1 /* 2 * Copyright (c) 2004 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 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. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 /* 35 * Copyright (c) 1982, 1986, 1991, 1993 36 * The Regents of the University of California. All rights reserved. 37 * (c) UNIX System Laboratories, Inc. 38 * All or some portions of this file are derived from material licensed 39 * to the University of California by American Telephone and Telegraph 40 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 41 * the permission of UNIX System Laboratories, Inc. 42 * 43 * Redistribution and use in source and binary forms, with or without 44 * modification, are permitted provided that the following conditions 45 * are met: 46 * 1. Redistributions of source code must retain the above copyright 47 * notice, this list of conditions and the following disclaimer. 48 * 2. Redistributions in binary form must reproduce the above copyright 49 * notice, this list of conditions and the following disclaimer in the 50 * documentation and/or other materials provided with the distribution. 51 * 3. All advertising materials mentioning features or use of this software 52 * must display the following acknowledgement: 53 * This product includes software developed by the University of 54 * California, Berkeley and its contributors. 55 * 4. Neither the name of the University nor the names of its contributors 56 * may be used to endorse or promote products derived from this software 57 * without specific prior written permission. 58 * 59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 69 * SUCH DAMAGE. 70 * 71 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 72 * $FreeBSD: src/sys/kern/kern_timeout.c,v 1.59.2.1 2001/11/13 18:24:52 archie Exp $ 73 * $DragonFly: src/sys/kern/kern_timeout.c,v 1.27 2007/11/14 18:27:52 swildner Exp $ 74 */ 75 /* 76 * DRAGONFLY BGL STATUS 77 * 78 * All the API functions should be MP safe. 79 * 80 * The callback functions will be flagged as being MP safe if the 81 * timeout structure is initialized with callout_init_mp() instead of 82 * callout_init(). 83 * 84 * The helper threads cannot be made preempt-capable until after we 85 * clean up all the uses of splsoftclock() and related interlocks (which 86 * require the related functions to be MP safe as well). 87 */ 88 /* 89 * The callout mechanism is based on the work of Adam M. Costello and 90 * George Varghese, published in a technical report entitled "Redesigning 91 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 92 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 93 * used in this implementation was published by G. Varghese and T. Lauck in 94 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 95 * the Efficient Implementation of a Timer Facility" in the Proceedings of 96 * the 11th ACM Annual Symposium on Operating Systems Principles, 97 * Austin, Texas Nov 1987. 98 * 99 * The per-cpu augmentation was done by Matthew Dillon. 100 */ 101 102 #include <sys/param.h> 103 #include <sys/systm.h> 104 #include <sys/callout.h> 105 #include <sys/kernel.h> 106 #include <sys/interrupt.h> 107 #include <sys/thread.h> 108 109 #include <sys/thread2.h> 110 #include <sys/mplock2.h> 111 112 #ifndef MAX_SOFTCLOCK_STEPS 113 #define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */ 114 #endif 115 116 117 struct softclock_pcpu { 118 struct callout_tailq *callwheel; 119 struct callout * volatile next; 120 struct callout *running;/* currently running callout */ 121 int softticks; /* softticks index */ 122 int curticks; /* per-cpu ticks counter */ 123 int isrunning; 124 struct thread thread; 125 126 }; 127 128 typedef struct softclock_pcpu *softclock_pcpu_t; 129 130 /* 131 * TODO: 132 * allocate more timeout table slots when table overflows. 133 */ 134 static MALLOC_DEFINE(M_CALLOUT, "callout", "callout structures"); 135 static int callwheelsize; 136 static int callwheelbits; 137 static int callwheelmask; 138 static struct softclock_pcpu softclock_pcpu_ary[MAXCPU]; 139 140 static void softclock_handler(void *arg); 141 142 static void 143 swi_softclock_setup(void *arg) 144 { 145 int cpu; 146 int i; 147 148 /* 149 * Figure out how large a callwheel we need. It must be a power of 2. 150 */ 151 callwheelsize = 1; 152 callwheelbits = 0; 153 while (callwheelsize < ncallout) { 154 callwheelsize <<= 1; 155 ++callwheelbits; 156 } 157 callwheelmask = callwheelsize - 1; 158 159 /* 160 * Initialize per-cpu data structures. 161 */ 162 for (cpu = 0; cpu < ncpus; ++cpu) { 163 softclock_pcpu_t sc; 164 165 sc = &softclock_pcpu_ary[cpu]; 166 167 sc->callwheel = kmalloc(sizeof(*sc->callwheel) * callwheelsize, 168 M_CALLOUT, M_WAITOK|M_ZERO); 169 for (i = 0; i < callwheelsize; ++i) 170 TAILQ_INIT(&sc->callwheel[i]); 171 172 /* 173 * Mark the softclock handler as being an interrupt thread 174 * even though it really isn't, but do not allow it to 175 * preempt other threads (do not assign td_preemptable). 176 * 177 * Kernel code now assumes that callouts do not preempt 178 * the cpu they were scheduled on. 179 */ 180 lwkt_create(softclock_handler, sc, NULL, 181 &sc->thread, TDF_NOSTART | TDF_INTTHREAD, 182 cpu, "softclock %d", cpu); 183 } 184 } 185 186 /* 187 * Must occur after ncpus has been initialized. 188 */ 189 SYSINIT(softclock_setup, SI_BOOT2_SOFTCLOCK, SI_ORDER_SECOND, 190 swi_softclock_setup, NULL); 191 192 /* 193 * This routine is called from the hardclock() (basically a FASTint/IPI) on 194 * each cpu in the system. sc->curticks is this cpu's notion of the timebase. 195 * It IS NOT NECESSARILY SYNCHRONIZED WITH 'ticks'! sc->softticks is where 196 * the callwheel is currently indexed. 197 * 198 * WARNING! The MP lock is not necessarily held on call, nor can it be 199 * safely obtained. 200 * 201 * sc->softticks is adjusted by either this routine or our helper thread 202 * depending on whether the helper thread is running or not. 203 */ 204 void 205 hardclock_softtick(globaldata_t gd) 206 { 207 softclock_pcpu_t sc; 208 209 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 210 ++sc->curticks; 211 if (sc->isrunning) 212 return; 213 if (sc->softticks == sc->curticks) { 214 /* 215 * in sync, only wakeup the thread if there is something to 216 * do. 217 */ 218 if (TAILQ_FIRST(&sc->callwheel[sc->softticks & callwheelmask])) 219 { 220 sc->isrunning = 1; 221 lwkt_schedule(&sc->thread); 222 } else { 223 ++sc->softticks; 224 } 225 } else { 226 /* 227 * out of sync, wakeup the thread unconditionally so it can 228 * catch up. 229 */ 230 sc->isrunning = 1; 231 lwkt_schedule(&sc->thread); 232 } 233 } 234 235 /* 236 * This procedure is the main loop of our per-cpu helper thread. The 237 * sc->isrunning flag prevents us from racing hardclock_softtick() and 238 * a critical section is sufficient to interlock sc->curticks and protect 239 * us from remote IPI's / list removal. 240 * 241 * The thread starts with the MP lock released and not in a critical 242 * section. The loop itself is MP safe while individual callbacks 243 * may or may not be, so we obtain or release the MP lock as appropriate. 244 */ 245 static void 246 softclock_handler(void *arg) 247 { 248 softclock_pcpu_t sc; 249 struct callout *c; 250 struct callout_tailq *bucket; 251 void (*c_func)(void *); 252 void *c_arg; 253 #ifdef SMP 254 int mpsafe = 1; 255 #endif 256 257 /* 258 * Run the callout thread at the same priority as other kernel 259 * threads so it can be round-robined. 260 */ 261 /*lwkt_setpri_self(TDPRI_SOFT_NORM);*/ 262 263 sc = arg; 264 crit_enter(); 265 loop: 266 while (sc->softticks != (int)(sc->curticks + 1)) { 267 bucket = &sc->callwheel[sc->softticks & callwheelmask]; 268 269 for (c = TAILQ_FIRST(bucket); c; c = sc->next) { 270 if (c->c_time != sc->softticks) { 271 sc->next = TAILQ_NEXT(c, c_links.tqe); 272 continue; 273 } 274 #ifdef SMP 275 if (c->c_flags & CALLOUT_MPSAFE) { 276 if (mpsafe == 0) { 277 mpsafe = 1; 278 rel_mplock(); 279 } 280 } else { 281 /* 282 * The request might be removed while we 283 * are waiting to get the MP lock. If it 284 * was removed sc->next will point to the 285 * next valid request or NULL, loop up. 286 */ 287 if (mpsafe) { 288 mpsafe = 0; 289 sc->next = c; 290 get_mplock(); 291 if (c != sc->next) 292 continue; 293 } 294 } 295 #endif 296 sc->next = TAILQ_NEXT(c, c_links.tqe); 297 TAILQ_REMOVE(bucket, c, c_links.tqe); 298 299 sc->running = c; 300 c_func = c->c_func; 301 c_arg = c->c_arg; 302 c->c_func = NULL; 303 KKASSERT(c->c_flags & CALLOUT_DID_INIT); 304 c->c_flags &= ~CALLOUT_PENDING; 305 crit_exit(); 306 c_func(c_arg); 307 crit_enter(); 308 sc->running = NULL; 309 /* NOTE: list may have changed */ 310 } 311 ++sc->softticks; 312 } 313 sc->isrunning = 0; 314 lwkt_deschedule_self(&sc->thread); /* == curthread */ 315 lwkt_switch(); 316 goto loop; 317 /* NOT REACHED */ 318 } 319 320 /* 321 * New interface; clients allocate their own callout structures. 322 * 323 * callout_reset() - establish or change a timeout 324 * callout_stop() - disestablish a timeout 325 * callout_init() - initialize a callout structure so that it can 326 * safely be passed to callout_reset() and callout_stop() 327 * callout_init_mp() - same but any installed functions must be MP safe. 328 * 329 * <sys/callout.h> defines three convenience macros: 330 * 331 * callout_active() - returns truth if callout has not been serviced 332 * callout_pending() - returns truth if callout is still waiting for timeout 333 * callout_deactivate() - marks the callout as having been serviced 334 */ 335 336 /* 337 * Start or restart a timeout. Install the callout structure in the 338 * callwheel. Callers may legally pass any value, even if 0 or negative, 339 * but since the sc->curticks index may have already been processed a 340 * minimum timeout of 1 tick will be enforced. 341 * 342 * The callout is installed on and will be processed on the current cpu's 343 * callout wheel. 344 * 345 * WARNING! This function may be called from any cpu but the caller must 346 * serialize callout_stop() and callout_reset() calls on the passed 347 * structure regardless of cpu. 348 */ 349 void 350 callout_reset(struct callout *c, int to_ticks, void (*ftn)(void *), 351 void *arg) 352 { 353 softclock_pcpu_t sc; 354 globaldata_t gd; 355 356 #ifdef INVARIANTS 357 if ((c->c_flags & CALLOUT_DID_INIT) == 0) { 358 callout_init(c); 359 kprintf( 360 "callout_reset(%p) from %p: callout was not initialized\n", 361 c, ((int **)&c)[-1]); 362 print_backtrace(-1); 363 } 364 #endif 365 gd = mycpu; 366 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 367 crit_enter_gd(gd); 368 369 if (c->c_flags & CALLOUT_ACTIVE) 370 callout_stop(c); 371 372 if (to_ticks <= 0) 373 to_ticks = 1; 374 375 c->c_arg = arg; 376 c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING); 377 c->c_func = ftn; 378 c->c_time = sc->curticks + to_ticks; 379 #ifdef SMP 380 c->c_gd = gd; 381 #endif 382 383 TAILQ_INSERT_TAIL(&sc->callwheel[c->c_time & callwheelmask], 384 c, c_links.tqe); 385 crit_exit_gd(gd); 386 } 387 388 /* 389 * Stop a running timer. WARNING! If called on a cpu other then the one 390 * the callout was started on this function will liveloop on its IPI to 391 * the target cpu to process the request. It is possible for the callout 392 * to execute in that case. 393 * 394 * WARNING! This function may be called from any cpu but the caller must 395 * serialize callout_stop() and callout_reset() calls on the passed 396 * structure regardless of cpu. 397 * 398 * WARNING! This routine may be called from an IPI 399 * 400 * WARNING! This function can return while it's c_func is still running 401 * in the callout thread, a secondary check may be needed. 402 * Use callout_stop_sync() to wait for any callout function to 403 * complete before returning, being sure that no deadlock is 404 * possible if you do. 405 */ 406 int 407 callout_stop(struct callout *c) 408 { 409 globaldata_t gd = mycpu; 410 #ifdef SMP 411 globaldata_t tgd; 412 #endif 413 softclock_pcpu_t sc; 414 415 #ifdef INVARIANTS 416 if ((c->c_flags & CALLOUT_DID_INIT) == 0) { 417 callout_init(c); 418 kprintf( 419 "callout_stop(%p) from %p: callout was not initialized\n", 420 c, ((int **)&c)[-1]); 421 print_backtrace(-1); 422 } 423 #endif 424 crit_enter_gd(gd); 425 426 /* 427 * Don't attempt to delete a callout that's not on the queue. The 428 * callout may not have a cpu assigned to it. Callers do not have 429 * to be on the issuing cpu but must still serialize access to the 430 * callout structure. 431 * 432 * We are not cpu-localized here and cannot safely modify the 433 * flags field in the callout structure. Note that most of the 434 * time CALLOUT_ACTIVE will be 0 if CALLOUT_PENDING is also 0. 435 * 436 * If we race another cpu's dispatch of this callout it is possible 437 * for CALLOUT_ACTIVE to be set with CALLOUT_PENDING unset. This 438 * will cause us to fall through and synchronize with the other 439 * cpu. 440 */ 441 if ((c->c_flags & CALLOUT_PENDING) == 0) { 442 #ifdef SMP 443 if ((c->c_flags & CALLOUT_ACTIVE) == 0) { 444 crit_exit_gd(gd); 445 return (0); 446 } 447 if (c->c_gd == NULL || c->c_gd == gd) { 448 c->c_flags &= ~CALLOUT_ACTIVE; 449 crit_exit_gd(gd); 450 return (0); 451 } 452 /* fall-through to the cpu-localization code. */ 453 #else 454 c->c_flags &= ~CALLOUT_ACTIVE; 455 crit_exit_gd(gd); 456 return (0); 457 #endif 458 } 459 #ifdef SMP 460 if ((tgd = c->c_gd) != gd) { 461 /* 462 * If the callout is owned by a different CPU we have to 463 * execute the function synchronously on the target cpu. 464 */ 465 int seq; 466 467 cpu_ccfence(); /* don't let tgd alias c_gd */ 468 seq = lwkt_send_ipiq(tgd, (void *)callout_stop, c); 469 lwkt_wait_ipiq(tgd, seq); 470 } else 471 #endif 472 { 473 /* 474 * If the callout is owned by the same CPU we can 475 * process it directly, but if we are racing our helper 476 * thread (sc->next), we have to adjust sc->next. The 477 * race is interlocked by a critical section. 478 */ 479 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 480 481 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING); 482 if (sc->next == c) 483 sc->next = TAILQ_NEXT(c, c_links.tqe); 484 485 TAILQ_REMOVE(&sc->callwheel[c->c_time & callwheelmask], 486 c, c_links.tqe); 487 c->c_func = NULL; 488 } 489 crit_exit_gd(gd); 490 return (1); 491 } 492 493 /* 494 * Issue a callout_stop() and ensure that any callout race completes 495 * before returning. Does NOT de-initialized the callout. 496 */ 497 void 498 callout_stop_sync(struct callout *c) 499 { 500 softclock_pcpu_t sc; 501 502 if (c->c_flags & CALLOUT_DID_INIT) { 503 callout_stop(c); 504 #ifdef SMP 505 if (c->c_gd) { 506 sc = &softclock_pcpu_ary[c->c_gd->gd_cpuid]; 507 if (sc->running == c) { 508 while (sc->running == c) 509 tsleep(&sc->running, 0, "crace", 1); 510 } 511 } 512 #else 513 sc = &softclock_pcpu_ary[0]; 514 if (sc->running == c) { 515 while (sc->running == c) 516 tsleep(&sc->running, 0, "crace", 1); 517 } 518 #endif 519 KKASSERT((c->c_flags & (CALLOUT_PENDING|CALLOUT_ACTIVE)) == 0); 520 } 521 } 522 523 /* 524 * Terminate a callout 525 * 526 * This function will stop any pending callout and also block while the 527 * callout's function is running. It should only be used in cases where 528 * no deadlock is possible (due to the callout function acquiring locks 529 * that the current caller of callout_terminate() already holds), when 530 * the caller is ready to destroy the callout structure. 531 * 532 * This function clears the CALLOUT_DID_INIT flag. 533 * 534 * lwkt_token locks are ok. 535 */ 536 void 537 callout_terminate(struct callout *c) 538 { 539 softclock_pcpu_t sc; 540 541 if (c->c_flags & CALLOUT_DID_INIT) { 542 callout_stop(c); 543 #ifdef SMP 544 sc = &softclock_pcpu_ary[c->c_gd->gd_cpuid]; 545 #else 546 sc = &softclock_pcpu_ary[0]; 547 #endif 548 if (sc->running == c) { 549 while (sc->running == c) 550 tsleep(&sc->running, 0, "crace", 1); 551 } 552 KKASSERT((c->c_flags & (CALLOUT_PENDING|CALLOUT_ACTIVE)) == 0); 553 c->c_flags &= ~CALLOUT_DID_INIT; 554 } 555 } 556 557 /* 558 * Prepare a callout structure for use by callout_reset() and/or 559 * callout_stop(). The MP version of this routine requires that the callback 560 * function installed by callout_reset() be MP safe. 561 * 562 * The init functions can be called from any cpu and do not have to be 563 * called from the cpu that the timer will eventually run on. 564 */ 565 void 566 callout_init(struct callout *c) 567 { 568 bzero(c, sizeof *c); 569 c->c_flags = CALLOUT_DID_INIT; 570 } 571 572 void 573 callout_init_mp(struct callout *c) 574 { 575 callout_init(c); 576 c->c_flags |= CALLOUT_MPSAFE; 577 } 578 579 /* What, are you joking? This is nuts! -Matt */ 580 #if 0 581 #ifdef APM_FIXUP_CALLTODO 582 /* 583 * Adjust the kernel calltodo timeout list. This routine is used after 584 * an APM resume to recalculate the calltodo timer list values with the 585 * number of hz's we have been sleeping. The next hardclock() will detect 586 * that there are fired timers and run softclock() to execute them. 587 * 588 * Please note, I have not done an exhaustive analysis of what code this 589 * might break. I am motivated to have my select()'s and alarm()'s that 590 * have expired during suspend firing upon resume so that the applications 591 * which set the timer can do the maintanence the timer was for as close 592 * as possible to the originally intended time. Testing this code for a 593 * week showed that resuming from a suspend resulted in 22 to 25 timers 594 * firing, which seemed independant on whether the suspend was 2 hours or 595 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu> 596 */ 597 void 598 adjust_timeout_calltodo(struct timeval *time_change) 599 { 600 struct callout *p; 601 unsigned long delta_ticks; 602 603 /* 604 * How many ticks were we asleep? 605 * (stolen from tvtohz()). 606 */ 607 608 /* Don't do anything */ 609 if (time_change->tv_sec < 0) 610 return; 611 else if (time_change->tv_sec <= LONG_MAX / 1000000) 612 delta_ticks = (time_change->tv_sec * 1000000 + 613 time_change->tv_usec + (tick - 1)) / tick + 1; 614 else if (time_change->tv_sec <= LONG_MAX / hz) 615 delta_ticks = time_change->tv_sec * hz + 616 (time_change->tv_usec + (tick - 1)) / tick + 1; 617 else 618 delta_ticks = LONG_MAX; 619 620 if (delta_ticks > INT_MAX) 621 delta_ticks = INT_MAX; 622 623 /* 624 * Now rip through the timer calltodo list looking for timers 625 * to expire. 626 */ 627 628 /* don't collide with softclock() */ 629 crit_enter(); 630 for (p = calltodo.c_next; p != NULL; p = p->c_next) { 631 p->c_time -= delta_ticks; 632 633 /* Break if the timer had more time on it than delta_ticks */ 634 if (p->c_time > 0) 635 break; 636 637 /* take back the ticks the timer didn't use (p->c_time <= 0) */ 638 delta_ticks = -p->c_time; 639 } 640 crit_exit(); 641 642 return; 643 } 644 #endif /* APM_FIXUP_CALLTODO */ 645 #endif 646 647