1 /* $OpenBSD: kern_timeout.c,v 1.99 2024/08/11 00:49:34 dlg Exp $ */ 2 /* 3 * Copyright (c) 2001 Thomas Nordin <nordin@openbsd.org> 4 * Copyright (c) 2000-2001 Artur Grabowski <art@openbsd.org> 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. The name of the author may not be used to endorse or promote products 14 * derived from this software without specific prior written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, 17 * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY 18 * AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL 19 * THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 20 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 21 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; 22 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 23 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR 24 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF 25 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 */ 27 28 #include <sys/param.h> 29 #include <sys/systm.h> 30 #include <sys/kthread.h> 31 #include <sys/proc.h> 32 #include <sys/timeout.h> 33 #include <sys/mutex.h> 34 #include <sys/kernel.h> 35 #include <sys/queue.h> /* _Q_INVALIDATE */ 36 #include <sys/sysctl.h> 37 #include <sys/witness.h> 38 39 #ifdef DDB 40 #include <machine/db_machdep.h> 41 #include <ddb/db_interface.h> 42 #include <ddb/db_sym.h> 43 #include <ddb/db_output.h> 44 #endif 45 46 #include "kcov.h" 47 #if NKCOV > 0 48 #include <sys/kcov.h> 49 #endif 50 51 /* 52 * Locks used to protect global variables in this file: 53 * 54 * I immutable after initialization 55 * T timeout_mutex 56 */ 57 struct mutex timeout_mutex = MUTEX_INITIALIZER(IPL_HIGH); 58 59 void *softclock_si; /* [I] softclock() interrupt handle */ 60 struct timeoutstat tostat; /* [T] statistics and totals */ 61 62 /* 63 * Timeouts are kept in a hierarchical timing wheel. The to_time is the value 64 * of the global variable "ticks" when the timeout should be called. There are 65 * four levels with 256 buckets each. 66 */ 67 #define WHEELCOUNT 4 68 #define WHEELSIZE 256 69 #define WHEELMASK 255 70 #define WHEELBITS 8 71 #define BUCKETS (WHEELCOUNT * WHEELSIZE) 72 73 struct circq timeout_wheel[BUCKETS]; /* [T] Tick-based timeouts */ 74 struct circq timeout_wheel_kc[BUCKETS]; /* [T] Clock-based timeouts */ 75 struct circq timeout_new; /* [T] New, unscheduled timeouts */ 76 struct circq timeout_todo; /* [T] Due or needs rescheduling */ 77 struct circq timeout_proc; /* [T] Due + needs process context */ 78 #ifdef MULTIPROCESSOR 79 struct circq timeout_proc_mp; /* [T] Process ctx + no kernel lock */ 80 #endif 81 82 time_t timeout_level_width[WHEELCOUNT]; /* [I] Wheel level width (seconds) */ 83 struct timespec tick_ts; /* [I] Length of a tick (1/hz secs) */ 84 85 struct kclock { 86 struct timespec kc_lastscan; /* [T] Clock time at last wheel scan */ 87 struct timespec kc_late; /* [T] Late if due prior */ 88 struct timespec kc_offset; /* [T] Offset from primary kclock */ 89 } timeout_kclock[KCLOCK_MAX]; 90 91 #define MASKWHEEL(wheel, time) (((time) >> ((wheel)*WHEELBITS)) & WHEELMASK) 92 93 #define BUCKET(rel, abs) \ 94 (timeout_wheel[ \ 95 ((rel) <= (1 << (2*WHEELBITS))) \ 96 ? ((rel) <= (1 << WHEELBITS)) \ 97 ? MASKWHEEL(0, (abs)) \ 98 : MASKWHEEL(1, (abs)) + WHEELSIZE \ 99 : ((rel) <= (1 << (3*WHEELBITS))) \ 100 ? MASKWHEEL(2, (abs)) + 2*WHEELSIZE \ 101 : MASKWHEEL(3, (abs)) + 3*WHEELSIZE]) 102 103 #define MOVEBUCKET(wheel, time) \ 104 CIRCQ_CONCAT(&timeout_todo, \ 105 &timeout_wheel[MASKWHEEL((wheel), (time)) + (wheel)*WHEELSIZE]) 106 107 /* 108 * Circular queue definitions. 109 */ 110 111 #define CIRCQ_INIT(elem) do { \ 112 (elem)->next = (elem); \ 113 (elem)->prev = (elem); \ 114 } while (0) 115 116 #define CIRCQ_INSERT_TAIL(list, elem) do { \ 117 (elem)->prev = (list)->prev; \ 118 (elem)->next = (list); \ 119 (list)->prev->next = (elem); \ 120 (list)->prev = (elem); \ 121 tostat.tos_pending++; \ 122 } while (0) 123 124 #define CIRCQ_CONCAT(fst, snd) do { \ 125 if (!CIRCQ_EMPTY(snd)) { \ 126 (fst)->prev->next = (snd)->next;\ 127 (snd)->next->prev = (fst)->prev;\ 128 (snd)->prev->next = (fst); \ 129 (fst)->prev = (snd)->prev; \ 130 CIRCQ_INIT(snd); \ 131 } \ 132 } while (0) 133 134 #define CIRCQ_REMOVE(elem) do { \ 135 (elem)->next->prev = (elem)->prev; \ 136 (elem)->prev->next = (elem)->next; \ 137 _Q_INVALIDATE((elem)->prev); \ 138 _Q_INVALIDATE((elem)->next); \ 139 tostat.tos_pending--; \ 140 } while (0) 141 142 #define CIRCQ_FIRST(elem) ((elem)->next) 143 144 #define CIRCQ_EMPTY(elem) (CIRCQ_FIRST(elem) == (elem)) 145 146 #define CIRCQ_FOREACH(elem, list) \ 147 for ((elem) = CIRCQ_FIRST(list); \ 148 (elem) != (list); \ 149 (elem) = CIRCQ_FIRST(elem)) 150 151 #ifdef WITNESS 152 struct lock_object timeout_sleeplock_obj = { 153 .lo_name = "timeout", 154 .lo_flags = LO_WITNESS | LO_INITIALIZED | LO_SLEEPABLE | 155 (LO_CLASS_RWLOCK << LO_CLASSSHIFT) 156 }; 157 struct lock_object timeout_spinlock_obj = { 158 .lo_name = "timeout", 159 .lo_flags = LO_WITNESS | LO_INITIALIZED | 160 (LO_CLASS_MUTEX << LO_CLASSSHIFT) 161 }; 162 struct lock_type timeout_sleeplock_type = { 163 .lt_name = "timeout" 164 }; 165 struct lock_type timeout_spinlock_type = { 166 .lt_name = "timeout" 167 }; 168 #define TIMEOUT_LOCK_OBJ(needsproc) \ 169 ((needsproc) ? &timeout_sleeplock_obj : &timeout_spinlock_obj) 170 #endif 171 172 void softclock(void *); 173 void softclock_create_thread(void *); 174 void softclock_process_kclock_timeout(struct timeout *, int); 175 void softclock_process_tick_timeout(struct timeout *, int); 176 void softclock_thread(void *); 177 #ifdef MULTIPROCESSOR 178 void softclock_thread_mp(void *); 179 #endif 180 void timeout_barrier_timeout(void *); 181 uint32_t timeout_bucket(const struct timeout *); 182 uint32_t timeout_maskwheel(uint32_t, const struct timespec *); 183 void timeout_run(struct timeout *); 184 185 /* 186 * The first thing in a struct timeout is its struct circq, so we 187 * can get back from a pointer to the latter to a pointer to the 188 * whole timeout with just a cast. 189 */ 190 static inline struct timeout * 191 timeout_from_circq(struct circq *p) 192 { 193 return ((struct timeout *)(p)); 194 } 195 196 static inline void 197 timeout_sync_order(int needsproc) 198 { 199 WITNESS_CHECKORDER(TIMEOUT_LOCK_OBJ(needsproc), LOP_NEWORDER, NULL); 200 } 201 202 static inline void 203 timeout_sync_enter(int needsproc) 204 { 205 timeout_sync_order(needsproc); 206 WITNESS_LOCK(TIMEOUT_LOCK_OBJ(needsproc), 0); 207 } 208 209 static inline void 210 timeout_sync_leave(int needsproc) 211 { 212 WITNESS_UNLOCK(TIMEOUT_LOCK_OBJ(needsproc), 0); 213 } 214 215 /* 216 * Some of the "math" in here is a bit tricky. 217 * 218 * We have to beware of wrapping ints. 219 * We use the fact that any element added to the queue must be added with a 220 * positive time. That means that any element `to' on the queue cannot be 221 * scheduled to timeout further in time than INT_MAX, but to->to_time can 222 * be positive or negative so comparing it with anything is dangerous. 223 * The only way we can use the to->to_time value in any predictable way 224 * is when we calculate how far in the future `to' will timeout - 225 * "to->to_time - ticks". The result will always be positive for future 226 * timeouts and 0 or negative for due timeouts. 227 */ 228 229 void 230 timeout_startup(void) 231 { 232 int b, level; 233 234 CIRCQ_INIT(&timeout_new); 235 CIRCQ_INIT(&timeout_todo); 236 CIRCQ_INIT(&timeout_proc); 237 #ifdef MULTIPROCESSOR 238 CIRCQ_INIT(&timeout_proc_mp); 239 #endif 240 for (b = 0; b < nitems(timeout_wheel); b++) 241 CIRCQ_INIT(&timeout_wheel[b]); 242 for (b = 0; b < nitems(timeout_wheel_kc); b++) 243 CIRCQ_INIT(&timeout_wheel_kc[b]); 244 245 for (level = 0; level < nitems(timeout_level_width); level++) 246 timeout_level_width[level] = 2 << (level * WHEELBITS); 247 NSEC_TO_TIMESPEC(tick_nsec, &tick_ts); 248 } 249 250 void 251 timeout_proc_init(void) 252 { 253 softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL); 254 if (softclock_si == NULL) 255 panic("%s: unable to register softclock interrupt", __func__); 256 257 WITNESS_INIT(&timeout_sleeplock_obj, &timeout_sleeplock_type); 258 WITNESS_INIT(&timeout_spinlock_obj, &timeout_spinlock_type); 259 260 kthread_create_deferred(softclock_create_thread, NULL); 261 } 262 263 void 264 timeout_set(struct timeout *new, void (*fn)(void *), void *arg) 265 { 266 timeout_set_flags(new, fn, arg, KCLOCK_NONE, 0); 267 } 268 269 void 270 timeout_set_flags(struct timeout *to, void (*fn)(void *), void *arg, int kclock, 271 int flags) 272 { 273 KASSERT(!ISSET(flags, ~(TIMEOUT_PROC | TIMEOUT_MPSAFE))); 274 KASSERT(kclock >= KCLOCK_NONE && kclock < KCLOCK_MAX); 275 276 to->to_func = fn; 277 to->to_arg = arg; 278 to->to_kclock = kclock; 279 to->to_flags = flags | TIMEOUT_INITIALIZED; 280 281 /* For now, only process context timeouts may be marked MP-safe. */ 282 if (ISSET(to->to_flags, TIMEOUT_MPSAFE)) 283 KASSERT(ISSET(to->to_flags, TIMEOUT_PROC)); 284 } 285 286 void 287 timeout_set_proc(struct timeout *new, void (*fn)(void *), void *arg) 288 { 289 timeout_set_flags(new, fn, arg, KCLOCK_NONE, TIMEOUT_PROC); 290 } 291 292 int 293 timeout_add(struct timeout *new, int to_ticks) 294 { 295 int old_time; 296 int ret = 1; 297 298 KASSERT(ISSET(new->to_flags, TIMEOUT_INITIALIZED)); 299 KASSERT(new->to_kclock == KCLOCK_NONE); 300 KASSERT(to_ticks >= 0); 301 302 mtx_enter(&timeout_mutex); 303 304 /* Initialize the time here, it won't change. */ 305 old_time = new->to_time; 306 new->to_time = to_ticks + ticks; 307 CLR(new->to_flags, TIMEOUT_TRIGGERED); 308 309 /* 310 * If this timeout already is scheduled and now is moved 311 * earlier, reschedule it now. Otherwise leave it in place 312 * and let it be rescheduled later. 313 */ 314 if (ISSET(new->to_flags, TIMEOUT_ONQUEUE)) { 315 if (new->to_time - ticks < old_time - ticks) { 316 CIRCQ_REMOVE(&new->to_list); 317 CIRCQ_INSERT_TAIL(&timeout_new, &new->to_list); 318 } 319 tostat.tos_readded++; 320 ret = 0; 321 } else { 322 SET(new->to_flags, TIMEOUT_ONQUEUE); 323 CIRCQ_INSERT_TAIL(&timeout_new, &new->to_list); 324 } 325 #if NKCOV > 0 326 if (!kcov_cold) 327 new->to_process = curproc->p_p; 328 #endif 329 tostat.tos_added++; 330 mtx_leave(&timeout_mutex); 331 332 return ret; 333 } 334 335 static inline int 336 timeout_add_ticks(struct timeout *to, uint64_t to_ticks, int notzero) 337 { 338 if (to_ticks > INT_MAX) 339 to_ticks = INT_MAX; 340 else if (to_ticks == 0 && notzero) 341 to_ticks = 1; 342 343 return timeout_add(to, (int)to_ticks); 344 } 345 346 int 347 timeout_add_tv(struct timeout *to, const struct timeval *tv) 348 { 349 uint64_t to_ticks; 350 351 to_ticks = (uint64_t)hz * tv->tv_sec + tv->tv_usec / tick; 352 353 return timeout_add_ticks(to, to_ticks, tv->tv_usec > 0); 354 } 355 356 int 357 timeout_add_sec(struct timeout *to, int secs) 358 { 359 uint64_t to_ticks; 360 361 to_ticks = (uint64_t)hz * secs; 362 363 return timeout_add_ticks(to, to_ticks, 1); 364 } 365 366 int 367 timeout_add_msec(struct timeout *to, uint64_t msecs) 368 { 369 uint64_t to_ticks; 370 371 to_ticks = msecs * 1000 / tick; 372 373 return timeout_add_ticks(to, to_ticks, msecs > 0); 374 } 375 376 int 377 timeout_add_usec(struct timeout *to, uint64_t usecs) 378 { 379 uint64_t to_ticks; 380 381 to_ticks = usecs / tick; 382 383 return timeout_add_ticks(to, to_ticks, usecs > 0); 384 } 385 386 int 387 timeout_add_nsec(struct timeout *to, uint64_t nsecs) 388 { 389 uint64_t to_ticks; 390 391 to_ticks = nsecs / (tick * 1000); 392 393 return timeout_add_ticks(to, to_ticks, nsecs > 0); 394 } 395 396 int 397 timeout_abs_ts(struct timeout *to, const struct timespec *abstime) 398 { 399 struct timespec old_abstime; 400 int ret = 1; 401 402 mtx_enter(&timeout_mutex); 403 404 KASSERT(ISSET(to->to_flags, TIMEOUT_INITIALIZED)); 405 KASSERT(to->to_kclock == KCLOCK_UPTIME); 406 407 old_abstime = to->to_abstime; 408 to->to_abstime = *abstime; 409 CLR(to->to_flags, TIMEOUT_TRIGGERED); 410 411 if (ISSET(to->to_flags, TIMEOUT_ONQUEUE)) { 412 if (timespeccmp(abstime, &old_abstime, <)) { 413 CIRCQ_REMOVE(&to->to_list); 414 CIRCQ_INSERT_TAIL(&timeout_new, &to->to_list); 415 } 416 tostat.tos_readded++; 417 ret = 0; 418 } else { 419 SET(to->to_flags, TIMEOUT_ONQUEUE); 420 CIRCQ_INSERT_TAIL(&timeout_new, &to->to_list); 421 } 422 #if NKCOV > 0 423 if (!kcov_cold) 424 to->to_process = curproc->p_p; 425 #endif 426 tostat.tos_added++; 427 428 mtx_leave(&timeout_mutex); 429 430 return ret; 431 } 432 433 int 434 timeout_del(struct timeout *to) 435 { 436 int ret = 0; 437 438 mtx_enter(&timeout_mutex); 439 if (ISSET(to->to_flags, TIMEOUT_ONQUEUE)) { 440 CIRCQ_REMOVE(&to->to_list); 441 CLR(to->to_flags, TIMEOUT_ONQUEUE); 442 tostat.tos_cancelled++; 443 ret = 1; 444 } 445 CLR(to->to_flags, TIMEOUT_TRIGGERED); 446 tostat.tos_deleted++; 447 mtx_leave(&timeout_mutex); 448 449 return ret; 450 } 451 452 int 453 timeout_del_barrier(struct timeout *to) 454 { 455 int removed; 456 457 timeout_sync_order(ISSET(to->to_flags, TIMEOUT_PROC)); 458 459 removed = timeout_del(to); 460 if (!removed) 461 timeout_barrier(to); 462 463 return removed; 464 } 465 466 void 467 timeout_barrier(struct timeout *to) 468 { 469 struct timeout barrier; 470 struct cond c; 471 int flags; 472 473 flags = to->to_flags & (TIMEOUT_PROC | TIMEOUT_MPSAFE); 474 timeout_sync_order(ISSET(flags, TIMEOUT_PROC)); 475 476 timeout_set_flags(&barrier, timeout_barrier_timeout, &c, KCLOCK_NONE, 477 flags); 478 barrier.to_process = curproc->p_p; 479 cond_init(&c); 480 481 mtx_enter(&timeout_mutex); 482 483 barrier.to_time = ticks; 484 SET(barrier.to_flags, TIMEOUT_ONQUEUE); 485 if (ISSET(flags, TIMEOUT_PROC)) { 486 #ifdef MULTIPROCESSOR 487 if (ISSET(flags, TIMEOUT_MPSAFE)) 488 CIRCQ_INSERT_TAIL(&timeout_proc_mp, &barrier.to_list); 489 else 490 #endif 491 CIRCQ_INSERT_TAIL(&timeout_proc, &barrier.to_list); 492 } else 493 CIRCQ_INSERT_TAIL(&timeout_todo, &barrier.to_list); 494 495 mtx_leave(&timeout_mutex); 496 497 if (ISSET(flags, TIMEOUT_PROC)) { 498 #ifdef MULTIPROCESSOR 499 if (ISSET(flags, TIMEOUT_MPSAFE)) 500 wakeup_one(&timeout_proc_mp); 501 else 502 #endif 503 wakeup_one(&timeout_proc); 504 } else 505 softintr_schedule(softclock_si); 506 507 cond_wait(&c, "tmobar"); 508 } 509 510 void 511 timeout_barrier_timeout(void *arg) 512 { 513 struct cond *c = arg; 514 515 cond_signal(c); 516 } 517 518 uint32_t 519 timeout_bucket(const struct timeout *to) 520 { 521 struct timespec diff, shifted_abstime; 522 struct kclock *kc; 523 uint32_t level; 524 525 KASSERT(to->to_kclock == KCLOCK_UPTIME); 526 kc = &timeout_kclock[to->to_kclock]; 527 528 KASSERT(timespeccmp(&kc->kc_lastscan, &to->to_abstime, <)); 529 timespecsub(&to->to_abstime, &kc->kc_lastscan, &diff); 530 for (level = 0; level < nitems(timeout_level_width) - 1; level++) { 531 if (diff.tv_sec < timeout_level_width[level]) 532 break; 533 } 534 timespecadd(&to->to_abstime, &kc->kc_offset, &shifted_abstime); 535 return level * WHEELSIZE + timeout_maskwheel(level, &shifted_abstime); 536 } 537 538 /* 539 * Hash the absolute time into a bucket on a given level of the wheel. 540 * 541 * The complete hash is 32 bits. The upper 25 bits are seconds, the 542 * lower 7 bits are nanoseconds. tv_nsec is a positive value less 543 * than one billion so we need to divide it to isolate the desired 544 * bits. We can't just shift it. 545 * 546 * The level is used to isolate an 8-bit portion of the hash. The 547 * resulting number indicates which bucket the absolute time belongs 548 * in on the given level of the wheel. 549 */ 550 uint32_t 551 timeout_maskwheel(uint32_t level, const struct timespec *abstime) 552 { 553 uint32_t hi, lo; 554 555 hi = abstime->tv_sec << 7; 556 lo = abstime->tv_nsec / 7812500; 557 558 return ((hi | lo) >> (level * WHEELBITS)) & WHEELMASK; 559 } 560 561 /* 562 * This is called from hardclock() on the primary CPU at the start of 563 * every tick. 564 */ 565 void 566 timeout_hardclock_update(void) 567 { 568 struct timespec elapsed, now; 569 struct kclock *kc; 570 struct timespec *lastscan = &timeout_kclock[KCLOCK_UPTIME].kc_lastscan; 571 int b, done, first, i, last, level, need_softclock = 1, off; 572 573 mtx_enter(&timeout_mutex); 574 575 MOVEBUCKET(0, ticks); 576 if (MASKWHEEL(0, ticks) == 0) { 577 MOVEBUCKET(1, ticks); 578 if (MASKWHEEL(1, ticks) == 0) { 579 MOVEBUCKET(2, ticks); 580 if (MASKWHEEL(2, ticks) == 0) 581 MOVEBUCKET(3, ticks); 582 } 583 } 584 585 /* 586 * Dump the buckets that expired while we were away. 587 * 588 * If the elapsed time has exceeded a level's limit then we need 589 * to dump every bucket in the level. We have necessarily completed 590 * a lap of that level, too, so we need to process buckets in the 591 * next level. 592 * 593 * Otherwise we need to compare indices: if the index of the first 594 * expired bucket is greater than that of the last then we have 595 * completed a lap of the level and need to process buckets in the 596 * next level. 597 */ 598 nanouptime(&now); 599 timespecsub(&now, lastscan, &elapsed); 600 for (level = 0; level < nitems(timeout_level_width); level++) { 601 first = timeout_maskwheel(level, lastscan); 602 if (elapsed.tv_sec >= timeout_level_width[level]) { 603 last = (first == 0) ? WHEELSIZE - 1 : first - 1; 604 done = 0; 605 } else { 606 last = timeout_maskwheel(level, &now); 607 done = first <= last; 608 } 609 off = level * WHEELSIZE; 610 for (b = first;; b = (b + 1) % WHEELSIZE) { 611 CIRCQ_CONCAT(&timeout_todo, &timeout_wheel_kc[off + b]); 612 if (b == last) 613 break; 614 } 615 if (done) 616 break; 617 } 618 619 /* 620 * Update the cached state for each kclock. 621 */ 622 for (i = 0; i < nitems(timeout_kclock); i++) { 623 kc = &timeout_kclock[i]; 624 timespecadd(&now, &kc->kc_offset, &kc->kc_lastscan); 625 timespecsub(&kc->kc_lastscan, &tick_ts, &kc->kc_late); 626 } 627 628 if (CIRCQ_EMPTY(&timeout_new) && CIRCQ_EMPTY(&timeout_todo)) 629 need_softclock = 0; 630 631 mtx_leave(&timeout_mutex); 632 633 if (need_softclock) 634 softintr_schedule(softclock_si); 635 } 636 637 void 638 timeout_run(struct timeout *to) 639 { 640 void (*fn)(void *); 641 void *arg; 642 int needsproc; 643 644 MUTEX_ASSERT_LOCKED(&timeout_mutex); 645 646 CLR(to->to_flags, TIMEOUT_ONQUEUE); 647 SET(to->to_flags, TIMEOUT_TRIGGERED); 648 649 fn = to->to_func; 650 arg = to->to_arg; 651 needsproc = ISSET(to->to_flags, TIMEOUT_PROC); 652 #if NKCOV > 0 653 struct process *kcov_process = to->to_process; 654 #endif 655 656 mtx_leave(&timeout_mutex); 657 timeout_sync_enter(needsproc); 658 #if NKCOV > 0 659 kcov_remote_enter(KCOV_REMOTE_COMMON, kcov_process); 660 #endif 661 fn(arg); 662 #if NKCOV > 0 663 kcov_remote_leave(KCOV_REMOTE_COMMON, kcov_process); 664 #endif 665 timeout_sync_leave(needsproc); 666 mtx_enter(&timeout_mutex); 667 } 668 669 void 670 softclock_process_kclock_timeout(struct timeout *to, int new) 671 { 672 struct kclock *kc = &timeout_kclock[to->to_kclock]; 673 674 if (timespeccmp(&to->to_abstime, &kc->kc_lastscan, >)) { 675 tostat.tos_scheduled++; 676 if (!new) 677 tostat.tos_rescheduled++; 678 CIRCQ_INSERT_TAIL(&timeout_wheel_kc[timeout_bucket(to)], 679 &to->to_list); 680 return; 681 } 682 if (!new && timespeccmp(&to->to_abstime, &kc->kc_late, <=)) 683 tostat.tos_late++; 684 if (ISSET(to->to_flags, TIMEOUT_PROC)) { 685 #ifdef MULTIPROCESSOR 686 if (ISSET(to->to_flags, TIMEOUT_MPSAFE)) 687 CIRCQ_INSERT_TAIL(&timeout_proc_mp, &to->to_list); 688 else 689 #endif 690 CIRCQ_INSERT_TAIL(&timeout_proc, &to->to_list); 691 return; 692 } 693 timeout_run(to); 694 tostat.tos_run_softclock++; 695 } 696 697 void 698 softclock_process_tick_timeout(struct timeout *to, int new) 699 { 700 int delta = to->to_time - ticks; 701 702 if (delta > 0) { 703 tostat.tos_scheduled++; 704 if (!new) 705 tostat.tos_rescheduled++; 706 CIRCQ_INSERT_TAIL(&BUCKET(delta, to->to_time), &to->to_list); 707 return; 708 } 709 if (!new && delta < 0) 710 tostat.tos_late++; 711 if (ISSET(to->to_flags, TIMEOUT_PROC)) { 712 #ifdef MULTIPROCESSOR 713 if (ISSET(to->to_flags, TIMEOUT_MPSAFE)) 714 CIRCQ_INSERT_TAIL(&timeout_proc_mp, &to->to_list); 715 else 716 #endif 717 CIRCQ_INSERT_TAIL(&timeout_proc, &to->to_list); 718 return; 719 } 720 timeout_run(to); 721 tostat.tos_run_softclock++; 722 } 723 724 /* 725 * Timeouts are processed here instead of timeout_hardclock_update() 726 * to avoid doing any more work at IPL_CLOCK than absolutely necessary. 727 * Down here at IPL_SOFTCLOCK other interrupts can be serviced promptly 728 * so the system remains responsive even if there is a surge of timeouts. 729 */ 730 void 731 softclock(void *arg) 732 { 733 struct timeout *first_new, *to; 734 int needsproc, new; 735 #ifdef MULTIPROCESSOR 736 int need_proc_mp; 737 #endif 738 739 first_new = NULL; 740 new = 0; 741 742 mtx_enter(&timeout_mutex); 743 if (!CIRCQ_EMPTY(&timeout_new)) 744 first_new = timeout_from_circq(CIRCQ_FIRST(&timeout_new)); 745 CIRCQ_CONCAT(&timeout_todo, &timeout_new); 746 while (!CIRCQ_EMPTY(&timeout_todo)) { 747 to = timeout_from_circq(CIRCQ_FIRST(&timeout_todo)); 748 CIRCQ_REMOVE(&to->to_list); 749 if (to == first_new) 750 new = 1; 751 if (to->to_kclock == KCLOCK_NONE) 752 softclock_process_tick_timeout(to, new); 753 else if (to->to_kclock == KCLOCK_UPTIME) 754 softclock_process_kclock_timeout(to, new); 755 else { 756 panic("%s: invalid to_clock: %d", 757 __func__, to->to_kclock); 758 } 759 } 760 tostat.tos_softclocks++; 761 needsproc = !CIRCQ_EMPTY(&timeout_proc); 762 #ifdef MULTIPROCESSOR 763 need_proc_mp = !CIRCQ_EMPTY(&timeout_proc_mp); 764 #endif 765 mtx_leave(&timeout_mutex); 766 767 if (needsproc) 768 wakeup(&timeout_proc); 769 #ifdef MULTIPROCESSOR 770 if (need_proc_mp) 771 wakeup(&timeout_proc_mp); 772 #endif 773 } 774 775 void 776 softclock_create_thread(void *arg) 777 { 778 if (kthread_create(softclock_thread, NULL, NULL, "softclock")) 779 panic("fork softclock"); 780 #ifdef MULTIPROCESSOR 781 if (kthread_create(softclock_thread_mp, NULL, NULL, "softclockmp")) 782 panic("kthread_create softclock_thread_mp"); 783 #endif 784 } 785 786 void 787 softclock_thread(void *arg) 788 { 789 CPU_INFO_ITERATOR cii; 790 struct cpu_info *ci; 791 struct timeout *to; 792 int s; 793 794 KERNEL_ASSERT_LOCKED(); 795 796 /* Be conservative for the moment */ 797 CPU_INFO_FOREACH(cii, ci) { 798 if (CPU_IS_PRIMARY(ci)) 799 break; 800 } 801 KASSERT(ci != NULL); 802 sched_peg_curproc(ci); 803 804 s = splsoftclock(); 805 mtx_enter(&timeout_mutex); 806 for (;;) { 807 while (!CIRCQ_EMPTY(&timeout_proc)) { 808 to = timeout_from_circq(CIRCQ_FIRST(&timeout_proc)); 809 CIRCQ_REMOVE(&to->to_list); 810 timeout_run(to); 811 tostat.tos_run_thread++; 812 } 813 tostat.tos_thread_wakeups++; 814 msleep_nsec(&timeout_proc, &timeout_mutex, PSWP, "tmoslp", 815 INFSLP); 816 } 817 splx(s); 818 } 819 820 #ifdef MULTIPROCESSOR 821 void 822 softclock_thread_mp(void *arg) 823 { 824 struct timeout *to; 825 826 KERNEL_ASSERT_LOCKED(); 827 KERNEL_UNLOCK(); 828 829 mtx_enter(&timeout_mutex); 830 for (;;) { 831 while (!CIRCQ_EMPTY(&timeout_proc_mp)) { 832 to = timeout_from_circq(CIRCQ_FIRST(&timeout_proc_mp)); 833 CIRCQ_REMOVE(&to->to_list); 834 timeout_run(to); 835 tostat.tos_run_thread++; 836 } 837 tostat.tos_thread_wakeups++; 838 msleep_nsec(&timeout_proc_mp, &timeout_mutex, PSWP, "tmoslp", 839 INFSLP); 840 } 841 } 842 #endif /* MULTIPROCESSOR */ 843 844 #ifndef SMALL_KERNEL 845 void 846 timeout_adjust_ticks(int adj) 847 { 848 struct timeout *to; 849 struct circq *p; 850 int new_ticks, b; 851 852 /* adjusting the monotonic clock backwards would be a Bad Thing */ 853 if (adj <= 0) 854 return; 855 856 mtx_enter(&timeout_mutex); 857 new_ticks = ticks + adj; 858 for (b = 0; b < nitems(timeout_wheel); b++) { 859 p = CIRCQ_FIRST(&timeout_wheel[b]); 860 while (p != &timeout_wheel[b]) { 861 to = timeout_from_circq(p); 862 p = CIRCQ_FIRST(p); 863 864 /* when moving a timeout forward need to reinsert it */ 865 if (to->to_time - ticks < adj) 866 to->to_time = new_ticks; 867 CIRCQ_REMOVE(&to->to_list); 868 CIRCQ_INSERT_TAIL(&timeout_todo, &to->to_list); 869 } 870 } 871 ticks = new_ticks; 872 mtx_leave(&timeout_mutex); 873 } 874 #endif 875 876 int 877 timeout_sysctl(void *oldp, size_t *oldlenp, void *newp, size_t newlen) 878 { 879 struct timeoutstat status; 880 881 mtx_enter(&timeout_mutex); 882 memcpy(&status, &tostat, sizeof(status)); 883 mtx_leave(&timeout_mutex); 884 885 return sysctl_rdstruct(oldp, oldlenp, newp, &status, sizeof(status)); 886 } 887 888 #ifdef DDB 889 const char *db_kclock(int); 890 void db_show_callout_bucket(struct circq *); 891 void db_show_timeout(struct timeout *, struct circq *); 892 const char *db_timespec(const struct timespec *); 893 894 const char * 895 db_kclock(int kclock) 896 { 897 switch (kclock) { 898 case KCLOCK_UPTIME: 899 return "uptime"; 900 default: 901 return "invalid"; 902 } 903 } 904 905 const char * 906 db_timespec(const struct timespec *ts) 907 { 908 static char buf[32]; 909 struct timespec tmp, zero; 910 911 if (ts->tv_sec >= 0) { 912 snprintf(buf, sizeof(buf), "%lld.%09ld", 913 ts->tv_sec, ts->tv_nsec); 914 return buf; 915 } 916 917 timespecclear(&zero); 918 timespecsub(&zero, ts, &tmp); 919 snprintf(buf, sizeof(buf), "-%lld.%09ld", tmp.tv_sec, tmp.tv_nsec); 920 return buf; 921 } 922 923 void 924 db_show_callout_bucket(struct circq *bucket) 925 { 926 struct circq *p; 927 928 CIRCQ_FOREACH(p, bucket) 929 db_show_timeout(timeout_from_circq(p), bucket); 930 } 931 932 void 933 db_show_timeout(struct timeout *to, struct circq *bucket) 934 { 935 struct timespec remaining; 936 struct kclock *kc; 937 char buf[8]; 938 db_expr_t offset; 939 struct circq *wheel; 940 char *name, *where; 941 int width = sizeof(long) * 2; 942 943 db_find_sym_and_offset((vaddr_t)to->to_func, &name, &offset); 944 name = name ? name : "?"; 945 if (bucket == &timeout_new) 946 where = "new"; 947 else if (bucket == &timeout_todo) 948 where = "softint"; 949 else if (bucket == &timeout_proc) 950 where = "thread"; 951 #ifdef MULTIPROCESSOR 952 else if (bucket == &timeout_proc_mp) 953 where = "thread-mp"; 954 #endif 955 else { 956 if (to->to_kclock == KCLOCK_UPTIME) 957 wheel = timeout_wheel_kc; 958 else if (to->to_kclock == KCLOCK_NONE) 959 wheel = timeout_wheel; 960 else 961 goto invalid; 962 snprintf(buf, sizeof(buf), "%3ld/%1ld", 963 (bucket - wheel) % WHEELSIZE, 964 (bucket - wheel) / WHEELSIZE); 965 where = buf; 966 } 967 if (to->to_kclock == KCLOCK_UPTIME) { 968 kc = &timeout_kclock[to->to_kclock]; 969 timespecsub(&to->to_abstime, &kc->kc_lastscan, &remaining); 970 db_printf("%20s %8s %9s 0x%0*lx %s\n", 971 db_timespec(&remaining), db_kclock(to->to_kclock), where, 972 width, (ulong)to->to_arg, name); 973 } else if (to->to_kclock == KCLOCK_NONE) { 974 db_printf("%20d %8s %9s 0x%0*lx %s\n", 975 to->to_time - ticks, "ticks", where, 976 width, (ulong)to->to_arg, name); 977 } else 978 goto invalid; 979 return; 980 981 invalid: 982 db_printf("%s: timeout 0x%p: invalid to_kclock: %d", 983 __func__, to, to->to_kclock); 984 } 985 986 void 987 db_show_callout(db_expr_t addr, int haddr, db_expr_t count, char *modif) 988 { 989 struct kclock *kc; 990 int width = sizeof(long) * 2 + 2; 991 int b, i; 992 993 db_printf("%20s %8s\n", "lastscan", "clock"); 994 db_printf("%20d %8s\n", ticks, "ticks"); 995 for (i = 0; i < nitems(timeout_kclock); i++) { 996 kc = &timeout_kclock[i]; 997 db_printf("%20s %8s\n", 998 db_timespec(&kc->kc_lastscan), db_kclock(i)); 999 } 1000 db_printf("\n"); 1001 db_printf("%20s %8s %9s %*s %s\n", 1002 "remaining", "clock", "wheel", width, "arg", "func"); 1003 db_show_callout_bucket(&timeout_new); 1004 db_show_callout_bucket(&timeout_todo); 1005 db_show_callout_bucket(&timeout_proc); 1006 #ifdef MULTIPROCESSOR 1007 db_show_callout_bucket(&timeout_proc_mp); 1008 #endif 1009 for (b = 0; b < nitems(timeout_wheel); b++) 1010 db_show_callout_bucket(&timeout_wheel[b]); 1011 for (b = 0; b < nitems(timeout_wheel_kc); b++) 1012 db_show_callout_bucket(&timeout_wheel_kc[b]); 1013 } 1014 #endif 1015