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 <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/callout.h> 43 #include <sys/condvar.h> 44 #include <sys/kernel.h> 45 #include <sys/ktr.h> 46 #include <sys/lock.h> 47 #include <sys/mutex.h> 48 #include <sys/proc.h> 49 #include <sys/sysctl.h> 50 51 static int avg_depth; 52 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0, 53 "Average number of items examined per softclock call. Units = 1/1000"); 54 static int avg_gcalls; 55 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0, 56 "Average number of Giant callouts made per softclock call. Units = 1/1000"); 57 static int avg_mtxcalls; 58 SYSCTL_INT(_debug, OID_AUTO, to_avg_mtxcalls, CTLFLAG_RD, &avg_mtxcalls, 0, 59 "Average number of mtx callouts made per softclock call. Units = 1/1000"); 60 static int avg_mpcalls; 61 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0, 62 "Average number of MP callouts made per softclock call. Units = 1/1000"); 63 /* 64 * TODO: 65 * allocate more timeout table slots when table overflows. 66 */ 67 68 /* Exported to machdep.c and/or kern_clock.c. */ 69 struct callout *callout; 70 struct callout_list callfree; 71 int callwheelsize, callwheelbits, callwheelmask; 72 struct callout_tailq *callwheel; 73 int softticks; /* Like ticks, but for softclock(). */ 74 struct mtx callout_lock; 75 76 static struct callout *nextsoftcheck; /* Next callout to be checked. */ 77 78 /** 79 * Locked by callout_lock: 80 * curr_callout - If a callout is in progress, it is curr_callout. 81 * If curr_callout is non-NULL, threads waiting in 82 * callout_drain() will be woken up as soon as the 83 * relevant callout completes. 84 * curr_cancelled - Changing to 1 with both callout_lock and c_mtx held 85 * guarantees that the current callout will not run. 86 * The softclock() function sets this to 0 before it 87 * drops callout_lock to acquire c_mtx, and it calls 88 * the handler only if curr_cancelled is still 0 after 89 * c_mtx is successfully acquired. 90 * callout_wait - If a thread is waiting in callout_drain(), then 91 * callout_wait is nonzero. Set only when 92 * curr_callout is non-NULL. 93 */ 94 static struct callout *curr_callout; 95 static int curr_cancelled; 96 static int callout_wait; 97 98 /* 99 * kern_timeout_callwheel_alloc() - kernel low level callwheel initialization 100 * 101 * This code is called very early in the kernel initialization sequence, 102 * and may be called more then once. 103 */ 104 caddr_t 105 kern_timeout_callwheel_alloc(caddr_t v) 106 { 107 /* 108 * Calculate callout wheel size 109 */ 110 for (callwheelsize = 1, callwheelbits = 0; 111 callwheelsize < ncallout; 112 callwheelsize <<= 1, ++callwheelbits) 113 ; 114 callwheelmask = callwheelsize - 1; 115 116 callout = (struct callout *)v; 117 v = (caddr_t)(callout + ncallout); 118 callwheel = (struct callout_tailq *)v; 119 v = (caddr_t)(callwheel + callwheelsize); 120 return(v); 121 } 122 123 /* 124 * kern_timeout_callwheel_init() - initialize previously reserved callwheel 125 * space. 126 * 127 * This code is called just once, after the space reserved for the 128 * callout wheel has been finalized. 129 */ 130 void 131 kern_timeout_callwheel_init(void) 132 { 133 int i; 134 135 SLIST_INIT(&callfree); 136 for (i = 0; i < ncallout; i++) { 137 callout_init(&callout[i], 0); 138 callout[i].c_flags = CALLOUT_LOCAL_ALLOC; 139 SLIST_INSERT_HEAD(&callfree, &callout[i], c_links.sle); 140 } 141 for (i = 0; i < callwheelsize; i++) { 142 TAILQ_INIT(&callwheel[i]); 143 } 144 mtx_init(&callout_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE); 145 } 146 147 /* 148 * The callout mechanism is based on the work of Adam M. Costello and 149 * George Varghese, published in a technical report entitled "Redesigning 150 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 151 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 152 * used in this implementation was published by G. Varghese and T. Lauck in 153 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 154 * the Efficient Implementation of a Timer Facility" in the Proceedings of 155 * the 11th ACM Annual Symposium on Operating Systems Principles, 156 * Austin, Texas Nov 1987. 157 */ 158 159 /* 160 * Software (low priority) clock interrupt. 161 * Run periodic events from timeout queue. 162 */ 163 void 164 softclock(void *dummy) 165 { 166 struct callout *c; 167 struct callout_tailq *bucket; 168 int curticks; 169 int steps; /* #steps since we last allowed interrupts */ 170 int depth; 171 int mpcalls; 172 int mtxcalls; 173 int gcalls; 174 #ifdef DIAGNOSTIC 175 struct bintime bt1, bt2; 176 struct timespec ts2; 177 static uint64_t maxdt = 36893488147419102LL; /* 2 msec */ 178 static timeout_t *lastfunc; 179 #endif 180 181 #ifndef MAX_SOFTCLOCK_STEPS 182 #define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */ 183 #endif /* MAX_SOFTCLOCK_STEPS */ 184 185 mpcalls = 0; 186 mtxcalls = 0; 187 gcalls = 0; 188 depth = 0; 189 steps = 0; 190 mtx_lock_spin(&callout_lock); 191 while (softticks != ticks) { 192 softticks++; 193 /* 194 * softticks may be modified by hard clock, so cache 195 * it while we work on a given bucket. 196 */ 197 curticks = softticks; 198 bucket = &callwheel[curticks & callwheelmask]; 199 c = TAILQ_FIRST(bucket); 200 while (c) { 201 depth++; 202 if (c->c_time != curticks) { 203 c = TAILQ_NEXT(c, c_links.tqe); 204 ++steps; 205 if (steps >= MAX_SOFTCLOCK_STEPS) { 206 nextsoftcheck = c; 207 /* Give interrupts a chance. */ 208 mtx_unlock_spin(&callout_lock); 209 ; /* nothing */ 210 mtx_lock_spin(&callout_lock); 211 c = nextsoftcheck; 212 steps = 0; 213 } 214 } else { 215 void (*c_func)(void *); 216 void *c_arg; 217 struct mtx *c_mtx; 218 int c_flags; 219 220 nextsoftcheck = TAILQ_NEXT(c, c_links.tqe); 221 TAILQ_REMOVE(bucket, c, c_links.tqe); 222 c_func = c->c_func; 223 c_arg = c->c_arg; 224 c_mtx = c->c_mtx; 225 c_flags = c->c_flags; 226 if (c->c_flags & CALLOUT_LOCAL_ALLOC) { 227 c->c_func = NULL; 228 c->c_flags = CALLOUT_LOCAL_ALLOC; 229 SLIST_INSERT_HEAD(&callfree, c, 230 c_links.sle); 231 curr_callout = NULL; 232 } else { 233 c->c_flags = 234 (c->c_flags & ~CALLOUT_PENDING); 235 curr_callout = c; 236 } 237 curr_cancelled = 0; 238 mtx_unlock_spin(&callout_lock); 239 if (c_mtx != NULL) { 240 mtx_lock(c_mtx); 241 /* 242 * The callout may have been cancelled 243 * while we switched locks. 244 */ 245 if (curr_cancelled) { 246 mtx_unlock(c_mtx); 247 goto skip; 248 } 249 /* The callout cannot be stopped now. */ 250 curr_cancelled = 1; 251 252 if (c_mtx == &Giant) { 253 gcalls++; 254 CTR1(KTR_CALLOUT, "callout %p", 255 c_func); 256 } else { 257 mtxcalls++; 258 CTR1(KTR_CALLOUT, 259 "callout mtx %p", 260 c_func); 261 } 262 } else { 263 mpcalls++; 264 CTR1(KTR_CALLOUT, "callout mpsafe %p", 265 c_func); 266 } 267 #ifdef DIAGNOSTIC 268 binuptime(&bt1); 269 #endif 270 THREAD_NO_SLEEPING(); 271 c_func(c_arg); 272 THREAD_SLEEPING_OK(); 273 #ifdef DIAGNOSTIC 274 binuptime(&bt2); 275 bintime_sub(&bt2, &bt1); 276 if (bt2.frac > maxdt) { 277 if (lastfunc != c_func || 278 bt2.frac > maxdt * 2) { 279 bintime2timespec(&bt2, &ts2); 280 printf( 281 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n", 282 c_func, c_arg, 283 (intmax_t)ts2.tv_sec, 284 ts2.tv_nsec); 285 } 286 maxdt = bt2.frac; 287 lastfunc = c_func; 288 } 289 #endif 290 if ((c_flags & CALLOUT_RETURNUNLOCKED) == 0) 291 mtx_unlock(c_mtx); 292 skip: 293 mtx_lock_spin(&callout_lock); 294 curr_callout = NULL; 295 if (callout_wait) { 296 /* 297 * There is someone waiting 298 * for the callout to complete. 299 */ 300 wakeup(&callout_wait); 301 callout_wait = 0; 302 } 303 steps = 0; 304 c = nextsoftcheck; 305 } 306 } 307 } 308 avg_depth += (depth * 1000 - avg_depth) >> 8; 309 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8; 310 avg_mtxcalls += (mtxcalls * 1000 - avg_mtxcalls) >> 8; 311 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8; 312 nextsoftcheck = NULL; 313 mtx_unlock_spin(&callout_lock); 314 } 315 316 /* 317 * timeout -- 318 * Execute a function after a specified length of time. 319 * 320 * untimeout -- 321 * Cancel previous timeout function call. 322 * 323 * callout_handle_init -- 324 * Initialize a handle so that using it with untimeout is benign. 325 * 326 * See AT&T BCI Driver Reference Manual for specification. This 327 * implementation differs from that one in that although an 328 * identification value is returned from timeout, the original 329 * arguments to timeout as well as the identifier are used to 330 * identify entries for untimeout. 331 */ 332 struct callout_handle 333 timeout(ftn, arg, to_ticks) 334 timeout_t *ftn; 335 void *arg; 336 int to_ticks; 337 { 338 struct callout *new; 339 struct callout_handle handle; 340 341 mtx_lock_spin(&callout_lock); 342 343 /* Fill in the next free callout structure. */ 344 new = SLIST_FIRST(&callfree); 345 if (new == NULL) 346 /* XXX Attempt to malloc first */ 347 panic("timeout table full"); 348 SLIST_REMOVE_HEAD(&callfree, c_links.sle); 349 350 callout_reset(new, to_ticks, ftn, arg); 351 352 handle.callout = new; 353 mtx_unlock_spin(&callout_lock); 354 return (handle); 355 } 356 357 void 358 untimeout(ftn, arg, handle) 359 timeout_t *ftn; 360 void *arg; 361 struct callout_handle handle; 362 { 363 364 /* 365 * Check for a handle that was initialized 366 * by callout_handle_init, but never used 367 * for a real timeout. 368 */ 369 if (handle.callout == NULL) 370 return; 371 372 mtx_lock_spin(&callout_lock); 373 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg) 374 callout_stop(handle.callout); 375 mtx_unlock_spin(&callout_lock); 376 } 377 378 void 379 callout_handle_init(struct callout_handle *handle) 380 { 381 handle->callout = NULL; 382 } 383 384 /* 385 * New interface; clients allocate their own callout structures. 386 * 387 * callout_reset() - establish or change a timeout 388 * callout_stop() - disestablish a timeout 389 * callout_init() - initialize a callout structure so that it can 390 * safely be passed to callout_reset() and callout_stop() 391 * 392 * <sys/callout.h> defines three convenience macros: 393 * 394 * callout_active() - returns truth if callout has not been stopped, 395 * drained, or deactivated since the last time the callout was 396 * reset. 397 * callout_pending() - returns truth if callout is still waiting for timeout 398 * callout_deactivate() - marks the callout as having been serviced 399 */ 400 int 401 callout_reset(c, to_ticks, ftn, arg) 402 struct callout *c; 403 int to_ticks; 404 void (*ftn)(void *); 405 void *arg; 406 { 407 int cancelled = 0; 408 409 #ifdef notyet /* Some callers of timeout() do not hold Giant. */ 410 if (c->c_mtx != NULL) 411 mtx_assert(c->c_mtx, MA_OWNED); 412 #endif 413 414 mtx_lock_spin(&callout_lock); 415 if (c == curr_callout) { 416 /* 417 * We're being asked to reschedule a callout which is 418 * currently in progress. If there is a mutex then we 419 * can cancel the callout if it has not really started. 420 */ 421 if (c->c_mtx != NULL && !curr_cancelled) 422 cancelled = curr_cancelled = 1; 423 if (callout_wait) { 424 /* 425 * Someone has called callout_drain to kill this 426 * callout. Don't reschedule. 427 */ 428 mtx_unlock_spin(&callout_lock); 429 return (cancelled); 430 } 431 } 432 if (c->c_flags & CALLOUT_PENDING) { 433 if (nextsoftcheck == c) { 434 nextsoftcheck = TAILQ_NEXT(c, c_links.tqe); 435 } 436 TAILQ_REMOVE(&callwheel[c->c_time & callwheelmask], c, 437 c_links.tqe); 438 439 cancelled = 1; 440 441 /* 442 * Part of the normal "stop a pending callout" process 443 * is to clear the CALLOUT_ACTIVE and CALLOUT_PENDING 444 * flags. We're not going to bother doing that here, 445 * because we're going to be setting those flags ten lines 446 * after this point, and we're holding callout_lock 447 * between now and then. 448 */ 449 } 450 451 /* 452 * We could unlock callout_lock here and lock it again before the 453 * TAILQ_INSERT_TAIL, but there's no point since doing this setup 454 * doesn't take much time. 455 */ 456 if (to_ticks <= 0) 457 to_ticks = 1; 458 459 c->c_arg = arg; 460 c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING); 461 c->c_func = ftn; 462 c->c_time = ticks + to_ticks; 463 TAILQ_INSERT_TAIL(&callwheel[c->c_time & callwheelmask], 464 c, c_links.tqe); 465 mtx_unlock_spin(&callout_lock); 466 467 return (cancelled); 468 } 469 470 int 471 _callout_stop_safe(c, safe) 472 struct callout *c; 473 int safe; 474 { 475 int use_mtx; 476 477 if (!safe && c->c_mtx != NULL) { 478 #ifdef notyet /* Some callers do not hold Giant for Giant-locked callouts. */ 479 mtx_assert(c->c_mtx, MA_OWNED); 480 use_mtx = 1; 481 #else 482 use_mtx = mtx_owned(c->c_mtx); 483 #endif 484 } else { 485 use_mtx = 0; 486 } 487 488 mtx_lock_spin(&callout_lock); 489 /* 490 * If the callout isn't pending, it's not on the queue, so 491 * don't attempt to remove it from the queue. We can try to 492 * stop it by other means however. 493 */ 494 if (!(c->c_flags & CALLOUT_PENDING)) { 495 c->c_flags &= ~CALLOUT_ACTIVE; 496 497 /* 498 * If it wasn't on the queue and it isn't the current 499 * callout, then we can't stop it, so just bail. 500 */ 501 if (c != curr_callout) { 502 mtx_unlock_spin(&callout_lock); 503 return (0); 504 } 505 506 if (safe) { 507 /* 508 * The current callout is running (or just 509 * about to run) and blocking is allowed, so 510 * just wait for the current invocation to 511 * finish. 512 */ 513 while (c == curr_callout) { 514 callout_wait = 1; 515 msleep_spin(&callout_wait, &callout_lock, 516 "codrain", 0); 517 } 518 } else if (use_mtx && !curr_cancelled) { 519 /* 520 * The current callout is waiting for it's 521 * mutex which we hold. Cancel the callout 522 * and return. After our caller drops the 523 * mutex, the callout will be skipped in 524 * softclock(). 525 */ 526 curr_cancelled = 1; 527 mtx_unlock_spin(&callout_lock); 528 return (1); 529 } 530 mtx_unlock_spin(&callout_lock); 531 return (0); 532 } 533 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING); 534 535 if (nextsoftcheck == c) { 536 nextsoftcheck = TAILQ_NEXT(c, c_links.tqe); 537 } 538 TAILQ_REMOVE(&callwheel[c->c_time & callwheelmask], c, c_links.tqe); 539 540 if (c->c_flags & CALLOUT_LOCAL_ALLOC) { 541 c->c_func = NULL; 542 SLIST_INSERT_HEAD(&callfree, c, c_links.sle); 543 } 544 mtx_unlock_spin(&callout_lock); 545 return (1); 546 } 547 548 void 549 callout_init(c, mpsafe) 550 struct callout *c; 551 int mpsafe; 552 { 553 bzero(c, sizeof *c); 554 if (mpsafe) { 555 c->c_mtx = NULL; 556 c->c_flags = CALLOUT_RETURNUNLOCKED; 557 } else { 558 c->c_mtx = &Giant; 559 c->c_flags = 0; 560 } 561 } 562 563 void 564 callout_init_mtx(c, mtx, flags) 565 struct callout *c; 566 struct mtx *mtx; 567 int flags; 568 { 569 bzero(c, sizeof *c); 570 c->c_mtx = mtx; 571 KASSERT((flags & ~CALLOUT_RETURNUNLOCKED) == 0, 572 ("callout_init_mtx: bad flags %d", flags)); 573 /* CALLOUT_RETURNUNLOCKED makes no sense without a mutex. */ 574 KASSERT(mtx != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0, 575 ("callout_init_mtx: CALLOUT_RETURNUNLOCKED with no mutex")); 576 c->c_flags = flags & CALLOUT_RETURNUNLOCKED; 577 } 578 579 #ifdef APM_FIXUP_CALLTODO 580 /* 581 * Adjust the kernel calltodo timeout list. This routine is used after 582 * an APM resume to recalculate the calltodo timer list values with the 583 * number of hz's we have been sleeping. The next hardclock() will detect 584 * that there are fired timers and run softclock() to execute them. 585 * 586 * Please note, I have not done an exhaustive analysis of what code this 587 * might break. I am motivated to have my select()'s and alarm()'s that 588 * have expired during suspend firing upon resume so that the applications 589 * which set the timer can do the maintanence the timer was for as close 590 * as possible to the originally intended time. Testing this code for a 591 * week showed that resuming from a suspend resulted in 22 to 25 timers 592 * firing, which seemed independant on whether the suspend was 2 hours or 593 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu> 594 */ 595 void 596 adjust_timeout_calltodo(time_change) 597 struct timeval *time_change; 598 { 599 register struct callout *p; 600 unsigned long delta_ticks; 601 602 /* 603 * How many ticks were we asleep? 604 * (stolen from tvtohz()). 605 */ 606 607 /* Don't do anything */ 608 if (time_change->tv_sec < 0) 609 return; 610 else if (time_change->tv_sec <= LONG_MAX / 1000000) 611 delta_ticks = (time_change->tv_sec * 1000000 + 612 time_change->tv_usec + (tick - 1)) / tick + 1; 613 else if (time_change->tv_sec <= LONG_MAX / hz) 614 delta_ticks = time_change->tv_sec * hz + 615 (time_change->tv_usec + (tick - 1)) / tick + 1; 616 else 617 delta_ticks = LONG_MAX; 618 619 if (delta_ticks > INT_MAX) 620 delta_ticks = INT_MAX; 621 622 /* 623 * Now rip through the timer calltodo list looking for timers 624 * to expire. 625 */ 626 627 /* don't collide with softclock() */ 628 mtx_lock_spin(&callout_lock); 629 for (p = calltodo.c_next; p != NULL; p = p->c_next) { 630 p->c_time -= delta_ticks; 631 632 /* Break if the timer had more time on it than delta_ticks */ 633 if (p->c_time > 0) 634 break; 635 636 /* take back the ticks the timer didn't use (p->c_time <= 0) */ 637 delta_ticks = -p->c_time; 638 } 639 mtx_unlock_spin(&callout_lock); 640 641 return; 642 } 643 #endif /* APM_FIXUP_CALLTODO */ 644