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. Neither the name of the University nor the names of its contributors 52 * may be used to endorse or promote products derived from this software 53 * without specific prior written permission. 54 * 55 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 56 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 57 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 58 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 59 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 60 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 61 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 62 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 63 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 64 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 65 * SUCH DAMAGE. 66 * 67 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 68 * $FreeBSD: src/sys/kern/kern_timeout.c,v 1.59.2.1 2001/11/13 18:24:52 archie Exp $ 69 */ 70 /* 71 * DRAGONFLY BGL STATUS 72 * 73 * All the API functions should be MP safe. 74 * 75 * The callback functions will be flagged as being MP safe if the 76 * timeout structure is initialized with callout_init_mp() instead of 77 * callout_init(). 78 * 79 * The helper threads cannot be made preempt-capable until after we 80 * clean up all the uses of splsoftclock() and related interlocks (which 81 * require the related functions to be MP safe as well). 82 */ 83 /* 84 * The callout mechanism is based on the work of Adam M. Costello and 85 * George Varghese, published in a technical report entitled "Redesigning 86 * the BSD Callout and Timer Facilities" and modified slightly for inclusion 87 * in FreeBSD by Justin T. Gibbs. The original work on the data structures 88 * used in this implementation was published by G. Varghese and T. Lauck in 89 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for 90 * the Efficient Implementation of a Timer Facility" in the Proceedings of 91 * the 11th ACM Annual Symposium on Operating Systems Principles, 92 * Austin, Texas Nov 1987. 93 * 94 * The per-cpu augmentation was done by Matthew Dillon. 95 */ 96 97 #include <sys/param.h> 98 #include <sys/systm.h> 99 #include <sys/callout.h> 100 #include <sys/kernel.h> 101 #include <sys/interrupt.h> 102 #include <sys/thread.h> 103 104 #include <sys/thread2.h> 105 #include <sys/mplock2.h> 106 107 #ifndef MAX_SOFTCLOCK_STEPS 108 #define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */ 109 #endif 110 111 112 struct softclock_pcpu { 113 struct callout_tailq *callwheel; 114 struct callout * volatile next; 115 struct callout *running;/* currently running callout */ 116 int softticks; /* softticks index */ 117 int curticks; /* per-cpu ticks counter */ 118 int isrunning; 119 struct thread thread; 120 121 }; 122 123 typedef struct softclock_pcpu *softclock_pcpu_t; 124 125 /* 126 * TODO: 127 * allocate more timeout table slots when table overflows. 128 */ 129 static MALLOC_DEFINE(M_CALLOUT, "callout", "callout structures"); 130 static int callwheelsize; 131 static int callwheelmask; 132 static struct softclock_pcpu softclock_pcpu_ary[MAXCPU]; 133 134 static void softclock_handler(void *arg); 135 static void slotimer_callback(void *arg); 136 137 static void 138 swi_softclock_setup(void *arg) 139 { 140 int cpu; 141 int i; 142 int target; 143 144 /* 145 * Figure out how large a callwheel we need. It must be a power of 2. 146 * 147 * ncallout is primarily based on available memory, don't explode 148 * the allocations if the system has a lot of cpus. 149 */ 150 target = ncallout / ncpus + 16; 151 152 callwheelsize = 1; 153 while (callwheelsize < target) 154 callwheelsize <<= 1; 155 callwheelmask = callwheelsize - 1; 156 157 /* 158 * Initialize per-cpu data structures. 159 */ 160 for (cpu = 0; cpu < ncpus; ++cpu) { 161 softclock_pcpu_t sc; 162 163 sc = &softclock_pcpu_ary[cpu]; 164 165 sc->callwheel = kmalloc(sizeof(*sc->callwheel) * callwheelsize, 166 M_CALLOUT, M_WAITOK|M_ZERO); 167 for (i = 0; i < callwheelsize; ++i) 168 TAILQ_INIT(&sc->callwheel[i]); 169 170 /* 171 * Mark the softclock handler as being an interrupt thread 172 * even though it really isn't, but do not allow it to 173 * preempt other threads (do not assign td_preemptable). 174 * 175 * Kernel code now assumes that callouts do not preempt 176 * the cpu they were scheduled on. 177 */ 178 lwkt_create(softclock_handler, sc, NULL, 179 &sc->thread, TDF_NOSTART | TDF_INTTHREAD, 180 cpu, "softclock %d", cpu); 181 } 182 } 183 184 /* 185 * Must occur after ncpus has been initialized. 186 */ 187 SYSINIT(softclock_setup, SI_BOOT2_SOFTCLOCK, SI_ORDER_SECOND, 188 swi_softclock_setup, NULL); 189 190 /* 191 * This routine is called from the hardclock() (basically a FASTint/IPI) on 192 * each cpu in the system. sc->curticks is this cpu's notion of the timebase. 193 * It IS NOT NECESSARILY SYNCHRONIZED WITH 'ticks'! sc->softticks is where 194 * the callwheel is currently indexed. 195 * 196 * WARNING! The MP lock is not necessarily held on call, nor can it be 197 * safely obtained. 198 * 199 * sc->softticks is adjusted by either this routine or our helper thread 200 * depending on whether the helper thread is running or not. 201 */ 202 void 203 hardclock_softtick(globaldata_t gd) 204 { 205 softclock_pcpu_t sc; 206 207 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 208 ++sc->curticks; 209 if (sc->isrunning) 210 return; 211 if (sc->softticks == sc->curticks) { 212 /* 213 * in sync, only wakeup the thread if there is something to 214 * do. 215 */ 216 if (TAILQ_FIRST(&sc->callwheel[sc->softticks & callwheelmask])) 217 { 218 sc->isrunning = 1; 219 lwkt_schedule(&sc->thread); 220 } else { 221 ++sc->softticks; 222 } 223 } else { 224 /* 225 * out of sync, wakeup the thread unconditionally so it can 226 * catch up. 227 */ 228 sc->isrunning = 1; 229 lwkt_schedule(&sc->thread); 230 } 231 } 232 233 /* 234 * This procedure is the main loop of our per-cpu helper thread. The 235 * sc->isrunning flag prevents us from racing hardclock_softtick() and 236 * a critical section is sufficient to interlock sc->curticks and protect 237 * us from remote IPI's / list removal. 238 * 239 * The thread starts with the MP lock released and not in a critical 240 * section. The loop itself is MP safe while individual callbacks 241 * may or may not be, so we obtain or release the MP lock as appropriate. 242 */ 243 static void 244 softclock_handler(void *arg) 245 { 246 softclock_pcpu_t sc; 247 struct callout *c; 248 struct callout_tailq *bucket; 249 struct callout slotimer; 250 void (*c_func)(void *); 251 void *c_arg; 252 int mpsafe = 1; 253 254 /* 255 * Setup pcpu slow clocks which we want to run from the callout 256 * thread. 257 */ 258 callout_init_mp(&slotimer); 259 callout_reset(&slotimer, hz * 10, slotimer_callback, &slotimer); 260 261 /* 262 * Run the callout thread at the same priority as other kernel 263 * threads so it can be round-robined. 264 */ 265 /*lwkt_setpri_self(TDPRI_SOFT_NORM);*/ 266 267 sc = arg; 268 crit_enter(); 269 loop: 270 while (sc->softticks != (int)(sc->curticks + 1)) { 271 bucket = &sc->callwheel[sc->softticks & callwheelmask]; 272 273 for (c = TAILQ_FIRST(bucket); c; c = sc->next) { 274 if (c->c_time != sc->softticks) { 275 sc->next = TAILQ_NEXT(c, c_links.tqe); 276 continue; 277 } 278 if (c->c_flags & CALLOUT_MPSAFE) { 279 if (mpsafe == 0) { 280 mpsafe = 1; 281 rel_mplock(); 282 } 283 } else { 284 /* 285 * The request might be removed while we 286 * are waiting to get the MP lock. If it 287 * was removed sc->next will point to the 288 * next valid request or NULL, loop up. 289 */ 290 if (mpsafe) { 291 mpsafe = 0; 292 sc->next = c; 293 get_mplock(); 294 if (c != sc->next) 295 continue; 296 } 297 } 298 sc->next = TAILQ_NEXT(c, c_links.tqe); 299 TAILQ_REMOVE(bucket, c, c_links.tqe); 300 301 sc->running = c; 302 c_func = c->c_func; 303 c_arg = c->c_arg; 304 c->c_func = NULL; 305 KKASSERT(c->c_flags & CALLOUT_DID_INIT); 306 c->c_flags &= ~CALLOUT_PENDING; 307 crit_exit(); 308 c_func(c_arg); 309 crit_enter(); 310 sc->running = NULL; 311 /* NOTE: list may have changed */ 312 } 313 ++sc->softticks; 314 } 315 316 /* 317 * Don't leave us holding the MP lock when we deschedule ourselves. 318 */ 319 if (mpsafe == 0) { 320 mpsafe = 1; 321 rel_mplock(); 322 } 323 sc->isrunning = 0; 324 lwkt_deschedule_self(&sc->thread); /* == curthread */ 325 lwkt_switch(); 326 goto loop; 327 /* NOT REACHED */ 328 } 329 330 /* 331 * A very slow system cleanup timer (10 second interval), 332 * per-cpu. 333 */ 334 void 335 slotimer_callback(void *arg) 336 { 337 struct callout *c = arg; 338 339 slab_cleanup(); 340 callout_reset(c, hz * 10, slotimer_callback, c); 341 } 342 343 /* 344 * New interface; clients allocate their own callout structures. 345 * 346 * callout_reset() - establish or change a timeout 347 * callout_stop() - disestablish a timeout 348 * callout_init() - initialize a callout structure so that it can 349 * safely be passed to callout_reset() and callout_stop() 350 * callout_init_mp() - same but any installed functions must be MP safe. 351 * 352 * <sys/callout.h> defines three convenience macros: 353 * 354 * callout_active() - returns truth if callout has not been serviced 355 * callout_pending() - returns truth if callout is still waiting for timeout 356 * callout_deactivate() - marks the callout as having been serviced 357 */ 358 359 /* 360 * Start or restart a timeout. Install the callout structure in the 361 * callwheel. Callers may legally pass any value, even if 0 or negative, 362 * but since the sc->curticks index may have already been processed a 363 * minimum timeout of 1 tick will be enforced. 364 * 365 * The callout is installed on and will be processed on the current cpu's 366 * callout wheel. 367 * 368 * WARNING! This function may be called from any cpu but the caller must 369 * serialize callout_stop() and callout_reset() calls on the passed 370 * structure regardless of cpu. 371 */ 372 void 373 callout_reset(struct callout *c, int to_ticks, void (*ftn)(void *), 374 void *arg) 375 { 376 softclock_pcpu_t sc; 377 globaldata_t gd; 378 379 #ifdef INVARIANTS 380 if ((c->c_flags & CALLOUT_DID_INIT) == 0) { 381 callout_init(c); 382 kprintf( 383 "callout_reset(%p) from %p: callout was not initialized\n", 384 c, ((int **)&c)[-1]); 385 print_backtrace(-1); 386 } 387 #endif 388 gd = mycpu; 389 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 390 crit_enter_gd(gd); 391 392 if (c->c_flags & CALLOUT_ACTIVE) 393 callout_stop(c); 394 395 if (to_ticks <= 0) 396 to_ticks = 1; 397 398 c->c_arg = arg; 399 c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING); 400 c->c_func = ftn; 401 c->c_time = sc->curticks + to_ticks; 402 c->c_gd = gd; 403 404 TAILQ_INSERT_TAIL(&sc->callwheel[c->c_time & callwheelmask], 405 c, c_links.tqe); 406 crit_exit_gd(gd); 407 } 408 409 struct callout_remote_arg { 410 struct callout *c; 411 void (*ftn)(void *); 412 void *arg; 413 int to_ticks; 414 }; 415 416 static void 417 callout_reset_ipi(void *arg) 418 { 419 struct callout_remote_arg *rmt = arg; 420 421 callout_reset(rmt->c, rmt->to_ticks, rmt->ftn, rmt->arg); 422 } 423 424 void 425 callout_reset_bycpu(struct callout *c, int to_ticks, void (*ftn)(void *), 426 void *arg, int cpuid) 427 { 428 KASSERT(cpuid >= 0 && cpuid < ncpus, ("invalid cpuid %d", cpuid)); 429 430 if (cpuid == mycpuid) { 431 callout_reset(c, to_ticks, ftn, arg); 432 } else { 433 struct globaldata *target_gd; 434 struct callout_remote_arg rmt; 435 int seq; 436 437 rmt.c = c; 438 rmt.ftn = ftn; 439 rmt.arg = arg; 440 rmt.to_ticks = to_ticks; 441 442 target_gd = globaldata_find(cpuid); 443 444 seq = lwkt_send_ipiq(target_gd, callout_reset_ipi, &rmt); 445 lwkt_wait_ipiq(target_gd, seq); 446 } 447 } 448 449 /* 450 * Stop a running timer. WARNING! If called on a cpu other then the one 451 * the callout was started on this function will liveloop on its IPI to 452 * the target cpu to process the request. It is possible for the callout 453 * to execute in that case. 454 * 455 * WARNING! This function may be called from any cpu but the caller must 456 * serialize callout_stop() and callout_reset() calls on the passed 457 * structure regardless of cpu. 458 * 459 * WARNING! This routine may be called from an IPI 460 * 461 * WARNING! This function can return while it's c_func is still running 462 * in the callout thread, a secondary check may be needed. 463 * Use callout_stop_sync() to wait for any callout function to 464 * complete before returning, being sure that no deadlock is 465 * possible if you do. 466 */ 467 int 468 callout_stop(struct callout *c) 469 { 470 globaldata_t gd = mycpu; 471 globaldata_t tgd; 472 softclock_pcpu_t sc; 473 474 #ifdef INVARIANTS 475 if ((c->c_flags & CALLOUT_DID_INIT) == 0) { 476 callout_init(c); 477 kprintf( 478 "callout_stop(%p) from %p: callout was not initialized\n", 479 c, ((int **)&c)[-1]); 480 print_backtrace(-1); 481 } 482 #endif 483 crit_enter_gd(gd); 484 485 /* 486 * Don't attempt to delete a callout that's not on the queue. The 487 * callout may not have a cpu assigned to it. Callers do not have 488 * to be on the issuing cpu but must still serialize access to the 489 * callout structure. 490 * 491 * We are not cpu-localized here and cannot safely modify the 492 * flags field in the callout structure. Note that most of the 493 * time CALLOUT_ACTIVE will be 0 if CALLOUT_PENDING is also 0. 494 * 495 * If we race another cpu's dispatch of this callout it is possible 496 * for CALLOUT_ACTIVE to be set with CALLOUT_PENDING unset. This 497 * will cause us to fall through and synchronize with the other 498 * cpu. 499 */ 500 if ((c->c_flags & CALLOUT_PENDING) == 0) { 501 if ((c->c_flags & CALLOUT_ACTIVE) == 0) { 502 crit_exit_gd(gd); 503 return (0); 504 } 505 if (c->c_gd == NULL || c->c_gd == gd) { 506 c->c_flags &= ~CALLOUT_ACTIVE; 507 crit_exit_gd(gd); 508 return (0); 509 } 510 } 511 if ((tgd = c->c_gd) != gd) { 512 /* 513 * If the callout is owned by a different CPU we have to 514 * execute the function synchronously on the target cpu. 515 */ 516 int seq; 517 518 cpu_ccfence(); /* don't let tgd alias c_gd */ 519 seq = lwkt_send_ipiq(tgd, (void *)callout_stop, c); 520 lwkt_wait_ipiq(tgd, seq); 521 } else { 522 /* 523 * If the callout is owned by the same CPU we can 524 * process it directly, but if we are racing our helper 525 * thread (sc->next), we have to adjust sc->next. The 526 * race is interlocked by a critical section. 527 */ 528 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 529 530 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING); 531 if (sc->next == c) 532 sc->next = TAILQ_NEXT(c, c_links.tqe); 533 534 TAILQ_REMOVE(&sc->callwheel[c->c_time & callwheelmask], 535 c, c_links.tqe); 536 c->c_func = NULL; 537 } 538 crit_exit_gd(gd); 539 return (1); 540 } 541 542 /* 543 * Issue a callout_stop() and ensure that any callout race completes 544 * before returning. Does NOT de-initialized the callout. 545 */ 546 void 547 callout_stop_sync(struct callout *c) 548 { 549 softclock_pcpu_t sc; 550 551 while (c->c_flags & CALLOUT_DID_INIT) { 552 callout_stop(c); 553 if (c->c_gd) { 554 sc = &softclock_pcpu_ary[c->c_gd->gd_cpuid]; 555 if (sc->running == c) { 556 while (sc->running == c) 557 tsleep(&sc->running, 0, "crace", 1); 558 } 559 } 560 if ((c->c_flags & (CALLOUT_PENDING | CALLOUT_ACTIVE)) == 0) 561 break; 562 kprintf("Warning: %s: callout race\n", curthread->td_comm); 563 } 564 } 565 566 /* 567 * Terminate a callout 568 * 569 * This function will stop any pending callout and also block while the 570 * callout's function is running. It should only be used in cases where 571 * no deadlock is possible (due to the callout function acquiring locks 572 * that the current caller of callout_terminate() already holds), when 573 * the caller is ready to destroy the callout structure. 574 * 575 * This function clears the CALLOUT_DID_INIT flag. 576 * 577 * lwkt_token locks are ok. 578 */ 579 void 580 callout_terminate(struct callout *c) 581 { 582 softclock_pcpu_t sc; 583 584 if (c->c_flags & CALLOUT_DID_INIT) { 585 callout_stop(c); 586 if (c->c_gd) { 587 sc = &softclock_pcpu_ary[c->c_gd->gd_cpuid]; 588 if (sc->running == c) { 589 while (sc->running == c) 590 tsleep(&sc->running, 0, "crace", 1); 591 } 592 } 593 KKASSERT((c->c_flags & (CALLOUT_PENDING|CALLOUT_ACTIVE)) == 0); 594 c->c_flags &= ~CALLOUT_DID_INIT; 595 } 596 } 597 598 /* 599 * Prepare a callout structure for use by callout_reset() and/or 600 * callout_stop(). The MP version of this routine requires that the callback 601 * function installed by callout_reset() be MP safe. 602 * 603 * The init functions can be called from any cpu and do not have to be 604 * called from the cpu that the timer will eventually run on. 605 */ 606 void 607 callout_init(struct callout *c) 608 { 609 bzero(c, sizeof *c); 610 c->c_flags = CALLOUT_DID_INIT; 611 } 612 613 void 614 callout_init_mp(struct callout *c) 615 { 616 callout_init(c); 617 c->c_flags |= CALLOUT_MPSAFE; 618 } 619