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 u_int callwheelsize, callwheelmask; 134 135 /* 136 * The callout cpu exec entities represent informations necessary for 137 * describing the state of callouts currently running on the CPU and the ones 138 * necessary for migrating callouts to the new callout cpu. In particular, 139 * the first entry of the array cc_exec_entity holds informations for callout 140 * running in SWI thread context, while the second one holds informations 141 * for callout running directly from hardware interrupt context. 142 * The cached informations are very important for deferring migration when 143 * the migrating callout is already running. 144 */ 145 struct cc_exec { 146 struct callout *cc_curr; 147 void (*cc_drain)(void *); 148 void *cc_last_func; 149 void *cc_last_arg; 150 #ifdef SMP 151 void (*ce_migration_func)(void *); 152 void *ce_migration_arg; 153 sbintime_t ce_migration_time; 154 sbintime_t ce_migration_prec; 155 int ce_migration_cpu; 156 #endif 157 bool cc_cancel; 158 bool cc_waiting; 159 }; 160 161 /* 162 * There is one struct callout_cpu per cpu, holding all relevant 163 * state for the callout processing thread on the individual CPU. 164 */ 165 struct callout_cpu { 166 struct mtx_padalign cc_lock; 167 struct cc_exec cc_exec_entity[2]; 168 struct callout *cc_next; 169 struct callout *cc_callout; 170 struct callout_list *cc_callwheel; 171 struct callout_tailq cc_expireq; 172 struct callout_slist cc_callfree; 173 sbintime_t cc_firstevent; 174 sbintime_t cc_lastscan; 175 void *cc_cookie; 176 u_int cc_bucket; 177 u_int cc_inited; 178 #ifdef KTR 179 char cc_ktr_event_name[20]; 180 #endif 181 }; 182 183 #define callout_migrating(c) ((c)->c_iflags & CALLOUT_DFRMIGRATION) 184 185 #define cc_exec_curr(cc, dir) cc->cc_exec_entity[dir].cc_curr 186 #define cc_exec_last_func(cc, dir) cc->cc_exec_entity[dir].cc_last_func 187 #define cc_exec_last_arg(cc, dir) cc->cc_exec_entity[dir].cc_last_arg 188 #define cc_exec_drain(cc, dir) cc->cc_exec_entity[dir].cc_drain 189 #define cc_exec_next(cc) cc->cc_next 190 #define cc_exec_cancel(cc, dir) cc->cc_exec_entity[dir].cc_cancel 191 #define cc_exec_waiting(cc, dir) cc->cc_exec_entity[dir].cc_waiting 192 #ifdef SMP 193 #define cc_migration_func(cc, dir) cc->cc_exec_entity[dir].ce_migration_func 194 #define cc_migration_arg(cc, dir) cc->cc_exec_entity[dir].ce_migration_arg 195 #define cc_migration_cpu(cc, dir) cc->cc_exec_entity[dir].ce_migration_cpu 196 #define cc_migration_time(cc, dir) cc->cc_exec_entity[dir].ce_migration_time 197 #define cc_migration_prec(cc, dir) cc->cc_exec_entity[dir].ce_migration_prec 198 199 struct callout_cpu cc_cpu[MAXCPU]; 200 #define CPUBLOCK MAXCPU 201 #define CC_CPU(cpu) (&cc_cpu[(cpu)]) 202 #define CC_SELF() CC_CPU(PCPU_GET(cpuid)) 203 #else 204 struct callout_cpu cc_cpu; 205 #define CC_CPU(cpu) &cc_cpu 206 #define CC_SELF() &cc_cpu 207 #endif 208 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock) 209 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock) 210 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED) 211 212 static int timeout_cpu; 213 214 static void callout_cpu_init(struct callout_cpu *cc, int cpu); 215 static void softclock_call_cc(struct callout *c, struct callout_cpu *cc, 216 #ifdef CALLOUT_PROFILING 217 int *mpcalls, int *lockcalls, int *gcalls, 218 #endif 219 int direct); 220 221 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures"); 222 223 /** 224 * Locked by cc_lock: 225 * cc_curr - If a callout is in progress, it is cc_curr. 226 * If cc_curr is non-NULL, threads waiting in 227 * callout_drain() will be woken up as soon as the 228 * relevant callout completes. 229 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held 230 * guarantees that the current callout will not run. 231 * The softclock() function sets this to 0 before it 232 * drops callout_lock to acquire c_lock, and it calls 233 * the handler only if curr_cancelled is still 0 after 234 * cc_lock is successfully acquired. 235 * cc_waiting - If a thread is waiting in callout_drain(), then 236 * callout_wait is nonzero. Set only when 237 * cc_curr is non-NULL. 238 */ 239 240 /* 241 * Resets the execution entity tied to a specific callout cpu. 242 */ 243 static void 244 cc_cce_cleanup(struct callout_cpu *cc, int direct) 245 { 246 247 cc_exec_curr(cc, direct) = NULL; 248 cc_exec_cancel(cc, direct) = false; 249 cc_exec_waiting(cc, direct) = false; 250 #ifdef SMP 251 cc_migration_cpu(cc, direct) = CPUBLOCK; 252 cc_migration_time(cc, direct) = 0; 253 cc_migration_prec(cc, direct) = 0; 254 cc_migration_func(cc, direct) = NULL; 255 cc_migration_arg(cc, direct) = NULL; 256 #endif 257 } 258 259 /* 260 * Checks if migration is requested by a specific callout cpu. 261 */ 262 static int 263 cc_cce_migrating(struct callout_cpu *cc, int direct) 264 { 265 266 #ifdef SMP 267 return (cc_migration_cpu(cc, direct) != CPUBLOCK); 268 #else 269 return (0); 270 #endif 271 } 272 273 /* 274 * Kernel low level callwheel initialization 275 * called on the BSP during kernel startup. 276 */ 277 static void 278 callout_callwheel_init(void *dummy) 279 { 280 struct callout_cpu *cc; 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 memset(CC_CPU(curcpu), 0, sizeof(cc_cpu)); 289 ncallout = imin(16 + maxproc + maxfiles, 18508); 290 TUNABLE_INT_FETCH("kern.ncallout", &ncallout); 291 292 /* 293 * Calculate callout wheel size, should be next power of two higher 294 * than 'ncallout'. 295 */ 296 callwheelsize = 1 << fls(ncallout); 297 callwheelmask = callwheelsize - 1; 298 299 /* 300 * Fetch whether we're pinning the swi's or not. 301 */ 302 TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi); 303 TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi); 304 305 /* 306 * Only BSP handles timeout(9) and receives a preallocation. 307 * 308 * XXX: Once all timeout(9) consumers are converted this can 309 * be removed. 310 */ 311 timeout_cpu = PCPU_GET(cpuid); 312 cc = CC_CPU(timeout_cpu); 313 cc->cc_callout = malloc(ncallout * sizeof(struct callout), 314 M_CALLOUT, M_WAITOK); 315 callout_cpu_init(cc, timeout_cpu); 316 } 317 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL); 318 319 /* 320 * Initialize the per-cpu callout structures. 321 */ 322 static void 323 callout_cpu_init(struct callout_cpu *cc, int cpu) 324 { 325 struct callout *c; 326 int i; 327 328 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE); 329 SLIST_INIT(&cc->cc_callfree); 330 cc->cc_inited = 1; 331 cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) * 332 callwheelsize, M_CALLOUT, 333 DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK); 334 for (i = 0; i < callwheelsize; i++) 335 LIST_INIT(&cc->cc_callwheel[i]); 336 TAILQ_INIT(&cc->cc_expireq); 337 cc->cc_firstevent = SBT_MAX; 338 for (i = 0; i < 2; i++) 339 cc_cce_cleanup(cc, i); 340 #ifdef KTR 341 snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name), 342 "callwheel cpu %d", cpu); 343 #endif 344 if (cc->cc_callout == NULL) /* Only BSP handles timeout(9) */ 345 return; 346 for (i = 0; i < ncallout; i++) { 347 c = &cc->cc_callout[i]; 348 callout_init(c, 0); 349 c->c_iflags = CALLOUT_LOCAL_ALLOC; 350 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 351 } 352 } 353 354 #ifdef SMP 355 /* 356 * Switches the cpu tied to a specific callout. 357 * The function expects a locked incoming callout cpu and returns with 358 * locked outcoming callout cpu. 359 */ 360 static struct callout_cpu * 361 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu) 362 { 363 struct callout_cpu *new_cc; 364 365 MPASS(c != NULL && cc != NULL); 366 CC_LOCK_ASSERT(cc); 367 368 /* 369 * Avoid interrupts and preemption firing after the callout cpu 370 * is blocked in order to avoid deadlocks as the new thread 371 * may be willing to acquire the callout cpu lock. 372 */ 373 c->c_cpu = CPUBLOCK; 374 spinlock_enter(); 375 CC_UNLOCK(cc); 376 new_cc = CC_CPU(new_cpu); 377 CC_LOCK(new_cc); 378 spinlock_exit(); 379 c->c_cpu = new_cpu; 380 return (new_cc); 381 } 382 #endif 383 384 /* 385 * Start standard softclock thread. 386 */ 387 static void 388 start_softclock(void *dummy) 389 { 390 struct callout_cpu *cc; 391 char name[MAXCOMLEN]; 392 #ifdef SMP 393 int cpu; 394 struct intr_event *ie; 395 #endif 396 397 cc = CC_CPU(timeout_cpu); 398 snprintf(name, sizeof(name), "clock (%d)", timeout_cpu); 399 if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK, 400 INTR_MPSAFE, &cc->cc_cookie)) 401 panic("died while creating standard software ithreads"); 402 if (pin_default_swi && 403 (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) { 404 printf("%s: timeout clock couldn't be pinned to cpu %d\n", 405 __func__, 406 timeout_cpu); 407 } 408 409 #ifdef SMP 410 CPU_FOREACH(cpu) { 411 if (cpu == timeout_cpu) 412 continue; 413 cc = CC_CPU(cpu); 414 cc->cc_callout = NULL; /* Only BSP handles timeout(9). */ 415 callout_cpu_init(cc, cpu); 416 snprintf(name, sizeof(name), "clock (%d)", cpu); 417 ie = NULL; 418 if (swi_add(&ie, name, softclock, cc, SWI_CLOCK, 419 INTR_MPSAFE, &cc->cc_cookie)) 420 panic("died while creating standard software ithreads"); 421 if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) { 422 printf("%s: per-cpu clock couldn't be pinned to " 423 "cpu %d\n", 424 __func__, 425 cpu); 426 } 427 } 428 #endif 429 } 430 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL); 431 432 #define CC_HASH_SHIFT 8 433 434 static inline u_int 435 callout_hash(sbintime_t sbt) 436 { 437 438 return (sbt >> (32 - CC_HASH_SHIFT)); 439 } 440 441 static inline u_int 442 callout_get_bucket(sbintime_t sbt) 443 { 444 445 return (callout_hash(sbt) & callwheelmask); 446 } 447 448 void 449 callout_process(sbintime_t now) 450 { 451 struct callout *tmp, *tmpn; 452 struct callout_cpu *cc; 453 struct callout_list *sc; 454 sbintime_t first, last, max, tmp_max; 455 uint32_t lookahead; 456 u_int firstb, lastb, nowb; 457 #ifdef CALLOUT_PROFILING 458 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0; 459 #endif 460 461 cc = CC_SELF(); 462 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET); 463 464 /* Compute the buckets of the last scan and present times. */ 465 firstb = callout_hash(cc->cc_lastscan); 466 cc->cc_lastscan = now; 467 nowb = callout_hash(now); 468 469 /* Compute the last bucket and minimum time of the bucket after it. */ 470 if (nowb == firstb) 471 lookahead = (SBT_1S / 16); 472 else if (nowb - firstb == 1) 473 lookahead = (SBT_1S / 8); 474 else 475 lookahead = (SBT_1S / 2); 476 first = last = now; 477 first += (lookahead / 2); 478 last += lookahead; 479 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT)); 480 lastb = callout_hash(last) - 1; 481 max = last; 482 483 /* 484 * Check if we wrapped around the entire wheel from the last scan. 485 * In case, we need to scan entirely the wheel for pending callouts. 486 */ 487 if (lastb - firstb >= callwheelsize) { 488 lastb = firstb + callwheelsize - 1; 489 if (nowb - firstb >= callwheelsize) 490 nowb = lastb; 491 } 492 493 /* Iterate callwheel from firstb to nowb and then up to lastb. */ 494 do { 495 sc = &cc->cc_callwheel[firstb & callwheelmask]; 496 tmp = LIST_FIRST(sc); 497 while (tmp != NULL) { 498 /* Run the callout if present time within allowed. */ 499 if (tmp->c_time <= now) { 500 /* 501 * Consumer told us the callout may be run 502 * directly from hardware interrupt context. 503 */ 504 if (tmp->c_iflags & CALLOUT_DIRECT) { 505 #ifdef CALLOUT_PROFILING 506 ++depth_dir; 507 #endif 508 cc_exec_next(cc) = 509 LIST_NEXT(tmp, c_links.le); 510 cc->cc_bucket = firstb & callwheelmask; 511 LIST_REMOVE(tmp, c_links.le); 512 softclock_call_cc(tmp, cc, 513 #ifdef CALLOUT_PROFILING 514 &mpcalls_dir, &lockcalls_dir, NULL, 515 #endif 516 1); 517 tmp = cc_exec_next(cc); 518 cc_exec_next(cc) = NULL; 519 } else { 520 tmpn = LIST_NEXT(tmp, c_links.le); 521 LIST_REMOVE(tmp, c_links.le); 522 TAILQ_INSERT_TAIL(&cc->cc_expireq, 523 tmp, c_links.tqe); 524 tmp->c_iflags |= CALLOUT_PROCESSED; 525 tmp = tmpn; 526 } 527 continue; 528 } 529 /* Skip events from distant future. */ 530 if (tmp->c_time >= max) 531 goto next; 532 /* 533 * Event minimal time is bigger than present maximal 534 * time, so it cannot be aggregated. 535 */ 536 if (tmp->c_time > last) { 537 lastb = nowb; 538 goto next; 539 } 540 /* Update first and last time, respecting this event. */ 541 if (tmp->c_time < first) 542 first = tmp->c_time; 543 tmp_max = tmp->c_time + tmp->c_precision; 544 if (tmp_max < last) 545 last = tmp_max; 546 next: 547 tmp = LIST_NEXT(tmp, c_links.le); 548 } 549 /* Proceed with the next bucket. */ 550 firstb++; 551 /* 552 * Stop if we looked after present time and found 553 * some event we can't execute at now. 554 * Stop if we looked far enough into the future. 555 */ 556 } while (((int)(firstb - lastb)) <= 0); 557 cc->cc_firstevent = last; 558 #ifndef NO_EVENTTIMERS 559 cpu_new_callout(curcpu, last, first); 560 #endif 561 #ifdef CALLOUT_PROFILING 562 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8; 563 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8; 564 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8; 565 #endif 566 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET); 567 /* 568 * swi_sched acquires the thread lock, so we don't want to call it 569 * with cc_lock held; incorrect locking order. 570 */ 571 if (!TAILQ_EMPTY(&cc->cc_expireq)) 572 swi_sched(cc->cc_cookie, 0); 573 } 574 575 static struct callout_cpu * 576 callout_lock(struct callout *c) 577 { 578 struct callout_cpu *cc; 579 int cpu; 580 581 for (;;) { 582 cpu = c->c_cpu; 583 #ifdef SMP 584 if (cpu == CPUBLOCK) { 585 while (c->c_cpu == CPUBLOCK) 586 cpu_spinwait(); 587 continue; 588 } 589 #endif 590 cc = CC_CPU(cpu); 591 CC_LOCK(cc); 592 if (cpu == c->c_cpu) 593 break; 594 CC_UNLOCK(cc); 595 } 596 return (cc); 597 } 598 599 static void 600 callout_cc_add(struct callout *c, struct callout_cpu *cc, 601 sbintime_t sbt, sbintime_t precision, void (*func)(void *), 602 void *arg, int cpu, int flags) 603 { 604 int bucket; 605 606 CC_LOCK_ASSERT(cc); 607 if (sbt < cc->cc_lastscan) 608 sbt = cc->cc_lastscan; 609 c->c_arg = arg; 610 c->c_iflags |= CALLOUT_PENDING; 611 c->c_iflags &= ~CALLOUT_PROCESSED; 612 c->c_flags |= CALLOUT_ACTIVE; 613 if (flags & C_DIRECT_EXEC) 614 c->c_iflags |= CALLOUT_DIRECT; 615 c->c_func = func; 616 c->c_time = sbt; 617 c->c_precision = precision; 618 bucket = callout_get_bucket(c->c_time); 619 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x", 620 c, (int)(c->c_precision >> 32), 621 (u_int)(c->c_precision & 0xffffffff)); 622 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le); 623 if (cc->cc_bucket == bucket) 624 cc_exec_next(cc) = c; 625 #ifndef NO_EVENTTIMERS 626 /* 627 * Inform the eventtimers(4) subsystem there's a new callout 628 * that has been inserted, but only if really required. 629 */ 630 if (SBT_MAX - c->c_time < c->c_precision) 631 c->c_precision = SBT_MAX - c->c_time; 632 sbt = c->c_time + c->c_precision; 633 if (sbt < cc->cc_firstevent) { 634 cc->cc_firstevent = sbt; 635 cpu_new_callout(cpu, sbt, c->c_time); 636 } 637 #endif 638 } 639 640 static void 641 callout_cc_del(struct callout *c, struct callout_cpu *cc) 642 { 643 644 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0) 645 return; 646 c->c_func = NULL; 647 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 648 } 649 650 static void 651 softclock_call_cc(struct callout *c, struct callout_cpu *cc, 652 #ifdef CALLOUT_PROFILING 653 int *mpcalls, int *lockcalls, int *gcalls, 654 #endif 655 int direct) 656 { 657 struct rm_priotracker tracker; 658 void (*c_func)(void *); 659 void *c_arg; 660 struct lock_class *class; 661 struct lock_object *c_lock; 662 uintptr_t lock_status; 663 int c_iflags; 664 #ifdef SMP 665 struct callout_cpu *new_cc; 666 void (*new_func)(void *); 667 void *new_arg; 668 int flags, new_cpu; 669 sbintime_t new_prec, new_time; 670 #endif 671 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 672 sbintime_t sbt1, sbt2; 673 struct timespec ts2; 674 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */ 675 static timeout_t *lastfunc; 676 #endif 677 678 KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING, 679 ("softclock_call_cc: pend %p %x", c, c->c_iflags)); 680 KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE, 681 ("softclock_call_cc: act %p %x", c, c->c_flags)); 682 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL; 683 lock_status = 0; 684 if (c->c_flags & CALLOUT_SHAREDLOCK) { 685 if (class == &lock_class_rm) 686 lock_status = (uintptr_t)&tracker; 687 else 688 lock_status = 1; 689 } 690 c_lock = c->c_lock; 691 c_func = c->c_func; 692 c_arg = c->c_arg; 693 c_iflags = c->c_iflags; 694 if (c->c_iflags & CALLOUT_LOCAL_ALLOC) 695 c->c_iflags = CALLOUT_LOCAL_ALLOC; 696 else 697 c->c_iflags &= ~CALLOUT_PENDING; 698 699 cc_exec_curr(cc, direct) = c; 700 cc_exec_last_func(cc, direct) = c_func; 701 cc_exec_last_arg(cc, direct) = c_arg; 702 cc_exec_cancel(cc, direct) = false; 703 cc_exec_drain(cc, direct) = NULL; 704 CC_UNLOCK(cc); 705 if (c_lock != NULL) { 706 class->lc_lock(c_lock, lock_status); 707 /* 708 * The callout may have been cancelled 709 * while we switched locks. 710 */ 711 if (cc_exec_cancel(cc, direct)) { 712 class->lc_unlock(c_lock); 713 goto skip; 714 } 715 /* The callout cannot be stopped now. */ 716 cc_exec_cancel(cc, direct) = true; 717 if (c_lock == &Giant.lock_object) { 718 #ifdef CALLOUT_PROFILING 719 (*gcalls)++; 720 #endif 721 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p", 722 c, c_func, c_arg); 723 } else { 724 #ifdef CALLOUT_PROFILING 725 (*lockcalls)++; 726 #endif 727 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p", 728 c, c_func, c_arg); 729 } 730 } else { 731 #ifdef CALLOUT_PROFILING 732 (*mpcalls)++; 733 #endif 734 CTR3(KTR_CALLOUT, "callout %p func %p arg %p", 735 c, c_func, c_arg); 736 } 737 KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running", 738 "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct); 739 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 740 sbt1 = sbinuptime(); 741 #endif 742 THREAD_NO_SLEEPING(); 743 SDT_PROBE1(callout_execute, , , callout__start, c); 744 c_func(c_arg); 745 SDT_PROBE1(callout_execute, , , callout__end, c); 746 THREAD_SLEEPING_OK(); 747 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 748 sbt2 = sbinuptime(); 749 sbt2 -= sbt1; 750 if (sbt2 > maxdt) { 751 if (lastfunc != c_func || sbt2 > maxdt * 2) { 752 ts2 = sbttots(sbt2); 753 printf( 754 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n", 755 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec); 756 } 757 maxdt = sbt2; 758 lastfunc = c_func; 759 } 760 #endif 761 KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle"); 762 CTR1(KTR_CALLOUT, "callout %p finished", c); 763 if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0) 764 class->lc_unlock(c_lock); 765 skip: 766 CC_LOCK(cc); 767 KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr")); 768 cc_exec_curr(cc, direct) = NULL; 769 if (cc_exec_drain(cc, direct)) { 770 void (*drain)(void *); 771 772 drain = cc_exec_drain(cc, direct); 773 cc_exec_drain(cc, direct) = NULL; 774 CC_UNLOCK(cc); 775 drain(c_arg); 776 CC_LOCK(cc); 777 } 778 if (cc_exec_waiting(cc, direct)) { 779 /* 780 * There is someone waiting for the 781 * callout to complete. 782 * If the callout was scheduled for 783 * migration just cancel it. 784 */ 785 if (cc_cce_migrating(cc, direct)) { 786 cc_cce_cleanup(cc, direct); 787 788 /* 789 * It should be assert here that the callout is not 790 * destroyed but that is not easy. 791 */ 792 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 793 } 794 cc_exec_waiting(cc, direct) = false; 795 CC_UNLOCK(cc); 796 wakeup(&cc_exec_waiting(cc, direct)); 797 CC_LOCK(cc); 798 } else if (cc_cce_migrating(cc, direct)) { 799 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0, 800 ("Migrating legacy callout %p", c)); 801 #ifdef SMP 802 /* 803 * If the callout was scheduled for 804 * migration just perform it now. 805 */ 806 new_cpu = cc_migration_cpu(cc, direct); 807 new_time = cc_migration_time(cc, direct); 808 new_prec = cc_migration_prec(cc, direct); 809 new_func = cc_migration_func(cc, direct); 810 new_arg = cc_migration_arg(cc, direct); 811 cc_cce_cleanup(cc, direct); 812 813 /* 814 * It should be assert here that the callout is not destroyed 815 * but that is not easy. 816 * 817 * As first thing, handle deferred callout stops. 818 */ 819 if (!callout_migrating(c)) { 820 CTR3(KTR_CALLOUT, 821 "deferred cancelled %p func %p arg %p", 822 c, new_func, new_arg); 823 callout_cc_del(c, cc); 824 return; 825 } 826 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 827 828 new_cc = callout_cpu_switch(c, cc, new_cpu); 829 flags = (direct) ? C_DIRECT_EXEC : 0; 830 callout_cc_add(c, new_cc, new_time, new_prec, new_func, 831 new_arg, new_cpu, flags); 832 CC_UNLOCK(new_cc); 833 CC_LOCK(cc); 834 #else 835 panic("migration should not happen"); 836 #endif 837 } 838 /* 839 * If the current callout is locally allocated (from 840 * timeout(9)) then put it on the freelist. 841 * 842 * Note: we need to check the cached copy of c_iflags because 843 * if it was not local, then it's not safe to deref the 844 * callout pointer. 845 */ 846 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 || 847 c->c_iflags == CALLOUT_LOCAL_ALLOC, 848 ("corrupted callout")); 849 if (c_iflags & CALLOUT_LOCAL_ALLOC) 850 callout_cc_del(c, cc); 851 } 852 853 /* 854 * The callout mechanism is based on the work of Adam M. Costello and 855 * George Varghese, published in a technical report entitled "Redesigning 856 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 857 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 858 * used in this implementation was published by G. Varghese and T. Lauck in 859 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 860 * the Efficient Implementation of a Timer Facility" in the Proceedings of 861 * the 11th ACM Annual Symposium on Operating Systems Principles, 862 * Austin, Texas Nov 1987. 863 */ 864 865 /* 866 * Software (low priority) clock interrupt. 867 * Run periodic events from timeout queue. 868 */ 869 void 870 softclock(void *arg) 871 { 872 struct callout_cpu *cc; 873 struct callout *c; 874 #ifdef CALLOUT_PROFILING 875 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0; 876 #endif 877 878 cc = (struct callout_cpu *)arg; 879 CC_LOCK(cc); 880 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) { 881 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 882 softclock_call_cc(c, cc, 883 #ifdef CALLOUT_PROFILING 884 &mpcalls, &lockcalls, &gcalls, 885 #endif 886 0); 887 #ifdef CALLOUT_PROFILING 888 ++depth; 889 #endif 890 } 891 #ifdef CALLOUT_PROFILING 892 avg_depth += (depth * 1000 - avg_depth) >> 8; 893 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8; 894 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8; 895 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8; 896 #endif 897 CC_UNLOCK(cc); 898 } 899 900 /* 901 * timeout -- 902 * Execute a function after a specified length of time. 903 * 904 * untimeout -- 905 * Cancel previous timeout function call. 906 * 907 * callout_handle_init -- 908 * Initialize a handle so that using it with untimeout is benign. 909 * 910 * See AT&T BCI Driver Reference Manual for specification. This 911 * implementation differs from that one in that although an 912 * identification value is returned from timeout, the original 913 * arguments to timeout as well as the identifier are used to 914 * identify entries for untimeout. 915 */ 916 struct callout_handle 917 timeout(timeout_t *ftn, void *arg, int to_ticks) 918 { 919 struct callout_cpu *cc; 920 struct callout *new; 921 struct callout_handle handle; 922 923 cc = CC_CPU(timeout_cpu); 924 CC_LOCK(cc); 925 /* Fill in the next free callout structure. */ 926 new = SLIST_FIRST(&cc->cc_callfree); 927 if (new == NULL) 928 /* XXX Attempt to malloc first */ 929 panic("timeout table full"); 930 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle); 931 callout_reset(new, to_ticks, ftn, arg); 932 handle.callout = new; 933 CC_UNLOCK(cc); 934 935 return (handle); 936 } 937 938 void 939 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle) 940 { 941 struct callout_cpu *cc; 942 943 /* 944 * Check for a handle that was initialized 945 * by callout_handle_init, but never used 946 * for a real timeout. 947 */ 948 if (handle.callout == NULL) 949 return; 950 951 cc = callout_lock(handle.callout); 952 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg) 953 callout_stop(handle.callout); 954 CC_UNLOCK(cc); 955 } 956 957 void 958 callout_handle_init(struct callout_handle *handle) 959 { 960 handle->callout = NULL; 961 } 962 963 void 964 callout_when(sbintime_t sbt, sbintime_t precision, int flags, 965 sbintime_t *res, sbintime_t *prec_res) 966 { 967 sbintime_t to_sbt, to_pr; 968 969 if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) { 970 *res = sbt; 971 *prec_res = precision; 972 return; 973 } 974 if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt) 975 sbt = tick_sbt; 976 if ((flags & C_HARDCLOCK) != 0 || 977 #ifdef NO_EVENTTIMERS 978 sbt >= sbt_timethreshold) { 979 to_sbt = getsbinuptime(); 980 981 /* Add safety belt for the case of hz > 1000. */ 982 to_sbt += tc_tick_sbt - tick_sbt; 983 #else 984 sbt >= sbt_tickthreshold) { 985 /* 986 * Obtain the time of the last hardclock() call on 987 * this CPU directly from the kern_clocksource.c. 988 * This value is per-CPU, but it is equal for all 989 * active ones. 990 */ 991 #ifdef __LP64__ 992 to_sbt = DPCPU_GET(hardclocktime); 993 #else 994 spinlock_enter(); 995 to_sbt = DPCPU_GET(hardclocktime); 996 spinlock_exit(); 997 #endif 998 #endif 999 if (cold && to_sbt == 0) 1000 to_sbt = sbinuptime(); 1001 if ((flags & C_HARDCLOCK) == 0) 1002 to_sbt += tick_sbt; 1003 } else 1004 to_sbt = sbinuptime(); 1005 if (SBT_MAX - to_sbt < sbt) 1006 to_sbt = SBT_MAX; 1007 else 1008 to_sbt += sbt; 1009 *res = to_sbt; 1010 to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp : 1011 sbt >> C_PRELGET(flags)); 1012 *prec_res = to_pr > precision ? to_pr : precision; 1013 } 1014 1015 /* 1016 * New interface; clients allocate their own callout structures. 1017 * 1018 * callout_reset() - establish or change a timeout 1019 * callout_stop() - disestablish a timeout 1020 * callout_init() - initialize a callout structure so that it can 1021 * safely be passed to callout_reset() and callout_stop() 1022 * 1023 * <sys/callout.h> defines three convenience macros: 1024 * 1025 * callout_active() - returns truth if callout has not been stopped, 1026 * drained, or deactivated since the last time the callout was 1027 * reset. 1028 * callout_pending() - returns truth if callout is still waiting for timeout 1029 * callout_deactivate() - marks the callout as having been serviced 1030 */ 1031 int 1032 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec, 1033 void (*ftn)(void *), void *arg, int cpu, int flags) 1034 { 1035 sbintime_t to_sbt, precision; 1036 struct callout_cpu *cc; 1037 int cancelled, direct; 1038 int ignore_cpu=0; 1039 1040 cancelled = 0; 1041 if (cpu == -1) { 1042 ignore_cpu = 1; 1043 } else if ((cpu >= MAXCPU) || 1044 ((CC_CPU(cpu))->cc_inited == 0)) { 1045 /* Invalid CPU spec */ 1046 panic("Invalid CPU in callout %d", cpu); 1047 } 1048 callout_when(sbt, prec, flags, &to_sbt, &precision); 1049 1050 /* 1051 * This flag used to be added by callout_cc_add, but the 1052 * first time you call this we could end up with the 1053 * wrong direct flag if we don't do it before we add. 1054 */ 1055 if (flags & C_DIRECT_EXEC) { 1056 direct = 1; 1057 } else { 1058 direct = 0; 1059 } 1060 KASSERT(!direct || c->c_lock == NULL, 1061 ("%s: direct callout %p has lock", __func__, c)); 1062 cc = callout_lock(c); 1063 /* 1064 * Don't allow migration of pre-allocated callouts lest they 1065 * become unbalanced or handle the case where the user does 1066 * not care. 1067 */ 1068 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) || 1069 ignore_cpu) { 1070 cpu = c->c_cpu; 1071 } 1072 1073 if (cc_exec_curr(cc, direct) == c) { 1074 /* 1075 * We're being asked to reschedule a callout which is 1076 * currently in progress. If there is a lock then we 1077 * can cancel the callout if it has not really started. 1078 */ 1079 if (c->c_lock != NULL && !cc_exec_cancel(cc, direct)) 1080 cancelled = cc_exec_cancel(cc, direct) = true; 1081 if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) { 1082 /* 1083 * Someone has called callout_drain to kill this 1084 * callout. Don't reschedule. 1085 */ 1086 CTR4(KTR_CALLOUT, "%s %p func %p arg %p", 1087 cancelled ? "cancelled" : "failed to cancel", 1088 c, c->c_func, c->c_arg); 1089 CC_UNLOCK(cc); 1090 return (cancelled); 1091 } 1092 #ifdef SMP 1093 if (callout_migrating(c)) { 1094 /* 1095 * This only occurs when a second callout_reset_sbt_on 1096 * is made after a previous one moved it into 1097 * deferred migration (below). Note we do *not* change 1098 * the prev_cpu even though the previous target may 1099 * be different. 1100 */ 1101 cc_migration_cpu(cc, direct) = cpu; 1102 cc_migration_time(cc, direct) = to_sbt; 1103 cc_migration_prec(cc, direct) = precision; 1104 cc_migration_func(cc, direct) = ftn; 1105 cc_migration_arg(cc, direct) = arg; 1106 cancelled = 1; 1107 CC_UNLOCK(cc); 1108 return (cancelled); 1109 } 1110 #endif 1111 } 1112 if (c->c_iflags & CALLOUT_PENDING) { 1113 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1114 if (cc_exec_next(cc) == c) 1115 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1116 LIST_REMOVE(c, c_links.le); 1117 } else { 1118 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1119 } 1120 cancelled = 1; 1121 c->c_iflags &= ~ CALLOUT_PENDING; 1122 c->c_flags &= ~ CALLOUT_ACTIVE; 1123 } 1124 1125 #ifdef SMP 1126 /* 1127 * If the callout must migrate try to perform it immediately. 1128 * If the callout is currently running, just defer the migration 1129 * to a more appropriate moment. 1130 */ 1131 if (c->c_cpu != cpu) { 1132 if (cc_exec_curr(cc, direct) == c) { 1133 /* 1134 * Pending will have been removed since we are 1135 * actually executing the callout on another 1136 * CPU. That callout should be waiting on the 1137 * lock the caller holds. If we set both 1138 * active/and/pending after we return and the 1139 * lock on the executing callout proceeds, it 1140 * will then see pending is true and return. 1141 * At the return from the actual callout execution 1142 * the migration will occur in softclock_call_cc 1143 * and this new callout will be placed on the 1144 * new CPU via a call to callout_cpu_switch() which 1145 * will get the lock on the right CPU followed 1146 * by a call callout_cc_add() which will add it there. 1147 * (see above in softclock_call_cc()). 1148 */ 1149 cc_migration_cpu(cc, direct) = cpu; 1150 cc_migration_time(cc, direct) = to_sbt; 1151 cc_migration_prec(cc, direct) = precision; 1152 cc_migration_func(cc, direct) = ftn; 1153 cc_migration_arg(cc, direct) = arg; 1154 c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING); 1155 c->c_flags |= CALLOUT_ACTIVE; 1156 CTR6(KTR_CALLOUT, 1157 "migration of %p func %p arg %p in %d.%08x to %u deferred", 1158 c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1159 (u_int)(to_sbt & 0xffffffff), cpu); 1160 CC_UNLOCK(cc); 1161 return (cancelled); 1162 } 1163 cc = callout_cpu_switch(c, cc, cpu); 1164 } 1165 #endif 1166 1167 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags); 1168 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x", 1169 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1170 (u_int)(to_sbt & 0xffffffff)); 1171 CC_UNLOCK(cc); 1172 1173 return (cancelled); 1174 } 1175 1176 /* 1177 * Common idioms that can be optimized in the future. 1178 */ 1179 int 1180 callout_schedule_on(struct callout *c, int to_ticks, int cpu) 1181 { 1182 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu); 1183 } 1184 1185 int 1186 callout_schedule(struct callout *c, int to_ticks) 1187 { 1188 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu); 1189 } 1190 1191 int 1192 _callout_stop_safe(struct callout *c, int flags, void (*drain)(void *)) 1193 { 1194 struct callout_cpu *cc, *old_cc; 1195 struct lock_class *class; 1196 int direct, sq_locked, use_lock; 1197 int cancelled, not_on_a_list; 1198 1199 if ((flags & CS_DRAIN) != 0) 1200 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock, 1201 "calling %s", __func__); 1202 1203 /* 1204 * Some old subsystems don't hold Giant while running a callout_stop(), 1205 * so just discard this check for the moment. 1206 */ 1207 if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) { 1208 if (c->c_lock == &Giant.lock_object) 1209 use_lock = mtx_owned(&Giant); 1210 else { 1211 use_lock = 1; 1212 class = LOCK_CLASS(c->c_lock); 1213 class->lc_assert(c->c_lock, LA_XLOCKED); 1214 } 1215 } else 1216 use_lock = 0; 1217 if (c->c_iflags & CALLOUT_DIRECT) { 1218 direct = 1; 1219 } else { 1220 direct = 0; 1221 } 1222 sq_locked = 0; 1223 old_cc = NULL; 1224 again: 1225 cc = callout_lock(c); 1226 1227 if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) == 1228 (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) && 1229 ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) { 1230 /* 1231 * Special case where this slipped in while we 1232 * were migrating *as* the callout is about to 1233 * execute. The caller probably holds the lock 1234 * the callout wants. 1235 * 1236 * Get rid of the migration first. Then set 1237 * the flag that tells this code *not* to 1238 * try to remove it from any lists (its not 1239 * on one yet). When the callout wheel runs, 1240 * it will ignore this callout. 1241 */ 1242 c->c_iflags &= ~CALLOUT_PENDING; 1243 c->c_flags &= ~CALLOUT_ACTIVE; 1244 not_on_a_list = 1; 1245 } else { 1246 not_on_a_list = 0; 1247 } 1248 1249 /* 1250 * If the callout was migrating while the callout cpu lock was 1251 * dropped, just drop the sleepqueue lock and check the states 1252 * again. 1253 */ 1254 if (sq_locked != 0 && cc != old_cc) { 1255 #ifdef SMP 1256 CC_UNLOCK(cc); 1257 sleepq_release(&cc_exec_waiting(old_cc, direct)); 1258 sq_locked = 0; 1259 old_cc = NULL; 1260 goto again; 1261 #else 1262 panic("migration should not happen"); 1263 #endif 1264 } 1265 1266 /* 1267 * If the callout is running, try to stop it or drain it. 1268 */ 1269 if (cc_exec_curr(cc, direct) == c) { 1270 /* 1271 * Succeed we to stop it or not, we must clear the 1272 * active flag - this is what API users expect. If we're 1273 * draining and the callout is currently executing, first wait 1274 * until it finishes. 1275 */ 1276 if ((flags & CS_DRAIN) == 0) 1277 c->c_flags &= ~CALLOUT_ACTIVE; 1278 1279 if ((flags & CS_DRAIN) != 0) { 1280 /* 1281 * The current callout is running (or just 1282 * about to run) and blocking is allowed, so 1283 * just wait for the current invocation to 1284 * finish. 1285 */ 1286 while (cc_exec_curr(cc, direct) == c) { 1287 /* 1288 * Use direct calls to sleepqueue interface 1289 * instead of cv/msleep in order to avoid 1290 * a LOR between cc_lock and sleepqueue 1291 * chain spinlocks. This piece of code 1292 * emulates a msleep_spin() call actually. 1293 * 1294 * If we already have the sleepqueue chain 1295 * locked, then we can safely block. If we 1296 * don't already have it locked, however, 1297 * we have to drop the cc_lock to lock 1298 * it. This opens several races, so we 1299 * restart at the beginning once we have 1300 * both locks. If nothing has changed, then 1301 * we will end up back here with sq_locked 1302 * set. 1303 */ 1304 if (!sq_locked) { 1305 CC_UNLOCK(cc); 1306 sleepq_lock( 1307 &cc_exec_waiting(cc, direct)); 1308 sq_locked = 1; 1309 old_cc = cc; 1310 goto again; 1311 } 1312 1313 /* 1314 * Migration could be cancelled here, but 1315 * as long as it is still not sure when it 1316 * will be packed up, just let softclock() 1317 * take care of it. 1318 */ 1319 cc_exec_waiting(cc, direct) = true; 1320 DROP_GIANT(); 1321 CC_UNLOCK(cc); 1322 sleepq_add( 1323 &cc_exec_waiting(cc, direct), 1324 &cc->cc_lock.lock_object, "codrain", 1325 SLEEPQ_SLEEP, 0); 1326 sleepq_wait( 1327 &cc_exec_waiting(cc, direct), 1328 0); 1329 sq_locked = 0; 1330 old_cc = NULL; 1331 1332 /* Reacquire locks previously released. */ 1333 PICKUP_GIANT(); 1334 CC_LOCK(cc); 1335 } 1336 c->c_flags &= ~CALLOUT_ACTIVE; 1337 } else if (use_lock && 1338 !cc_exec_cancel(cc, direct) && (drain == NULL)) { 1339 1340 /* 1341 * The current callout is waiting for its 1342 * lock which we hold. Cancel the callout 1343 * and return. After our caller drops the 1344 * lock, the callout will be skipped in 1345 * softclock(). This *only* works with a 1346 * callout_stop() *not* callout_drain() or 1347 * callout_async_drain(). 1348 */ 1349 cc_exec_cancel(cc, direct) = true; 1350 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1351 c, c->c_func, c->c_arg); 1352 KASSERT(!cc_cce_migrating(cc, direct), 1353 ("callout wrongly scheduled for migration")); 1354 if (callout_migrating(c)) { 1355 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1356 #ifdef SMP 1357 cc_migration_cpu(cc, direct) = CPUBLOCK; 1358 cc_migration_time(cc, direct) = 0; 1359 cc_migration_prec(cc, direct) = 0; 1360 cc_migration_func(cc, direct) = NULL; 1361 cc_migration_arg(cc, direct) = NULL; 1362 #endif 1363 } 1364 CC_UNLOCK(cc); 1365 KASSERT(!sq_locked, ("sleepqueue chain locked")); 1366 return (1); 1367 } else if (callout_migrating(c)) { 1368 /* 1369 * The callout is currently being serviced 1370 * and the "next" callout is scheduled at 1371 * its completion with a migration. We remove 1372 * the migration flag so it *won't* get rescheduled, 1373 * but we can't stop the one thats running so 1374 * we return 0. 1375 */ 1376 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1377 #ifdef SMP 1378 /* 1379 * We can't call cc_cce_cleanup here since 1380 * if we do it will remove .ce_curr and 1381 * its still running. This will prevent a 1382 * reschedule of the callout when the 1383 * execution completes. 1384 */ 1385 cc_migration_cpu(cc, direct) = CPUBLOCK; 1386 cc_migration_time(cc, direct) = 0; 1387 cc_migration_prec(cc, direct) = 0; 1388 cc_migration_func(cc, direct) = NULL; 1389 cc_migration_arg(cc, direct) = NULL; 1390 #endif 1391 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p", 1392 c, c->c_func, c->c_arg); 1393 if (drain) { 1394 cc_exec_drain(cc, direct) = drain; 1395 } 1396 CC_UNLOCK(cc); 1397 return ((flags & CS_EXECUTING) != 0); 1398 } 1399 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1400 c, c->c_func, c->c_arg); 1401 if (drain) { 1402 cc_exec_drain(cc, direct) = drain; 1403 } 1404 KASSERT(!sq_locked, ("sleepqueue chain still locked")); 1405 cancelled = ((flags & CS_EXECUTING) != 0); 1406 } else 1407 cancelled = 1; 1408 1409 if (sq_locked) 1410 sleepq_release(&cc_exec_waiting(cc, direct)); 1411 1412 if ((c->c_iflags & CALLOUT_PENDING) == 0) { 1413 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1414 c, c->c_func, c->c_arg); 1415 /* 1416 * For not scheduled and not executing callout return 1417 * negative value. 1418 */ 1419 if (cc_exec_curr(cc, direct) != c) 1420 cancelled = -1; 1421 CC_UNLOCK(cc); 1422 return (cancelled); 1423 } 1424 1425 c->c_iflags &= ~CALLOUT_PENDING; 1426 c->c_flags &= ~CALLOUT_ACTIVE; 1427 1428 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1429 c, c->c_func, c->c_arg); 1430 if (not_on_a_list == 0) { 1431 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1432 if (cc_exec_next(cc) == c) 1433 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1434 LIST_REMOVE(c, c_links.le); 1435 } else { 1436 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1437 } 1438 } 1439 callout_cc_del(c, cc); 1440 CC_UNLOCK(cc); 1441 return (cancelled); 1442 } 1443 1444 void 1445 callout_init(struct callout *c, int mpsafe) 1446 { 1447 bzero(c, sizeof *c); 1448 if (mpsafe) { 1449 c->c_lock = NULL; 1450 c->c_iflags = CALLOUT_RETURNUNLOCKED; 1451 } else { 1452 c->c_lock = &Giant.lock_object; 1453 c->c_iflags = 0; 1454 } 1455 c->c_cpu = timeout_cpu; 1456 } 1457 1458 void 1459 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags) 1460 { 1461 bzero(c, sizeof *c); 1462 c->c_lock = lock; 1463 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0, 1464 ("callout_init_lock: bad flags %d", flags)); 1465 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0, 1466 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock")); 1467 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & 1468 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class", 1469 __func__)); 1470 c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK); 1471 c->c_cpu = timeout_cpu; 1472 } 1473 1474 #ifdef APM_FIXUP_CALLTODO 1475 /* 1476 * Adjust the kernel calltodo timeout list. This routine is used after 1477 * an APM resume to recalculate the calltodo timer list values with the 1478 * number of hz's we have been sleeping. The next hardclock() will detect 1479 * that there are fired timers and run softclock() to execute them. 1480 * 1481 * Please note, I have not done an exhaustive analysis of what code this 1482 * might break. I am motivated to have my select()'s and alarm()'s that 1483 * have expired during suspend firing upon resume so that the applications 1484 * which set the timer can do the maintanence the timer was for as close 1485 * as possible to the originally intended time. Testing this code for a 1486 * week showed that resuming from a suspend resulted in 22 to 25 timers 1487 * firing, which seemed independent on whether the suspend was 2 hours or 1488 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu> 1489 */ 1490 void 1491 adjust_timeout_calltodo(struct timeval *time_change) 1492 { 1493 struct callout *p; 1494 unsigned long delta_ticks; 1495 1496 /* 1497 * How many ticks were we asleep? 1498 * (stolen from tvtohz()). 1499 */ 1500 1501 /* Don't do anything */ 1502 if (time_change->tv_sec < 0) 1503 return; 1504 else if (time_change->tv_sec <= LONG_MAX / 1000000) 1505 delta_ticks = howmany(time_change->tv_sec * 1000000 + 1506 time_change->tv_usec, tick) + 1; 1507 else if (time_change->tv_sec <= LONG_MAX / hz) 1508 delta_ticks = time_change->tv_sec * hz + 1509 howmany(time_change->tv_usec, tick) + 1; 1510 else 1511 delta_ticks = LONG_MAX; 1512 1513 if (delta_ticks > INT_MAX) 1514 delta_ticks = INT_MAX; 1515 1516 /* 1517 * Now rip through the timer calltodo list looking for timers 1518 * to expire. 1519 */ 1520 1521 /* don't collide with softclock() */ 1522 CC_LOCK(cc); 1523 for (p = calltodo.c_next; p != NULL; p = p->c_next) { 1524 p->c_time -= delta_ticks; 1525 1526 /* Break if the timer had more time on it than delta_ticks */ 1527 if (p->c_time > 0) 1528 break; 1529 1530 /* take back the ticks the timer didn't use (p->c_time <= 0) */ 1531 delta_ticks = -p->c_time; 1532 } 1533 CC_UNLOCK(cc); 1534 1535 return; 1536 } 1537 #endif /* APM_FIXUP_CALLTODO */ 1538 1539 static int 1540 flssbt(sbintime_t sbt) 1541 { 1542 1543 sbt += (uint64_t)sbt >> 1; 1544 if (sizeof(long) >= sizeof(sbintime_t)) 1545 return (flsl(sbt)); 1546 if (sbt >= SBT_1S) 1547 return (flsl(((uint64_t)sbt) >> 32) + 32); 1548 return (flsl(sbt)); 1549 } 1550 1551 /* 1552 * Dump immediate statistic snapshot of the scheduled callouts. 1553 */ 1554 static int 1555 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS) 1556 { 1557 struct callout *tmp; 1558 struct callout_cpu *cc; 1559 struct callout_list *sc; 1560 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t; 1561 int ct[64], cpr[64], ccpbk[32]; 1562 int error, val, i, count, tcum, pcum, maxc, c, medc; 1563 #ifdef SMP 1564 int cpu; 1565 #endif 1566 1567 val = 0; 1568 error = sysctl_handle_int(oidp, &val, 0, req); 1569 if (error != 0 || req->newptr == NULL) 1570 return (error); 1571 count = maxc = 0; 1572 st = spr = maxt = maxpr = 0; 1573 bzero(ccpbk, sizeof(ccpbk)); 1574 bzero(ct, sizeof(ct)); 1575 bzero(cpr, sizeof(cpr)); 1576 now = sbinuptime(); 1577 #ifdef SMP 1578 CPU_FOREACH(cpu) { 1579 cc = CC_CPU(cpu); 1580 #else 1581 cc = CC_CPU(timeout_cpu); 1582 #endif 1583 CC_LOCK(cc); 1584 for (i = 0; i < callwheelsize; i++) { 1585 sc = &cc->cc_callwheel[i]; 1586 c = 0; 1587 LIST_FOREACH(tmp, sc, c_links.le) { 1588 c++; 1589 t = tmp->c_time - now; 1590 if (t < 0) 1591 t = 0; 1592 st += t / SBT_1US; 1593 spr += tmp->c_precision / SBT_1US; 1594 if (t > maxt) 1595 maxt = t; 1596 if (tmp->c_precision > maxpr) 1597 maxpr = tmp->c_precision; 1598 ct[flssbt(t)]++; 1599 cpr[flssbt(tmp->c_precision)]++; 1600 } 1601 if (c > maxc) 1602 maxc = c; 1603 ccpbk[fls(c + c / 2)]++; 1604 count += c; 1605 } 1606 CC_UNLOCK(cc); 1607 #ifdef SMP 1608 } 1609 #endif 1610 1611 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++) 1612 tcum += ct[i]; 1613 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1614 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++) 1615 pcum += cpr[i]; 1616 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1617 for (i = 0, c = 0; i < 32 && c < count / 2; i++) 1618 c += ccpbk[i]; 1619 medc = (i >= 2) ? (1 << (i - 2)) : 0; 1620 1621 printf("Scheduled callouts statistic snapshot:\n"); 1622 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n", 1623 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT); 1624 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n", 1625 medc, 1626 count / callwheelsize / mp_ncpus, 1627 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000, 1628 maxc); 1629 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1630 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32, 1631 (st / count) / 1000000, (st / count) % 1000000, 1632 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32); 1633 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1634 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32, 1635 (spr / count) / 1000000, (spr / count) % 1000000, 1636 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32); 1637 printf(" Distribution: \tbuckets\t time\t tcum\t" 1638 " prec\t pcum\n"); 1639 for (i = 0, tcum = pcum = 0; i < 64; i++) { 1640 if (ct[i] == 0 && cpr[i] == 0) 1641 continue; 1642 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0; 1643 tcum += ct[i]; 1644 pcum += cpr[i]; 1645 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n", 1646 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32, 1647 i - 1 - (32 - CC_HASH_SHIFT), 1648 ct[i], tcum, cpr[i], pcum); 1649 } 1650 return (error); 1651 } 1652 SYSCTL_PROC(_kern, OID_AUTO, callout_stat, 1653 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1654 0, 0, sysctl_kern_callout_stat, "I", 1655 "Dump immediate statistic snapshot of the scheduled callouts"); 1656 1657 #ifdef DDB 1658 static void 1659 _show_callout(struct callout *c) 1660 { 1661 1662 db_printf("callout %p\n", c); 1663 #define C_DB_PRINTF(f, e) db_printf(" %s = " f "\n", #e, c->e); 1664 db_printf(" &c_links = %p\n", &(c->c_links)); 1665 C_DB_PRINTF("%" PRId64, c_time); 1666 C_DB_PRINTF("%" PRId64, c_precision); 1667 C_DB_PRINTF("%p", c_arg); 1668 C_DB_PRINTF("%p", c_func); 1669 C_DB_PRINTF("%p", c_lock); 1670 C_DB_PRINTF("%#x", c_flags); 1671 C_DB_PRINTF("%#x", c_iflags); 1672 C_DB_PRINTF("%d", c_cpu); 1673 #undef C_DB_PRINTF 1674 } 1675 1676 DB_SHOW_COMMAND(callout, db_show_callout) 1677 { 1678 1679 if (!have_addr) { 1680 db_printf("usage: show callout <struct callout *>\n"); 1681 return; 1682 } 1683 1684 _show_callout((struct callout *)addr); 1685 } 1686 1687 static void 1688 _show_last_callout(int cpu, int direct, const char *dirstr) 1689 { 1690 struct callout_cpu *cc; 1691 void *func, *arg; 1692 1693 cc = CC_CPU(cpu); 1694 func = cc_exec_last_func(cc, direct); 1695 arg = cc_exec_last_arg(cc, direct); 1696 db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func); 1697 db_printsym((db_expr_t)func, DB_STGY_ANY); 1698 db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg); 1699 } 1700 1701 DB_SHOW_COMMAND(callout_last, db_show_callout_last) 1702 { 1703 int cpu, last; 1704 1705 if (have_addr) { 1706 if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) { 1707 db_printf("no such cpu: %d\n", (int)addr); 1708 return; 1709 } 1710 cpu = last = addr; 1711 } else { 1712 cpu = 0; 1713 last = mp_maxid; 1714 } 1715 1716 while (cpu <= last) { 1717 if (!CPU_ABSENT(cpu)) { 1718 _show_last_callout(cpu, 0, ""); 1719 _show_last_callout(cpu, 1, " direct"); 1720 } 1721 cpu++; 1722 } 1723 } 1724 #endif /* DDB */ 1725