1 /*- 2 * Copyright (c) 1982, 1986, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 35 */ 36 37 #include <sys/cdefs.h> 38 __FBSDID("$FreeBSD$"); 39 40 #include "opt_callout_profiling.h" 41 #if defined(__arm__) 42 #include "opt_timer.h" 43 #endif 44 #include "opt_rss.h" 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/bus.h> 49 #include <sys/callout.h> 50 #include <sys/file.h> 51 #include <sys/interrupt.h> 52 #include <sys/kernel.h> 53 #include <sys/ktr.h> 54 #include <sys/lock.h> 55 #include <sys/malloc.h> 56 #include <sys/mutex.h> 57 #include <sys/proc.h> 58 #include <sys/sdt.h> 59 #include <sys/sleepqueue.h> 60 #include <sys/sysctl.h> 61 #include <sys/smp.h> 62 63 #ifdef SMP 64 #include <machine/cpu.h> 65 #endif 66 67 #ifndef NO_EVENTTIMERS 68 DPCPU_DECLARE(sbintime_t, hardclocktime); 69 #endif 70 71 SDT_PROVIDER_DEFINE(callout_execute); 72 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *"); 73 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *"); 74 75 #ifdef CALLOUT_PROFILING 76 static int avg_depth; 77 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0, 78 "Average number of items examined per softclock call. Units = 1/1000"); 79 static int avg_gcalls; 80 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0, 81 "Average number of Giant callouts made per softclock call. Units = 1/1000"); 82 static int avg_lockcalls; 83 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0, 84 "Average number of lock callouts made per softclock call. Units = 1/1000"); 85 static int avg_mpcalls; 86 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0, 87 "Average number of MP callouts made per softclock call. Units = 1/1000"); 88 static int avg_depth_dir; 89 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0, 90 "Average number of direct callouts examined per callout_process call. " 91 "Units = 1/1000"); 92 static int avg_lockcalls_dir; 93 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD, 94 &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per " 95 "callout_process call. Units = 1/1000"); 96 static int avg_mpcalls_dir; 97 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir, 98 0, "Average number of MP direct callouts made per callout_process call. " 99 "Units = 1/1000"); 100 #endif 101 102 static int ncallout; 103 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0, 104 "Number of entries in callwheel and size of timeout() preallocation"); 105 106 #ifdef RSS 107 static int pin_default_swi = 1; 108 static int pin_pcpu_swi = 1; 109 #else 110 static int pin_default_swi = 0; 111 static int pin_pcpu_swi = 0; 112 #endif 113 114 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi, 115 0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)"); 116 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi, 117 0, "Pin the per-CPU swis (except PCPU 0, which is also default"); 118 119 /* 120 * TODO: 121 * allocate more timeout table slots when table overflows. 122 */ 123 u_int callwheelsize, callwheelmask; 124 125 /* 126 * The callout cpu exec entities represent informations necessary for 127 * describing the state of callouts currently running on the CPU and the ones 128 * necessary for migrating callouts to the new callout cpu. In particular, 129 * the first entry of the array cc_exec_entity holds informations for callout 130 * running in SWI thread context, while the second one holds informations 131 * for callout running directly from hardware interrupt context. 132 * The cached informations are very important for deferring migration when 133 * the migrating callout is already running. 134 */ 135 struct cc_exec { 136 struct callout *cc_curr; 137 void (*cc_drain)(void *); 138 #ifdef SMP 139 void (*ce_migration_func)(void *); 140 void *ce_migration_arg; 141 int ce_migration_cpu; 142 sbintime_t ce_migration_time; 143 sbintime_t ce_migration_prec; 144 #endif 145 bool cc_cancel; 146 bool cc_waiting; 147 }; 148 149 /* 150 * There is one struct callout_cpu per cpu, holding all relevant 151 * state for the callout processing thread on the individual CPU. 152 */ 153 struct callout_cpu { 154 struct mtx_padalign cc_lock; 155 struct cc_exec cc_exec_entity[2]; 156 struct callout *cc_next; 157 struct callout *cc_callout; 158 struct callout_list *cc_callwheel; 159 struct callout_tailq cc_expireq; 160 struct callout_slist cc_callfree; 161 sbintime_t cc_firstevent; 162 sbintime_t cc_lastscan; 163 void *cc_cookie; 164 u_int cc_bucket; 165 u_int cc_inited; 166 char cc_ktr_event_name[20]; 167 }; 168 169 #define callout_migrating(c) ((c)->c_iflags & CALLOUT_DFRMIGRATION) 170 171 #define cc_exec_curr(cc, dir) cc->cc_exec_entity[dir].cc_curr 172 #define cc_exec_drain(cc, dir) cc->cc_exec_entity[dir].cc_drain 173 #define cc_exec_next(cc) cc->cc_next 174 #define cc_exec_cancel(cc, dir) cc->cc_exec_entity[dir].cc_cancel 175 #define cc_exec_waiting(cc, dir) cc->cc_exec_entity[dir].cc_waiting 176 #ifdef SMP 177 #define cc_migration_func(cc, dir) cc->cc_exec_entity[dir].ce_migration_func 178 #define cc_migration_arg(cc, dir) cc->cc_exec_entity[dir].ce_migration_arg 179 #define cc_migration_cpu(cc, dir) cc->cc_exec_entity[dir].ce_migration_cpu 180 #define cc_migration_time(cc, dir) cc->cc_exec_entity[dir].ce_migration_time 181 #define cc_migration_prec(cc, dir) cc->cc_exec_entity[dir].ce_migration_prec 182 183 struct callout_cpu cc_cpu[MAXCPU]; 184 #define CPUBLOCK MAXCPU 185 #define CC_CPU(cpu) (&cc_cpu[(cpu)]) 186 #define CC_SELF() CC_CPU(PCPU_GET(cpuid)) 187 #else 188 struct callout_cpu cc_cpu; 189 #define CC_CPU(cpu) &cc_cpu 190 #define CC_SELF() &cc_cpu 191 #endif 192 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock) 193 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock) 194 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED) 195 196 static int timeout_cpu; 197 198 static void callout_cpu_init(struct callout_cpu *cc, int cpu); 199 static void softclock_call_cc(struct callout *c, struct callout_cpu *cc, 200 #ifdef CALLOUT_PROFILING 201 int *mpcalls, int *lockcalls, int *gcalls, 202 #endif 203 int direct); 204 205 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures"); 206 207 /** 208 * Locked by cc_lock: 209 * cc_curr - If a callout is in progress, it is cc_curr. 210 * If cc_curr is non-NULL, threads waiting in 211 * callout_drain() will be woken up as soon as the 212 * relevant callout completes. 213 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held 214 * guarantees that the current callout will not run. 215 * The softclock() function sets this to 0 before it 216 * drops callout_lock to acquire c_lock, and it calls 217 * the handler only if curr_cancelled is still 0 after 218 * cc_lock is successfully acquired. 219 * cc_waiting - If a thread is waiting in callout_drain(), then 220 * callout_wait is nonzero. Set only when 221 * cc_curr is non-NULL. 222 */ 223 224 /* 225 * Resets the execution entity tied to a specific callout cpu. 226 */ 227 static void 228 cc_cce_cleanup(struct callout_cpu *cc, int direct) 229 { 230 231 cc_exec_curr(cc, direct) = NULL; 232 cc_exec_cancel(cc, direct) = false; 233 cc_exec_waiting(cc, direct) = false; 234 #ifdef SMP 235 cc_migration_cpu(cc, direct) = CPUBLOCK; 236 cc_migration_time(cc, direct) = 0; 237 cc_migration_prec(cc, direct) = 0; 238 cc_migration_func(cc, direct) = NULL; 239 cc_migration_arg(cc, direct) = NULL; 240 #endif 241 } 242 243 /* 244 * Checks if migration is requested by a specific callout cpu. 245 */ 246 static int 247 cc_cce_migrating(struct callout_cpu *cc, int direct) 248 { 249 250 #ifdef SMP 251 return (cc_migration_cpu(cc, direct) != CPUBLOCK); 252 #else 253 return (0); 254 #endif 255 } 256 257 /* 258 * Kernel low level callwheel initialization 259 * called on cpu0 during kernel startup. 260 */ 261 static void 262 callout_callwheel_init(void *dummy) 263 { 264 struct callout_cpu *cc; 265 266 /* 267 * Calculate the size of the callout wheel and the preallocated 268 * timeout() structures. 269 * XXX: Clip callout to result of previous function of maxusers 270 * maximum 384. This is still huge, but acceptable. 271 */ 272 memset(CC_CPU(0), 0, sizeof(cc_cpu)); 273 ncallout = imin(16 + maxproc + maxfiles, 18508); 274 TUNABLE_INT_FETCH("kern.ncallout", &ncallout); 275 276 /* 277 * Calculate callout wheel size, should be next power of two higher 278 * than 'ncallout'. 279 */ 280 callwheelsize = 1 << fls(ncallout); 281 callwheelmask = callwheelsize - 1; 282 283 /* 284 * Fetch whether we're pinning the swi's or not. 285 */ 286 TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi); 287 TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi); 288 289 /* 290 * Only cpu0 handles timeout(9) and receives a preallocation. 291 * 292 * XXX: Once all timeout(9) consumers are converted this can 293 * be removed. 294 */ 295 timeout_cpu = PCPU_GET(cpuid); 296 cc = CC_CPU(timeout_cpu); 297 cc->cc_callout = malloc(ncallout * sizeof(struct callout), 298 M_CALLOUT, M_WAITOK); 299 callout_cpu_init(cc, timeout_cpu); 300 } 301 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL); 302 303 /* 304 * Initialize the per-cpu callout structures. 305 */ 306 static void 307 callout_cpu_init(struct callout_cpu *cc, int cpu) 308 { 309 struct callout *c; 310 int i; 311 312 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE); 313 SLIST_INIT(&cc->cc_callfree); 314 cc->cc_inited = 1; 315 cc->cc_callwheel = malloc(sizeof(struct callout_list) * callwheelsize, 316 M_CALLOUT, M_WAITOK); 317 for (i = 0; i < callwheelsize; i++) 318 LIST_INIT(&cc->cc_callwheel[i]); 319 TAILQ_INIT(&cc->cc_expireq); 320 cc->cc_firstevent = SBT_MAX; 321 for (i = 0; i < 2; i++) 322 cc_cce_cleanup(cc, i); 323 snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name), 324 "callwheel cpu %d", cpu); 325 if (cc->cc_callout == NULL) /* Only cpu0 handles timeout(9) */ 326 return; 327 for (i = 0; i < ncallout; i++) { 328 c = &cc->cc_callout[i]; 329 callout_init(c, 0); 330 c->c_iflags = CALLOUT_LOCAL_ALLOC; 331 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 332 } 333 } 334 335 #ifdef SMP 336 /* 337 * Switches the cpu tied to a specific callout. 338 * The function expects a locked incoming callout cpu and returns with 339 * locked outcoming callout cpu. 340 */ 341 static struct callout_cpu * 342 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu) 343 { 344 struct callout_cpu *new_cc; 345 346 MPASS(c != NULL && cc != NULL); 347 CC_LOCK_ASSERT(cc); 348 349 /* 350 * Avoid interrupts and preemption firing after the callout cpu 351 * is blocked in order to avoid deadlocks as the new thread 352 * may be willing to acquire the callout cpu lock. 353 */ 354 c->c_cpu = CPUBLOCK; 355 spinlock_enter(); 356 CC_UNLOCK(cc); 357 new_cc = CC_CPU(new_cpu); 358 CC_LOCK(new_cc); 359 spinlock_exit(); 360 c->c_cpu = new_cpu; 361 return (new_cc); 362 } 363 #endif 364 365 /* 366 * Start standard softclock thread. 367 */ 368 static void 369 start_softclock(void *dummy) 370 { 371 struct callout_cpu *cc; 372 char name[MAXCOMLEN]; 373 #ifdef SMP 374 int cpu; 375 struct intr_event *ie; 376 #endif 377 378 cc = CC_CPU(timeout_cpu); 379 snprintf(name, sizeof(name), "clock (%d)", timeout_cpu); 380 if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK, 381 INTR_MPSAFE, &cc->cc_cookie)) 382 panic("died while creating standard software ithreads"); 383 if (pin_default_swi && 384 (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) { 385 printf("%s: timeout clock couldn't be pinned to cpu %d\n", 386 __func__, 387 timeout_cpu); 388 } 389 390 #ifdef SMP 391 CPU_FOREACH(cpu) { 392 if (cpu == timeout_cpu) 393 continue; 394 cc = CC_CPU(cpu); 395 cc->cc_callout = NULL; /* Only cpu0 handles timeout(9). */ 396 callout_cpu_init(cc, cpu); 397 snprintf(name, sizeof(name), "clock (%d)", cpu); 398 ie = NULL; 399 if (swi_add(&ie, name, softclock, cc, SWI_CLOCK, 400 INTR_MPSAFE, &cc->cc_cookie)) 401 panic("died while creating standard software ithreads"); 402 if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) { 403 printf("%s: per-cpu clock couldn't be pinned to " 404 "cpu %d\n", 405 __func__, 406 cpu); 407 } 408 } 409 #endif 410 } 411 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL); 412 413 #define CC_HASH_SHIFT 8 414 415 static inline u_int 416 callout_hash(sbintime_t sbt) 417 { 418 419 return (sbt >> (32 - CC_HASH_SHIFT)); 420 } 421 422 static inline u_int 423 callout_get_bucket(sbintime_t sbt) 424 { 425 426 return (callout_hash(sbt) & callwheelmask); 427 } 428 429 void 430 callout_process(sbintime_t now) 431 { 432 struct callout *tmp, *tmpn; 433 struct callout_cpu *cc; 434 struct callout_list *sc; 435 sbintime_t first, last, max, tmp_max; 436 uint32_t lookahead; 437 u_int firstb, lastb, nowb; 438 #ifdef CALLOUT_PROFILING 439 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0; 440 #endif 441 442 cc = CC_SELF(); 443 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET); 444 445 /* Compute the buckets of the last scan and present times. */ 446 firstb = callout_hash(cc->cc_lastscan); 447 cc->cc_lastscan = now; 448 nowb = callout_hash(now); 449 450 /* Compute the last bucket and minimum time of the bucket after it. */ 451 if (nowb == firstb) 452 lookahead = (SBT_1S / 16); 453 else if (nowb - firstb == 1) 454 lookahead = (SBT_1S / 8); 455 else 456 lookahead = (SBT_1S / 2); 457 first = last = now; 458 first += (lookahead / 2); 459 last += lookahead; 460 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT)); 461 lastb = callout_hash(last) - 1; 462 max = last; 463 464 /* 465 * Check if we wrapped around the entire wheel from the last scan. 466 * In case, we need to scan entirely the wheel for pending callouts. 467 */ 468 if (lastb - firstb >= callwheelsize) { 469 lastb = firstb + callwheelsize - 1; 470 if (nowb - firstb >= callwheelsize) 471 nowb = lastb; 472 } 473 474 /* Iterate callwheel from firstb to nowb and then up to lastb. */ 475 do { 476 sc = &cc->cc_callwheel[firstb & callwheelmask]; 477 tmp = LIST_FIRST(sc); 478 while (tmp != NULL) { 479 /* Run the callout if present time within allowed. */ 480 if (tmp->c_time <= now) { 481 /* 482 * Consumer told us the callout may be run 483 * directly from hardware interrupt context. 484 */ 485 if (tmp->c_iflags & CALLOUT_DIRECT) { 486 #ifdef CALLOUT_PROFILING 487 ++depth_dir; 488 #endif 489 cc_exec_next(cc) = 490 LIST_NEXT(tmp, c_links.le); 491 cc->cc_bucket = firstb & callwheelmask; 492 LIST_REMOVE(tmp, c_links.le); 493 softclock_call_cc(tmp, cc, 494 #ifdef CALLOUT_PROFILING 495 &mpcalls_dir, &lockcalls_dir, NULL, 496 #endif 497 1); 498 tmp = cc_exec_next(cc); 499 cc_exec_next(cc) = NULL; 500 } else { 501 tmpn = LIST_NEXT(tmp, c_links.le); 502 LIST_REMOVE(tmp, c_links.le); 503 TAILQ_INSERT_TAIL(&cc->cc_expireq, 504 tmp, c_links.tqe); 505 tmp->c_iflags |= CALLOUT_PROCESSED; 506 tmp = tmpn; 507 } 508 continue; 509 } 510 /* Skip events from distant future. */ 511 if (tmp->c_time >= max) 512 goto next; 513 /* 514 * Event minimal time is bigger than present maximal 515 * time, so it cannot be aggregated. 516 */ 517 if (tmp->c_time > last) { 518 lastb = nowb; 519 goto next; 520 } 521 /* Update first and last time, respecting this event. */ 522 if (tmp->c_time < first) 523 first = tmp->c_time; 524 tmp_max = tmp->c_time + tmp->c_precision; 525 if (tmp_max < last) 526 last = tmp_max; 527 next: 528 tmp = LIST_NEXT(tmp, c_links.le); 529 } 530 /* Proceed with the next bucket. */ 531 firstb++; 532 /* 533 * Stop if we looked after present time and found 534 * some event we can't execute at now. 535 * Stop if we looked far enough into the future. 536 */ 537 } while (((int)(firstb - lastb)) <= 0); 538 cc->cc_firstevent = last; 539 #ifndef NO_EVENTTIMERS 540 cpu_new_callout(curcpu, last, first); 541 #endif 542 #ifdef CALLOUT_PROFILING 543 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8; 544 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8; 545 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8; 546 #endif 547 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET); 548 /* 549 * swi_sched acquires the thread lock, so we don't want to call it 550 * with cc_lock held; incorrect locking order. 551 */ 552 if (!TAILQ_EMPTY(&cc->cc_expireq)) 553 swi_sched(cc->cc_cookie, 0); 554 } 555 556 static struct callout_cpu * 557 callout_lock(struct callout *c) 558 { 559 struct callout_cpu *cc; 560 int cpu; 561 562 for (;;) { 563 cpu = c->c_cpu; 564 #ifdef SMP 565 if (cpu == CPUBLOCK) { 566 while (c->c_cpu == CPUBLOCK) 567 cpu_spinwait(); 568 continue; 569 } 570 #endif 571 cc = CC_CPU(cpu); 572 CC_LOCK(cc); 573 if (cpu == c->c_cpu) 574 break; 575 CC_UNLOCK(cc); 576 } 577 return (cc); 578 } 579 580 static void 581 callout_cc_add(struct callout *c, struct callout_cpu *cc, 582 sbintime_t sbt, sbintime_t precision, void (*func)(void *), 583 void *arg, int cpu, int flags) 584 { 585 int bucket; 586 587 CC_LOCK_ASSERT(cc); 588 if (sbt < cc->cc_lastscan) 589 sbt = cc->cc_lastscan; 590 c->c_arg = arg; 591 c->c_iflags |= CALLOUT_PENDING; 592 c->c_iflags &= ~CALLOUT_PROCESSED; 593 c->c_flags |= CALLOUT_ACTIVE; 594 if (flags & C_DIRECT_EXEC) 595 c->c_iflags |= CALLOUT_DIRECT; 596 c->c_func = func; 597 c->c_time = sbt; 598 c->c_precision = precision; 599 bucket = callout_get_bucket(c->c_time); 600 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x", 601 c, (int)(c->c_precision >> 32), 602 (u_int)(c->c_precision & 0xffffffff)); 603 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le); 604 if (cc->cc_bucket == bucket) 605 cc_exec_next(cc) = c; 606 #ifndef NO_EVENTTIMERS 607 /* 608 * Inform the eventtimers(4) subsystem there's a new callout 609 * that has been inserted, but only if really required. 610 */ 611 if (SBT_MAX - c->c_time < c->c_precision) 612 c->c_precision = SBT_MAX - c->c_time; 613 sbt = c->c_time + c->c_precision; 614 if (sbt < cc->cc_firstevent) { 615 cc->cc_firstevent = sbt; 616 cpu_new_callout(cpu, sbt, c->c_time); 617 } 618 #endif 619 } 620 621 static void 622 callout_cc_del(struct callout *c, struct callout_cpu *cc) 623 { 624 625 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0) 626 return; 627 c->c_func = NULL; 628 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 629 } 630 631 static void 632 softclock_call_cc(struct callout *c, struct callout_cpu *cc, 633 #ifdef CALLOUT_PROFILING 634 int *mpcalls, int *lockcalls, int *gcalls, 635 #endif 636 int direct) 637 { 638 struct rm_priotracker tracker; 639 void (*c_func)(void *); 640 void *c_arg; 641 struct lock_class *class; 642 struct lock_object *c_lock; 643 uintptr_t lock_status; 644 int c_iflags; 645 #ifdef SMP 646 struct callout_cpu *new_cc; 647 void (*new_func)(void *); 648 void *new_arg; 649 int flags, new_cpu; 650 sbintime_t new_prec, new_time; 651 #endif 652 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 653 sbintime_t sbt1, sbt2; 654 struct timespec ts2; 655 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */ 656 static timeout_t *lastfunc; 657 #endif 658 659 KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING, 660 ("softclock_call_cc: pend %p %x", c, c->c_iflags)); 661 KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE, 662 ("softclock_call_cc: act %p %x", c, c->c_flags)); 663 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL; 664 lock_status = 0; 665 if (c->c_flags & CALLOUT_SHAREDLOCK) { 666 if (class == &lock_class_rm) 667 lock_status = (uintptr_t)&tracker; 668 else 669 lock_status = 1; 670 } 671 c_lock = c->c_lock; 672 c_func = c->c_func; 673 c_arg = c->c_arg; 674 c_iflags = c->c_iflags; 675 if (c->c_iflags & CALLOUT_LOCAL_ALLOC) 676 c->c_iflags = CALLOUT_LOCAL_ALLOC; 677 else 678 c->c_iflags &= ~CALLOUT_PENDING; 679 680 cc_exec_curr(cc, direct) = c; 681 cc_exec_cancel(cc, direct) = false; 682 cc_exec_drain(cc, direct) = NULL; 683 CC_UNLOCK(cc); 684 if (c_lock != NULL) { 685 class->lc_lock(c_lock, lock_status); 686 /* 687 * The callout may have been cancelled 688 * while we switched locks. 689 */ 690 if (cc_exec_cancel(cc, direct)) { 691 class->lc_unlock(c_lock); 692 goto skip; 693 } 694 /* The callout cannot be stopped now. */ 695 cc_exec_cancel(cc, direct) = true; 696 if (c_lock == &Giant.lock_object) { 697 #ifdef CALLOUT_PROFILING 698 (*gcalls)++; 699 #endif 700 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p", 701 c, c_func, c_arg); 702 } else { 703 #ifdef CALLOUT_PROFILING 704 (*lockcalls)++; 705 #endif 706 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p", 707 c, c_func, c_arg); 708 } 709 } else { 710 #ifdef CALLOUT_PROFILING 711 (*mpcalls)++; 712 #endif 713 CTR3(KTR_CALLOUT, "callout %p func %p arg %p", 714 c, c_func, c_arg); 715 } 716 KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running", 717 "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct); 718 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 719 sbt1 = sbinuptime(); 720 #endif 721 THREAD_NO_SLEEPING(); 722 SDT_PROBE1(callout_execute, , , callout__start, c); 723 c_func(c_arg); 724 SDT_PROBE1(callout_execute, , , callout__end, c); 725 THREAD_SLEEPING_OK(); 726 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 727 sbt2 = sbinuptime(); 728 sbt2 -= sbt1; 729 if (sbt2 > maxdt) { 730 if (lastfunc != c_func || sbt2 > maxdt * 2) { 731 ts2 = sbttots(sbt2); 732 printf( 733 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n", 734 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec); 735 } 736 maxdt = sbt2; 737 lastfunc = c_func; 738 } 739 #endif 740 KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle"); 741 CTR1(KTR_CALLOUT, "callout %p finished", c); 742 if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0) 743 class->lc_unlock(c_lock); 744 skip: 745 CC_LOCK(cc); 746 KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr")); 747 cc_exec_curr(cc, direct) = NULL; 748 if (cc_exec_drain(cc, direct)) { 749 void (*drain)(void *); 750 751 drain = cc_exec_drain(cc, direct); 752 cc_exec_drain(cc, direct) = NULL; 753 CC_UNLOCK(cc); 754 drain(c_arg); 755 CC_LOCK(cc); 756 } 757 if (cc_exec_waiting(cc, direct)) { 758 /* 759 * There is someone waiting for the 760 * callout to complete. 761 * If the callout was scheduled for 762 * migration just cancel it. 763 */ 764 if (cc_cce_migrating(cc, direct)) { 765 cc_cce_cleanup(cc, direct); 766 767 /* 768 * It should be assert here that the callout is not 769 * destroyed but that is not easy. 770 */ 771 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 772 } 773 cc_exec_waiting(cc, direct) = false; 774 CC_UNLOCK(cc); 775 wakeup(&cc_exec_waiting(cc, direct)); 776 CC_LOCK(cc); 777 } else if (cc_cce_migrating(cc, direct)) { 778 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0, 779 ("Migrating legacy callout %p", c)); 780 #ifdef SMP 781 /* 782 * If the callout was scheduled for 783 * migration just perform it now. 784 */ 785 new_cpu = cc_migration_cpu(cc, direct); 786 new_time = cc_migration_time(cc, direct); 787 new_prec = cc_migration_prec(cc, direct); 788 new_func = cc_migration_func(cc, direct); 789 new_arg = cc_migration_arg(cc, direct); 790 cc_cce_cleanup(cc, direct); 791 792 /* 793 * It should be assert here that the callout is not destroyed 794 * but that is not easy. 795 * 796 * As first thing, handle deferred callout stops. 797 */ 798 if (!callout_migrating(c)) { 799 CTR3(KTR_CALLOUT, 800 "deferred cancelled %p func %p arg %p", 801 c, new_func, new_arg); 802 callout_cc_del(c, cc); 803 return; 804 } 805 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 806 807 new_cc = callout_cpu_switch(c, cc, new_cpu); 808 flags = (direct) ? C_DIRECT_EXEC : 0; 809 callout_cc_add(c, new_cc, new_time, new_prec, new_func, 810 new_arg, new_cpu, flags); 811 CC_UNLOCK(new_cc); 812 CC_LOCK(cc); 813 #else 814 panic("migration should not happen"); 815 #endif 816 } 817 /* 818 * If the current callout is locally allocated (from 819 * timeout(9)) then put it on the freelist. 820 * 821 * Note: we need to check the cached copy of c_iflags because 822 * if it was not local, then it's not safe to deref the 823 * callout pointer. 824 */ 825 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 || 826 c->c_iflags == CALLOUT_LOCAL_ALLOC, 827 ("corrupted callout")); 828 if (c_iflags & CALLOUT_LOCAL_ALLOC) 829 callout_cc_del(c, cc); 830 } 831 832 /* 833 * The callout mechanism is based on the work of Adam M. Costello and 834 * George Varghese, published in a technical report entitled "Redesigning 835 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 836 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 837 * used in this implementation was published by G. Varghese and T. Lauck in 838 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 839 * the Efficient Implementation of a Timer Facility" in the Proceedings of 840 * the 11th ACM Annual Symposium on Operating Systems Principles, 841 * Austin, Texas Nov 1987. 842 */ 843 844 /* 845 * Software (low priority) clock interrupt. 846 * Run periodic events from timeout queue. 847 */ 848 void 849 softclock(void *arg) 850 { 851 struct callout_cpu *cc; 852 struct callout *c; 853 #ifdef CALLOUT_PROFILING 854 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0; 855 #endif 856 857 cc = (struct callout_cpu *)arg; 858 CC_LOCK(cc); 859 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) { 860 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 861 softclock_call_cc(c, cc, 862 #ifdef CALLOUT_PROFILING 863 &mpcalls, &lockcalls, &gcalls, 864 #endif 865 0); 866 #ifdef CALLOUT_PROFILING 867 ++depth; 868 #endif 869 } 870 #ifdef CALLOUT_PROFILING 871 avg_depth += (depth * 1000 - avg_depth) >> 8; 872 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8; 873 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8; 874 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8; 875 #endif 876 CC_UNLOCK(cc); 877 } 878 879 /* 880 * timeout -- 881 * Execute a function after a specified length of time. 882 * 883 * untimeout -- 884 * Cancel previous timeout function call. 885 * 886 * callout_handle_init -- 887 * Initialize a handle so that using it with untimeout is benign. 888 * 889 * See AT&T BCI Driver Reference Manual for specification. This 890 * implementation differs from that one in that although an 891 * identification value is returned from timeout, the original 892 * arguments to timeout as well as the identifier are used to 893 * identify entries for untimeout. 894 */ 895 struct callout_handle 896 timeout(timeout_t *ftn, void *arg, int to_ticks) 897 { 898 struct callout_cpu *cc; 899 struct callout *new; 900 struct callout_handle handle; 901 902 cc = CC_CPU(timeout_cpu); 903 CC_LOCK(cc); 904 /* Fill in the next free callout structure. */ 905 new = SLIST_FIRST(&cc->cc_callfree); 906 if (new == NULL) 907 /* XXX Attempt to malloc first */ 908 panic("timeout table full"); 909 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle); 910 callout_reset(new, to_ticks, ftn, arg); 911 handle.callout = new; 912 CC_UNLOCK(cc); 913 914 return (handle); 915 } 916 917 void 918 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle) 919 { 920 struct callout_cpu *cc; 921 922 /* 923 * Check for a handle that was initialized 924 * by callout_handle_init, but never used 925 * for a real timeout. 926 */ 927 if (handle.callout == NULL) 928 return; 929 930 cc = callout_lock(handle.callout); 931 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg) 932 callout_stop(handle.callout); 933 CC_UNLOCK(cc); 934 } 935 936 void 937 callout_handle_init(struct callout_handle *handle) 938 { 939 handle->callout = NULL; 940 } 941 942 /* 943 * New interface; clients allocate their own callout structures. 944 * 945 * callout_reset() - establish or change a timeout 946 * callout_stop() - disestablish a timeout 947 * callout_init() - initialize a callout structure so that it can 948 * safely be passed to callout_reset() and callout_stop() 949 * 950 * <sys/callout.h> defines three convenience macros: 951 * 952 * callout_active() - returns truth if callout has not been stopped, 953 * drained, or deactivated since the last time the callout was 954 * reset. 955 * callout_pending() - returns truth if callout is still waiting for timeout 956 * callout_deactivate() - marks the callout as having been serviced 957 */ 958 int 959 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t precision, 960 void (*ftn)(void *), void *arg, int cpu, int flags) 961 { 962 sbintime_t to_sbt, pr; 963 struct callout_cpu *cc; 964 int cancelled, direct; 965 int ignore_cpu=0; 966 967 cancelled = 0; 968 if (cpu == -1) { 969 ignore_cpu = 1; 970 } else if ((cpu >= MAXCPU) || 971 ((CC_CPU(cpu))->cc_inited == 0)) { 972 /* Invalid CPU spec */ 973 panic("Invalid CPU in callout %d", cpu); 974 } 975 if (flags & C_ABSOLUTE) { 976 to_sbt = sbt; 977 } else { 978 if ((flags & C_HARDCLOCK) && (sbt < tick_sbt)) 979 sbt = tick_sbt; 980 if ((flags & C_HARDCLOCK) || 981 #ifdef NO_EVENTTIMERS 982 sbt >= sbt_timethreshold) { 983 to_sbt = getsbinuptime(); 984 985 /* Add safety belt for the case of hz > 1000. */ 986 to_sbt += tc_tick_sbt - tick_sbt; 987 #else 988 sbt >= sbt_tickthreshold) { 989 /* 990 * Obtain the time of the last hardclock() call on 991 * this CPU directly from the kern_clocksource.c. 992 * This value is per-CPU, but it is equal for all 993 * active ones. 994 */ 995 #ifdef __LP64__ 996 to_sbt = DPCPU_GET(hardclocktime); 997 #else 998 spinlock_enter(); 999 to_sbt = DPCPU_GET(hardclocktime); 1000 spinlock_exit(); 1001 #endif 1002 #endif 1003 if ((flags & C_HARDCLOCK) == 0) 1004 to_sbt += tick_sbt; 1005 } else 1006 to_sbt = sbinuptime(); 1007 if (SBT_MAX - to_sbt < sbt) 1008 to_sbt = SBT_MAX; 1009 else 1010 to_sbt += sbt; 1011 pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp : 1012 sbt >> C_PRELGET(flags)); 1013 if (pr > precision) 1014 precision = pr; 1015 } 1016 /* 1017 * This flag used to be added by callout_cc_add, but the 1018 * first time you call this we could end up with the 1019 * wrong direct flag if we don't do it before we add. 1020 */ 1021 if (flags & C_DIRECT_EXEC) { 1022 direct = 1; 1023 } else { 1024 direct = 0; 1025 } 1026 KASSERT(!direct || c->c_lock == NULL, 1027 ("%s: direct callout %p has lock", __func__, c)); 1028 cc = callout_lock(c); 1029 /* 1030 * Don't allow migration of pre-allocated callouts lest they 1031 * become unbalanced or handle the case where the user does 1032 * not care. 1033 */ 1034 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) || 1035 ignore_cpu) { 1036 cpu = c->c_cpu; 1037 } 1038 1039 if (cc_exec_curr(cc, direct) == c) { 1040 /* 1041 * We're being asked to reschedule a callout which is 1042 * currently in progress. If there is a lock then we 1043 * can cancel the callout if it has not really started. 1044 */ 1045 if (c->c_lock != NULL && !cc_exec_cancel(cc, direct)) 1046 cancelled = cc_exec_cancel(cc, direct) = true; 1047 if (cc_exec_waiting(cc, direct)) { 1048 /* 1049 * Someone has called callout_drain to kill this 1050 * callout. Don't reschedule. 1051 */ 1052 CTR4(KTR_CALLOUT, "%s %p func %p arg %p", 1053 cancelled ? "cancelled" : "failed to cancel", 1054 c, c->c_func, c->c_arg); 1055 CC_UNLOCK(cc); 1056 return (cancelled); 1057 } 1058 #ifdef SMP 1059 if (callout_migrating(c)) { 1060 /* 1061 * This only occurs when a second callout_reset_sbt_on 1062 * is made after a previous one moved it into 1063 * deferred migration (below). Note we do *not* change 1064 * the prev_cpu even though the previous target may 1065 * be different. 1066 */ 1067 cc_migration_cpu(cc, direct) = cpu; 1068 cc_migration_time(cc, direct) = to_sbt; 1069 cc_migration_prec(cc, direct) = precision; 1070 cc_migration_func(cc, direct) = ftn; 1071 cc_migration_arg(cc, direct) = arg; 1072 cancelled = 1; 1073 CC_UNLOCK(cc); 1074 return (cancelled); 1075 } 1076 #endif 1077 } 1078 if (c->c_iflags & CALLOUT_PENDING) { 1079 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1080 if (cc_exec_next(cc) == c) 1081 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1082 LIST_REMOVE(c, c_links.le); 1083 } else { 1084 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1085 } 1086 cancelled = 1; 1087 c->c_iflags &= ~ CALLOUT_PENDING; 1088 c->c_flags &= ~ CALLOUT_ACTIVE; 1089 } 1090 1091 #ifdef SMP 1092 /* 1093 * If the callout must migrate try to perform it immediately. 1094 * If the callout is currently running, just defer the migration 1095 * to a more appropriate moment. 1096 */ 1097 if (c->c_cpu != cpu) { 1098 if (cc_exec_curr(cc, direct) == c) { 1099 /* 1100 * Pending will have been removed since we are 1101 * actually executing the callout on another 1102 * CPU. That callout should be waiting on the 1103 * lock the caller holds. If we set both 1104 * active/and/pending after we return and the 1105 * lock on the executing callout proceeds, it 1106 * will then see pending is true and return. 1107 * At the return from the actual callout execution 1108 * the migration will occur in softclock_call_cc 1109 * and this new callout will be placed on the 1110 * new CPU via a call to callout_cpu_switch() which 1111 * will get the lock on the right CPU followed 1112 * by a call callout_cc_add() which will add it there. 1113 * (see above in softclock_call_cc()). 1114 */ 1115 cc_migration_cpu(cc, direct) = cpu; 1116 cc_migration_time(cc, direct) = to_sbt; 1117 cc_migration_prec(cc, direct) = precision; 1118 cc_migration_func(cc, direct) = ftn; 1119 cc_migration_arg(cc, direct) = arg; 1120 c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING); 1121 c->c_flags |= CALLOUT_ACTIVE; 1122 CTR6(KTR_CALLOUT, 1123 "migration of %p func %p arg %p in %d.%08x to %u deferred", 1124 c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1125 (u_int)(to_sbt & 0xffffffff), cpu); 1126 CC_UNLOCK(cc); 1127 return (cancelled); 1128 } 1129 cc = callout_cpu_switch(c, cc, cpu); 1130 } 1131 #endif 1132 1133 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags); 1134 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x", 1135 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1136 (u_int)(to_sbt & 0xffffffff)); 1137 CC_UNLOCK(cc); 1138 1139 return (cancelled); 1140 } 1141 1142 /* 1143 * Common idioms that can be optimized in the future. 1144 */ 1145 int 1146 callout_schedule_on(struct callout *c, int to_ticks, int cpu) 1147 { 1148 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu); 1149 } 1150 1151 int 1152 callout_schedule(struct callout *c, int to_ticks) 1153 { 1154 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu); 1155 } 1156 1157 int 1158 _callout_stop_safe(struct callout *c, int safe, void (*drain)(void *)) 1159 { 1160 struct callout_cpu *cc, *old_cc; 1161 struct lock_class *class; 1162 int direct, sq_locked, use_lock; 1163 int not_on_a_list; 1164 1165 if (safe) 1166 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock, 1167 "calling %s", __func__); 1168 1169 /* 1170 * Some old subsystems don't hold Giant while running a callout_stop(), 1171 * so just discard this check for the moment. 1172 */ 1173 if (!safe && c->c_lock != NULL) { 1174 if (c->c_lock == &Giant.lock_object) 1175 use_lock = mtx_owned(&Giant); 1176 else { 1177 use_lock = 1; 1178 class = LOCK_CLASS(c->c_lock); 1179 class->lc_assert(c->c_lock, LA_XLOCKED); 1180 } 1181 } else 1182 use_lock = 0; 1183 if (c->c_iflags & CALLOUT_DIRECT) { 1184 direct = 1; 1185 } else { 1186 direct = 0; 1187 } 1188 sq_locked = 0; 1189 old_cc = NULL; 1190 again: 1191 cc = callout_lock(c); 1192 1193 if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) == 1194 (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) && 1195 ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) { 1196 /* 1197 * Special case where this slipped in while we 1198 * were migrating *as* the callout is about to 1199 * execute. The caller probably holds the lock 1200 * the callout wants. 1201 * 1202 * Get rid of the migration first. Then set 1203 * the flag that tells this code *not* to 1204 * try to remove it from any lists (its not 1205 * on one yet). When the callout wheel runs, 1206 * it will ignore this callout. 1207 */ 1208 c->c_iflags &= ~CALLOUT_PENDING; 1209 c->c_flags &= ~CALLOUT_ACTIVE; 1210 not_on_a_list = 1; 1211 } else { 1212 not_on_a_list = 0; 1213 } 1214 1215 /* 1216 * If the callout was migrating while the callout cpu lock was 1217 * dropped, just drop the sleepqueue lock and check the states 1218 * again. 1219 */ 1220 if (sq_locked != 0 && cc != old_cc) { 1221 #ifdef SMP 1222 CC_UNLOCK(cc); 1223 sleepq_release(&cc_exec_waiting(old_cc, direct)); 1224 sq_locked = 0; 1225 old_cc = NULL; 1226 goto again; 1227 #else 1228 panic("migration should not happen"); 1229 #endif 1230 } 1231 1232 /* 1233 * If the callout isn't pending, it's not on the queue, so 1234 * don't attempt to remove it from the queue. We can try to 1235 * stop it by other means however. 1236 */ 1237 if (!(c->c_iflags & CALLOUT_PENDING)) { 1238 /* 1239 * If it wasn't on the queue and it isn't the current 1240 * callout, then we can't stop it, so just bail. 1241 * It probably has already been run (if locking 1242 * is properly done). You could get here if the caller 1243 * calls stop twice in a row for example. The second 1244 * call would fall here without CALLOUT_ACTIVE set. 1245 */ 1246 c->c_flags &= ~CALLOUT_ACTIVE; 1247 if (cc_exec_curr(cc, direct) != c) { 1248 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1249 c, c->c_func, c->c_arg); 1250 CC_UNLOCK(cc); 1251 if (sq_locked) 1252 sleepq_release(&cc_exec_waiting(cc, direct)); 1253 return (-1); 1254 } 1255 1256 if (safe) { 1257 /* 1258 * The current callout is running (or just 1259 * about to run) and blocking is allowed, so 1260 * just wait for the current invocation to 1261 * finish. 1262 */ 1263 while (cc_exec_curr(cc, direct) == c) { 1264 /* 1265 * Use direct calls to sleepqueue interface 1266 * instead of cv/msleep in order to avoid 1267 * a LOR between cc_lock and sleepqueue 1268 * chain spinlocks. This piece of code 1269 * emulates a msleep_spin() call actually. 1270 * 1271 * If we already have the sleepqueue chain 1272 * locked, then we can safely block. If we 1273 * don't already have it locked, however, 1274 * we have to drop the cc_lock to lock 1275 * it. This opens several races, so we 1276 * restart at the beginning once we have 1277 * both locks. If nothing has changed, then 1278 * we will end up back here with sq_locked 1279 * set. 1280 */ 1281 if (!sq_locked) { 1282 CC_UNLOCK(cc); 1283 sleepq_lock( 1284 &cc_exec_waiting(cc, direct)); 1285 sq_locked = 1; 1286 old_cc = cc; 1287 goto again; 1288 } 1289 1290 /* 1291 * Migration could be cancelled here, but 1292 * as long as it is still not sure when it 1293 * will be packed up, just let softclock() 1294 * take care of it. 1295 */ 1296 cc_exec_waiting(cc, direct) = true; 1297 DROP_GIANT(); 1298 CC_UNLOCK(cc); 1299 sleepq_add( 1300 &cc_exec_waiting(cc, direct), 1301 &cc->cc_lock.lock_object, "codrain", 1302 SLEEPQ_SLEEP, 0); 1303 sleepq_wait( 1304 &cc_exec_waiting(cc, direct), 1305 0); 1306 sq_locked = 0; 1307 old_cc = NULL; 1308 1309 /* Reacquire locks previously released. */ 1310 PICKUP_GIANT(); 1311 CC_LOCK(cc); 1312 } 1313 } else if (use_lock && 1314 !cc_exec_cancel(cc, direct) && (drain == NULL)) { 1315 1316 /* 1317 * The current callout is waiting for its 1318 * lock which we hold. Cancel the callout 1319 * and return. After our caller drops the 1320 * lock, the callout will be skipped in 1321 * softclock(). This *only* works with a 1322 * callout_stop() *not* callout_drain() or 1323 * callout_async_drain(). 1324 */ 1325 cc_exec_cancel(cc, direct) = true; 1326 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1327 c, c->c_func, c->c_arg); 1328 KASSERT(!cc_cce_migrating(cc, direct), 1329 ("callout wrongly scheduled for migration")); 1330 if (callout_migrating(c)) { 1331 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1332 #ifdef SMP 1333 cc_migration_cpu(cc, direct) = CPUBLOCK; 1334 cc_migration_time(cc, direct) = 0; 1335 cc_migration_prec(cc, direct) = 0; 1336 cc_migration_func(cc, direct) = NULL; 1337 cc_migration_arg(cc, direct) = NULL; 1338 #endif 1339 } 1340 CC_UNLOCK(cc); 1341 KASSERT(!sq_locked, ("sleepqueue chain locked")); 1342 return (1); 1343 } else if (callout_migrating(c)) { 1344 /* 1345 * The callout is currently being serviced 1346 * and the "next" callout is scheduled at 1347 * its completion with a migration. We remove 1348 * the migration flag so it *won't* get rescheduled, 1349 * but we can't stop the one thats running so 1350 * we return 0. 1351 */ 1352 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1353 #ifdef SMP 1354 /* 1355 * We can't call cc_cce_cleanup here since 1356 * if we do it will remove .ce_curr and 1357 * its still running. This will prevent a 1358 * reschedule of the callout when the 1359 * execution completes. 1360 */ 1361 cc_migration_cpu(cc, direct) = CPUBLOCK; 1362 cc_migration_time(cc, direct) = 0; 1363 cc_migration_prec(cc, direct) = 0; 1364 cc_migration_func(cc, direct) = NULL; 1365 cc_migration_arg(cc, direct) = NULL; 1366 #endif 1367 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p", 1368 c, c->c_func, c->c_arg); 1369 if (drain) { 1370 cc_exec_drain(cc, direct) = drain; 1371 } 1372 CC_UNLOCK(cc); 1373 return (0); 1374 } 1375 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1376 c, c->c_func, c->c_arg); 1377 if (drain) { 1378 cc_exec_drain(cc, direct) = drain; 1379 } 1380 CC_UNLOCK(cc); 1381 KASSERT(!sq_locked, ("sleepqueue chain still locked")); 1382 return (0); 1383 } 1384 if (sq_locked) 1385 sleepq_release(&cc_exec_waiting(cc, direct)); 1386 1387 c->c_iflags &= ~CALLOUT_PENDING; 1388 c->c_flags &= ~CALLOUT_ACTIVE; 1389 1390 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1391 c, c->c_func, c->c_arg); 1392 if (not_on_a_list == 0) { 1393 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1394 if (cc_exec_next(cc) == c) 1395 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1396 LIST_REMOVE(c, c_links.le); 1397 } else { 1398 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1399 } 1400 } 1401 callout_cc_del(c, cc); 1402 CC_UNLOCK(cc); 1403 return (1); 1404 } 1405 1406 void 1407 callout_init(struct callout *c, int mpsafe) 1408 { 1409 bzero(c, sizeof *c); 1410 if (mpsafe) { 1411 c->c_lock = NULL; 1412 c->c_iflags = CALLOUT_RETURNUNLOCKED; 1413 } else { 1414 c->c_lock = &Giant.lock_object; 1415 c->c_iflags = 0; 1416 } 1417 c->c_cpu = timeout_cpu; 1418 } 1419 1420 void 1421 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags) 1422 { 1423 bzero(c, sizeof *c); 1424 c->c_lock = lock; 1425 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0, 1426 ("callout_init_lock: bad flags %d", flags)); 1427 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0, 1428 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock")); 1429 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & 1430 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class", 1431 __func__)); 1432 c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK); 1433 c->c_cpu = timeout_cpu; 1434 } 1435 1436 #ifdef APM_FIXUP_CALLTODO 1437 /* 1438 * Adjust the kernel calltodo timeout list. This routine is used after 1439 * an APM resume to recalculate the calltodo timer list values with the 1440 * number of hz's we have been sleeping. The next hardclock() will detect 1441 * that there are fired timers and run softclock() to execute them. 1442 * 1443 * Please note, I have not done an exhaustive analysis of what code this 1444 * might break. I am motivated to have my select()'s and alarm()'s that 1445 * have expired during suspend firing upon resume so that the applications 1446 * which set the timer can do the maintanence the timer was for as close 1447 * as possible to the originally intended time. Testing this code for a 1448 * week showed that resuming from a suspend resulted in 22 to 25 timers 1449 * firing, which seemed independant on whether the suspend was 2 hours or 1450 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu> 1451 */ 1452 void 1453 adjust_timeout_calltodo(struct timeval *time_change) 1454 { 1455 register struct callout *p; 1456 unsigned long delta_ticks; 1457 1458 /* 1459 * How many ticks were we asleep? 1460 * (stolen from tvtohz()). 1461 */ 1462 1463 /* Don't do anything */ 1464 if (time_change->tv_sec < 0) 1465 return; 1466 else if (time_change->tv_sec <= LONG_MAX / 1000000) 1467 delta_ticks = (time_change->tv_sec * 1000000 + 1468 time_change->tv_usec + (tick - 1)) / tick + 1; 1469 else if (time_change->tv_sec <= LONG_MAX / hz) 1470 delta_ticks = time_change->tv_sec * hz + 1471 (time_change->tv_usec + (tick - 1)) / tick + 1; 1472 else 1473 delta_ticks = LONG_MAX; 1474 1475 if (delta_ticks > INT_MAX) 1476 delta_ticks = INT_MAX; 1477 1478 /* 1479 * Now rip through the timer calltodo list looking for timers 1480 * to expire. 1481 */ 1482 1483 /* don't collide with softclock() */ 1484 CC_LOCK(cc); 1485 for (p = calltodo.c_next; p != NULL; p = p->c_next) { 1486 p->c_time -= delta_ticks; 1487 1488 /* Break if the timer had more time on it than delta_ticks */ 1489 if (p->c_time > 0) 1490 break; 1491 1492 /* take back the ticks the timer didn't use (p->c_time <= 0) */ 1493 delta_ticks = -p->c_time; 1494 } 1495 CC_UNLOCK(cc); 1496 1497 return; 1498 } 1499 #endif /* APM_FIXUP_CALLTODO */ 1500 1501 static int 1502 flssbt(sbintime_t sbt) 1503 { 1504 1505 sbt += (uint64_t)sbt >> 1; 1506 if (sizeof(long) >= sizeof(sbintime_t)) 1507 return (flsl(sbt)); 1508 if (sbt >= SBT_1S) 1509 return (flsl(((uint64_t)sbt) >> 32) + 32); 1510 return (flsl(sbt)); 1511 } 1512 1513 /* 1514 * Dump immediate statistic snapshot of the scheduled callouts. 1515 */ 1516 static int 1517 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS) 1518 { 1519 struct callout *tmp; 1520 struct callout_cpu *cc; 1521 struct callout_list *sc; 1522 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t; 1523 int ct[64], cpr[64], ccpbk[32]; 1524 int error, val, i, count, tcum, pcum, maxc, c, medc; 1525 #ifdef SMP 1526 int cpu; 1527 #endif 1528 1529 val = 0; 1530 error = sysctl_handle_int(oidp, &val, 0, req); 1531 if (error != 0 || req->newptr == NULL) 1532 return (error); 1533 count = maxc = 0; 1534 st = spr = maxt = maxpr = 0; 1535 bzero(ccpbk, sizeof(ccpbk)); 1536 bzero(ct, sizeof(ct)); 1537 bzero(cpr, sizeof(cpr)); 1538 now = sbinuptime(); 1539 #ifdef SMP 1540 CPU_FOREACH(cpu) { 1541 cc = CC_CPU(cpu); 1542 #else 1543 cc = CC_CPU(timeout_cpu); 1544 #endif 1545 CC_LOCK(cc); 1546 for (i = 0; i < callwheelsize; i++) { 1547 sc = &cc->cc_callwheel[i]; 1548 c = 0; 1549 LIST_FOREACH(tmp, sc, c_links.le) { 1550 c++; 1551 t = tmp->c_time - now; 1552 if (t < 0) 1553 t = 0; 1554 st += t / SBT_1US; 1555 spr += tmp->c_precision / SBT_1US; 1556 if (t > maxt) 1557 maxt = t; 1558 if (tmp->c_precision > maxpr) 1559 maxpr = tmp->c_precision; 1560 ct[flssbt(t)]++; 1561 cpr[flssbt(tmp->c_precision)]++; 1562 } 1563 if (c > maxc) 1564 maxc = c; 1565 ccpbk[fls(c + c / 2)]++; 1566 count += c; 1567 } 1568 CC_UNLOCK(cc); 1569 #ifdef SMP 1570 } 1571 #endif 1572 1573 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++) 1574 tcum += ct[i]; 1575 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1576 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++) 1577 pcum += cpr[i]; 1578 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1579 for (i = 0, c = 0; i < 32 && c < count / 2; i++) 1580 c += ccpbk[i]; 1581 medc = (i >= 2) ? (1 << (i - 2)) : 0; 1582 1583 printf("Scheduled callouts statistic snapshot:\n"); 1584 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n", 1585 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT); 1586 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n", 1587 medc, 1588 count / callwheelsize / mp_ncpus, 1589 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000, 1590 maxc); 1591 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1592 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32, 1593 (st / count) / 1000000, (st / count) % 1000000, 1594 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32); 1595 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1596 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32, 1597 (spr / count) / 1000000, (spr / count) % 1000000, 1598 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32); 1599 printf(" Distribution: \tbuckets\t time\t tcum\t" 1600 " prec\t pcum\n"); 1601 for (i = 0, tcum = pcum = 0; i < 64; i++) { 1602 if (ct[i] == 0 && cpr[i] == 0) 1603 continue; 1604 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0; 1605 tcum += ct[i]; 1606 pcum += cpr[i]; 1607 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n", 1608 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32, 1609 i - 1 - (32 - CC_HASH_SHIFT), 1610 ct[i], tcum, cpr[i], pcum); 1611 } 1612 return (error); 1613 } 1614 SYSCTL_PROC(_kern, OID_AUTO, callout_stat, 1615 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1616 0, 0, sysctl_kern_callout_stat, "I", 1617 "Dump immediate statistic snapshot of the scheduled callouts"); 1618