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 #if defined(__arm__) 45 #include "opt_timer.h" 46 #endif 47 #include "opt_rss.h" 48 49 #include <sys/param.h> 50 #include <sys/systm.h> 51 #include <sys/bus.h> 52 #include <sys/callout.h> 53 #include <sys/domainset.h> 54 #include <sys/file.h> 55 #include <sys/interrupt.h> 56 #include <sys/kernel.h> 57 #include <sys/ktr.h> 58 #include <sys/lock.h> 59 #include <sys/malloc.h> 60 #include <sys/mutex.h> 61 #include <sys/proc.h> 62 #include <sys/sdt.h> 63 #include <sys/sleepqueue.h> 64 #include <sys/sysctl.h> 65 #include <sys/smp.h> 66 67 #ifdef DDB 68 #include <ddb/ddb.h> 69 #include <ddb/db_sym.h> 70 #include <machine/_inttypes.h> 71 #endif 72 73 #ifdef SMP 74 #include <machine/cpu.h> 75 #endif 76 77 #ifndef NO_EVENTTIMERS 78 DPCPU_DECLARE(sbintime_t, hardclocktime); 79 #endif 80 81 SDT_PROVIDER_DEFINE(callout_execute); 82 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *"); 83 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *"); 84 85 #ifdef CALLOUT_PROFILING 86 static int avg_depth; 87 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0, 88 "Average number of items examined per softclock call. Units = 1/1000"); 89 static int avg_gcalls; 90 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0, 91 "Average number of Giant callouts made per softclock call. Units = 1/1000"); 92 static int avg_lockcalls; 93 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0, 94 "Average number of lock callouts made per softclock call. Units = 1/1000"); 95 static int avg_mpcalls; 96 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0, 97 "Average number of MP callouts made per softclock call. Units = 1/1000"); 98 static int avg_depth_dir; 99 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0, 100 "Average number of direct callouts examined per callout_process call. " 101 "Units = 1/1000"); 102 static int avg_lockcalls_dir; 103 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD, 104 &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per " 105 "callout_process call. Units = 1/1000"); 106 static int avg_mpcalls_dir; 107 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir, 108 0, "Average number of MP direct callouts made per callout_process call. " 109 "Units = 1/1000"); 110 #endif 111 112 static int ncallout; 113 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0, 114 "Number of entries in callwheel and size of timeout() preallocation"); 115 116 #ifdef RSS 117 static int pin_default_swi = 1; 118 static int pin_pcpu_swi = 1; 119 #else 120 static int pin_default_swi = 0; 121 static int pin_pcpu_swi = 0; 122 #endif 123 124 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi, 125 0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)"); 126 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi, 127 0, "Pin the per-CPU swis (except PCPU 0, which is also default"); 128 129 /* 130 * TODO: 131 * allocate more timeout table slots when table overflows. 132 */ 133 static u_int __read_mostly callwheelsize; 134 static u_int __read_mostly callwheelmask; 135 136 /* 137 * The callout cpu exec entities represent informations necessary for 138 * describing the state of callouts currently running on the CPU and the ones 139 * necessary for migrating callouts to the new callout cpu. In particular, 140 * the first entry of the array cc_exec_entity holds informations for callout 141 * running in SWI thread context, while the second one holds informations 142 * for callout running directly from hardware interrupt context. 143 * The cached informations are very important for deferring migration when 144 * the migrating callout is already running. 145 */ 146 struct cc_exec { 147 struct callout *cc_curr; 148 callout_func_t *cc_drain; 149 void *cc_last_func; 150 void *cc_last_arg; 151 #ifdef SMP 152 callout_func_t *ce_migration_func; 153 void *ce_migration_arg; 154 sbintime_t ce_migration_time; 155 sbintime_t ce_migration_prec; 156 int ce_migration_cpu; 157 #endif 158 bool cc_cancel; 159 bool cc_waiting; 160 }; 161 162 /* 163 * There is one struct callout_cpu per cpu, holding all relevant 164 * state for the callout processing thread on the individual CPU. 165 */ 166 struct callout_cpu { 167 struct mtx_padalign cc_lock; 168 struct cc_exec cc_exec_entity[2]; 169 struct callout *cc_next; 170 struct callout_list *cc_callwheel; 171 struct callout_tailq cc_expireq; 172 sbintime_t cc_firstevent; 173 sbintime_t cc_lastscan; 174 void *cc_cookie; 175 u_int cc_bucket; 176 u_int cc_inited; 177 #ifdef KTR 178 char cc_ktr_event_name[20]; 179 #endif 180 }; 181 182 #define callout_migrating(c) ((c)->c_iflags & CALLOUT_DFRMIGRATION) 183 184 #define cc_exec_curr(cc, dir) cc->cc_exec_entity[dir].cc_curr 185 #define cc_exec_last_func(cc, dir) cc->cc_exec_entity[dir].cc_last_func 186 #define cc_exec_last_arg(cc, dir) cc->cc_exec_entity[dir].cc_last_arg 187 #define cc_exec_drain(cc, dir) cc->cc_exec_entity[dir].cc_drain 188 #define cc_exec_next(cc) cc->cc_next 189 #define cc_exec_cancel(cc, dir) cc->cc_exec_entity[dir].cc_cancel 190 #define cc_exec_waiting(cc, dir) cc->cc_exec_entity[dir].cc_waiting 191 #ifdef SMP 192 #define cc_migration_func(cc, dir) cc->cc_exec_entity[dir].ce_migration_func 193 #define cc_migration_arg(cc, dir) cc->cc_exec_entity[dir].ce_migration_arg 194 #define cc_migration_cpu(cc, dir) cc->cc_exec_entity[dir].ce_migration_cpu 195 #define cc_migration_time(cc, dir) cc->cc_exec_entity[dir].ce_migration_time 196 #define cc_migration_prec(cc, dir) cc->cc_exec_entity[dir].ce_migration_prec 197 198 struct callout_cpu cc_cpu[MAXCPU]; 199 #define CPUBLOCK MAXCPU 200 #define CC_CPU(cpu) (&cc_cpu[(cpu)]) 201 #define CC_SELF() CC_CPU(PCPU_GET(cpuid)) 202 #else 203 struct callout_cpu cc_cpu; 204 #define CC_CPU(cpu) &cc_cpu 205 #define CC_SELF() &cc_cpu 206 #endif 207 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock) 208 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock) 209 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED) 210 211 static int __read_mostly cc_default_cpu; 212 213 static void callout_cpu_init(struct callout_cpu *cc, int cpu); 214 static void softclock_call_cc(struct callout *c, struct callout_cpu *cc, 215 #ifdef CALLOUT_PROFILING 216 int *mpcalls, int *lockcalls, int *gcalls, 217 #endif 218 int direct); 219 220 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures"); 221 222 /** 223 * Locked by cc_lock: 224 * cc_curr - If a callout is in progress, it is cc_curr. 225 * If cc_curr is non-NULL, threads waiting in 226 * callout_drain() will be woken up as soon as the 227 * relevant callout completes. 228 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held 229 * guarantees that the current callout will not run. 230 * The softclock() function sets this to 0 before it 231 * drops callout_lock to acquire c_lock, and it calls 232 * the handler only if curr_cancelled is still 0 after 233 * cc_lock is successfully acquired. 234 * cc_waiting - If a thread is waiting in callout_drain(), then 235 * callout_wait is nonzero. Set only when 236 * cc_curr is non-NULL. 237 */ 238 239 /* 240 * Resets the execution entity tied to a specific callout cpu. 241 */ 242 static void 243 cc_cce_cleanup(struct callout_cpu *cc, int direct) 244 { 245 246 cc_exec_curr(cc, direct) = NULL; 247 cc_exec_cancel(cc, direct) = false; 248 cc_exec_waiting(cc, direct) = false; 249 #ifdef SMP 250 cc_migration_cpu(cc, direct) = CPUBLOCK; 251 cc_migration_time(cc, direct) = 0; 252 cc_migration_prec(cc, direct) = 0; 253 cc_migration_func(cc, direct) = NULL; 254 cc_migration_arg(cc, direct) = NULL; 255 #endif 256 } 257 258 /* 259 * Checks if migration is requested by a specific callout cpu. 260 */ 261 static int 262 cc_cce_migrating(struct callout_cpu *cc, int direct) 263 { 264 265 #ifdef SMP 266 return (cc_migration_cpu(cc, direct) != CPUBLOCK); 267 #else 268 return (0); 269 #endif 270 } 271 272 /* 273 * Kernel low level callwheel initialization 274 * called on the BSP during kernel startup. 275 */ 276 static void 277 callout_callwheel_init(void *dummy) 278 { 279 struct callout_cpu *cc; 280 int cpu; 281 282 /* 283 * Calculate the size of the callout wheel and the preallocated 284 * timeout() structures. 285 * XXX: Clip callout to result of previous function of maxusers 286 * maximum 384. This is still huge, but acceptable. 287 */ 288 ncallout = imin(16 + maxproc + maxfiles, 18508); 289 TUNABLE_INT_FETCH("kern.ncallout", &ncallout); 290 291 /* 292 * Calculate callout wheel size, should be next power of two higher 293 * than 'ncallout'. 294 */ 295 callwheelsize = 1 << fls(ncallout); 296 callwheelmask = callwheelsize - 1; 297 298 /* 299 * Fetch whether we're pinning the swi's or not. 300 */ 301 TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi); 302 TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi); 303 304 /* 305 * Initialize callout wheels. The software interrupt threads 306 * are created later. 307 */ 308 cc_default_cpu = PCPU_GET(cpuid); 309 CPU_FOREACH(cpu) { 310 cc = CC_CPU(cpu); 311 callout_cpu_init(cc, cpu); 312 } 313 } 314 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL); 315 316 /* 317 * Initialize the per-cpu callout structures. 318 */ 319 static void 320 callout_cpu_init(struct callout_cpu *cc, int cpu) 321 { 322 int i; 323 324 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE); 325 cc->cc_inited = 1; 326 cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) * 327 callwheelsize, M_CALLOUT, 328 DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK); 329 for (i = 0; i < callwheelsize; i++) 330 LIST_INIT(&cc->cc_callwheel[i]); 331 TAILQ_INIT(&cc->cc_expireq); 332 cc->cc_firstevent = SBT_MAX; 333 for (i = 0; i < 2; i++) 334 cc_cce_cleanup(cc, i); 335 #ifdef KTR 336 snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name), 337 "callwheel cpu %d", cpu); 338 #endif 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 */ 347 static struct callout_cpu * 348 callout_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 softclock threads. 373 */ 374 static void 375 start_softclock(void *dummy) 376 { 377 struct callout_cpu *cc; 378 char name[MAXCOMLEN]; 379 int cpu; 380 bool pin_swi; 381 struct intr_event *ie; 382 383 CPU_FOREACH(cpu) { 384 cc = CC_CPU(cpu); 385 snprintf(name, sizeof(name), "clock (%d)", cpu); 386 ie = NULL; 387 if (swi_add(&ie, name, softclock, cc, SWI_CLOCK, 388 INTR_MPSAFE, &cc->cc_cookie)) 389 panic("died while creating standard software ithreads"); 390 if (cpu == cc_default_cpu) 391 pin_swi = pin_default_swi; 392 else 393 pin_swi = pin_pcpu_swi; 394 if (pin_swi && (intr_event_bind(ie, cpu) != 0)) { 395 printf("%s: %s clock couldn't be pinned to cpu %d\n", 396 __func__, 397 cpu == cc_default_cpu ? "default" : "per-cpu", 398 cpu); 399 } 400 } 401 } 402 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL); 403 404 #define CC_HASH_SHIFT 8 405 406 static inline u_int 407 callout_hash(sbintime_t sbt) 408 { 409 410 return (sbt >> (32 - CC_HASH_SHIFT)); 411 } 412 413 static inline u_int 414 callout_get_bucket(sbintime_t sbt) 415 { 416 417 return (callout_hash(sbt) & callwheelmask); 418 } 419 420 void 421 callout_process(sbintime_t now) 422 { 423 struct callout *tmp, *tmpn; 424 struct callout_cpu *cc; 425 struct callout_list *sc; 426 sbintime_t first, last, max, tmp_max; 427 uint32_t lookahead; 428 u_int firstb, lastb, nowb; 429 #ifdef CALLOUT_PROFILING 430 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0; 431 #endif 432 433 cc = CC_SELF(); 434 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET); 435 436 /* Compute the buckets of the last scan and present times. */ 437 firstb = callout_hash(cc->cc_lastscan); 438 cc->cc_lastscan = now; 439 nowb = callout_hash(now); 440 441 /* Compute the last bucket and minimum time of the bucket after it. */ 442 if (nowb == firstb) 443 lookahead = (SBT_1S / 16); 444 else if (nowb - firstb == 1) 445 lookahead = (SBT_1S / 8); 446 else 447 lookahead = (SBT_1S / 2); 448 first = last = now; 449 first += (lookahead / 2); 450 last += lookahead; 451 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT)); 452 lastb = callout_hash(last) - 1; 453 max = last; 454 455 /* 456 * Check if we wrapped around the entire wheel from the last scan. 457 * In case, we need to scan entirely the wheel for pending callouts. 458 */ 459 if (lastb - firstb >= callwheelsize) { 460 lastb = firstb + callwheelsize - 1; 461 if (nowb - firstb >= callwheelsize) 462 nowb = lastb; 463 } 464 465 /* Iterate callwheel from firstb to nowb and then up to lastb. */ 466 do { 467 sc = &cc->cc_callwheel[firstb & callwheelmask]; 468 tmp = LIST_FIRST(sc); 469 while (tmp != NULL) { 470 /* Run the callout if present time within allowed. */ 471 if (tmp->c_time <= now) { 472 /* 473 * Consumer told us the callout may be run 474 * directly from hardware interrupt context. 475 */ 476 if (tmp->c_iflags & CALLOUT_DIRECT) { 477 #ifdef CALLOUT_PROFILING 478 ++depth_dir; 479 #endif 480 cc_exec_next(cc) = 481 LIST_NEXT(tmp, c_links.le); 482 cc->cc_bucket = firstb & callwheelmask; 483 LIST_REMOVE(tmp, c_links.le); 484 softclock_call_cc(tmp, cc, 485 #ifdef CALLOUT_PROFILING 486 &mpcalls_dir, &lockcalls_dir, NULL, 487 #endif 488 1); 489 tmp = cc_exec_next(cc); 490 cc_exec_next(cc) = NULL; 491 } else { 492 tmpn = LIST_NEXT(tmp, c_links.le); 493 LIST_REMOVE(tmp, c_links.le); 494 TAILQ_INSERT_TAIL(&cc->cc_expireq, 495 tmp, c_links.tqe); 496 tmp->c_iflags |= CALLOUT_PROCESSED; 497 tmp = tmpn; 498 } 499 continue; 500 } 501 /* Skip events from distant future. */ 502 if (tmp->c_time >= max) 503 goto next; 504 /* 505 * Event minimal time is bigger than present maximal 506 * time, so it cannot be aggregated. 507 */ 508 if (tmp->c_time > last) { 509 lastb = nowb; 510 goto next; 511 } 512 /* Update first and last time, respecting this event. */ 513 if (tmp->c_time < first) 514 first = tmp->c_time; 515 tmp_max = tmp->c_time + tmp->c_precision; 516 if (tmp_max < last) 517 last = tmp_max; 518 next: 519 tmp = LIST_NEXT(tmp, c_links.le); 520 } 521 /* Proceed with the next bucket. */ 522 firstb++; 523 /* 524 * Stop if we looked after present time and found 525 * some event we can't execute at now. 526 * Stop if we looked far enough into the future. 527 */ 528 } while (((int)(firstb - lastb)) <= 0); 529 cc->cc_firstevent = last; 530 #ifndef NO_EVENTTIMERS 531 cpu_new_callout(curcpu, last, first); 532 #endif 533 #ifdef CALLOUT_PROFILING 534 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8; 535 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8; 536 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8; 537 #endif 538 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET); 539 /* 540 * swi_sched acquires the thread lock, so we don't want to call it 541 * with cc_lock held; incorrect locking order. 542 */ 543 if (!TAILQ_EMPTY(&cc->cc_expireq)) 544 swi_sched(cc->cc_cookie, 0); 545 } 546 547 static struct callout_cpu * 548 callout_lock(struct callout *c) 549 { 550 struct callout_cpu *cc; 551 int cpu; 552 553 for (;;) { 554 cpu = c->c_cpu; 555 #ifdef SMP 556 if (cpu == CPUBLOCK) { 557 while (c->c_cpu == CPUBLOCK) 558 cpu_spinwait(); 559 continue; 560 } 561 #endif 562 cc = CC_CPU(cpu); 563 CC_LOCK(cc); 564 if (cpu == c->c_cpu) 565 break; 566 CC_UNLOCK(cc); 567 } 568 return (cc); 569 } 570 571 static void 572 callout_cc_add(struct callout *c, struct callout_cpu *cc, 573 sbintime_t sbt, sbintime_t precision, void (*func)(void *), 574 void *arg, int cpu, int flags) 575 { 576 int bucket; 577 578 CC_LOCK_ASSERT(cc); 579 if (sbt < cc->cc_lastscan) 580 sbt = cc->cc_lastscan; 581 c->c_arg = arg; 582 c->c_iflags |= CALLOUT_PENDING; 583 c->c_iflags &= ~CALLOUT_PROCESSED; 584 c->c_flags |= CALLOUT_ACTIVE; 585 if (flags & C_DIRECT_EXEC) 586 c->c_iflags |= CALLOUT_DIRECT; 587 c->c_func = func; 588 c->c_time = sbt; 589 c->c_precision = precision; 590 bucket = callout_get_bucket(c->c_time); 591 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x", 592 c, (int)(c->c_precision >> 32), 593 (u_int)(c->c_precision & 0xffffffff)); 594 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le); 595 if (cc->cc_bucket == bucket) 596 cc_exec_next(cc) = c; 597 #ifndef NO_EVENTTIMERS 598 /* 599 * Inform the eventtimers(4) subsystem there's a new callout 600 * that has been inserted, but only if really required. 601 */ 602 if (SBT_MAX - c->c_time < c->c_precision) 603 c->c_precision = SBT_MAX - c->c_time; 604 sbt = c->c_time + c->c_precision; 605 if (sbt < cc->cc_firstevent) { 606 cc->cc_firstevent = sbt; 607 cpu_new_callout(cpu, sbt, c->c_time); 608 } 609 #endif 610 } 611 612 static void 613 softclock_call_cc(struct callout *c, struct callout_cpu *cc, 614 #ifdef CALLOUT_PROFILING 615 int *mpcalls, int *lockcalls, int *gcalls, 616 #endif 617 int direct) 618 { 619 struct rm_priotracker tracker; 620 callout_func_t *c_func, *drain; 621 void *c_arg; 622 struct lock_class *class; 623 struct lock_object *c_lock; 624 uintptr_t lock_status; 625 int c_iflags; 626 #ifdef SMP 627 struct callout_cpu *new_cc; 628 callout_func_t *new_func; 629 void *new_arg; 630 int flags, new_cpu; 631 sbintime_t new_prec, new_time; 632 #endif 633 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 634 sbintime_t sbt1, sbt2; 635 struct timespec ts2; 636 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */ 637 static callout_func_t *lastfunc; 638 #endif 639 640 KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING, 641 ("softclock_call_cc: pend %p %x", c, c->c_iflags)); 642 KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE, 643 ("softclock_call_cc: act %p %x", c, c->c_flags)); 644 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL; 645 lock_status = 0; 646 if (c->c_flags & CALLOUT_SHAREDLOCK) { 647 if (class == &lock_class_rm) 648 lock_status = (uintptr_t)&tracker; 649 else 650 lock_status = 1; 651 } 652 c_lock = c->c_lock; 653 c_func = c->c_func; 654 c_arg = c->c_arg; 655 c_iflags = c->c_iflags; 656 c->c_iflags &= ~CALLOUT_PENDING; 657 658 cc_exec_curr(cc, direct) = c; 659 cc_exec_last_func(cc, direct) = c_func; 660 cc_exec_last_arg(cc, direct) = c_arg; 661 cc_exec_cancel(cc, direct) = false; 662 cc_exec_drain(cc, direct) = NULL; 663 CC_UNLOCK(cc); 664 if (c_lock != NULL) { 665 class->lc_lock(c_lock, lock_status); 666 /* 667 * The callout may have been cancelled 668 * while we switched locks. 669 */ 670 if (cc_exec_cancel(cc, direct)) { 671 class->lc_unlock(c_lock); 672 goto skip; 673 } 674 /* The callout cannot be stopped now. */ 675 cc_exec_cancel(cc, direct) = true; 676 if (c_lock == &Giant.lock_object) { 677 #ifdef CALLOUT_PROFILING 678 (*gcalls)++; 679 #endif 680 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p", 681 c, c_func, c_arg); 682 } else { 683 #ifdef CALLOUT_PROFILING 684 (*lockcalls)++; 685 #endif 686 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p", 687 c, c_func, c_arg); 688 } 689 } else { 690 #ifdef CALLOUT_PROFILING 691 (*mpcalls)++; 692 #endif 693 CTR3(KTR_CALLOUT, "callout %p func %p arg %p", 694 c, c_func, c_arg); 695 } 696 KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running", 697 "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct); 698 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 699 sbt1 = sbinuptime(); 700 #endif 701 THREAD_NO_SLEEPING(); 702 SDT_PROBE1(callout_execute, , , callout__start, c); 703 c_func(c_arg); 704 SDT_PROBE1(callout_execute, , , callout__end, c); 705 THREAD_SLEEPING_OK(); 706 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 707 sbt2 = sbinuptime(); 708 sbt2 -= sbt1; 709 if (sbt2 > maxdt) { 710 if (lastfunc != c_func || sbt2 > maxdt * 2) { 711 ts2 = sbttots(sbt2); 712 printf( 713 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n", 714 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec); 715 } 716 maxdt = sbt2; 717 lastfunc = c_func; 718 } 719 #endif 720 KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle"); 721 CTR1(KTR_CALLOUT, "callout %p finished", c); 722 if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0) 723 class->lc_unlock(c_lock); 724 skip: 725 CC_LOCK(cc); 726 KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr")); 727 cc_exec_curr(cc, direct) = NULL; 728 if (cc_exec_drain(cc, direct)) { 729 drain = cc_exec_drain(cc, direct); 730 cc_exec_drain(cc, direct) = NULL; 731 CC_UNLOCK(cc); 732 drain(c_arg); 733 CC_LOCK(cc); 734 } 735 if (cc_exec_waiting(cc, direct)) { 736 /* 737 * There is someone waiting for the 738 * callout to complete. 739 * If the callout was scheduled for 740 * migration just cancel it. 741 */ 742 if (cc_cce_migrating(cc, direct)) { 743 cc_cce_cleanup(cc, direct); 744 745 /* 746 * It should be assert here that the callout is not 747 * destroyed but that is not easy. 748 */ 749 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 750 } 751 cc_exec_waiting(cc, direct) = false; 752 CC_UNLOCK(cc); 753 wakeup(&cc_exec_waiting(cc, direct)); 754 CC_LOCK(cc); 755 } else if (cc_cce_migrating(cc, direct)) { 756 #ifdef SMP 757 /* 758 * If the callout was scheduled for 759 * migration just perform it now. 760 */ 761 new_cpu = cc_migration_cpu(cc, direct); 762 new_time = cc_migration_time(cc, direct); 763 new_prec = cc_migration_prec(cc, direct); 764 new_func = cc_migration_func(cc, direct); 765 new_arg = cc_migration_arg(cc, direct); 766 cc_cce_cleanup(cc, direct); 767 768 /* 769 * It should be assert here that the callout is not destroyed 770 * but that is not easy. 771 * 772 * As first thing, handle deferred callout stops. 773 */ 774 if (!callout_migrating(c)) { 775 CTR3(KTR_CALLOUT, 776 "deferred cancelled %p func %p arg %p", 777 c, new_func, new_arg); 778 return; 779 } 780 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 781 782 new_cc = callout_cpu_switch(c, cc, new_cpu); 783 flags = (direct) ? C_DIRECT_EXEC : 0; 784 callout_cc_add(c, new_cc, new_time, new_prec, new_func, 785 new_arg, new_cpu, flags); 786 CC_UNLOCK(new_cc); 787 CC_LOCK(cc); 788 #else 789 panic("migration should not happen"); 790 #endif 791 } 792 } 793 794 /* 795 * The callout mechanism is based on the work of Adam M. Costello and 796 * George Varghese, published in a technical report entitled "Redesigning 797 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 798 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 799 * used in this implementation was published by G. Varghese and T. Lauck in 800 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 801 * the Efficient Implementation of a Timer Facility" in the Proceedings of 802 * the 11th ACM Annual Symposium on Operating Systems Principles, 803 * Austin, Texas Nov 1987. 804 */ 805 806 /* 807 * Software (low priority) clock interrupt. 808 * Run periodic events from timeout queue. 809 */ 810 void 811 softclock(void *arg) 812 { 813 struct callout_cpu *cc; 814 struct callout *c; 815 #ifdef CALLOUT_PROFILING 816 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0; 817 #endif 818 819 cc = (struct callout_cpu *)arg; 820 CC_LOCK(cc); 821 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) { 822 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 823 softclock_call_cc(c, cc, 824 #ifdef CALLOUT_PROFILING 825 &mpcalls, &lockcalls, &gcalls, 826 #endif 827 0); 828 #ifdef CALLOUT_PROFILING 829 ++depth; 830 #endif 831 } 832 #ifdef CALLOUT_PROFILING 833 avg_depth += (depth * 1000 - avg_depth) >> 8; 834 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8; 835 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8; 836 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8; 837 #endif 838 CC_UNLOCK(cc); 839 } 840 841 void 842 callout_when(sbintime_t sbt, sbintime_t precision, int flags, 843 sbintime_t *res, sbintime_t *prec_res) 844 { 845 sbintime_t to_sbt, to_pr; 846 847 if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) { 848 *res = sbt; 849 *prec_res = precision; 850 return; 851 } 852 if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt) 853 sbt = tick_sbt; 854 if ((flags & C_HARDCLOCK) != 0 || 855 #ifdef NO_EVENTTIMERS 856 sbt >= sbt_timethreshold) { 857 to_sbt = getsbinuptime(); 858 859 /* Add safety belt for the case of hz > 1000. */ 860 to_sbt += tc_tick_sbt - tick_sbt; 861 #else 862 sbt >= sbt_tickthreshold) { 863 /* 864 * Obtain the time of the last hardclock() call on 865 * this CPU directly from the kern_clocksource.c. 866 * This value is per-CPU, but it is equal for all 867 * active ones. 868 */ 869 #ifdef __LP64__ 870 to_sbt = DPCPU_GET(hardclocktime); 871 #else 872 spinlock_enter(); 873 to_sbt = DPCPU_GET(hardclocktime); 874 spinlock_exit(); 875 #endif 876 #endif 877 if (cold && to_sbt == 0) 878 to_sbt = sbinuptime(); 879 if ((flags & C_HARDCLOCK) == 0) 880 to_sbt += tick_sbt; 881 } else 882 to_sbt = sbinuptime(); 883 if (SBT_MAX - to_sbt < sbt) 884 to_sbt = SBT_MAX; 885 else 886 to_sbt += sbt; 887 *res = to_sbt; 888 to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp : 889 sbt >> C_PRELGET(flags)); 890 *prec_res = to_pr > precision ? to_pr : precision; 891 } 892 893 /* 894 * New interface; clients allocate their own callout structures. 895 * 896 * callout_reset() - establish or change a timeout 897 * callout_stop() - disestablish a timeout 898 * callout_init() - initialize a callout structure so that it can 899 * safely be passed to callout_reset() and callout_stop() 900 * 901 * <sys/callout.h> defines three convenience macros: 902 * 903 * callout_active() - returns truth if callout has not been stopped, 904 * drained, or deactivated since the last time the callout was 905 * reset. 906 * callout_pending() - returns truth if callout is still waiting for timeout 907 * callout_deactivate() - marks the callout as having been serviced 908 */ 909 int 910 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec, 911 callout_func_t *ftn, void *arg, int cpu, int flags) 912 { 913 sbintime_t to_sbt, precision; 914 struct callout_cpu *cc; 915 int cancelled, direct; 916 int ignore_cpu=0; 917 918 cancelled = 0; 919 if (cpu == -1) { 920 ignore_cpu = 1; 921 } else if ((cpu >= MAXCPU) || 922 ((CC_CPU(cpu))->cc_inited == 0)) { 923 /* Invalid CPU spec */ 924 panic("Invalid CPU in callout %d", cpu); 925 } 926 callout_when(sbt, prec, flags, &to_sbt, &precision); 927 928 /* 929 * This flag used to be added by callout_cc_add, but the 930 * first time you call this we could end up with the 931 * wrong direct flag if we don't do it before we add. 932 */ 933 if (flags & C_DIRECT_EXEC) { 934 direct = 1; 935 } else { 936 direct = 0; 937 } 938 KASSERT(!direct || c->c_lock == NULL, 939 ("%s: direct callout %p has lock", __func__, c)); 940 cc = callout_lock(c); 941 /* 942 * Don't allow migration if the user does not care. 943 */ 944 if (ignore_cpu) { 945 cpu = c->c_cpu; 946 } 947 948 if (cc_exec_curr(cc, direct) == c) { 949 /* 950 * We're being asked to reschedule a callout which is 951 * currently in progress. If there is a lock then we 952 * can cancel the callout if it has not really started. 953 */ 954 if (c->c_lock != NULL && !cc_exec_cancel(cc, direct)) 955 cancelled = cc_exec_cancel(cc, direct) = true; 956 if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) { 957 /* 958 * Someone has called callout_drain to kill this 959 * callout. Don't reschedule. 960 */ 961 CTR4(KTR_CALLOUT, "%s %p func %p arg %p", 962 cancelled ? "cancelled" : "failed to cancel", 963 c, c->c_func, c->c_arg); 964 CC_UNLOCK(cc); 965 return (cancelled); 966 } 967 #ifdef SMP 968 if (callout_migrating(c)) { 969 /* 970 * This only occurs when a second callout_reset_sbt_on 971 * is made after a previous one moved it into 972 * deferred migration (below). Note we do *not* change 973 * the prev_cpu even though the previous target may 974 * be different. 975 */ 976 cc_migration_cpu(cc, direct) = cpu; 977 cc_migration_time(cc, direct) = to_sbt; 978 cc_migration_prec(cc, direct) = precision; 979 cc_migration_func(cc, direct) = ftn; 980 cc_migration_arg(cc, direct) = arg; 981 cancelled = 1; 982 CC_UNLOCK(cc); 983 return (cancelled); 984 } 985 #endif 986 } 987 if (c->c_iflags & CALLOUT_PENDING) { 988 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 989 if (cc_exec_next(cc) == c) 990 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 991 LIST_REMOVE(c, c_links.le); 992 } else { 993 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 994 } 995 cancelled = 1; 996 c->c_iflags &= ~ CALLOUT_PENDING; 997 c->c_flags &= ~ CALLOUT_ACTIVE; 998 } 999 1000 #ifdef SMP 1001 /* 1002 * If the callout must migrate try to perform it immediately. 1003 * If the callout is currently running, just defer the migration 1004 * to a more appropriate moment. 1005 */ 1006 if (c->c_cpu != cpu) { 1007 if (cc_exec_curr(cc, direct) == c) { 1008 /* 1009 * Pending will have been removed since we are 1010 * actually executing the callout on another 1011 * CPU. That callout should be waiting on the 1012 * lock the caller holds. If we set both 1013 * active/and/pending after we return and the 1014 * lock on the executing callout proceeds, it 1015 * will then see pending is true and return. 1016 * At the return from the actual callout execution 1017 * the migration will occur in softclock_call_cc 1018 * and this new callout will be placed on the 1019 * new CPU via a call to callout_cpu_switch() which 1020 * will get the lock on the right CPU followed 1021 * by a call callout_cc_add() which will add it there. 1022 * (see above in softclock_call_cc()). 1023 */ 1024 cc_migration_cpu(cc, direct) = cpu; 1025 cc_migration_time(cc, direct) = to_sbt; 1026 cc_migration_prec(cc, direct) = precision; 1027 cc_migration_func(cc, direct) = ftn; 1028 cc_migration_arg(cc, direct) = arg; 1029 c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING); 1030 c->c_flags |= CALLOUT_ACTIVE; 1031 CTR6(KTR_CALLOUT, 1032 "migration of %p func %p arg %p in %d.%08x to %u deferred", 1033 c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1034 (u_int)(to_sbt & 0xffffffff), cpu); 1035 CC_UNLOCK(cc); 1036 return (cancelled); 1037 } 1038 cc = callout_cpu_switch(c, cc, cpu); 1039 } 1040 #endif 1041 1042 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags); 1043 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x", 1044 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1045 (u_int)(to_sbt & 0xffffffff)); 1046 CC_UNLOCK(cc); 1047 1048 return (cancelled); 1049 } 1050 1051 /* 1052 * Common idioms that can be optimized in the future. 1053 */ 1054 int 1055 callout_schedule_on(struct callout *c, int to_ticks, int cpu) 1056 { 1057 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu); 1058 } 1059 1060 int 1061 callout_schedule(struct callout *c, int to_ticks) 1062 { 1063 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu); 1064 } 1065 1066 int 1067 _callout_stop_safe(struct callout *c, int flags, callout_func_t *drain) 1068 { 1069 struct callout_cpu *cc, *old_cc; 1070 struct lock_class *class; 1071 int direct, sq_locked, use_lock; 1072 int cancelled, not_on_a_list; 1073 1074 if ((flags & CS_DRAIN) != 0) 1075 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock, 1076 "calling %s", __func__); 1077 1078 KASSERT((flags & CS_DRAIN) == 0 || drain == NULL, 1079 ("Cannot set drain callback and CS_DRAIN flag at the same time")); 1080 1081 /* 1082 * Some old subsystems don't hold Giant while running a callout_stop(), 1083 * so just discard this check for the moment. 1084 */ 1085 if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) { 1086 if (c->c_lock == &Giant.lock_object) 1087 use_lock = mtx_owned(&Giant); 1088 else { 1089 use_lock = 1; 1090 class = LOCK_CLASS(c->c_lock); 1091 class->lc_assert(c->c_lock, LA_XLOCKED); 1092 } 1093 } else 1094 use_lock = 0; 1095 if (c->c_iflags & CALLOUT_DIRECT) { 1096 direct = 1; 1097 } else { 1098 direct = 0; 1099 } 1100 sq_locked = 0; 1101 old_cc = NULL; 1102 again: 1103 cc = callout_lock(c); 1104 1105 if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) == 1106 (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) && 1107 ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) { 1108 /* 1109 * Special case where this slipped in while we 1110 * were migrating *as* the callout is about to 1111 * execute. The caller probably holds the lock 1112 * the callout wants. 1113 * 1114 * Get rid of the migration first. Then set 1115 * the flag that tells this code *not* to 1116 * try to remove it from any lists (its not 1117 * on one yet). When the callout wheel runs, 1118 * it will ignore this callout. 1119 */ 1120 c->c_iflags &= ~CALLOUT_PENDING; 1121 c->c_flags &= ~CALLOUT_ACTIVE; 1122 not_on_a_list = 1; 1123 } else { 1124 not_on_a_list = 0; 1125 } 1126 1127 /* 1128 * If the callout was migrating while the callout cpu lock was 1129 * dropped, just drop the sleepqueue lock and check the states 1130 * again. 1131 */ 1132 if (sq_locked != 0 && cc != old_cc) { 1133 #ifdef SMP 1134 CC_UNLOCK(cc); 1135 sleepq_release(&cc_exec_waiting(old_cc, direct)); 1136 sq_locked = 0; 1137 old_cc = NULL; 1138 goto again; 1139 #else 1140 panic("migration should not happen"); 1141 #endif 1142 } 1143 1144 /* 1145 * If the callout is running, try to stop it or drain it. 1146 */ 1147 if (cc_exec_curr(cc, direct) == c) { 1148 /* 1149 * Succeed we to stop it or not, we must clear the 1150 * active flag - this is what API users expect. If we're 1151 * draining and the callout is currently executing, first wait 1152 * until it finishes. 1153 */ 1154 if ((flags & CS_DRAIN) == 0) 1155 c->c_flags &= ~CALLOUT_ACTIVE; 1156 1157 if ((flags & CS_DRAIN) != 0) { 1158 /* 1159 * The current callout is running (or just 1160 * about to run) and blocking is allowed, so 1161 * just wait for the current invocation to 1162 * finish. 1163 */ 1164 while (cc_exec_curr(cc, direct) == c) { 1165 /* 1166 * Use direct calls to sleepqueue interface 1167 * instead of cv/msleep in order to avoid 1168 * a LOR between cc_lock and sleepqueue 1169 * chain spinlocks. This piece of code 1170 * emulates a msleep_spin() call actually. 1171 * 1172 * If we already have the sleepqueue chain 1173 * locked, then we can safely block. If we 1174 * don't already have it locked, however, 1175 * we have to drop the cc_lock to lock 1176 * it. This opens several races, so we 1177 * restart at the beginning once we have 1178 * both locks. If nothing has changed, then 1179 * we will end up back here with sq_locked 1180 * set. 1181 */ 1182 if (!sq_locked) { 1183 CC_UNLOCK(cc); 1184 sleepq_lock( 1185 &cc_exec_waiting(cc, direct)); 1186 sq_locked = 1; 1187 old_cc = cc; 1188 goto again; 1189 } 1190 1191 /* 1192 * Migration could be cancelled here, but 1193 * as long as it is still not sure when it 1194 * will be packed up, just let softclock() 1195 * take care of it. 1196 */ 1197 cc_exec_waiting(cc, direct) = true; 1198 DROP_GIANT(); 1199 CC_UNLOCK(cc); 1200 sleepq_add( 1201 &cc_exec_waiting(cc, direct), 1202 &cc->cc_lock.lock_object, "codrain", 1203 SLEEPQ_SLEEP, 0); 1204 sleepq_wait( 1205 &cc_exec_waiting(cc, direct), 1206 0); 1207 sq_locked = 0; 1208 old_cc = NULL; 1209 1210 /* Reacquire locks previously released. */ 1211 PICKUP_GIANT(); 1212 CC_LOCK(cc); 1213 } 1214 c->c_flags &= ~CALLOUT_ACTIVE; 1215 } else if (use_lock && 1216 !cc_exec_cancel(cc, direct) && (drain == NULL)) { 1217 1218 /* 1219 * The current callout is waiting for its 1220 * lock which we hold. Cancel the callout 1221 * and return. After our caller drops the 1222 * lock, the callout will be skipped in 1223 * softclock(). This *only* works with a 1224 * callout_stop() *not* callout_drain() or 1225 * callout_async_drain(). 1226 */ 1227 cc_exec_cancel(cc, direct) = true; 1228 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1229 c, c->c_func, c->c_arg); 1230 KASSERT(!cc_cce_migrating(cc, direct), 1231 ("callout wrongly scheduled for migration")); 1232 if (callout_migrating(c)) { 1233 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1234 #ifdef SMP 1235 cc_migration_cpu(cc, direct) = CPUBLOCK; 1236 cc_migration_time(cc, direct) = 0; 1237 cc_migration_prec(cc, direct) = 0; 1238 cc_migration_func(cc, direct) = NULL; 1239 cc_migration_arg(cc, direct) = NULL; 1240 #endif 1241 } 1242 CC_UNLOCK(cc); 1243 KASSERT(!sq_locked, ("sleepqueue chain locked")); 1244 return (1); 1245 } else if (callout_migrating(c)) { 1246 /* 1247 * The callout is currently being serviced 1248 * and the "next" callout is scheduled at 1249 * its completion with a migration. We remove 1250 * the migration flag so it *won't* get rescheduled, 1251 * but we can't stop the one thats running so 1252 * we return 0. 1253 */ 1254 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1255 #ifdef SMP 1256 /* 1257 * We can't call cc_cce_cleanup here since 1258 * if we do it will remove .ce_curr and 1259 * its still running. This will prevent a 1260 * reschedule of the callout when the 1261 * execution completes. 1262 */ 1263 cc_migration_cpu(cc, direct) = CPUBLOCK; 1264 cc_migration_time(cc, direct) = 0; 1265 cc_migration_prec(cc, direct) = 0; 1266 cc_migration_func(cc, direct) = NULL; 1267 cc_migration_arg(cc, direct) = NULL; 1268 #endif 1269 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p", 1270 c, c->c_func, c->c_arg); 1271 if (drain) { 1272 KASSERT(cc_exec_drain(cc, direct) == NULL, 1273 ("callout drain function already set to %p", 1274 cc_exec_drain(cc, direct))); 1275 cc_exec_drain(cc, direct) = drain; 1276 } 1277 CC_UNLOCK(cc); 1278 return ((flags & CS_EXECUTING) != 0); 1279 } else { 1280 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1281 c, c->c_func, c->c_arg); 1282 if (drain) { 1283 KASSERT(cc_exec_drain(cc, direct) == NULL, 1284 ("callout drain function already set to %p", 1285 cc_exec_drain(cc, direct))); 1286 cc_exec_drain(cc, direct) = drain; 1287 } 1288 } 1289 KASSERT(!sq_locked, ("sleepqueue chain still locked")); 1290 cancelled = ((flags & CS_EXECUTING) != 0); 1291 } else 1292 cancelled = 1; 1293 1294 if (sq_locked) 1295 sleepq_release(&cc_exec_waiting(cc, direct)); 1296 1297 if ((c->c_iflags & CALLOUT_PENDING) == 0) { 1298 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1299 c, c->c_func, c->c_arg); 1300 /* 1301 * For not scheduled and not executing callout return 1302 * negative value. 1303 */ 1304 if (cc_exec_curr(cc, direct) != c) 1305 cancelled = -1; 1306 CC_UNLOCK(cc); 1307 return (cancelled); 1308 } 1309 1310 c->c_iflags &= ~CALLOUT_PENDING; 1311 c->c_flags &= ~CALLOUT_ACTIVE; 1312 1313 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1314 c, c->c_func, c->c_arg); 1315 if (not_on_a_list == 0) { 1316 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1317 if (cc_exec_next(cc) == c) 1318 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1319 LIST_REMOVE(c, c_links.le); 1320 } else { 1321 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1322 } 1323 } 1324 CC_UNLOCK(cc); 1325 return (cancelled); 1326 } 1327 1328 void 1329 callout_init(struct callout *c, int mpsafe) 1330 { 1331 bzero(c, sizeof *c); 1332 if (mpsafe) { 1333 c->c_lock = NULL; 1334 c->c_iflags = CALLOUT_RETURNUNLOCKED; 1335 } else { 1336 c->c_lock = &Giant.lock_object; 1337 c->c_iflags = 0; 1338 } 1339 c->c_cpu = cc_default_cpu; 1340 } 1341 1342 void 1343 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags) 1344 { 1345 bzero(c, sizeof *c); 1346 c->c_lock = lock; 1347 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0, 1348 ("callout_init_lock: bad flags %d", flags)); 1349 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0, 1350 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock")); 1351 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & 1352 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class", 1353 __func__)); 1354 c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK); 1355 c->c_cpu = cc_default_cpu; 1356 } 1357 1358 #ifdef APM_FIXUP_CALLTODO 1359 /* 1360 * Adjust the kernel calltodo timeout list. This routine is used after 1361 * an APM resume to recalculate the calltodo timer list values with the 1362 * number of hz's we have been sleeping. The next hardclock() will detect 1363 * that there are fired timers and run softclock() to execute them. 1364 * 1365 * Please note, I have not done an exhaustive analysis of what code this 1366 * might break. I am motivated to have my select()'s and alarm()'s that 1367 * have expired during suspend firing upon resume so that the applications 1368 * which set the timer can do the maintanence the timer was for as close 1369 * as possible to the originally intended time. Testing this code for a 1370 * week showed that resuming from a suspend resulted in 22 to 25 timers 1371 * firing, which seemed independent on whether the suspend was 2 hours or 1372 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu> 1373 */ 1374 void 1375 adjust_timeout_calltodo(struct timeval *time_change) 1376 { 1377 struct callout *p; 1378 unsigned long delta_ticks; 1379 1380 /* 1381 * How many ticks were we asleep? 1382 * (stolen from tvtohz()). 1383 */ 1384 1385 /* Don't do anything */ 1386 if (time_change->tv_sec < 0) 1387 return; 1388 else if (time_change->tv_sec <= LONG_MAX / 1000000) 1389 delta_ticks = howmany(time_change->tv_sec * 1000000 + 1390 time_change->tv_usec, tick) + 1; 1391 else if (time_change->tv_sec <= LONG_MAX / hz) 1392 delta_ticks = time_change->tv_sec * hz + 1393 howmany(time_change->tv_usec, tick) + 1; 1394 else 1395 delta_ticks = LONG_MAX; 1396 1397 if (delta_ticks > INT_MAX) 1398 delta_ticks = INT_MAX; 1399 1400 /* 1401 * Now rip through the timer calltodo list looking for timers 1402 * to expire. 1403 */ 1404 1405 /* don't collide with softclock() */ 1406 CC_LOCK(cc); 1407 for (p = calltodo.c_next; p != NULL; p = p->c_next) { 1408 p->c_time -= delta_ticks; 1409 1410 /* Break if the timer had more time on it than delta_ticks */ 1411 if (p->c_time > 0) 1412 break; 1413 1414 /* take back the ticks the timer didn't use (p->c_time <= 0) */ 1415 delta_ticks = -p->c_time; 1416 } 1417 CC_UNLOCK(cc); 1418 1419 return; 1420 } 1421 #endif /* APM_FIXUP_CALLTODO */ 1422 1423 static int 1424 flssbt(sbintime_t sbt) 1425 { 1426 1427 sbt += (uint64_t)sbt >> 1; 1428 if (sizeof(long) >= sizeof(sbintime_t)) 1429 return (flsl(sbt)); 1430 if (sbt >= SBT_1S) 1431 return (flsl(((uint64_t)sbt) >> 32) + 32); 1432 return (flsl(sbt)); 1433 } 1434 1435 /* 1436 * Dump immediate statistic snapshot of the scheduled callouts. 1437 */ 1438 static int 1439 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS) 1440 { 1441 struct callout *tmp; 1442 struct callout_cpu *cc; 1443 struct callout_list *sc; 1444 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t; 1445 int ct[64], cpr[64], ccpbk[32]; 1446 int error, val, i, count, tcum, pcum, maxc, c, medc; 1447 int cpu; 1448 1449 val = 0; 1450 error = sysctl_handle_int(oidp, &val, 0, req); 1451 if (error != 0 || req->newptr == NULL) 1452 return (error); 1453 count = maxc = 0; 1454 st = spr = maxt = maxpr = 0; 1455 bzero(ccpbk, sizeof(ccpbk)); 1456 bzero(ct, sizeof(ct)); 1457 bzero(cpr, sizeof(cpr)); 1458 now = sbinuptime(); 1459 CPU_FOREACH(cpu) { 1460 cc = CC_CPU(cpu); 1461 CC_LOCK(cc); 1462 for (i = 0; i < callwheelsize; i++) { 1463 sc = &cc->cc_callwheel[i]; 1464 c = 0; 1465 LIST_FOREACH(tmp, sc, c_links.le) { 1466 c++; 1467 t = tmp->c_time - now; 1468 if (t < 0) 1469 t = 0; 1470 st += t / SBT_1US; 1471 spr += tmp->c_precision / SBT_1US; 1472 if (t > maxt) 1473 maxt = t; 1474 if (tmp->c_precision > maxpr) 1475 maxpr = tmp->c_precision; 1476 ct[flssbt(t)]++; 1477 cpr[flssbt(tmp->c_precision)]++; 1478 } 1479 if (c > maxc) 1480 maxc = c; 1481 ccpbk[fls(c + c / 2)]++; 1482 count += c; 1483 } 1484 CC_UNLOCK(cc); 1485 } 1486 1487 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++) 1488 tcum += ct[i]; 1489 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1490 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++) 1491 pcum += cpr[i]; 1492 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1493 for (i = 0, c = 0; i < 32 && c < count / 2; i++) 1494 c += ccpbk[i]; 1495 medc = (i >= 2) ? (1 << (i - 2)) : 0; 1496 1497 printf("Scheduled callouts statistic snapshot:\n"); 1498 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n", 1499 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT); 1500 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n", 1501 medc, 1502 count / callwheelsize / mp_ncpus, 1503 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000, 1504 maxc); 1505 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1506 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32, 1507 (st / count) / 1000000, (st / count) % 1000000, 1508 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32); 1509 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1510 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32, 1511 (spr / count) / 1000000, (spr / count) % 1000000, 1512 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32); 1513 printf(" Distribution: \tbuckets\t time\t tcum\t" 1514 " prec\t pcum\n"); 1515 for (i = 0, tcum = pcum = 0; i < 64; i++) { 1516 if (ct[i] == 0 && cpr[i] == 0) 1517 continue; 1518 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0; 1519 tcum += ct[i]; 1520 pcum += cpr[i]; 1521 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n", 1522 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32, 1523 i - 1 - (32 - CC_HASH_SHIFT), 1524 ct[i], tcum, cpr[i], pcum); 1525 } 1526 return (error); 1527 } 1528 SYSCTL_PROC(_kern, OID_AUTO, callout_stat, 1529 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1530 0, 0, sysctl_kern_callout_stat, "I", 1531 "Dump immediate statistic snapshot of the scheduled callouts"); 1532 1533 #ifdef DDB 1534 static void 1535 _show_callout(struct callout *c) 1536 { 1537 1538 db_printf("callout %p\n", c); 1539 #define C_DB_PRINTF(f, e) db_printf(" %s = " f "\n", #e, c->e); 1540 db_printf(" &c_links = %p\n", &(c->c_links)); 1541 C_DB_PRINTF("%" PRId64, c_time); 1542 C_DB_PRINTF("%" PRId64, c_precision); 1543 C_DB_PRINTF("%p", c_arg); 1544 C_DB_PRINTF("%p", c_func); 1545 C_DB_PRINTF("%p", c_lock); 1546 C_DB_PRINTF("%#x", c_flags); 1547 C_DB_PRINTF("%#x", c_iflags); 1548 C_DB_PRINTF("%d", c_cpu); 1549 #undef C_DB_PRINTF 1550 } 1551 1552 DB_SHOW_COMMAND(callout, db_show_callout) 1553 { 1554 1555 if (!have_addr) { 1556 db_printf("usage: show callout <struct callout *>\n"); 1557 return; 1558 } 1559 1560 _show_callout((struct callout *)addr); 1561 } 1562 1563 static void 1564 _show_last_callout(int cpu, int direct, const char *dirstr) 1565 { 1566 struct callout_cpu *cc; 1567 void *func, *arg; 1568 1569 cc = CC_CPU(cpu); 1570 func = cc_exec_last_func(cc, direct); 1571 arg = cc_exec_last_arg(cc, direct); 1572 db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func); 1573 db_printsym((db_expr_t)func, DB_STGY_ANY); 1574 db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg); 1575 } 1576 1577 DB_SHOW_COMMAND(callout_last, db_show_callout_last) 1578 { 1579 int cpu, last; 1580 1581 if (have_addr) { 1582 if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) { 1583 db_printf("no such cpu: %d\n", (int)addr); 1584 return; 1585 } 1586 cpu = last = addr; 1587 } else { 1588 cpu = 0; 1589 last = mp_maxid; 1590 } 1591 1592 while (cpu <= last) { 1593 if (!CPU_ABSENT(cpu)) { 1594 _show_last_callout(cpu, 0, ""); 1595 _show_last_callout(cpu, 1, " direct"); 1596 } 1597 cpu++; 1598 } 1599 } 1600 #endif /* DDB */ 1601