exec_next(cc) cc->cc_next 180#define cc_exec_cancel(cc, dir) cc->cc_exec_entity[dir].cc_cancel 181#define cc_exec_waiting(cc, dir) cc->cc_exec_entity[dir].cc_waiting 182#ifdef SMP 183#define cc_migration_func(cc, dir) cc->cc_exec_entity[dir].ce_migration_func 184#define cc_migration_arg(cc, dir) cc->cc_exec_entity[dir].ce_migration_arg 185#define cc_migration_cpu(cc, dir) cc->cc_exec_entity[dir].ce_migration_cpu 186#define cc_migration_time(cc, dir) cc->cc_exec_entity[dir].ce_migration_time 187#define cc_migration_prec(cc, dir) cc->cc_exec_entity[dir].ce_migration_prec 188 189struct callout_cpu cc_cpu[MAXCPU]; 190#define CPUBLOCK MAXCPU 191#define CC_CPU(cpu) (&cc_cpu[(cpu)]) 192#define CC_SELF() CC_CPU(PCPU_GET(cpuid)) 193#else 194struct callout_cpu cc_cpu; 195#define CC_CPU(cpu) &cc_cpu 196#define CC_SELF() &cc_cpu 197#endif 198#define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock) 199#define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock) 200#define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED) 201 202static int timeout_cpu; 203 204static void callout_cpu_init(struct callout_cpu *cc, int cpu); 205static void softclock_call_cc(struct callout *c, struct callout_cpu *cc, 206#ifdef CALLOUT_PROFILING 207 int *mpcalls, int *lockcalls, int *gcalls, 208#endif 209 int direct); 210 211static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures"); 212 213/** 214 * Locked by cc_lock: 215 * cc_curr - If a callout is in progress, it is cc_curr. 216 * If cc_curr is non-NULL, threads waiting in 217 * callout_drain() will be woken up as soon as the 218 * relevant callout completes. 219 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held 220 * guarantees that the current callout will not run. 221 * The softclock() function sets this to 0 before it 222 * drops callout_lock to acquire c_lock, and it calls 223 * the handler only if curr_cancelled is still 0 after 224 * cc_lock is successfully acquired. 225 * cc_waiting - If a thread is waiting in callout_drain(), then 226 * callout_wait is nonzero. Set only when 227 * cc_curr is non-NULL. 228 */ 229 230/* 231 * Resets the execution entity tied to a specific callout cpu. 232 */ 233static void 234cc_cce_cleanup(struct callout_cpu *cc, int direct) 235{ 236 237 cc_exec_curr(cc, direct) = NULL; 238 cc_exec_cancel(cc, direct) = false; 239 cc_exec_waiting(cc, direct) = false; 240#ifdef SMP 241 cc_migration_cpu(cc, direct) = CPUBLOCK; 242 cc_migration_time(cc, direct) = 0; 243 cc_migration_prec(cc, direct) = 0; 244 cc_migration_func(cc, direct) = NULL; 245 cc_migration_arg(cc, direct) = NULL; 246#endif 247} 248 249/* 250 * Checks if migration is requested by a specific callout cpu. 251 */ 252static int 253cc_cce_migrating(struct callout_cpu *cc, int direct) 254{ 255 256#ifdef SMP 257 return (cc_migration_cpu(cc, direct) != CPUBLOCK); 258#else 259 return (0); 260#endif 261} 262 263/* 264 * Kernel low level callwheel initialization 265 * called on cpu0 during kernel startup. 266 */ 267static void 268callout_callwheel_init(void *dummy) 269{ 270 struct callout_cpu *cc; 271 272 /* 273 * Calculate the size of the callout wheel and the preallocated 274 * timeout() structures. 275 * XXX: Clip callout to result of previous function of maxusers 276 * maximum 384. This is still huge, but acceptable. 277 */ 278 memset(CC_CPU(0), 0, sizeof(cc_cpu)); 279 ncallout = imin(16 + maxproc + maxfiles, 18508); 280 TUNABLE_INT_FETCH("kern.ncallout", &ncallout); 281 282 /* 283 * Calculate callout wheel size, should be next power of two higher 284 * than 'ncallout'. 285 */ 286 callwheelsize = 1 << fls(ncallout); 287 callwheelmask = callwheelsize - 1; 288 289 /* 290 * Fetch whether we're pinning the swi's or not. 291 */ 292 TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi); 293 TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi); 294 295 /* 296 * Only cpu0 handles timeout(9) and receives a preallocation. 297 * 298 * XXX: Once all timeout(9) consumers are converted this can 299 * be removed. 300 */ 301 timeout_cpu = PCPU_GET(cpuid); 302 cc = CC_CPU(timeout_cpu); 303 cc->cc_callout = malloc(ncallout * sizeof(struct callout), 304 M_CALLOUT, M_WAITOK); 305 callout_cpu_init(cc, timeout_cpu); 306} 307SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL); 308 309/* 310 * Initialize the per-cpu callout structures. 311 */ 312static void 313callout_cpu_init(struct callout_cpu *cc, int cpu) 314{ 315 struct callout *c; 316 int i; 317 318 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE); 319 SLIST_INIT(&cc->cc_callfree); 320 cc->cc_inited = 1; 321 cc->cc_callwheel = malloc(sizeof(struct callout_list) * callwheelsize, 322 M_CALLOUT, M_WAITOK); 323 for (i = 0; i < callwheelsize; i++) 324 LIST_INIT(&cc->cc_callwheel[i]); 325 TAILQ_INIT(&cc->cc_expireq); 326 cc->cc_firstevent = SBT_MAX; 327 for (i = 0; i < 2; i++) 328 cc_cce_cleanup(cc, i); 329 snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name), 330 "callwheel cpu %d", cpu); 331 if (cc->cc_callout == NULL) /* Only cpu0 handles timeout(9) */ 332 return; 333 for (i = 0; i < ncallout; i++) { 334 c = &cc->cc_callout[i]; 335 callout_init(c, 0); 336 c->c_iflags = CALLOUT_LOCAL_ALLOC; 337 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 338 } 339} 340 341#ifdef SMP 342/* 343 * Switches the cpu tied to a specific callout. 344 * The function expects a locked incoming callout cpu and returns with 345 * locked outcoming callout cpu. 346 */ 347static struct callout_cpu * 348callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu) 349{ 350 struct callout_cpu *new_cc; 351 352 MPASS(c != NULL && cc != NULL); 353 CC_LOCK_ASSERT(cc); 354 355 /* 356 * Avoid interrupts and preemption firing after the callout cpu 357 * is blocked in order to avoid deadlocks as the new thread 358 * may be willing to acquire the callout cpu lock. 359 */ 360 c->c_cpu = CPUBLOCK; 361 spinlock_enter(); 362 CC_UNLOCK(cc); 363 new_cc = CC_CPU(new_cpu); 364 CC_LOCK(new_cc); 365 spinlock_exit(); 366 c->c_cpu = new_cpu; 367 return (new_cc); 368} 369#endif 370 371/* 372 * Start standard softclock thread. 373 */ 374static void 375start_softclock(void *dummy) 376{ 377 struct callout_cpu *cc; 378 char name[MAXCOMLEN]; 379#ifdef SMP 380 int cpu; 381 struct intr_event *ie; 382#endif 383 384 cc = CC_CPU(timeout_cpu); 385 snprintf(name, sizeof(name), "clock (%d)", timeout_cpu); 386 if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK, 387 INTR_MPSAFE, &cc->cc_cookie)) 388 panic("died while creating standard software ithreads"); 389 if (pin_default_swi && 390 (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) { 391 printf("%s: timeout clock couldn't be pinned to cpu %d\n", 392 __func__, 393 timeout_cpu); 394 } 395 396#ifdef SMP 397 CPU_FOREACH(cpu) { 398 if (cpu == timeout_cpu) 399 continue; 400 cc = CC_CPU(cpu); 401 cc->cc_callout = NULL; /* Only cpu0 handles timeout(9). */ 402 callout_cpu_init(cc, cpu); 403 snprintf(name, sizeof(name), "clock (%d)", cpu); 404 ie = NULL; 405 if (swi_add(&ie, name, softclock, cc, SWI_CLOCK, 406 INTR_MPSAFE, &cc->cc_cookie)) 407 panic("died while creating standard software ithreads"); 408 if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) { 409 printf("%s: per-cpu clock couldn't be pinned to " 410 "cpu %d\n", 411 __func__, 412 cpu); 413 } 414 } 415#endif 416} 417SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL); 418 419#define CC_HASH_SHIFT 8 420 421static inline u_int 422callout_hash(sbintime_t sbt) 423{ 424 425 return (sbt >> (32 - CC_HASH_SHIFT)); 426} 427 428static inline u_int 429callout_get_bucket(sbintime_t sbt) 430{ 431 432 return (callout_hash(sbt) & callwheelmask); 433} 434 435void 436callout_process(sbintime_t now) 437{ 438 struct callout *tmp, *tmpn; 439 struct callout_cpu *cc; 440 struct callout_list *sc; 441 sbintime_t first, last, max, tmp_max; 442 uint32_t lookahead; 443 u_int firstb, lastb, nowb; 444#ifdef CALLOUT_PROFILING 445 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0; 446#endif 447 448 cc = CC_SELF(); 449 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET); 450 451 /* Compute the buckets of the last scan and present times. */ 452 firstb = callout_hash(cc->cc_lastscan); 453 cc->cc_lastscan = now; 454 nowb = callout_hash(now); 455 456 /* Compute the last bucket and minimum time of the bucket after it. */ 457 if (nowb == firstb) 458 lookahead = (SBT_1S / 16); 459 else if (nowb - firstb == 1) 460 lookahead = (SBT_1S / 8); 461 else 462 lookahead = (SBT_1S / 2); 463 first = last = now; 464 first += (lookahead / 2); 465 last += lookahead; 466 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT)); 467 lastb = callout_hash(last) - 1; 468 max = last; 469 470 /* 471 * Check if we wrapped around the entire wheel from the last scan. 472 * In case, we need to scan entirely the wheel for pending callouts. 473 */ 474 if (lastb - firstb >= callwheelsize) { 475 lastb = firstb + callwheelsize - 1; 476 if (nowb - firstb >= callwheelsize) 477 nowb = lastb; 478 } 479 480 /* Iterate callwheel from firstb to nowb and then up to lastb. */ 481 do { 482 sc = &cc->cc_callwheel[firstb & callwheelmask]; 483 tmp = LIST_FIRST(sc); 484 while (tmp != NULL) { 485 /* Run the callout if present time within allowed. */ 486 if (tmp->c_time <= now) { 487 /* 488 * Consumer told us the callout may be run 489 * directly from hardware interrupt context. 490 */ 491 if (tmp->c_iflags & CALLOUT_DIRECT) { 492#ifdef CALLOUT_PROFILING 493 ++depth_dir; 494#endif 495 cc_exec_next(cc) = 496 LIST_NEXT(tmp, c_links.le); 497 cc->cc_bucket = firstb & callwheelmask; 498 LIST_REMOVE(tmp, c_links.le); 499 softclock_call_cc(tmp, cc, 500#ifdef CALLOUT_PROFILING 501 &mpcalls_dir, &lockcalls_dir, NULL, 502#endif 503 1); 504 tmp = cc_exec_next(cc); 505 cc_exec_next(cc) = NULL; 506 } else { 507 tmpn = LIST_NEXT(tmp, c_links.le); 508 LIST_REMOVE(tmp, c_links.le); 509 TAILQ_INSERT_TAIL(&cc->cc_expireq, 510 tmp, c_links.tqe); 511 tmp->c_iflags |= CALLOUT_PROCESSED; 512 tmp = tmpn; 513 } 514 continue; 515 } 516 /* Skip events from distant future. */ 517 if (tmp->c_time >= max) 518 goto next; 519 /* 520 * Event minimal time is bigger than present maximal 521 * time, so it cannot be aggregated. 522 */ 523 if (tmp->c_time > last) { 524 lastb = nowb; 525 goto next; 526 } 527 /* Update first and last time, respecting this event. */ 528 if (tmp->c_time < first) 529 first = tmp->c_time; 530 tmp_max = tmp->c_time + tmp->c_precision; 531 if (tmp_max < last) 532 last = tmp_max; 533next: 534 tmp = LIST_NEXT(tmp, c_links.le); 535 } 536 /* Proceed with the next bucket. */ 537 firstb++; 538 /* 539 * Stop if we looked after present time and found 540 * some event we can't execute at now. 541 * Stop if we looked far enough into the future. 542 */ 543 } while (((int)(firstb - lastb)) <= 0); 544 cc->cc_firstevent = last; 545#ifndef NO_EVENTTIMERS 546 cpu_new_callout(curcpu, last, first); 547#endif 548#ifdef CALLOUT_PROFILING 549 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8; 550 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8; 551 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8; 552#endif 553 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET); 554 /* 555 * swi_sched acquires the thread lock, so we don't want to call it 556 * with cc_lock held; incorrect locking order. 557 */ 558 if (!TAILQ_EMPTY(&cc->cc_expireq)) 559 swi_sched(cc->cc_cookie, 0); 560} 561 562static struct callout_cpu * 563callout_lock(struct callout *c) 564{ 565 struct callout_cpu *cc; 566 int cpu; 567 568 for (;;) { 569 cpu = c->c_cpu; 570#ifdef SMP 571 if (cpu == CPUBLOCK) { 572 while (c->c_cpu == CPUBLOCK) 573 cpu_spinwait(); 574 continue; 575 } 576#endif 577 cc = CC_CPU(cpu); 578 CC_LOCK(cc); 579 if (cpu == c->c_cpu) 580 break; 581 CC_UNLOCK(cc); 582 } 583 return (cc); 584} 585 586static void 587callout_cc_add(struct callout *c, struct callout_cpu *cc, 588 sbintime_t sbt, sbintime_t precision, void (*func)(void *), 589 void *arg, int cpu, int flags) 590{ 591 int bucket; 592 593 CC_LOCK_ASSERT(cc); 594 if (sbt < cc->cc_lastscan) 595 sbt = cc->cc_lastscan; 596 c->c_arg = arg; 597 c->c_iflags |= CALLOUT_PENDING; 598 c->c_iflags &= ~CALLOUT_PROCESSED; 599 c->c_flags |= CALLOUT_ACTIVE; 600 if (flags & C_DIRECT_EXEC) 601 c->c_iflags |= CALLOUT_DIRECT; 602 c->c_func = func; 603 c->c_time = sbt; 604 c->c_precision = precision; 605 bucket = callout_get_bucket(c->c_time); 606 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x", 607 c, (int)(c->c_precision >> 32), 608 (u_int)(c->c_precision & 0xffffffff)); 609 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le); 610 if (cc->cc_bucket == bucket) 611 cc_exec_next(cc) = c; 612#ifndef NO_EVENTTIMERS 613 /* 614 * Inform the eventtimers(4) subsystem there's a new callout 615 * that has been inserted, but only if really required. 616 */ 617 if (SBT_MAX - c->c_time < c->c_precision) 618 c->c_precision = SBT_MAX - c->c_time; 619 sbt = c->c_time + c->c_precision; 620 if (sbt < cc->cc_firstevent) { 621 cc->cc_firstevent = sbt; 622 cpu_new_callout(cpu, sbt, c->c_time); 623 } 624#endif 625} 626 627static void 628callout_cc_del(struct callout *c, struct callout_cpu *cc) 629{ 630 631 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0) 632 return; 633 c->c_func = NULL; 634 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 635} 636 637static void 638softclock_call_cc(struct callout *c, struct callout_cpu *cc, 639#ifdef CALLOUT_PROFILING 640 int *mpcalls, int *lockcalls, int *gcalls, 641#endif 642 int direct) 643{ 644 struct rm_priotracker tracker; 645 callout_func_t *c_func, *drain; 646 void *c_arg; 647 struct lock_class *class; 648 struct lock_object *c_lock; 649 uintptr_t lock_status; 650 int c_iflags; 651#ifdef SMP 652 struct callout_cpu *new_cc; 653 callout_func_t *new_func; 654 void *new_arg; 655 int flags, new_cpu; 656 sbintime_t new_prec, new_time; 657#endif 658#if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 659 sbintime_t sbt1, sbt2; 660 struct timespec ts2; 661 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */ 662 static callout_func_t *lastfunc; 663#endif 664 665 KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING, 666 ("softclock_call_cc: pend %p %x", c, c->c_iflags)); 667 KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE, 668 ("softclock_call_cc: act %p %x", c, c->c_flags)); 669 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL; 670 lock_status = 0; 671 if (c->c_flags & CALLOUT_SHAREDLOCK) { 672 if (class == &lock_class_rm) 673 lock_status = (uintptr_t)&tracker; 674 else 675 lock_status = 1; 676 } 677 c_lock = c->c_lock; 678 c_func = c->c_func; 679 c_arg = c->c_arg; 680 c_iflags = c->c_iflags; 681 if (c->c_iflags & CALLOUT_LOCAL_ALLOC) 682 c->c_iflags = CALLOUT_LOCAL_ALLOC; 683 else 684 c->c_iflags &= ~CALLOUT_PENDING; 685 686 cc_exec_curr(cc, direct) = c; 687 cc_exec_cancel(cc, direct) = false; 688 cc_exec_drain(cc, direct) = NULL; 689 CC_UNLOCK(cc); 690 if (c_lock != NULL) { 691 class->lc_lock(c_lock, lock_status); 692 /* 693 * The callout may have been cancelled 694 * while we switched locks. 695 */ 696 if (cc_exec_cancel(cc, direct)) { 697 class->lc_unlock(c_lock); 698 goto skip; 699 } 700 /* The callout cannot be stopped now. */ 701 cc_exec_cancel(cc, direct) = true; 702 if (c_lock == &Giant.lock_object) { 703#ifdef CALLOUT_PROFILING 704 (*gcalls)++; 705#endif 706 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p", 707 c, c_func, c_arg); 708 } else { 709#ifdef CALLOUT_PROFILING 710 (*lockcalls)++; 711#endif 712 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p", 713 c, c_func, c_arg); 714 } 715 } else { 716#ifdef CALLOUT_PROFILING 717 (*mpcalls)++; 718#endif 719 CTR3(KTR_CALLOUT, "callout %p func %p arg %p", 720 c, c_func, c_arg); 721 } 722 KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running", 723 "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct); 724#if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 725 sbt1 = sbinuptime(); 726#endif 727 THREAD_NO_SLEEPING(); 728 SDT_PROBE1(callout_execute, , , callout__start, c); 729 c_func(c_arg); 730 SDT_PROBE1(callout_execute, , , callout__end, c); 731 THREAD_SLEEPING_OK(); 732#if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 733 sbt2 = sbinuptime(); 734 sbt2 -= sbt1; 735 if (sbt2 > maxdt) { 736 if (lastfunc != c_func || sbt2 > maxdt * 2) { 737 ts2 = sbttots(sbt2); 738 printf( 739 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n", 740 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec); 741 } 742 maxdt = sbt2; 743 lastfunc = c_func; 744 } 745#endif 746 KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle"); 747 CTR1(KTR_CALLOUT, "callout %p finished", c); 748 if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0) 749 class->lc_unlock(c_lock); 750skip: 751 CC_LOCK(cc); 752 KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr")); 753 cc_exec_curr(cc, direct) = NULL; 754 if (cc_exec_drain(cc, direct)) { 755 drain = cc_exec_drain(cc, direct); 756 cc_exec_drain(cc, direct) = NULL; 757 CC_UNLOCK(cc); 758 drain(c_arg); 759 CC_LOCK(cc); 760 } 761 if (cc_exec_waiting(cc, direct)) { 762 /* 763 * There is someone waiting for the 764 * callout to complete. 765 * If the callout was scheduled for 766 * migration just cancel it. 767 */ 768 if (cc_cce_migrating(cc, direct)) { 769 cc_cce_cleanup(cc, direct); 770 771 /* 772 * It should be assert here that the callout is not 773 * destroyed but that is not easy. 774 */ 775 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 776 } 777 cc_exec_waiting(cc, direct) = false; 778 CC_UNLOCK(cc); 779 wakeup(&cc_exec_waiting(cc, direct)); 780 CC_LOCK(cc); 781 } else if (cc_cce_migrating(cc, direct)) { 782 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0, 783 ("Migrating legacy callout %p", c)); 784#ifdef SMP 785 /* 786 * If the callout was scheduled for 787 * migration just perform it now. 788 */ 789 new_cpu = cc_migration_cpu(cc, direct); 790 new_time = cc_migration_time(cc, direct); 791 new_prec = cc_migration_prec(cc, direct); 792 new_func = cc_migration_func(cc, direct); 793 new_arg = cc_migration_arg(cc, direct); 794 cc_cce_cleanup(cc, direct); 795 796 /* 797 * It should be assert here that the callout is not destroyed 798 * but that is not easy. 799 * 800 * As first thing, handle deferred callout stops. 801 */ 802 if (!callout_migrating(c)) { 803 CTR3(KTR_CALLOUT, 804 "deferred cancelled %p func %p arg %p", 805 c, new_func, new_arg); 806 callout_cc_del(c, cc); 807 return; 808 } 809 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 810 811 new_cc = callout_cpu_switch(c, cc, new_cpu); 812 flags = (direct) ? C_DIRECT_EXEC : 0; 813 callout_cc_add(c, new_cc, new_time, new_prec, new_func, 814 new_arg, new_cpu, flags); 815 CC_UNLOCK(new_cc); 816 CC_LOCK(cc); 817#else 818 panic("migration should not happen"); 819#endif 820 } 821 /* 822 * If the current callout is locally allocated (from 823 * timeout(9)) then put it on the freelist. 824 * 825 * Note: we need to check the cached copy of c_iflags because 826 * if it was not local, then it's not safe to deref the 827 * callout pointer. 828 */ 829 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 || 830 c->c_iflags == CALLOUT_LOCAL_ALLOC, 831 ("corrupted callout")); 832 if (c_iflags & CALLOUT_LOCAL_ALLOC) 833 callout_cc_del(c, cc); 834} 835 836/* 837 * The callout mechanism is based on the work of Adam M. Costello and 838 * George Varghese, published in a technical report entitled "Redesigning 839 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 840 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 841 * used in this implementation was published by G. Varghese and T. Lauck in 842 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 843 * the Efficient Implementation of a Timer Facility" in the Proceedings of 844 * the 11th ACM Annual Symposium on Operating Systems Principles, 845 * Austin, Texas Nov 1987. 846 */ 847 848/* 849 * Software (low priority) clock interrupt. 850 * Run periodic events from timeout queue. 851 */ 852void 853softclock(void *arg) 854{ 855 struct callout_cpu *cc; 856 struct callout *c; 857#ifdef CALLOUT_PROFILING 858 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0; 859#endif 860 861 cc = (struct callout_cpu *)arg; 862 CC_LOCK(cc); 863 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) { 864 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 865 softclock_call_cc(c, cc, 866#ifdef CALLOUT_PROFILING 867 &mpcalls, &lockcalls, &gcalls, 868#endif 869 0); 870#ifdef CALLOUT_PROFILING 871 ++depth; 872#endif 873 } 874#ifdef CALLOUT_PROFILING 875 avg_depth += (depth * 1000 - avg_depth) >> 8; 876 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8; 877 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8; 878 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8; 879#endif 880 CC_UNLOCK(cc); 881} 882 883/* 884 * timeout -- 885 * Execute a function after a specified length of time. 886 * 887 * untimeout -- 888 * Cancel previous timeout function call. 889 * 890 * callout_handle_init -- 891 * Initialize a handle so that using it with untimeout is benign. 892 * 893 * See AT&T BCI Driver Reference Manual for specification. This 894 * implementation differs from that one in that although an 895 * identification value is returned from timeout, the original 896 * arguments to timeout as well as the identifier are used to 897 * identify entries for untimeout. 898 */ 899struct callout_handle 900timeout(timeout_t *ftn, void *arg, int to_ticks) 901{ 902 struct callout_cpu *cc; 903 struct callout *new; 904 struct callout_handle handle; 905 906 cc = CC_CPU(timeout_cpu); 907 CC_LOCK(cc); 908 /* Fill in the next free callout structure. */ 909 new = SLIST_FIRST(&cc->cc_callfree); 910 if (new == NULL) 911 /* XXX Attempt to malloc first */ 912 panic("timeout table full"); 913 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle); 914 callout_reset(new, to_ticks, ftn, arg); 915 handle.callout = new; 916 CC_UNLOCK(cc); 917 918 return (handle); 919} 920 921void 922untimeout(timeout_t *ftn, void *arg, struct callout_handle handle) 923{ 924 struct callout_cpu *cc; 925 926 /* 927 * Check for a handle that was initialized 928 * by callout_handle_init, but never used 929 * for a real timeout. 930 */ 931 if (handle.callout == NULL) 932 return; 933 934 cc = callout_lock(handle.callout); 935 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg) 936 callout_stop(handle.callout); 937 CC_UNLOCK(cc); 938} 939 940void 941callout_handle_init(struct callout_handle *handle) 942{ 943 handle->callout = NULL; 944} 945 946void 947callout_when(sbintime_t sbt, sbintime_t precision, int flags, 948 sbintime_t *res, sbintime_t *prec_res) 949{ 950 sbintime_t to_sbt, to_pr; 951 952 if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) { 953 *res = sbt; 954 *prec_res = precision; 955 return; 956 } 957 if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt) 958 sbt = tick_sbt; 959 if ((flags & C_HARDCLOCK) != 0 || 960#ifdef NO_EVENTTIMERS 961 sbt >= sbt_timethreshold) { 962 to_sbt = getsbinuptime(); 963 964 /* Add safety belt for the case of hz > 1000. */ 965 to_sbt += tc_tick_sbt - tick_sbt; 966#else 967 sbt >= sbt_tickthreshold) { 968 /* 969 * Obtain the time of the last hardclock() call on 970 * this CPU directly from the kern_clocksource.c. 971 * This value is per-CPU, but it is equal for all 972 * active ones. 973 */ 974#ifdef __LP64__ 975 to_sbt = DPCPU_GET(hardclocktime); 976#else 977 spinlock_enter(); 978 to_sbt = DPCPU_GET(hardclocktime); 979 spinlock_exit(); 980#endif 981#endif 982 if (cold && to_sbt == 0) 983 to_sbt = sbinuptime(); 984 if ((flags & C_HARDCLOCK) == 0) 985 to_sbt += tick_sbt; 986 } else 987 to_sbt = sbinuptime(); 988 if (SBT_MAX - to_sbt < sbt) 989 to_sbt = SBT_MAX; 990 else 991 to_sbt += sbt; 992 *res = to_sbt; 993 to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp : 994 sbt >> C_PRELGET(flags)); 995 *prec_res = to_pr > precision ? to_pr : precision; 996} 997 998/* 999 * New interface; clients allocate their own callout structures. 1000 * 1001 * callout_reset() - establish or change a timeout 1002 * callout_stop() - disestablish a timeout 1003 * callout_init() - initialize a callout structure so that it can 1004 * safely be passed to callout_reset() and callout_stop() 1005 * 1006 * <sys/callout.h> defines three convenience macros: 1007 * 1008 * callout_active() - returns truth if callout has not been stopped, 1009 * drained, or deactivated since the last time the callout was 1010 * reset. 1011 * callout_pending() - returns truth if callout is still waiting for timeout 1012 * callout_deactivate() - marks the callout as having been serviced 1013 */ 1014int 1015callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec, 1016 callout_func_t *ftn, void *arg, int cpu, int flags) 1017{ 1018 sbintime_t to_sbt, precision; 1019 struct callout_cpu *cc; 1020 int cancelled, direct; 1021 int ignore_cpu=0; 1022 1023 cancelled = 0; 1024 if (cpu == -1) { 1025 ignore_cpu = 1; 1026 } else if ((cpu >= MAXCPU) || 1027 ((CC_CPU(cpu))->cc_inited == 0)) { 1028 /* Invalid CPU spec */ 1029 panic("Invalid CPU in callout %d", cpu); 1030 } 1031 callout_when(sbt, prec, flags, &to_sbt, &precision); 1032 1033 /* 1034 * This flag used to be added by callout_cc_add, but the 1035 * first time you call this we could end up with the 1036 * wrong direct flag if we don't do it before we add. 1037 */ 1038 if (flags & C_DIRECT_EXEC) { 1039 direct = 1; 1040 } else { 1041 direct = 0; 1042 } 1043 KASSERT(!direct || c->c_lock == NULL, 1044 ("%s: direct callout %p has lock", __func__, c)); 1045 cc = callout_lock(c); 1046 /* 1047 * Don't allow migration of pre-allocated callouts lest they 1048 * become unbalanced or handle the case where the user does 1049 * not care. 1050 */ 1051 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) || 1052 ignore_cpu) { 1053 cpu = c->c_cpu; 1054 } 1055 1056 if (cc_exec_curr(cc, direct) == c) { 1057 /* 1058 * We're being asked to reschedule a callout which is 1059 * currently in progress. If there is a lock then we 1060 * can cancel the callout if it has not really started. 1061 */ 1062 if (c->c_lock != NULL && !cc_exec_cancel(cc, direct)) 1063 cancelled = cc_exec_cancel(cc, direct) = true; 1064 if (cc_exec_waiting(cc, direct)) { 1065 /* 1066 * Someone has called callout_drain to kill this 1067 * callout. Don't reschedule. 1068 */ 1069 CTR4(KTR_CALLOUT, "%s %p func %p arg %p", 1070 cancelled ? "cancelled" : "failed to cancel", 1071 c, c->c_func, c->c_arg); 1072 CC_UNLOCK(cc); 1073 return (cancelled); 1074 } 1075#ifdef SMP 1076 if (callout_migrating(c)) { 1077 /* 1078 * This only occurs when a second callout_reset_sbt_on 1079 * is made after a previous one moved it into 1080 * deferred migration (below). Note we do *not* change 1081 * the prev_cpu even though the previous target may 1082 * be different. 1083 */ 1084 cc_migration_cpu(cc, direct) = cpu; 1085 cc_migration_time(cc, direct) = to_sbt; 1086 cc_migration_prec(cc, direct) = precision; 1087 cc_migration_func(cc, direct) = ftn; 1088 cc_migration_arg(cc, direct) = arg; 1089 cancelled = 1; 1090 CC_UNLOCK(cc); 1091 return (cancelled); 1092 } 1093#endif 1094 } 1095 if (c->c_iflags & CALLOUT_PENDING) { 1096 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1097 if (cc_exec_next(cc) == c) 1098 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1099 LIST_REMOVE(c, c_links.le); 1100 } else { 1101 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1102 } 1103 cancelled = 1; 1104 c->c_iflags &= ~ CALLOUT_PENDING; 1105 c->c_flags &= ~ CALLOUT_ACTIVE; 1106 } 1107 1108#ifdef SMP 1109 /* 1110 * If the callout must migrate try to perform it immediately. 1111 * If the callout is currently running, just defer the migration 1112 * to a more appropriate moment. 1113 */ 1114 if (c->c_cpu != cpu) { 1115 if (cc_exec_curr(cc, direct) == c) { 1116 /* 1117 * Pending will have been removed since we are 1118 * actually executing the callout on another 1119 * CPU. That callout should be waiting on the 1120 * lock the caller holds. If we set both 1121 * active/and/pending after we return and the 1122 * lock on the executing callout proceeds, it 1123 * will then see pending is true and return. 1124 * At the return from the actual callout execution 1125 * the migration will occur in softclock_call_cc 1126 * and this new callout will be placed on the 1127 * new CPU via a call to callout_cpu_switch() which 1128 * will get the lock on the right CPU followed 1129 * by a call callout_cc_add() which will add it there. 1130 * (see above in softclock_call_cc()). 1131 */ 1132 cc_migration_cpu(cc, direct) = cpu; 1133 cc_migration_time(cc, direct) = to_sbt; 1134 cc_migration_prec(cc, direct) = precision; 1135 cc_migration_func(cc, direct) = ftn; 1136 cc_migration_arg(cc, direct) = arg; 1137 c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING); 1138 c->c_flags |= CALLOUT_ACTIVE; 1139 CTR6(KTR_CALLOUT, 1140 "migration of %p func %p arg %p in %d.%08x to %u deferred", 1141 c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1142 (u_int)(to_sbt & 0xffffffff), cpu); 1143 CC_UNLOCK(cc); 1144 return (cancelled); 1145 } 1146 cc = callout_cpu_switch(c, cc, cpu); 1147 } 1148#endif 1149 1150 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags); 1151 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x", 1152 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1153 (u_int)(to_sbt & 0xffffffff)); 1154 CC_UNLOCK(cc); 1155 1156 return (cancelled); 1157} 1158 1159/* 1160 * Common idioms that can be optimized in the future. 1161 */ 1162int 1163callout_schedule_on(struct callout *c, int to_ticks, int cpu) 1164{ 1165 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu); 1166} 1167 1168int 1169callout_schedule(struct callout *c, int to_ticks) 1170{ 1171 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu); 1172} 1173 1174int 1175_callout_stop_safe(struct callout *c, int flags, callout_func_t *drain) 1176{ 1177 struct callout_cpu *cc, *old_cc; 1178 struct lock_class *class; 1179 int direct, sq_locked, use_lock; 1180 int cancelled, not_on_a_list; 1181 1182 if ((flags & CS_DRAIN) != 0) 1183 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock, 1184 "calling %s", __func__); 1185 1186 /* 1187 * Some old subsystems don't hold Giant while running a callout_stop(), 1188 * so just discard this check for the moment. 1189 */ 1190 if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) { 1191 if (c->c_lock == &Giant.lock_object) 1192 use_lock = mtx_owned(&Giant); 1193 else { 1194 use_lock = 1; 1195 class = LOCK_CLASS(c->c_lock); 1196 class->lc_assert(c->c_lock, LA_XLOCKED); 1197 } 1198 } else 1199 use_lock = 0; 1200 if (c->c_iflags & CALLOUT_DIRECT) { 1201 direct = 1; 1202 } else { 1203 direct = 0; 1204 } 1205 sq_locked = 0; 1206 old_cc = NULL; 1207again: 1208 cc = callout_lock(c); 1209 1210 if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) == 1211 (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) && 1212 ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) { 1213 /* 1214 * Special case where this slipped in while we 1215 * were migrating *as* the callout is about to 1216 * execute. The caller probably holds the lock 1217 * the callout wants. 1218 * 1219 * Get rid of the migration first. Then set 1220 * the flag that tells this code *not* to 1221 * try to remove it from any lists (its not 1222 * on one yet). When the callout wheel runs, 1223 * it will ignore this callout. 1224 */ 1225 c->c_iflags &= ~CALLOUT_PENDING; 1226 c->c_flags &= ~CALLOUT_ACTIVE; 1227 not_on_a_list = 1; 1228 } else { 1229 not_on_a_list = 0; 1230 } 1231 1232 /* 1233 * If the callout was migrating while the callout cpu lock was 1234 * dropped, just drop the sleepqueue lock and check the states 1235 * again. 1236 */ 1237 if (sq_locked != 0 && cc != old_cc) { 1238#ifdef SMP 1239 CC_UNLOCK(cc); 1240 sleepq_release(&cc_exec_waiting(old_cc, direct)); 1241 sq_locked = 0; 1242 old_cc = NULL; 1243 goto again; 1244#else 1245 panic("migration should not happen"); 1246#endif 1247 } 1248 1249 /* 1250 * If the callout is running, try to stop it or drain it. 1251 */ 1252 if (cc_exec_curr(cc, direct) == c) { 1253 /* 1254 * Succeed we to stop it or not, we must clear the 1255 * active flag - this is what API users expect. If we're 1256 * draining and the callout is currently executing, first wait 1257 * until it finishes. 1258 */ 1259 if ((flags & CS_DRAIN) == 0) 1260 c->c_flags &= ~CALLOUT_ACTIVE; 1261 1262 if ((flags & CS_DRAIN) != 0) { 1263 /* 1264 * The current callout is running (or just 1265 * about to run) and blocking is allowed, so 1266 * just wait for the current invocation to 1267 * finish. 1268 */ 1269 while (cc_exec_curr(cc, direct) == c) { 1270 /* 1271 * Use direct calls to sleepqueue interface 1272 * instead of cv/msleep in order to avoid 1273 * a LOR between cc_lock and sleepqueue 1274 * chain spinlocks. This piece of code 1275 * emulates a msleep_spin() call actually. 1276 * 1277 * If we already have the sleepqueue chain 1278 * locked, then we can safely block. If we 1279 * don't already have it locked, however, 1280 * we have to drop the cc_lock to lock 1281 * it. This opens several races, so we 1282 * restart at the beginning once we have 1283 * both locks. If nothing has changed, then 1284 * we will end up back here with sq_locked 1285 * set. 1286 */ 1287 if (!sq_locked) { 1288 CC_UNLOCK(cc); 1289 sleepq_lock( 1290 &cc_exec_waiting(cc, direct)); 1291 sq_locked = 1; 1292 old_cc = cc; 1293 goto again; 1294 } 1295 1296 /* 1297 * Migration could be cancelled here, but 1298 * as long as it is still not sure when it 1299 * will be packed up, just let softclock() 1300 * take care of it. 1301 */ 1302 cc_exec_waiting(cc, direct) = true; 1303 DROP_GIANT(); 1304 CC_UNLOCK(cc); 1305 sleepq_add( 1306 &cc_exec_waiting(cc, direct), 1307 &cc->cc_lock.lock_object, "codrain", 1308 SLEEPQ_SLEEP, 0); 1309 sleepq_wait( 1310 &cc_exec_waiting(cc, direct), 1311 0); 1312 sq_locked = 0; 1313 old_cc = NULL; 1314 1315 /* Reacquire locks previously released. */ 1316 PICKUP_GIANT(); 1317 CC_LOCK(cc); 1318 } 1319 c->c_flags &= ~CALLOUT_ACTIVE; 1320 } else if (use_lock && 1321 !cc_exec_cancel(cc, direct) && (drain == NULL)) { 1322 1323 /* 1324 * The current callout is waiting for its 1325 * lock which we hold. Cancel the callout 1326 * and return. After our caller drops the 1327 * lock, the callout will be skipped in 1328 * softclock(). This *only* works with a 1329 * callout_stop() *not* callout_drain() or 1330 * callout_async_drain(). 1331 */ 1332 cc_exec_cancel(cc, direct) = true; 1333 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1334 c, c->c_func, c->c_arg); 1335 KASSERT(!cc_cce_migrating(cc, direct), 1336 ("callout wrongly scheduled for migration")); 1337 if (callout_migrating(c)) { 1338 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1339#ifdef SMP 1340 cc_migration_cpu(cc, direct) = CPUBLOCK; 1341 cc_migration_time(cc, direct) = 0; 1342 cc_migration_prec(cc, direct) = 0; 1343 cc_migration_func(cc, direct) = NULL; 1344 cc_migration_arg(cc, direct) = NULL; 1345#endif 1346 } 1347 CC_UNLOCK(cc); 1348 KASSERT(!sq_locked, ("sleepqueue chain locked")); 1349 return (1); 1350 } else if (callout_migrating(c)) { 1351 /* 1352 * The callout is currently being serviced 1353 * and the "next" callout is scheduled at 1354 * its completion with a migration. We remove 1355 * the migration flag so it *won't* get rescheduled, 1356 * but we can't stop the one thats running so 1357 * we return 0. 1358 */ 1359 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1360#ifdef SMP 1361 /* 1362 * We can't call cc_cce_cleanup here since 1363 * if we do it will remove .ce_curr and 1364 * its still running. This will prevent a 1365 * reschedule of the callout when the 1366 * execution completes. 1367 */ 1368 cc_migration_cpu(cc, direct) = CPUBLOCK; 1369 cc_migration_time(cc, direct) = 0; 1370 cc_migration_prec(cc, direct) = 0; 1371 cc_migration_func(cc, direct) = NULL; 1372 cc_migration_arg(cc, direct) = NULL; 1373#endif 1374 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p", 1375 c, c->c_func, c->c_arg); 1376 if (drain) { 1377 cc_exec_drain(cc, direct) = drain; 1378 } 1379 CC_UNLOCK(cc); 1380 return ((flags & CS_EXECUTING) != 0); 1381 } 1382 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1383 c, c->c_func, c->c_arg); 1384 if (drain) { 1385 cc_exec_drain(cc, direct) = drain; 1386 } 1387 KASSERT(!sq_locked, ("sleepqueue chain still locked")); 1388 cancelled = ((flags & CS_EXECUTING) != 0); 1389 } else 1390 cancelled = 1; 1391 1392 if (sq_locked) 1393 sleepq_release(&cc_exec_waiting(cc, direct)); 1394 1395 if ((c->c_iflags & CALLOUT_PENDING) == 0) { 1396 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1397 c, c->c_func, c->c_arg); 1398 /* 1399 * For not scheduled and not executing callout return 1400 * negative value. 1401 */ 1402 if (cc_exec_curr(cc, direct) != c) 1403 cancelled = -1; 1404 CC_UNLOCK(cc); 1405 return (cancelled); 1406 } 1407 1408 c->c_iflags &= ~CALLOUT_PENDING; 1409 c->c_flags &= ~CALLOUT_ACTIVE; 1410 1411 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1412 c, c->c_func, c->c_arg); 1413 if (not_on_a_list == 0) { 1414 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1415 if (cc_exec_next(cc) == c) 1416 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1417 LIST_REMOVE(c, c_links.le); 1418 } else { 1419 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1420 } 1421 } 1422 callout_cc_del(c, cc); 1423 CC_UNLOCK(cc); 1424 return (cancelled); 1425} 1426 1427void 1428callout_init(struct callout *c, int mpsafe) 1429{ 1430 bzero(c, sizeof *c); 1431 if (mpsafe) { 1432 c->c_lock = NULL; 1433 c->c_iflags = CALLOUT_RETURNUNLOCKED; 1434 } else { 1435 c->c_lock = &Giant.lock_object; 1436 c->c_iflags = 0; 1437 } 1438 c->c_cpu = timeout_cpu; 1439} 1440 1441void 1442_callout_init_lock(struct callout *c, struct lock_object *lock, int flags) 1443{ 1444 bzero(c, sizeof *c); 1445 c->c_lock = lock; 1446 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0, 1447 ("callout_init_lock: bad flags %d", flags)); 1448 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0, 1449 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock")); 1450 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & 1451 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class", 1452 __func__)); 1453 c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK); 1454 c->c_cpu = timeout_cpu; 1455} 1456 1457#ifdef APM_FIXUP_CALLTODO 1458/* 1459 * Adjust the kernel calltodo timeout list. This routine is used after 1460 * an APM resume to recalculate the calltodo timer list values with the 1461 * number of hz's we have been sleeping. The next hardclock() will detect 1462 * that there are fired timers and run softclock() to execute them. 1463 * 1464 * Please note, I have not done an exhaustive analysis of what code this 1465 * might break. I am motivated to have my select()'s and alarm()'s that 1466 * have expired during suspend firing upon resume so that the applications 1467 * which set the timer can do the maintanence the timer was for as close 1468 * as possible to the originally intended time. Testing this code for a 1469 * week showed that resuming from a suspend resulted in 22 to 25 timers 1470 * firing, which seemed independent on whether the suspend was 2 hours or 1471 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu> 1472 */ 1473void 1474adjust_timeout_calltodo(struct timeval *time_change) 1475{ 1476 struct callout *p; 1477 unsigned long delta_ticks; 1478 1479 /* 1480 * How many ticks were we asleep? 1481 * (stolen from tvtohz()). 1482 */ 1483 1484 /* Don't do anything */ 1485 if (time_change->tv_sec < 0) 1486 return; 1487 else if (time_change->tv_sec <= LONG_MAX / 1000000) 1488 delta_ticks = howmany(time_change->tv_sec * 1000000 + 1489 time_change->tv_usec, tick) + 1; 1490 else if (time_change->tv_sec <= LONG_MAX / hz) 1491 delta_ticks = time_change->tv_sec * hz + 1492 howmany(time_change->tv_usec, tick) + 1; 1493 else 1494 delta_ticks = LONG_MAX; 1495 1496 if (delta_ticks > INT_MAX) 1497 delta_ticks = INT_MAX; 1498 1499 /* 1500 * Now rip through the timer calltodo list looking for timers 1501 * to expire. 1502 */ 1503 1504 /* don't collide with softclock() */ 1505 CC_LOCK(cc); 1506 for (p = calltodo.c_next; p != NULL; p = p->c_next) { 1507 p->c_time -= delta_ticks; 1508 1509 /* Break if the timer had more time on it than delta_ticks */ 1510 if (p->c_time > 0) 1511 break; 1512 1513 /* take back the ticks the timer didn't use (p->c_time <= 0) */ 1514 delta_ticks = -p->c_time; 1515 } 1516 CC_UNLOCK(cc); 1517 1518 return; 1519} 1520#endif /* APM_FIXUP_CALLTODO */ 1521 1522static int 1523flssbt(sbintime_t sbt) 1524{ 1525 1526 sbt += (uint64_t)sbt >> 1; 1527 if (sizeof(long) >= sizeof(sbintime_t)) 1528 return (flsl(sbt)); 1529 if (sbt >= SBT_1S) 1530 return (flsl(((uint64_t)sbt) >> 32) + 32); 1531 return (flsl(sbt)); 1532} 1533 1534/* 1535 * Dump immediate statistic snapshot of the scheduled callouts. 1536 */ 1537static int 1538sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS) 1539{ 1540 struct callout *tmp; 1541 struct callout_cpu *cc; 1542 struct callout_list *sc; 1543 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t; 1544 int ct[64], cpr[64], ccpbk[32]; 1545 int error, val, i, count, tcum, pcum, maxc, c, medc; 1546#ifdef SMP 1547 int cpu; 1548#endif 1549 1550 val = 0; 1551 error = sysctl_handle_int(oidp, &val, 0, req); 1552 if (error != 0 || req->newptr == NULL) 1553 return (error); 1554 count = maxc = 0; 1555 st = spr = maxt = maxpr = 0; 1556 bzero(ccpbk, sizeof(ccpbk)); 1557 bzero(ct, sizeof(ct)); 1558 bzero(cpr, sizeof(cpr)); 1559 now = sbinuptime(); 1560#ifdef SMP 1561 CPU_FOREACH(cpu) { 1562 cc = CC_CPU(cpu); 1563#else 1564 cc = CC_CPU(timeout_cpu); 1565#endif 1566 CC_LOCK(cc); 1567 for (i = 0; i < callwheelsize; i++) { 1568 sc = &cc->cc_callwheel[i]; 1569 c = 0; 1570 LIST_FOREACH(tmp, sc, c_links.le) { 1571 c++; 1572 t = tmp->c_time - now; 1573 if (t < 0) 1574 t = 0; 1575 st += t / SBT_1US; 1576 spr += tmp->c_precision / SBT_1US; 1577 if (t > maxt) 1578 maxt = t; 1579 if (tmp->c_precision > maxpr) 1580 maxpr = tmp->c_precision; 1581 ct[flssbt(t)]++; 1582 cpr[flssbt(tmp->c_precision)]++; 1583 } 1584 if (c > maxc) 1585 maxc = c; 1586 ccpbk[fls(c + c / 2)]++; 1587 count += c; 1588 } 1589 CC_UNLOCK(cc); 1590#ifdef SMP 1591 } 1592#endif 1593 1594 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++) 1595 tcum += ct[i]; 1596 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1597 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++) 1598 pcum += cpr[i]; 1599 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1600 for (i = 0, c = 0; i < 32 && c < count / 2; i++) 1601 c += ccpbk[i]; 1602 medc = (i >= 2) ? (1 << (i - 2)) : 0; 1603 1604 printf("Scheduled callouts statistic snapshot:\n"); 1605 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n", 1606 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT); 1607 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n", 1608 medc, 1609 count / callwheelsize / mp_ncpus, 1610 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000, 1611 maxc); 1612 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1613 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32, 1614 (st / count) / 1000000, (st / count) % 1000000, 1615 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32); 1616 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1617 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32, 1618 (spr / count) / 1000000, (spr / count) % 1000000, 1619 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32); 1620 printf(" Distribution: \tbuckets\t time\t tcum\t" 1621 " prec\t pcum\n"); 1622 for (i = 0, tcum = pcum = 0; i < 64; i++) { 1623 if (ct[i] == 0 && cpr[i] == 0) 1624 continue; 1625 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0; 1626 tcum += ct[i]; 1627 pcum += cpr[i]; 1628 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n", 1629 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32, 1630 i - 1 - (32 - CC_HASH_SHIFT), 1631 ct[i], tcum, cpr[i], pcum); 1632 } 1633 return (error); 1634} 1635SYSCTL_PROC(_kern, OID_AUTO, callout_stat, 1636 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1637 0, 0, sysctl_kern_callout_stat, "I", 1638 "Dump immediate statistic snapshot of the scheduled callouts"); 1639 1640#ifdef DDB 1641static void 1642_show_callout(struct callout *c) 1643{ 1644 1645 db_printf("callout %p\n", c); 1646#define C_DB_PRINTF(f, e) db_printf(" %s = " f "\n", #e, c->e); 1647 db_printf(" &c_links = %p\n", &(c->c_links)); 1648 C_DB_PRINTF("%" PRId64, c_time); 1649 C_DB_PRINTF("%" PRId64, c_precision); 1650 C_DB_PRINTF("%p", c_arg); 1651 C_DB_PRINTF("%p", c_func); 1652 C_DB_PRINTF("%p", c_lock); 1653 C_DB_PRINTF("%#x", c_flags); 1654 C_DB_PRINTF("%#x", c_iflags); 1655 C_DB_PRINTF("%d", c_cpu); 1656#undef C_DB_PRINTF 1657} 1658 1659DB_SHOW_COMMAND(callout, db_show_callout) 1660{ 1661 1662 if (!have_addr) { 1663 db_printf("usage: show callout <struct callout *>\n"); 1664 return; 1665 } 1666 1667 _show_callout((struct callout *)addr); 1668} 1669#endif /* DDB */ 1670