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