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