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