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); 135 136 static void 137 swi_softclock_setup(void *arg) 138 { 139 int cpu; 140 int i; 141 int target; 142 143 /* 144 * Figure out how large a callwheel we need. It must be a power of 2. 145 * 146 * ncallout is primarily based on available memory, don't explode 147 * the allocations if the system has a lot of cpus. 148 */ 149 target = ncallout / ncpus + 16; 150 151 callwheelsize = 1; 152 while (callwheelsize < target) 153 callwheelsize <<= 1; 154 callwheelmask = callwheelsize - 1; 155 156 /* 157 * Initialize per-cpu data structures. 158 */ 159 for (cpu = 0; cpu < ncpus; ++cpu) { 160 softclock_pcpu_t sc; 161 162 sc = &softclock_pcpu_ary[cpu]; 163 164 sc->callwheel = kmalloc(sizeof(*sc->callwheel) * callwheelsize, 165 M_CALLOUT, M_WAITOK|M_ZERO); 166 for (i = 0; i < callwheelsize; ++i) 167 TAILQ_INIT(&sc->callwheel[i]); 168 169 /* 170 * Mark the softclock handler as being an interrupt thread 171 * even though it really isn't, but do not allow it to 172 * preempt other threads (do not assign td_preemptable). 173 * 174 * Kernel code now assumes that callouts do not preempt 175 * the cpu they were scheduled on. 176 */ 177 lwkt_create(softclock_handler, sc, NULL, 178 &sc->thread, TDF_NOSTART | TDF_INTTHREAD, 179 cpu, "softclock %d", cpu); 180 } 181 } 182 183 /* 184 * Must occur after ncpus has been initialized. 185 */ 186 SYSINIT(softclock_setup, SI_BOOT2_SOFTCLOCK, SI_ORDER_SECOND, 187 swi_softclock_setup, NULL); 188 189 /* 190 * This routine is called from the hardclock() (basically a FASTint/IPI) on 191 * each cpu in the system. sc->curticks is this cpu's notion of the timebase. 192 * It IS NOT NECESSARILY SYNCHRONIZED WITH 'ticks'! sc->softticks is where 193 * the callwheel is currently indexed. 194 * 195 * WARNING! The MP lock is not necessarily held on call, nor can it be 196 * safely obtained. 197 * 198 * sc->softticks is adjusted by either this routine or our helper thread 199 * depending on whether the helper thread is running or not. 200 */ 201 void 202 hardclock_softtick(globaldata_t gd) 203 { 204 softclock_pcpu_t sc; 205 206 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 207 ++sc->curticks; 208 if (sc->isrunning) 209 return; 210 if (sc->softticks == sc->curticks) { 211 /* 212 * in sync, only wakeup the thread if there is something to 213 * do. 214 */ 215 if (TAILQ_FIRST(&sc->callwheel[sc->softticks & callwheelmask])) 216 { 217 sc->isrunning = 1; 218 lwkt_schedule(&sc->thread); 219 } else { 220 ++sc->softticks; 221 } 222 } else { 223 /* 224 * out of sync, wakeup the thread unconditionally so it can 225 * catch up. 226 */ 227 sc->isrunning = 1; 228 lwkt_schedule(&sc->thread); 229 } 230 } 231 232 /* 233 * This procedure is the main loop of our per-cpu helper thread. The 234 * sc->isrunning flag prevents us from racing hardclock_softtick() and 235 * a critical section is sufficient to interlock sc->curticks and protect 236 * us from remote IPI's / list removal. 237 * 238 * The thread starts with the MP lock released and not in a critical 239 * section. The loop itself is MP safe while individual callbacks 240 * may or may not be, so we obtain or release the MP lock as appropriate. 241 */ 242 static void 243 softclock_handler(void *arg) 244 { 245 softclock_pcpu_t sc; 246 struct callout *c; 247 struct callout_tailq *bucket; 248 void (*c_func)(void *); 249 void *c_arg; 250 int mpsafe = 1; 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 if (c->c_flags & CALLOUT_MPSAFE) { 270 if (mpsafe == 0) { 271 mpsafe = 1; 272 rel_mplock(); 273 } 274 } else { 275 /* 276 * The request might be removed while we 277 * are waiting to get the MP lock. If it 278 * was removed sc->next will point to the 279 * next valid request or NULL, loop up. 280 */ 281 if (mpsafe) { 282 mpsafe = 0; 283 sc->next = c; 284 get_mplock(); 285 if (c != sc->next) 286 continue; 287 } 288 } 289 sc->next = TAILQ_NEXT(c, c_links.tqe); 290 TAILQ_REMOVE(bucket, c, c_links.tqe); 291 292 sc->running = c; 293 c_func = c->c_func; 294 c_arg = c->c_arg; 295 c->c_func = NULL; 296 KKASSERT(c->c_flags & CALLOUT_DID_INIT); 297 c->c_flags &= ~CALLOUT_PENDING; 298 crit_exit(); 299 c_func(c_arg); 300 crit_enter(); 301 sc->running = NULL; 302 /* NOTE: list may have changed */ 303 } 304 ++sc->softticks; 305 } 306 sc->isrunning = 0; 307 lwkt_deschedule_self(&sc->thread); /* == curthread */ 308 lwkt_switch(); 309 goto loop; 310 /* NOT REACHED */ 311 } 312 313 /* 314 * New interface; clients allocate their own callout structures. 315 * 316 * callout_reset() - establish or change a timeout 317 * callout_stop() - disestablish a timeout 318 * callout_init() - initialize a callout structure so that it can 319 * safely be passed to callout_reset() and callout_stop() 320 * callout_init_mp() - same but any installed functions must be MP safe. 321 * 322 * <sys/callout.h> defines three convenience macros: 323 * 324 * callout_active() - returns truth if callout has not been serviced 325 * callout_pending() - returns truth if callout is still waiting for timeout 326 * callout_deactivate() - marks the callout as having been serviced 327 */ 328 329 /* 330 * Start or restart a timeout. Install the callout structure in the 331 * callwheel. Callers may legally pass any value, even if 0 or negative, 332 * but since the sc->curticks index may have already been processed a 333 * minimum timeout of 1 tick will be enforced. 334 * 335 * The callout is installed on and will be processed on the current cpu's 336 * callout wheel. 337 * 338 * WARNING! This function may be called from any cpu but the caller must 339 * serialize callout_stop() and callout_reset() calls on the passed 340 * structure regardless of cpu. 341 */ 342 void 343 callout_reset(struct callout *c, int to_ticks, void (*ftn)(void *), 344 void *arg) 345 { 346 softclock_pcpu_t sc; 347 globaldata_t gd; 348 349 #ifdef INVARIANTS 350 if ((c->c_flags & CALLOUT_DID_INIT) == 0) { 351 callout_init(c); 352 kprintf( 353 "callout_reset(%p) from %p: callout was not initialized\n", 354 c, ((int **)&c)[-1]); 355 print_backtrace(-1); 356 } 357 #endif 358 gd = mycpu; 359 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 360 crit_enter_gd(gd); 361 362 if (c->c_flags & CALLOUT_ACTIVE) 363 callout_stop(c); 364 365 if (to_ticks <= 0) 366 to_ticks = 1; 367 368 c->c_arg = arg; 369 c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING); 370 c->c_func = ftn; 371 c->c_time = sc->curticks + to_ticks; 372 c->c_gd = gd; 373 374 TAILQ_INSERT_TAIL(&sc->callwheel[c->c_time & callwheelmask], 375 c, c_links.tqe); 376 crit_exit_gd(gd); 377 } 378 379 struct callout_remote_arg { 380 struct callout *c; 381 void (*ftn)(void *); 382 void *arg; 383 int to_ticks; 384 }; 385 386 static void 387 callout_reset_ipi(void *arg) 388 { 389 struct callout_remote_arg *rmt = arg; 390 391 callout_reset(rmt->c, rmt->to_ticks, rmt->ftn, rmt->arg); 392 } 393 394 void 395 callout_reset_bycpu(struct callout *c, int to_ticks, void (*ftn)(void *), 396 void *arg, int cpuid) 397 { 398 KASSERT(cpuid >= 0 && cpuid < ncpus, ("invalid cpuid %d", cpuid)); 399 400 if (cpuid == mycpuid) { 401 callout_reset(c, to_ticks, ftn, arg); 402 } else { 403 struct globaldata *target_gd; 404 struct callout_remote_arg rmt; 405 int seq; 406 407 rmt.c = c; 408 rmt.ftn = ftn; 409 rmt.arg = arg; 410 rmt.to_ticks = to_ticks; 411 412 target_gd = globaldata_find(cpuid); 413 414 seq = lwkt_send_ipiq(target_gd, callout_reset_ipi, &rmt); 415 lwkt_wait_ipiq(target_gd, seq); 416 } 417 } 418 419 /* 420 * Stop a running timer. WARNING! If called on a cpu other then the one 421 * the callout was started on this function will liveloop on its IPI to 422 * the target cpu to process the request. It is possible for the callout 423 * to execute in that case. 424 * 425 * WARNING! This function may be called from any cpu but the caller must 426 * serialize callout_stop() and callout_reset() calls on the passed 427 * structure regardless of cpu. 428 * 429 * WARNING! This routine may be called from an IPI 430 * 431 * WARNING! This function can return while it's c_func is still running 432 * in the callout thread, a secondary check may be needed. 433 * Use callout_stop_sync() to wait for any callout function to 434 * complete before returning, being sure that no deadlock is 435 * possible if you do. 436 */ 437 int 438 callout_stop(struct callout *c) 439 { 440 globaldata_t gd = mycpu; 441 globaldata_t tgd; 442 softclock_pcpu_t sc; 443 444 #ifdef INVARIANTS 445 if ((c->c_flags & CALLOUT_DID_INIT) == 0) { 446 callout_init(c); 447 kprintf( 448 "callout_stop(%p) from %p: callout was not initialized\n", 449 c, ((int **)&c)[-1]); 450 print_backtrace(-1); 451 } 452 #endif 453 crit_enter_gd(gd); 454 455 /* 456 * Don't attempt to delete a callout that's not on the queue. The 457 * callout may not have a cpu assigned to it. Callers do not have 458 * to be on the issuing cpu but must still serialize access to the 459 * callout structure. 460 * 461 * We are not cpu-localized here and cannot safely modify the 462 * flags field in the callout structure. Note that most of the 463 * time CALLOUT_ACTIVE will be 0 if CALLOUT_PENDING is also 0. 464 * 465 * If we race another cpu's dispatch of this callout it is possible 466 * for CALLOUT_ACTIVE to be set with CALLOUT_PENDING unset. This 467 * will cause us to fall through and synchronize with the other 468 * cpu. 469 */ 470 if ((c->c_flags & CALLOUT_PENDING) == 0) { 471 if ((c->c_flags & CALLOUT_ACTIVE) == 0) { 472 crit_exit_gd(gd); 473 return (0); 474 } 475 if (c->c_gd == NULL || c->c_gd == gd) { 476 c->c_flags &= ~CALLOUT_ACTIVE; 477 crit_exit_gd(gd); 478 return (0); 479 } 480 } 481 if ((tgd = c->c_gd) != gd) { 482 /* 483 * If the callout is owned by a different CPU we have to 484 * execute the function synchronously on the target cpu. 485 */ 486 int seq; 487 488 cpu_ccfence(); /* don't let tgd alias c_gd */ 489 seq = lwkt_send_ipiq(tgd, (void *)callout_stop, c); 490 lwkt_wait_ipiq(tgd, seq); 491 } else { 492 /* 493 * If the callout is owned by the same CPU we can 494 * process it directly, but if we are racing our helper 495 * thread (sc->next), we have to adjust sc->next. The 496 * race is interlocked by a critical section. 497 */ 498 sc = &softclock_pcpu_ary[gd->gd_cpuid]; 499 500 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING); 501 if (sc->next == c) 502 sc->next = TAILQ_NEXT(c, c_links.tqe); 503 504 TAILQ_REMOVE(&sc->callwheel[c->c_time & callwheelmask], 505 c, c_links.tqe); 506 c->c_func = NULL; 507 } 508 crit_exit_gd(gd); 509 return (1); 510 } 511 512 /* 513 * Issue a callout_stop() and ensure that any callout race completes 514 * before returning. Does NOT de-initialized the callout. 515 */ 516 void 517 callout_stop_sync(struct callout *c) 518 { 519 softclock_pcpu_t sc; 520 521 while (c->c_flags & CALLOUT_DID_INIT) { 522 callout_stop(c); 523 if (c->c_gd) { 524 sc = &softclock_pcpu_ary[c->c_gd->gd_cpuid]; 525 if (sc->running == c) { 526 while (sc->running == c) 527 tsleep(&sc->running, 0, "crace", 1); 528 } 529 } 530 if ((c->c_flags & (CALLOUT_PENDING | CALLOUT_ACTIVE)) == 0) 531 break; 532 kprintf("Warning: %s: callout race\n", curthread->td_comm); 533 } 534 } 535 536 /* 537 * Terminate a callout 538 * 539 * This function will stop any pending callout and also block while the 540 * callout's function is running. It should only be used in cases where 541 * no deadlock is possible (due to the callout function acquiring locks 542 * that the current caller of callout_terminate() already holds), when 543 * the caller is ready to destroy the callout structure. 544 * 545 * This function clears the CALLOUT_DID_INIT flag. 546 * 547 * lwkt_token locks are ok. 548 */ 549 void 550 callout_terminate(struct callout *c) 551 { 552 softclock_pcpu_t sc; 553 554 if (c->c_flags & CALLOUT_DID_INIT) { 555 callout_stop(c); 556 sc = &softclock_pcpu_ary[c->c_gd->gd_cpuid]; 557 if (sc->running == c) { 558 while (sc->running == c) 559 tsleep(&sc->running, 0, "crace", 1); 560 } 561 KKASSERT((c->c_flags & (CALLOUT_PENDING|CALLOUT_ACTIVE)) == 0); 562 c->c_flags &= ~CALLOUT_DID_INIT; 563 } 564 } 565 566 /* 567 * Prepare a callout structure for use by callout_reset() and/or 568 * callout_stop(). The MP version of this routine requires that the callback 569 * function installed by callout_reset() be MP safe. 570 * 571 * The init functions can be called from any cpu and do not have to be 572 * called from the cpu that the timer will eventually run on. 573 */ 574 void 575 callout_init(struct callout *c) 576 { 577 bzero(c, sizeof *c); 578 c->c_flags = CALLOUT_DID_INIT; 579 } 580 581 void 582 callout_init_mp(struct callout *c) 583 { 584 callout_init(c); 585 c->c_flags |= CALLOUT_MPSAFE; 586 } 587