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 * Copyright (c) 2010 Kip Macy. All rights reserved. 11 * Copyright (c) 2013 Patrick Kelsey. All rights reserved. 12 * Copyright (C) 2017 THL A29 Limited, a Tencent company. 13 * All rights reserved. 14 * 15 * Redistribution and use in source and binary forms, with or without 16 * modification, are permitted provided that the following conditions 17 * are met: 18 * 1. Redistributions of source code must retain the above copyright 19 * notice, this list of conditions and the following disclaimer. 20 * 2. Redistributions in binary form must reproduce the above copyright 21 * notice, this list of conditions and the following disclaimer in the 22 * documentation and/or other materials provided with the distribution. 23 * 4. Neither the name of the University nor the names of its contributors 24 * may be used to endorse or promote products derived from this software 25 * without specific prior written permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 37 * SUCH DAMAGE. 38 * 39 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 40 * 41 * Derived in part from libplebnet's pn_kern_timeout.c and libuinet's uinet_timecounter.c. 42 * 43 */ 44 45 #include <sys/cdefs.h> 46 __FBSDID("$FreeBSD$"); 47 48 #include "opt_callout_profiling.h" 49 #include "opt_ddb.h" 50 #if defined(__arm__) 51 #include "opt_timer.h" 52 #endif 53 #include "opt_rss.h" 54 55 #include <sys/param.h> 56 #include <sys/systm.h> 57 #include <sys/bus.h> 58 #include <sys/callout.h> 59 #include <sys/file.h> 60 #include <sys/interrupt.h> 61 #include <sys/kernel.h> 62 #include <sys/ktr.h> 63 #include <sys/lock.h> 64 #include <sys/malloc.h> 65 #include <sys/mutex.h> 66 #include <sys/proc.h> 67 #include <sys/sdt.h> 68 #include <sys/sleepqueue.h> 69 #include <sys/sysctl.h> 70 #include <sys/smp.h> 71 #include <sys/timetc.h> 72 73 #ifdef DDB 74 #include <ddb/ddb.h> 75 #include <machine/_inttypes.h> 76 #endif 77 78 #ifdef SMP 79 #include <machine/cpu.h> 80 #endif 81 82 #ifndef NO_EVENTTIMERS 83 DPCPU_DECLARE(sbintime_t, hardclocktime); 84 #endif 85 86 SDT_PROVIDER_DEFINE(callout_execute); 87 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *"); 88 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *"); 89 90 #ifdef CALLOUT_PROFILING 91 static int avg_depth; 92 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0, 93 "Average number of items examined per softclock call. Units = 1/1000"); 94 static int avg_gcalls; 95 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0, 96 "Average number of Giant callouts made per softclock call. Units = 1/1000"); 97 static int avg_lockcalls; 98 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0, 99 "Average number of lock callouts made per softclock call. Units = 1/1000"); 100 static int avg_mpcalls; 101 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0, 102 "Average number of MP callouts made per softclock call. Units = 1/1000"); 103 static int avg_depth_dir; 104 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0, 105 "Average number of direct callouts examined per callout_process call. " 106 "Units = 1/1000"); 107 static int avg_lockcalls_dir; 108 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD, 109 &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per " 110 "callout_process call. Units = 1/1000"); 111 static int avg_mpcalls_dir; 112 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir, 113 0, "Average number of MP direct callouts made per callout_process call. " 114 "Units = 1/1000"); 115 #endif 116 117 static int ncallout; 118 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0, 119 "Number of entries in callwheel and size of timeout() preallocation"); 120 121 #ifdef RSS 122 static int pin_default_swi = 1; 123 static int pin_pcpu_swi = 1; 124 #else 125 static int pin_default_swi = 0; 126 static int pin_pcpu_swi = 0; 127 #endif 128 129 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi, 130 0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)"); 131 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi, 132 0, "Pin the per-CPU swis (except PCPU 0, which is also default"); 133 134 #define sleepq_lock(w) do {} while(0) 135 #define sleepq_release(w) do {} while(0) 136 #define sleepq_add(a, b, c, d, e) do {} while(0) 137 #define sleepq_wait(w, p) do {} while(0) 138 139 /* 140 * TODO: 141 * allocate more timeout table slots when table overflows. 142 */ 143 u_int callwheelsize, callwheelmask; 144 145 /* 146 * The callout cpu exec entities represent informations necessary for 147 * describing the state of callouts currently running on the CPU and the ones 148 * necessary for migrating callouts to the new callout cpu. In particular, 149 * the first entry of the array cc_exec_entity holds informations for callout 150 * running in SWI thread context, while the second one holds informations 151 * for callout running directly from hardware interrupt context. 152 * The cached informations are very important for deferring migration when 153 * the migrating callout is already running. 154 */ 155 struct cc_exec { 156 struct callout *cc_curr; 157 void (*cc_drain)(void *); 158 #ifdef SMP 159 void (*ce_migration_func)(void *); 160 void *ce_migration_arg; 161 int ce_migration_cpu; 162 sbintime_t ce_migration_time; 163 sbintime_t ce_migration_prec; 164 #endif 165 bool cc_cancel; 166 bool cc_waiting; 167 }; 168 169 /* 170 * There is one struct callout_cpu per cpu, holding all relevant 171 * state for the callout processing thread on the individual CPU. 172 */ 173 struct callout_cpu { 174 struct mtx_padalign cc_lock; 175 struct cc_exec cc_exec_entity[2]; 176 struct callout *cc_next; 177 struct callout *cc_callout; 178 struct callout_list *cc_callwheel; 179 struct callout_tailq cc_expireq; 180 struct callout_slist cc_callfree; 181 sbintime_t cc_firstevent; 182 sbintime_t cc_lastscan; 183 void *cc_cookie; 184 u_int cc_bucket; 185 u_int cc_inited; 186 char cc_ktr_event_name[20]; 187 }; 188 189 #define callout_migrating(c) ((c)->c_iflags & CALLOUT_DFRMIGRATION) 190 191 #define cc_exec_curr(cc, dir) cc->cc_exec_entity[dir].cc_curr 192 #define cc_exec_drain(cc, dir) cc->cc_exec_entity[dir].cc_drain 193 #define cc_exec_next(cc) cc->cc_next 194 #define cc_exec_cancel(cc, dir) cc->cc_exec_entity[dir].cc_cancel 195 #define cc_exec_waiting(cc, dir) cc->cc_exec_entity[dir].cc_waiting 196 #ifdef SMP 197 #define cc_migration_func(cc, dir) cc->cc_exec_entity[dir].ce_migration_func 198 #define cc_migration_arg(cc, dir) cc->cc_exec_entity[dir].ce_migration_arg 199 #define cc_migration_cpu(cc, dir) cc->cc_exec_entity[dir].ce_migration_cpu 200 #define cc_migration_time(cc, dir) cc->cc_exec_entity[dir].ce_migration_time 201 #define cc_migration_prec(cc, dir) cc->cc_exec_entity[dir].ce_migration_prec 202 203 struct callout_cpu cc_cpu[MAXCPU]; 204 #define CPUBLOCK MAXCPU 205 #define CC_CPU(cpu) (&cc_cpu[(cpu)]) 206 #define CC_SELF() CC_CPU(PCPU_GET(cpuid)) 207 #else 208 struct callout_cpu cc_cpu; 209 #define CC_CPU(cpu) &cc_cpu 210 #define CC_SELF() &cc_cpu 211 #endif 212 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock) 213 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock) 214 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED) 215 216 static int timeout_cpu; 217 218 static void callout_cpu_init(struct callout_cpu *cc, int cpu); 219 static void softclock_call_cc(struct callout *c, struct callout_cpu *cc, 220 #ifdef CALLOUT_PROFILING 221 int *mpcalls, int *lockcalls, int *gcalls, 222 #endif 223 int direct); 224 225 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures"); 226 227 /** 228 * Locked by cc_lock: 229 * cc_curr - If a callout is in progress, it is cc_curr. 230 * If cc_curr is non-NULL, threads waiting in 231 * callout_drain() will be woken up as soon as the 232 * relevant callout completes. 233 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held 234 * guarantees that the current callout will not run. 235 * The softclock() function sets this to 0 before it 236 * drops callout_lock to acquire c_lock, and it calls 237 * the handler only if curr_cancelled is still 0 after 238 * cc_lock is successfully acquired. 239 * cc_waiting - If a thread is waiting in callout_drain(), then 240 * callout_wait is nonzero. Set only when 241 * cc_curr is non-NULL. 242 */ 243 244 /* 245 * Resets the execution entity tied to a specific callout cpu. 246 */ 247 static void 248 cc_cce_cleanup(struct callout_cpu *cc, int direct) 249 { 250 cc_exec_curr(cc, direct) = NULL; 251 cc_exec_cancel(cc, direct) = false; 252 cc_exec_waiting(cc, direct) = false; 253 #ifdef SMP 254 cc_migration_cpu(cc, direct) = CPUBLOCK; 255 cc_migration_time(cc, direct) = 0; 256 cc_migration_prec(cc, direct) = 0; 257 cc_migration_func(cc, direct) = NULL; 258 cc_migration_arg(cc, direct) = NULL; 259 #endif 260 } 261 262 /* 263 * Checks if migration is requested by a specific callout cpu. 264 */ 265 static int 266 cc_cce_migrating(struct callout_cpu *cc, int direct) 267 { 268 #ifdef SMP 269 return (cc_migration_cpu(cc, direct) != CPUBLOCK); 270 #else 271 return (0); 272 #endif 273 } 274 275 /* 276 * Kernel low level callwheel initialization 277 * called on cpu0 during kernel startup. 278 */ 279 static void 280 callout_callwheel_init(void *dummy) 281 { 282 struct callout_cpu *cc; 283 284 /* 285 * Calculate the size of the callout wheel and the preallocated 286 * timeout() structures. 287 * XXX: Clip callout to result of previous function of maxusers 288 * maximum 384. This is still huge, but acceptable. 289 */ 290 memset(CC_CPU(0), 0, sizeof(cc_cpu)); 291 ncallout = imin(16 + maxproc + maxfiles, 18508); 292 TUNABLE_INT_FETCH("kern.ncallout", &ncallout); 293 294 /* 295 * Calculate callout wheel size, should be next power of two higher 296 * than 'ncallout'. 297 */ 298 callwheelsize = 1 << fls(ncallout); 299 callwheelmask = callwheelsize - 1; 300 301 /* 302 * Fetch whether we're pinning the swi's or not. 303 */ 304 TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi); 305 TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi); 306 307 /* 308 * Only cpu0 handles timeout(9) and receives a preallocation. 309 * 310 * XXX: Once all timeout(9) consumers are converted this can 311 * be removed. 312 */ 313 timeout_cpu = PCPU_GET(cpuid); 314 cc = CC_CPU(timeout_cpu); 315 cc->cc_callout = malloc(ncallout * sizeof(struct callout), 316 M_CALLOUT, M_WAITOK); 317 callout_cpu_init(cc, timeout_cpu); 318 } 319 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL); 320 321 /* 322 * Initialize the per-cpu callout structures. 323 */ 324 static void 325 callout_cpu_init(struct callout_cpu *cc, int cpu) 326 { 327 struct callout *c; 328 int i; 329 330 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE); 331 SLIST_INIT(&cc->cc_callfree); 332 cc->cc_inited = 1; 333 cc->cc_callwheel = malloc(sizeof(struct callout_list) * callwheelsize, 334 M_CALLOUT, M_WAITOK); 335 for (i = 0; i < callwheelsize; i++) 336 LIST_INIT(&cc->cc_callwheel[i]); 337 TAILQ_INIT(&cc->cc_expireq); 338 cc->cc_firstevent = SBT_MAX; 339 for (i = 0; i < 2; i++) 340 cc_cce_cleanup(cc, i); 341 snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name), 342 "callwheel cpu %d", cpu); 343 if (cc->cc_callout == NULL) /* Only cpu0 handles timeout(9) */ 344 return; 345 for (i = 0; i < ncallout; i++) { 346 c = &cc->cc_callout[i]; 347 callout_init(c, 0); 348 c->c_iflags = CALLOUT_LOCAL_ALLOC; 349 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 350 } 351 } 352 353 #ifdef SMP 354 /* 355 * Switches the cpu tied to a specific callout. 356 * The function expects a locked incoming callout cpu and returns with 357 * locked outcoming callout cpu. 358 */ 359 static struct callout_cpu * 360 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu) 361 { 362 struct callout_cpu *new_cc; 363 364 MPASS(c != NULL && cc != NULL); 365 CC_LOCK_ASSERT(cc); 366 367 /* 368 * Avoid interrupts and preemption firing after the callout cpu 369 * is blocked in order to avoid deadlocks as the new thread 370 * may be willing to acquire the callout cpu lock. 371 */ 372 c->c_cpu = CPUBLOCK; 373 spinlock_enter(); 374 CC_UNLOCK(cc); 375 new_cc = CC_CPU(new_cpu); 376 CC_LOCK(new_cc); 377 spinlock_exit(); 378 c->c_cpu = new_cpu; 379 return (new_cc); 380 } 381 #endif 382 383 #ifndef FSTACK 384 /* 385 * Start standard softclock thread. 386 */ 387 static void 388 start_softclock(void *dummy) 389 { 390 struct callout_cpu *cc; 391 char name[MAXCOMLEN]; 392 #ifdef SMP 393 int cpu; 394 struct intr_event *ie; 395 #endif 396 397 cc = CC_CPU(timeout_cpu); 398 snprintf(name, sizeof(name), "clock (%d)", timeout_cpu); 399 if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK, 400 INTR_MPSAFE, &cc->cc_cookie)) 401 panic("died while creating standard software ithreads"); 402 if (pin_default_swi && 403 (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) { 404 printf("%s: timeout clock couldn't be pinned to cpu %d\n", 405 __func__, 406 timeout_cpu); 407 } 408 409 #ifdef SMP 410 CPU_FOREACH(cpu) { 411 if (cpu == timeout_cpu) 412 continue; 413 cc = CC_CPU(cpu); 414 cc->cc_callout = NULL; /* Only cpu0 handles timeout(9). */ 415 callout_cpu_init(cc, cpu); 416 snprintf(name, sizeof(name), "clock (%d)", cpu); 417 ie = NULL; 418 if (swi_add(&ie, name, softclock, cc, SWI_CLOCK, 419 INTR_MPSAFE, &cc->cc_cookie)) 420 panic("died while creating standard software ithreads"); 421 if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) { 422 printf("%s: per-cpu clock couldn't be pinned to " 423 "cpu %d\n", 424 __func__, 425 cpu); 426 } 427 } 428 #endif 429 } 430 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL); 431 #endif 432 433 #define CC_HASH_SHIFT 8 434 435 static inline u_int 436 callout_hash(sbintime_t sbt) 437 { 438 return (sbt >> (32 - CC_HASH_SHIFT)); 439 } 440 441 static inline u_int 442 callout_get_bucket(sbintime_t sbt) 443 { 444 return (callout_hash(sbt) & callwheelmask); 445 } 446 447 void 448 callout_process(sbintime_t now) 449 { 450 struct callout *tmp, *tmpn; 451 struct callout_cpu *cc; 452 struct callout_list *sc; 453 sbintime_t first, last, max, tmp_max; 454 uint32_t lookahead; 455 u_int firstb, lastb, nowb; 456 #ifdef CALLOUT_PROFILING 457 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0; 458 #endif 459 460 cc = CC_SELF(); 461 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET); 462 463 /* Compute the buckets of the last scan and present times. */ 464 firstb = callout_hash(cc->cc_lastscan); 465 cc->cc_lastscan = now; 466 nowb = callout_hash(now); 467 468 /* Compute the last bucket and minimum time of the bucket after it. */ 469 if (nowb == firstb) 470 lookahead = (SBT_1S / 16); 471 else if (nowb - firstb == 1) 472 lookahead = (SBT_1S / 8); 473 else 474 lookahead = (SBT_1S / 2); 475 first = last = now; 476 first += (lookahead / 2); 477 last += lookahead; 478 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT)); 479 lastb = callout_hash(last) - 1; 480 max = last; 481 482 /* 483 * Check if we wrapped around the entire wheel from the last scan. 484 * In case, we need to scan entirely the wheel for pending callouts. 485 */ 486 if (lastb - firstb >= callwheelsize) { 487 lastb = firstb + callwheelsize - 1; 488 if (nowb - firstb >= callwheelsize) 489 nowb = lastb; 490 } 491 492 /* Iterate callwheel from firstb to nowb and then up to lastb. */ 493 do { 494 sc = &cc->cc_callwheel[firstb & callwheelmask]; 495 tmp = LIST_FIRST(sc); 496 while (tmp != NULL) { 497 /* Run the callout if present time within allowed. */ 498 if (tmp->c_time <= now) { 499 /* 500 * Consumer told us the callout may be run 501 * directly from hardware interrupt context. 502 */ 503 if (tmp->c_iflags & CALLOUT_DIRECT) { 504 #ifdef CALLOUT_PROFILING 505 ++depth_dir; 506 #endif 507 cc_exec_next(cc) = 508 LIST_NEXT(tmp, c_links.le); 509 cc->cc_bucket = firstb & callwheelmask; 510 LIST_REMOVE(tmp, c_links.le); 511 softclock_call_cc(tmp, cc, 512 #ifdef CALLOUT_PROFILING 513 &mpcalls_dir, &lockcalls_dir, NULL, 514 #endif 515 1); 516 tmp = cc_exec_next(cc); 517 cc_exec_next(cc) = NULL; 518 } else { 519 tmpn = LIST_NEXT(tmp, c_links.le); 520 LIST_REMOVE(tmp, c_links.le); 521 TAILQ_INSERT_TAIL(&cc->cc_expireq, 522 tmp, c_links.tqe); 523 tmp->c_iflags |= CALLOUT_PROCESSED; 524 tmp = tmpn; 525 } 526 continue; 527 } 528 /* Skip events from distant future. */ 529 if (tmp->c_time >= max) 530 goto next; 531 /* 532 * Event minimal time is bigger than present maximal 533 * time, so it cannot be aggregated. 534 */ 535 if (tmp->c_time > last) { 536 lastb = nowb; 537 goto next; 538 } 539 /* Update first and last time, respecting this event. */ 540 if (tmp->c_time < first) 541 first = tmp->c_time; 542 tmp_max = tmp->c_time + tmp->c_precision; 543 if (tmp_max < last) 544 last = tmp_max; 545 next: 546 tmp = LIST_NEXT(tmp, c_links.le); 547 } 548 /* Proceed with the next bucket. */ 549 firstb++; 550 /* 551 * Stop if we looked after present time and found 552 * some event we can't execute at now. 553 * Stop if we looked far enough into the future. 554 */ 555 } while (((int)(firstb - lastb)) <= 0); 556 cc->cc_firstevent = last; 557 #ifndef NO_EVENTTIMERS 558 cpu_new_callout(curcpu, last, first); 559 #endif 560 #ifdef CALLOUT_PROFILING 561 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8; 562 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8; 563 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8; 564 #endif 565 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET); 566 /* 567 * swi_sched acquires the thread lock, so we don't want to call it 568 * with cc_lock held; incorrect locking order. 569 */ 570 if (!TAILQ_EMPTY(&cc->cc_expireq)) 571 #ifndef FSTACK 572 swi_sched(cc->cc_cookie, 0); 573 #else 574 softclock(cc); 575 #endif 576 } 577 578 static struct callout_cpu * 579 callout_lock(struct callout *c) 580 { 581 struct callout_cpu *cc; 582 int cpu; 583 584 for (;;) { 585 cpu = c->c_cpu; 586 #ifdef SMP 587 if (cpu == CPUBLOCK) { 588 while (c->c_cpu == CPUBLOCK) 589 cpu_spinwait(); 590 continue; 591 } 592 #endif 593 cc = CC_CPU(cpu); 594 CC_LOCK(cc); 595 if (cpu == c->c_cpu) 596 break; 597 CC_UNLOCK(cc); 598 } 599 return (cc); 600 } 601 602 static void 603 callout_cc_add(struct callout *c, struct callout_cpu *cc, 604 sbintime_t sbt, sbintime_t precision, void (*func)(void *), 605 void *arg, int cpu, int flags) 606 { 607 int bucket; 608 609 CC_LOCK_ASSERT(cc); 610 if (sbt < cc->cc_lastscan) 611 sbt = cc->cc_lastscan; 612 c->c_arg = arg; 613 c->c_iflags |= CALLOUT_PENDING; 614 c->c_iflags &= ~CALLOUT_PROCESSED; 615 c->c_flags |= CALLOUT_ACTIVE; 616 if (flags & C_DIRECT_EXEC) 617 c->c_iflags |= CALLOUT_DIRECT; 618 c->c_func = func; 619 c->c_time = sbt; 620 c->c_precision = precision; 621 bucket = callout_get_bucket(c->c_time); 622 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x", 623 c, (int)(c->c_precision >> 32), 624 (u_int)(c->c_precision & 0xffffffff)); 625 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le); 626 if (cc->cc_bucket == bucket) 627 cc_exec_next(cc) = c; 628 #ifndef NO_EVENTTIMERS 629 /* 630 * Inform the eventtimers(4) subsystem there's a new callout 631 * that has been inserted, but only if really required. 632 */ 633 if (SBT_MAX - c->c_time < c->c_precision) 634 c->c_precision = SBT_MAX - c->c_time; 635 sbt = c->c_time + c->c_precision; 636 if (sbt < cc->cc_firstevent) { 637 cc->cc_firstevent = sbt; 638 cpu_new_callout(cpu, sbt, c->c_time); 639 } 640 #endif 641 } 642 643 static void 644 callout_cc_del(struct callout *c, struct callout_cpu *cc) 645 { 646 647 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0) 648 return; 649 c->c_func = NULL; 650 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle); 651 } 652 653 static void 654 softclock_call_cc(struct callout *c, struct callout_cpu *cc, 655 #ifdef CALLOUT_PROFILING 656 int *mpcalls, int *lockcalls, int *gcalls, 657 #endif 658 int direct) 659 { 660 struct rm_priotracker tracker; 661 void (*c_func)(void *); 662 void *c_arg; 663 struct lock_class *class; 664 struct lock_object *c_lock; 665 uintptr_t lock_status; 666 int c_iflags; 667 #ifdef SMP 668 struct callout_cpu *new_cc; 669 void (*new_func)(void *); 670 void *new_arg; 671 int flags, new_cpu; 672 sbintime_t new_prec, new_time; 673 #endif 674 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 675 sbintime_t sbt1, sbt2; 676 struct timespec ts2; 677 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */ 678 static timeout_t *lastfunc; 679 #endif 680 681 KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING, 682 ("softclock_call_cc: pend %p %x", c, c->c_iflags)); 683 KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE, 684 ("softclock_call_cc: act %p %x", c, c->c_flags)); 685 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL; 686 lock_status = 0; 687 if (c->c_flags & CALLOUT_SHAREDLOCK) { 688 if (class == &lock_class_rm) 689 lock_status = (uintptr_t)&tracker; 690 else 691 lock_status = 1; 692 } 693 c_lock = c->c_lock; 694 c_func = c->c_func; 695 c_arg = c->c_arg; 696 c_iflags = c->c_iflags; 697 if (c->c_iflags & CALLOUT_LOCAL_ALLOC) 698 c->c_iflags = CALLOUT_LOCAL_ALLOC; 699 else 700 c->c_iflags &= ~CALLOUT_PENDING; 701 702 cc_exec_curr(cc, direct) = c; 703 cc_exec_cancel(cc, direct) = false; 704 cc_exec_drain(cc, direct) = NULL; 705 CC_UNLOCK(cc); 706 if (c_lock != NULL) { 707 class->lc_lock(c_lock, lock_status); 708 /* 709 * The callout may have been cancelled 710 * while we switched locks. 711 */ 712 if (cc_exec_cancel(cc, direct)) { 713 class->lc_unlock(c_lock); 714 goto skip; 715 } 716 /* The callout cannot be stopped now. */ 717 cc_exec_cancel(cc, direct) = true; 718 if (c_lock == &Giant.lock_object) { 719 #ifdef CALLOUT_PROFILING 720 (*gcalls)++; 721 #endif 722 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p", 723 c, c_func, c_arg); 724 } else { 725 #ifdef CALLOUT_PROFILING 726 (*lockcalls)++; 727 #endif 728 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p", 729 c, c_func, c_arg); 730 } 731 } else { 732 #ifdef CALLOUT_PROFILING 733 (*mpcalls)++; 734 #endif 735 CTR3(KTR_CALLOUT, "callout %p func %p arg %p", 736 c, c_func, c_arg); 737 } 738 KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running", 739 "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct); 740 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 741 sbt1 = sbinuptime(); 742 #endif 743 THREAD_NO_SLEEPING(); 744 SDT_PROBE1(callout_execute, , , callout__start, c); 745 c_func(c_arg); 746 SDT_PROBE1(callout_execute, , , callout__end, c); 747 THREAD_SLEEPING_OK(); 748 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING) 749 sbt2 = sbinuptime(); 750 sbt2 -= sbt1; 751 if (sbt2 > maxdt) { 752 if (lastfunc != c_func || sbt2 > maxdt * 2) { 753 ts2 = sbttots(sbt2); 754 printf( 755 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n", 756 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec); 757 } 758 maxdt = sbt2; 759 lastfunc = c_func; 760 } 761 #endif 762 KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle"); 763 CTR1(KTR_CALLOUT, "callout %p finished", c); 764 if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0) 765 class->lc_unlock(c_lock); 766 skip: 767 CC_LOCK(cc); 768 KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr")); 769 cc_exec_curr(cc, direct) = NULL; 770 if (cc_exec_drain(cc, direct)) { 771 void (*drain)(void *); 772 773 drain = cc_exec_drain(cc, direct); 774 cc_exec_drain(cc, direct) = NULL; 775 CC_UNLOCK(cc); 776 drain(c_arg); 777 CC_LOCK(cc); 778 } 779 if (cc_exec_waiting(cc, direct)) { 780 /* 781 * There is someone waiting for the 782 * callout to complete. 783 * If the callout was scheduled for 784 * migration just cancel it. 785 */ 786 if (cc_cce_migrating(cc, direct)) { 787 cc_cce_cleanup(cc, direct); 788 789 /* 790 * It should be assert here that the callout is not 791 * destroyed but that is not easy. 792 */ 793 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 794 } 795 cc_exec_waiting(cc, direct) = false; 796 CC_UNLOCK(cc); 797 wakeup(&cc_exec_waiting(cc, direct)); 798 CC_LOCK(cc); 799 } else if (cc_cce_migrating(cc, direct)) { 800 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0, 801 ("Migrating legacy callout %p", c)); 802 #ifdef SMP 803 /* 804 * If the callout was scheduled for 805 * migration just perform it now. 806 */ 807 new_cpu = cc_migration_cpu(cc, direct); 808 new_time = cc_migration_time(cc, direct); 809 new_prec = cc_migration_prec(cc, direct); 810 new_func = cc_migration_func(cc, direct); 811 new_arg = cc_migration_arg(cc, direct); 812 cc_cce_cleanup(cc, direct); 813 814 /* 815 * It should be assert here that the callout is not destroyed 816 * but that is not easy. 817 * 818 * As first thing, handle deferred callout stops. 819 */ 820 if (!callout_migrating(c)) { 821 CTR3(KTR_CALLOUT, 822 "deferred cancelled %p func %p arg %p", 823 c, new_func, new_arg); 824 callout_cc_del(c, cc); 825 return; 826 } 827 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 828 829 new_cc = callout_cpu_switch(c, cc, new_cpu); 830 flags = (direct) ? C_DIRECT_EXEC : 0; 831 callout_cc_add(c, new_cc, new_time, new_prec, new_func, 832 new_arg, new_cpu, flags); 833 CC_UNLOCK(new_cc); 834 CC_LOCK(cc); 835 #else 836 panic("migration should not happen"); 837 #endif 838 } 839 /* 840 * If the current callout is locally allocated (from 841 * timeout(9)) then put it on the freelist. 842 * 843 * Note: we need to check the cached copy of c_iflags because 844 * if it was not local, then it's not safe to deref the 845 * callout pointer. 846 */ 847 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 || 848 c->c_iflags == CALLOUT_LOCAL_ALLOC, 849 ("corrupted callout")); 850 if (c_iflags & CALLOUT_LOCAL_ALLOC) 851 callout_cc_del(c, cc); 852 } 853 854 /* 855 * The callout mechanism is based on the work of Adam M. Costello and 856 * George Varghese, published in a technical report entitled "Redesigning 857 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 858 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 859 * used in this implementation was published by G. Varghese and T. Lauck in 860 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 861 * the Efficient Implementation of a Timer Facility" in the Proceedings of 862 * the 11th ACM Annual Symposium on Operating Systems Principles, 863 * Austin, Texas Nov 1987. 864 */ 865 866 /* 867 * Software (low priority) clock interrupt. 868 * Run periodic events from timeout queue. 869 */ 870 void 871 softclock(void *arg) 872 { 873 struct callout_cpu *cc; 874 struct callout *c; 875 #ifdef CALLOUT_PROFILING 876 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0; 877 #endif 878 879 cc = (struct callout_cpu *)arg; 880 CC_LOCK(cc); 881 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) { 882 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 883 softclock_call_cc(c, cc, 884 #ifdef CALLOUT_PROFILING 885 &mpcalls, &lockcalls, &gcalls, 886 #endif 887 0); 888 #ifdef CALLOUT_PROFILING 889 ++depth; 890 #endif 891 } 892 #ifdef CALLOUT_PROFILING 893 avg_depth += (depth * 1000 - avg_depth) >> 8; 894 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8; 895 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8; 896 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8; 897 #endif 898 CC_UNLOCK(cc); 899 } 900 901 /* 902 * timeout -- 903 * Execute a function after a specified length of time. 904 * 905 * untimeout -- 906 * Cancel previous timeout function call. 907 * 908 * callout_handle_init -- 909 * Initialize a handle so that using it with untimeout is benign. 910 * 911 * See AT&T BCI Driver Reference Manual for specification. This 912 * implementation differs from that one in that although an 913 * identification value is returned from timeout, the original 914 * arguments to timeout as well as the identifier are used to 915 * identify entries for untimeout. 916 */ 917 struct callout_handle 918 timeout(timeout_t *ftn, void *arg, int to_ticks) 919 { 920 struct callout_cpu *cc; 921 struct callout *new; 922 struct callout_handle handle; 923 924 cc = CC_CPU(timeout_cpu); 925 CC_LOCK(cc); 926 /* Fill in the next free callout structure. */ 927 new = SLIST_FIRST(&cc->cc_callfree); 928 if (new == NULL) 929 /* XXX Attempt to malloc first */ 930 panic("timeout table full"); 931 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle); 932 callout_reset(new, to_ticks, ftn, arg); 933 handle.callout = new; 934 CC_UNLOCK(cc); 935 936 return (handle); 937 } 938 939 void 940 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle) 941 { 942 struct callout_cpu *cc; 943 944 /* 945 * Check for a handle that was initialized 946 * by callout_handle_init, but never used 947 * for a real timeout. 948 */ 949 if (handle.callout == NULL) 950 return; 951 952 cc = callout_lock(handle.callout); 953 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg) 954 callout_stop(handle.callout); 955 CC_UNLOCK(cc); 956 } 957 958 void 959 callout_handle_init(struct callout_handle *handle) 960 { 961 handle->callout = NULL; 962 } 963 964 /* 965 * New interface; clients allocate their own callout structures. 966 * 967 * callout_reset() - establish or change a timeout 968 * callout_stop() - disestablish a timeout 969 * callout_init() - initialize a callout structure so that it can 970 * safely be passed to callout_reset() and callout_stop() 971 * 972 * <sys/callout.h> defines three convenience macros: 973 * 974 * callout_active() - returns truth if callout has not been stopped, 975 * drained, or deactivated since the last time the callout was 976 * reset. 977 * callout_pending() - returns truth if callout is still waiting for timeout 978 * callout_deactivate() - marks the callout as having been serviced 979 */ 980 int 981 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t precision, 982 void (*ftn)(void *), void *arg, int cpu, int flags) 983 { 984 sbintime_t to_sbt, pr; 985 struct callout_cpu *cc; 986 int cancelled, direct; 987 int ignore_cpu=0; 988 989 cancelled = 0; 990 if (cpu == -1) { 991 ignore_cpu = 1; 992 } else if ((cpu >= MAXCPU) || 993 ((CC_CPU(cpu))->cc_inited == 0)) { 994 /* Invalid CPU spec */ 995 panic("Invalid CPU in callout %d", cpu); 996 } 997 if (flags & C_ABSOLUTE) { 998 to_sbt = sbt; 999 } else { 1000 if ((flags & C_HARDCLOCK) && (sbt < tick_sbt)) 1001 sbt = tick_sbt; 1002 if ((flags & C_HARDCLOCK) || 1003 #ifdef NO_EVENTTIMERS 1004 sbt >= sbt_timethreshold) { 1005 to_sbt = getsbinuptime(); 1006 1007 /* Add safety belt for the case of hz > 1000. */ 1008 to_sbt += tc_tick_sbt - tick_sbt; 1009 #else 1010 sbt >= sbt_tickthreshold) { 1011 /* 1012 * Obtain the time of the last hardclock() call on 1013 * this CPU directly from the kern_clocksource.c. 1014 * This value is per-CPU, but it is equal for all 1015 * active ones. 1016 */ 1017 #ifdef __LP64__ 1018 to_sbt = DPCPU_GET(hardclocktime); 1019 #else 1020 spinlock_enter(); 1021 to_sbt = DPCPU_GET(hardclocktime); 1022 spinlock_exit(); 1023 #endif 1024 #endif 1025 if ((flags & C_HARDCLOCK) == 0) 1026 to_sbt += tick_sbt; 1027 } else 1028 to_sbt = sbinuptime(); 1029 if (SBT_MAX - to_sbt < sbt) 1030 to_sbt = SBT_MAX; 1031 else 1032 to_sbt += sbt; 1033 pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp : 1034 sbt >> C_PRELGET(flags)); 1035 if (pr > precision) 1036 precision = pr; 1037 } 1038 /* 1039 * This flag used to be added by callout_cc_add, but the 1040 * first time you call this we could end up with the 1041 * wrong direct flag if we don't do it before we add. 1042 */ 1043 if (flags & C_DIRECT_EXEC) { 1044 direct = 1; 1045 } else { 1046 direct = 0; 1047 } 1048 KASSERT(!direct || c->c_lock == NULL, 1049 ("%s: direct callout %p has lock", __func__, c)); 1050 cc = callout_lock(c); 1051 /* 1052 * Don't allow migration of pre-allocated callouts lest they 1053 * become unbalanced or handle the case where the user does 1054 * not care. 1055 */ 1056 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) || 1057 ignore_cpu) { 1058 cpu = c->c_cpu; 1059 } 1060 1061 if (cc_exec_curr(cc, direct) == c) { 1062 /* 1063 * We're being asked to reschedule a callout which is 1064 * currently in progress. If there is a lock then we 1065 * can cancel the callout if it has not really started. 1066 */ 1067 if (c->c_lock != NULL && !cc_exec_cancel(cc, direct)) 1068 cancelled = cc_exec_cancel(cc, direct) = true; 1069 if (cc_exec_waiting(cc, direct)) { 1070 /* 1071 * Someone has called callout_drain to kill this 1072 * callout. Don't reschedule. 1073 */ 1074 CTR4(KTR_CALLOUT, "%s %p func %p arg %p", 1075 cancelled ? "cancelled" : "failed to cancel", 1076 c, c->c_func, c->c_arg); 1077 CC_UNLOCK(cc); 1078 return (cancelled); 1079 } 1080 #ifdef SMP 1081 if (callout_migrating(c)) { 1082 /* 1083 * This only occurs when a second callout_reset_sbt_on 1084 * is made after a previous one moved it into 1085 * deferred migration (below). Note we do *not* change 1086 * the prev_cpu even though the previous target may 1087 * be different. 1088 */ 1089 cc_migration_cpu(cc, direct) = cpu; 1090 cc_migration_time(cc, direct) = to_sbt; 1091 cc_migration_prec(cc, direct) = precision; 1092 cc_migration_func(cc, direct) = ftn; 1093 cc_migration_arg(cc, direct) = arg; 1094 cancelled = 1; 1095 CC_UNLOCK(cc); 1096 return (cancelled); 1097 } 1098 #endif 1099 } 1100 if (c->c_iflags & CALLOUT_PENDING) { 1101 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1102 if (cc_exec_next(cc) == c) 1103 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1104 LIST_REMOVE(c, c_links.le); 1105 } else { 1106 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1107 } 1108 cancelled = 1; 1109 c->c_iflags &= ~ CALLOUT_PENDING; 1110 c->c_flags &= ~ CALLOUT_ACTIVE; 1111 } 1112 1113 #ifdef SMP 1114 /* 1115 * If the callout must migrate try to perform it immediately. 1116 * If the callout is currently running, just defer the migration 1117 * to a more appropriate moment. 1118 */ 1119 if (c->c_cpu != cpu) { 1120 if (cc_exec_curr(cc, direct) == c) { 1121 /* 1122 * Pending will have been removed since we are 1123 * actually executing the callout on another 1124 * CPU. That callout should be waiting on the 1125 * lock the caller holds. If we set both 1126 * active/and/pending after we return and the 1127 * lock on the executing callout proceeds, it 1128 * will then see pending is true and return. 1129 * At the return from the actual callout execution 1130 * the migration will occur in softclock_call_cc 1131 * and this new callout will be placed on the 1132 * new CPU via a call to callout_cpu_switch() which 1133 * will get the lock on the right CPU followed 1134 * by a call callout_cc_add() which will add it there. 1135 * (see above in softclock_call_cc()). 1136 */ 1137 cc_migration_cpu(cc, direct) = cpu; 1138 cc_migration_time(cc, direct) = to_sbt; 1139 cc_migration_prec(cc, direct) = precision; 1140 cc_migration_func(cc, direct) = ftn; 1141 cc_migration_arg(cc, direct) = arg; 1142 c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING); 1143 c->c_flags |= CALLOUT_ACTIVE; 1144 CTR6(KTR_CALLOUT, 1145 "migration of %p func %p arg %p in %d.%08x to %u deferred", 1146 c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1147 (u_int)(to_sbt & 0xffffffff), cpu); 1148 CC_UNLOCK(cc); 1149 return (cancelled); 1150 } 1151 cc = callout_cpu_switch(c, cc, cpu); 1152 } 1153 #endif 1154 1155 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags); 1156 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x", 1157 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32), 1158 (u_int)(to_sbt & 0xffffffff)); 1159 CC_UNLOCK(cc); 1160 1161 return (cancelled); 1162 } 1163 1164 /* 1165 * Common idioms that can be optimized in the future. 1166 */ 1167 int 1168 callout_schedule_on(struct callout *c, int to_ticks, int cpu) 1169 { 1170 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu); 1171 } 1172 1173 int 1174 callout_schedule(struct callout *c, int to_ticks) 1175 { 1176 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu); 1177 } 1178 1179 int 1180 _callout_stop_safe(struct callout *c, int flags, void (*drain)(void *)) 1181 { 1182 struct callout_cpu *cc, *old_cc; 1183 struct lock_class *class; 1184 int direct, sq_locked, use_lock; 1185 int cancelled, not_on_a_list; 1186 1187 if ((flags & CS_DRAIN) != 0) 1188 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock, 1189 "calling %s", __func__); 1190 1191 /* 1192 * Some old subsystems don't hold Giant while running a callout_stop(), 1193 * so just discard this check for the moment. 1194 */ 1195 if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) { 1196 if (c->c_lock == &Giant.lock_object) 1197 use_lock = mtx_owned(&Giant); 1198 else { 1199 use_lock = 1; 1200 class = LOCK_CLASS(c->c_lock); 1201 class->lc_assert(c->c_lock, LA_XLOCKED); 1202 } 1203 } else 1204 use_lock = 0; 1205 if (c->c_iflags & CALLOUT_DIRECT) { 1206 direct = 1; 1207 } else { 1208 direct = 0; 1209 } 1210 sq_locked = 0; 1211 old_cc = NULL; 1212 again: 1213 cc = callout_lock(c); 1214 1215 if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) == 1216 (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) && 1217 ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) { 1218 /* 1219 * Special case where this slipped in while we 1220 * were migrating *as* the callout is about to 1221 * execute. The caller probably holds the lock 1222 * the callout wants. 1223 * 1224 * Get rid of the migration first. Then set 1225 * the flag that tells this code *not* to 1226 * try to remove it from any lists (its not 1227 * on one yet). When the callout wheel runs, 1228 * it will ignore this callout. 1229 */ 1230 c->c_iflags &= ~CALLOUT_PENDING; 1231 c->c_flags &= ~CALLOUT_ACTIVE; 1232 not_on_a_list = 1; 1233 } else { 1234 not_on_a_list = 0; 1235 } 1236 1237 /* 1238 * If the callout was migrating while the callout cpu lock was 1239 * dropped, just drop the sleepqueue lock and check the states 1240 * again. 1241 */ 1242 if (sq_locked != 0 && cc != old_cc) { 1243 #ifdef SMP 1244 CC_UNLOCK(cc); 1245 sleepq_release(&cc_exec_waiting(old_cc, direct)); 1246 sq_locked = 0; 1247 old_cc = NULL; 1248 goto again; 1249 #else 1250 panic("migration should not happen"); 1251 #endif 1252 } 1253 1254 /* 1255 * If the callout is running, try to stop it or drain it. 1256 */ 1257 if (cc_exec_curr(cc, direct) == c) { 1258 /* 1259 * Succeed we to stop it or not, we must clear the 1260 * active flag - this is what API users expect. 1261 */ 1262 c->c_flags &= ~CALLOUT_ACTIVE; 1263 1264 if ((flags & CS_DRAIN) != 0) { 1265 /* 1266 * The current callout is running (or just 1267 * about to run) and blocking is allowed, so 1268 * just wait for the current invocation to 1269 * finish. 1270 */ 1271 while (cc_exec_curr(cc, direct) == c) { 1272 /* 1273 * Use direct calls to sleepqueue interface 1274 * instead of cv/msleep in order to avoid 1275 * a LOR between cc_lock and sleepqueue 1276 * chain spinlocks. This piece of code 1277 * emulates a msleep_spin() call actually. 1278 * 1279 * If we already have the sleepqueue chain 1280 * locked, then we can safely block. If we 1281 * don't already have it locked, however, 1282 * we have to drop the cc_lock to lock 1283 * it. This opens several races, so we 1284 * restart at the beginning once we have 1285 * both locks. If nothing has changed, then 1286 * we will end up back here with sq_locked 1287 * set. 1288 */ 1289 if (!sq_locked) { 1290 CC_UNLOCK(cc); 1291 sleepq_lock( 1292 &cc_exec_waiting(cc, direct)); 1293 sq_locked = 1; 1294 old_cc = cc; 1295 goto again; 1296 } 1297 1298 /* 1299 * Migration could be cancelled here, but 1300 * as long as it is still not sure when it 1301 * will be packed up, just let softclock() 1302 * take care of it. 1303 */ 1304 cc_exec_waiting(cc, direct) = true; 1305 DROP_GIANT(); 1306 CC_UNLOCK(cc); 1307 sleepq_add( 1308 &cc_exec_waiting(cc, direct), 1309 &cc->cc_lock.lock_object, "codrain", 1310 SLEEPQ_SLEEP, 0); 1311 sleepq_wait( 1312 &cc_exec_waiting(cc, direct), 1313 0); 1314 sq_locked = 0; 1315 old_cc = NULL; 1316 1317 /* Reacquire locks previously released. */ 1318 PICKUP_GIANT(); 1319 CC_LOCK(cc); 1320 } 1321 } else if (use_lock && 1322 !cc_exec_cancel(cc, direct) && (drain == NULL)) { 1323 1324 /* 1325 * The current callout is waiting for its 1326 * lock which we hold. Cancel the callout 1327 * and return. After our caller drops the 1328 * lock, the callout will be skipped in 1329 * softclock(). This *only* works with a 1330 * callout_stop() *not* callout_drain() or 1331 * callout_async_drain(). 1332 */ 1333 cc_exec_cancel(cc, direct) = true; 1334 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1335 c, c->c_func, c->c_arg); 1336 KASSERT(!cc_cce_migrating(cc, direct), 1337 ("callout wrongly scheduled for migration")); 1338 if (callout_migrating(c)) { 1339 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1340 #ifdef SMP 1341 cc_migration_cpu(cc, direct) = CPUBLOCK; 1342 cc_migration_time(cc, direct) = 0; 1343 cc_migration_prec(cc, direct) = 0; 1344 cc_migration_func(cc, direct) = NULL; 1345 cc_migration_arg(cc, direct) = NULL; 1346 #endif 1347 } 1348 CC_UNLOCK(cc); 1349 KASSERT(!sq_locked, ("sleepqueue chain locked")); 1350 return (1); 1351 } else if (callout_migrating(c)) { 1352 /* 1353 * The callout is currently being serviced 1354 * and the "next" callout is scheduled at 1355 * its completion with a migration. We remove 1356 * the migration flag so it *won't* get rescheduled, 1357 * but we can't stop the one thats running so 1358 * we return 0. 1359 */ 1360 c->c_iflags &= ~CALLOUT_DFRMIGRATION; 1361 #ifdef SMP 1362 /* 1363 * We can't call cc_cce_cleanup here since 1364 * if we do it will remove .ce_curr and 1365 * its still running. This will prevent a 1366 * reschedule of the callout when the 1367 * execution completes. 1368 */ 1369 cc_migration_cpu(cc, direct) = CPUBLOCK; 1370 cc_migration_time(cc, direct) = 0; 1371 cc_migration_prec(cc, direct) = 0; 1372 cc_migration_func(cc, direct) = NULL; 1373 cc_migration_arg(cc, direct) = NULL; 1374 #endif 1375 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p", 1376 c, c->c_func, c->c_arg); 1377 if (drain) { 1378 cc_exec_drain(cc, direct) = drain; 1379 } 1380 CC_UNLOCK(cc); 1381 return ((flags & CS_EXECUTING) != 0); 1382 } 1383 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1384 c, c->c_func, c->c_arg); 1385 if (drain) { 1386 cc_exec_drain(cc, direct) = drain; 1387 } 1388 KASSERT(!sq_locked, ("sleepqueue chain still locked")); 1389 cancelled = ((flags & CS_EXECUTING) != 0); 1390 } else 1391 cancelled = 1; 1392 1393 if (sq_locked) 1394 sleepq_release(&cc_exec_waiting(cc, direct)); 1395 1396 if ((c->c_iflags & CALLOUT_PENDING) == 0) { 1397 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p", 1398 c, c->c_func, c->c_arg); 1399 /* 1400 * For not scheduled and not executing callout return 1401 * negative value. 1402 */ 1403 if (cc_exec_curr(cc, direct) != c) 1404 cancelled = -1; 1405 CC_UNLOCK(cc); 1406 return (cancelled); 1407 } 1408 1409 c->c_iflags &= ~CALLOUT_PENDING; 1410 c->c_flags &= ~CALLOUT_ACTIVE; 1411 1412 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p", 1413 c, c->c_func, c->c_arg); 1414 if (not_on_a_list == 0) { 1415 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) { 1416 if (cc_exec_next(cc) == c) 1417 cc_exec_next(cc) = LIST_NEXT(c, c_links.le); 1418 LIST_REMOVE(c, c_links.le); 1419 } else { 1420 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe); 1421 } 1422 } 1423 callout_cc_del(c, cc); 1424 CC_UNLOCK(cc); 1425 return (cancelled); 1426 } 1427 1428 void 1429 callout_init(struct callout *c, int mpsafe) 1430 { 1431 bzero(c, sizeof *c); 1432 if (mpsafe) { 1433 c->c_lock = NULL; 1434 c->c_iflags = CALLOUT_RETURNUNLOCKED; 1435 } else { 1436 c->c_lock = &Giant.lock_object; 1437 c->c_iflags = 0; 1438 } 1439 c->c_cpu = timeout_cpu; 1440 } 1441 1442 void 1443 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags) 1444 { 1445 bzero(c, sizeof *c); 1446 c->c_lock = lock; 1447 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0, 1448 ("callout_init_lock: bad flags %d", flags)); 1449 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0, 1450 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock")); 1451 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & 1452 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class", 1453 __func__)); 1454 c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK); 1455 c->c_cpu = timeout_cpu; 1456 } 1457 1458 #ifdef APM_FIXUP_CALLTODO 1459 /* 1460 * Adjust the kernel calltodo timeout list. This routine is used after 1461 * an APM resume to recalculate the calltodo timer list values with the 1462 * number of hz's we have been sleeping. The next hardclock() will detect 1463 * that there are fired timers and run softclock() to execute them. 1464 * 1465 * Please note, I have not done an exhaustive analysis of what code this 1466 * might break. I am motivated to have my select()'s and alarm()'s that 1467 * have expired during suspend firing upon resume so that the applications 1468 * which set the timer can do the maintanence the timer was for as close 1469 * as possible to the originally intended time. Testing this code for a 1470 * week showed that resuming from a suspend resulted in 22 to 25 timers 1471 * firing, which seemed independent on whether the suspend was 2 hours or 1472 * 2 days. Your milage may vary. - Ken Key <[email protected]> 1473 */ 1474 void 1475 adjust_timeout_calltodo(struct timeval *time_change) 1476 { 1477 register struct callout *p; 1478 unsigned long delta_ticks; 1479 1480 /* 1481 * How many ticks were we asleep? 1482 * (stolen from tvtohz()). 1483 */ 1484 1485 /* Don't do anything */ 1486 if (time_change->tv_sec < 0) 1487 return; 1488 else if (time_change->tv_sec <= LONG_MAX / 1000000) 1489 delta_ticks = howmany(time_change->tv_sec * 1000000 + 1490 time_change->tv_usec, tick) + 1; 1491 else if (time_change->tv_sec <= LONG_MAX / hz) 1492 delta_ticks = time_change->tv_sec * hz + 1493 howmany(time_change->tv_usec, tick) + 1; 1494 else 1495 delta_ticks = LONG_MAX; 1496 1497 if (delta_ticks > INT_MAX) 1498 delta_ticks = INT_MAX; 1499 1500 /* 1501 * Now rip through the timer calltodo list looking for timers 1502 * to expire. 1503 */ 1504 1505 /* don't collide with softclock() */ 1506 CC_LOCK(cc); 1507 for (p = calltodo.c_next; p != NULL; p = p->c_next) { 1508 p->c_time -= delta_ticks; 1509 1510 /* Break if the timer had more time on it than delta_ticks */ 1511 if (p->c_time > 0) 1512 break; 1513 1514 /* take back the ticks the timer didn't use (p->c_time <= 0) */ 1515 delta_ticks = -p->c_time; 1516 } 1517 CC_UNLOCK(cc); 1518 1519 return; 1520 } 1521 #endif /* APM_FIXUP_CALLTODO */ 1522 1523 static int 1524 flssbt(sbintime_t sbt) 1525 { 1526 1527 sbt += (uint64_t)sbt >> 1; 1528 if (sizeof(long) >= sizeof(sbintime_t)) 1529 return (flsl(sbt)); 1530 if (sbt >= SBT_1S) 1531 return (flsl(((uint64_t)sbt) >> 32) + 32); 1532 return (flsl(sbt)); 1533 } 1534 1535 /* 1536 * Dump immediate statistic snapshot of the scheduled callouts. 1537 */ 1538 static int 1539 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS) 1540 { 1541 struct callout *tmp; 1542 struct callout_cpu *cc; 1543 struct callout_list *sc; 1544 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t; 1545 int ct[64], cpr[64], ccpbk[32]; 1546 int error, val, i, count, tcum, pcum, maxc, c, medc; 1547 #ifdef SMP 1548 int cpu; 1549 #endif 1550 1551 val = 0; 1552 error = sysctl_handle_int(oidp, &val, 0, req); 1553 if (error != 0 || req->newptr == NULL) 1554 return (error); 1555 count = maxc = 0; 1556 st = spr = maxt = maxpr = 0; 1557 bzero(ccpbk, sizeof(ccpbk)); 1558 bzero(ct, sizeof(ct)); 1559 bzero(cpr, sizeof(cpr)); 1560 now = sbinuptime(); 1561 #ifdef SMP 1562 CPU_FOREACH(cpu) { 1563 cc = CC_CPU(cpu); 1564 #else 1565 cc = CC_CPU(timeout_cpu); 1566 #endif 1567 CC_LOCK(cc); 1568 for (i = 0; i < callwheelsize; i++) { 1569 sc = &cc->cc_callwheel[i]; 1570 c = 0; 1571 LIST_FOREACH(tmp, sc, c_links.le) { 1572 c++; 1573 t = tmp->c_time - now; 1574 if (t < 0) 1575 t = 0; 1576 st += t / SBT_1US; 1577 spr += tmp->c_precision / SBT_1US; 1578 if (t > maxt) 1579 maxt = t; 1580 if (tmp->c_precision > maxpr) 1581 maxpr = tmp->c_precision; 1582 ct[flssbt(t)]++; 1583 cpr[flssbt(tmp->c_precision)]++; 1584 } 1585 if (c > maxc) 1586 maxc = c; 1587 ccpbk[fls(c + c / 2)]++; 1588 count += c; 1589 } 1590 CC_UNLOCK(cc); 1591 #ifdef SMP 1592 } 1593 #endif 1594 1595 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++) 1596 tcum += ct[i]; 1597 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1598 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++) 1599 pcum += cpr[i]; 1600 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0; 1601 for (i = 0, c = 0; i < 32 && c < count / 2; i++) 1602 c += ccpbk[i]; 1603 medc = (i >= 2) ? (1 << (i - 2)) : 0; 1604 1605 printf("Scheduled callouts statistic snapshot:\n"); 1606 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n", 1607 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT); 1608 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n", 1609 medc, 1610 count / callwheelsize / mp_ncpus, 1611 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000, 1612 maxc); 1613 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1614 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32, 1615 (st / count) / 1000000, (st / count) % 1000000, 1616 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32); 1617 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n", 1618 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32, 1619 (spr / count) / 1000000, (spr / count) % 1000000, 1620 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32); 1621 printf(" Distribution: \tbuckets\t time\t tcum\t" 1622 " prec\t pcum\n"); 1623 for (i = 0, tcum = pcum = 0; i < 64; i++) { 1624 if (ct[i] == 0 && cpr[i] == 0) 1625 continue; 1626 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0; 1627 tcum += ct[i]; 1628 pcum += cpr[i]; 1629 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n", 1630 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32, 1631 i - 1 - (32 - CC_HASH_SHIFT), 1632 ct[i], tcum, cpr[i], pcum); 1633 } 1634 return (error); 1635 } 1636 SYSCTL_PROC(_kern, OID_AUTO, callout_stat, 1637 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1638 0, 0, sysctl_kern_callout_stat, "I", 1639 "Dump immediate statistic snapshot of the scheduled callouts"); 1640 1641 #ifdef FSTACK 1642 void ff_hardclock(void); 1643 1644 void 1645 ff_hardclock(void) 1646 { 1647 atomic_add_int(&ticks, 1); 1648 callout_process(getsbinuptime()); 1649 tc_ticktock(1); 1650 cpu_tick_calibration(); 1651 1652 #ifdef DEVICE_POLLING 1653 hardclock_device_poll(); /* this is very short and quick */ 1654 #endif /* DEVICE_POLLING */ 1655 } 1656 1657 static unsigned int 1658 ff_tc_get_timecount(struct timecounter *tc) 1659 { 1660 static u_int now; 1661 return (++now); 1662 } 1663 1664 static struct timecounter ff_timecounter = { 1665 ff_tc_get_timecount, 0, ~0u, 100, "ff_clock", 1 1666 }; 1667 1668 static void 1669 ff_tc_init(void) 1670 { 1671 ff_timecounter.tc_frequency = hz; 1672 tc_init(&ff_timecounter); 1673 } 1674 SYSINIT(ff_tc, SI_SUB_SMP, SI_ORDER_ANY, ff_tc_init, NULL); 1675 #endif 1676 1677 #ifdef DDB 1678 static void 1679 _show_callout(struct callout *c) 1680 { 1681 1682 db_printf("callout %p\n", c); 1683 #define C_DB_PRINTF(f, e) db_printf(" %s = " f "\n", #e, c->e); 1684 db_printf(" &c_links = %p\n", &(c->c_links)); 1685 C_DB_PRINTF("%" PRId64, c_time); 1686 C_DB_PRINTF("%" PRId64, c_precision); 1687 C_DB_PRINTF("%p", c_arg); 1688 C_DB_PRINTF("%p", c_func); 1689 C_DB_PRINTF("%p", c_lock); 1690 C_DB_PRINTF("%#x", c_flags); 1691 C_DB_PRINTF("%#x", c_iflags); 1692 C_DB_PRINTF("%d", c_cpu); 1693 #undef C_DB_PRINTF 1694 } 1695 1696 DB_SHOW_COMMAND(callout, db_show_callout) 1697 { 1698 1699 if (!have_addr) { 1700 db_printf("usage: show callout <struct callout *>\n"); 1701 return; 1702 } 1703 1704 _show_callout((struct callout *)addr); 1705 } 1706 #endif /* DDB */ 1707