sys/kern/kern_timeout.c

   1 /*-
   2  * Copyright (c) 1982, 1986, 1991, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * (c) UNIX System Laboratories, Inc.
   5  * All or some portions of this file are derived from material licensed
   6  * to the University of California by American Telephone and Telegraph
   7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
   8  * the permission of UNIX System Laboratories, Inc.
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  * 1. Redistributions of source code must retain the above copyright
  14  *    notice, this list of conditions and the following disclaimer.
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice, this list of conditions and the following disclaimer in the
  17  *    documentation and/or other materials provided with the distribution.
  18  * 4. Neither the name of the University nor the names of its contributors
  19  *    may be used to endorse or promote products derived from this software
  20  *    without specific prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  *      From: @(#)kern_clock.c  8.5 (Berkeley) 1/21/94
  35  */
  36
  37 #include <sys/cdefs.h>
  38 __FBSDID("$FreeBSD$");
  39
  40 #include "opt_kdtrace.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/systm.h>
  44 #include <sys/bus.h>
  45 #include <sys/callout.h>
  46 #include <sys/condvar.h>
  47 #include <sys/interrupt.h>
  48 #include <sys/kernel.h>
  49 #include <sys/ktr.h>
  50 #include <sys/lock.h>
  51 #include <sys/malloc.h>
  52 #include <sys/mutex.h>
  53 #include <sys/proc.h>
  54 #include <sys/sdt.h>
  55 #include <sys/sleepqueue.h>
  56 #include <sys/sysctl.h>
  57 #include <sys/smp.h>
  58
  59 #ifdef SMP
  60 #include <machine/cpu.h>
  61 #endif
  62
  63 SDT_PROVIDER_DEFINE(callout_execute);
  64 SDT_PROBE_DEFINE(callout_execute, kernel, , callout_start, callout-start);
  65 SDT_PROBE_ARGTYPE(callout_execute, kernel, , callout_start, 0,
  66     "struct callout *");
  67 SDT_PROBE_DEFINE(callout_execute, kernel, , callout_end, callout-end);
  68 SDT_PROBE_ARGTYPE(callout_execute, kernel, , callout_end, 0,
  69     "struct callout *");
  70
  71 static int avg_depth;
  72 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
  73     "Average number of items examined per softclock call. Units = 1/1000");
  74 static int avg_gcalls;
  75 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
  76     "Average number of Giant callouts made per softclock call. Units = 1/1000");
  77 static int avg_lockcalls;
  78 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
  79     "Average number of lock callouts made per softclock call. Units = 1/1000");
  80 static int avg_mpcalls;
  81 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
  82     "Average number of MP callouts made per softclock call. Units = 1/1000");
  83 /*
  84  * TODO:
  85  *      allocate more timeout table slots when table overflows.
  86  */
  87 int callwheelsize, callwheelbits, callwheelmask;
  88
  89 /*
  90  * The callout cpu migration entity represents informations necessary for
  91  * describing the migrating callout to the new callout cpu.
  92  * The cached informations are very important for deferring migration when
  93  * the migrating callout is already running.
  94  */
  95 struct cc_mig_ent {
  96 #ifdef SMP
  97         void    (*ce_migration_func)(void *);
  98         void    *ce_migration_arg;
  99         int     ce_migration_cpu;
 100         int     ce_migration_ticks;
 101 #endif
 102 };
 103
 104 /*
 105  * There is one struct callout_cpu per cpu, holding all relevant
 106  * state for the callout processing thread on the individual CPU.
 107  * In particular:
 108  *      cc_ticks is incremented once per tick in callout_cpu().
 109  *      It tracks the global 'ticks' but in a way that the individual
 110  *      threads should not worry about races in the order in which
 111  *      hardclock() and hardclock_cpu() run on the various CPUs.
 112  *      cc_softclock is advanced in callout_cpu() to point to the
 113  *      first entry in cc_callwheel that may need handling. In turn,
 114  *      a softclock() is scheduled so it can serve the various entries i
 115  *      such that cc_softclock <= i <= cc_ticks .
 116  *      XXX maybe cc_softclock and cc_ticks should be volatile ?
 117  *
 118  *      cc_ticks is also used in callout_reset_cpu() to determine
 119  *      when the callout should be served.
 120  */
 121 struct callout_cpu {
 122         struct cc_mig_ent       cc_migrating_entity;
 123         struct mtx              cc_lock;
 124         struct callout          *cc_callout;
 125         struct callout_tailq    *cc_callwheel;
 126         struct callout_list     cc_callfree;
 127         struct callout          *cc_next;
 128         struct callout          *cc_curr;
 129         void                    *cc_cookie;
 130         int                     cc_ticks;
 131         int                     cc_softticks;
 132         int                     cc_cancel;
 133         int                     cc_waiting;
 134         int                     cc_firsttick;
 135 };
 136
 137 #ifdef SMP
 138 #define cc_migration_func       cc_migrating_entity.ce_migration_func
 139 #define cc_migration_arg        cc_migrating_entity.ce_migration_arg
 140 #define cc_migration_cpu        cc_migrating_entity.ce_migration_cpu
 141 #define cc_migration_ticks      cc_migrating_entity.ce_migration_ticks
 142
 143 struct callout_cpu cc_cpu[MAXCPU];
 144 #define CPUBLOCK        MAXCPU
 145 #define CC_CPU(cpu)     (&cc_cpu[(cpu)])
 146 #define CC_SELF()       CC_CPU(PCPU_GET(cpuid))
 147 #else
 148 struct callout_cpu cc_cpu;
 149 #define CC_CPU(cpu)     &cc_cpu
 150 #define CC_SELF()       &cc_cpu
 151 #endif
 152 #define CC_LOCK(cc)     mtx_lock_spin(&(cc)->cc_lock)
 153 #define CC_UNLOCK(cc)   mtx_unlock_spin(&(cc)->cc_lock)
 154 #define CC_LOCK_ASSERT(cc)      mtx_assert(&(cc)->cc_lock, MA_OWNED)
 155
 156 static int timeout_cpu;
 157 void (*callout_new_inserted)(int cpu, int ticks) = NULL;
 158
 159 MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
 160
 161 /**
 162  * Locked by cc_lock:
 163  *   cc_curr         - If a callout is in progress, it is curr_callout.
 164  *                     If curr_callout is non-NULL, threads waiting in
 165  *                     callout_drain() will be woken up as soon as the
 166  *                     relevant callout completes.
 167  *   cc_cancel       - Changing to 1 with both callout_lock and c_lock held
 168  *                     guarantees that the current callout will not run.
 169  *                     The softclock() function sets this to 0 before it
 170  *                     drops callout_lock to acquire c_lock, and it calls
 171  *                     the handler only if curr_cancelled is still 0 after
 172  *                     c_lock is successfully acquired.
 173  *   cc_waiting      - If a thread is waiting in callout_drain(), then
 174  *                     callout_wait is nonzero.  Set only when
 175  *                     curr_callout is non-NULL.
 176  */
 177
 178 /*
 179  * Resets the migration entity tied to a specific callout cpu.
 180  */
 181 static void
 182 cc_cme_cleanup(struct callout_cpu *cc)
 183 {
 184
 185 #ifdef SMP
 186         cc->cc_migration_cpu = CPUBLOCK;
 187         cc->cc_migration_ticks = 0;
 188         cc->cc_migration_func = NULL;
 189         cc->cc_migration_arg = NULL;
 190 #endif
 191 }
 192
 193 /*
 194  * Checks if migration is requested by a specific callout cpu.
 195  */
 196 static int
 197 cc_cme_migrating(struct callout_cpu *cc)
 198 {
 199
 200 #ifdef SMP
 201         return (cc->cc_migration_cpu != CPUBLOCK);
 202 #else
 203         return (0);
 204 #endif
 205 }
 206
 207 /*
 208  * kern_timeout_callwheel_alloc() - kernel low level callwheel initialization
 209  *
 210  *      This code is called very early in the kernel initialization sequence,
 211  *      and may be called more then once.
 212  */
 213 caddr_t
 214 kern_timeout_callwheel_alloc(caddr_t v)
 215 {
 216         struct callout_cpu *cc;
 217
 218         timeout_cpu = PCPU_GET(cpuid);
 219         cc = CC_CPU(timeout_cpu);
 220         /*
 221          * Calculate callout wheel size
 222          */
 223         for (callwheelsize = 1, callwheelbits = 0;
 224              callwheelsize < ncallout;
 225              callwheelsize <<= 1, ++callwheelbits)
 226                 ;
 227         callwheelmask = callwheelsize - 1;
 228
 229         cc->cc_callout = (struct callout *)v;
 230         v = (caddr_t)(cc->cc_callout + ncallout);
 231         cc->cc_callwheel = (struct callout_tailq *)v;
 232         v = (caddr_t)(cc->cc_callwheel + callwheelsize);
 233         return(v);
 234 }
 235
 236 static void
 237 callout_cpu_init(struct callout_cpu *cc)
 238 {
 239         struct callout *c;
 240         int i;
 241
 242         mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
 243         SLIST_INIT(&cc->cc_callfree);
 244         for (i = 0; i < callwheelsize; i++) {
 245                 TAILQ_INIT(&cc->cc_callwheel[i]);
 246         }
 247         cc_cme_cleanup(cc);
 248         if (cc->cc_callout == NULL)
 249                 return;
 250         for (i = 0; i < ncallout; i++) {
 251                 c = &cc->cc_callout[i];
 252                 callout_init(c, 0);
 253                 c->c_flags = CALLOUT_LOCAL_ALLOC;
 254                 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
 255         }
 256 }
 257
 258 #ifdef SMP
 259 /*
 260  * Switches the cpu tied to a specific callout.
 261  * The function expects a locked incoming callout cpu and returns with
 262  * locked outcoming callout cpu.
 263  */
 264 static struct callout_cpu *
 265 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
 266 {
 267         struct callout_cpu *new_cc;
 268
 269         MPASS(c != NULL && cc != NULL);
 270         CC_LOCK_ASSERT(cc);
 271
 272         /*
 273          * Avoid interrupts and preemption firing after the callout cpu
 274          * is blocked in order to avoid deadlocks as the new thread
 275          * may be willing to acquire the callout cpu lock.
 276          */
 277         c->c_cpu = CPUBLOCK;
 278         spinlock_enter();
 279         CC_UNLOCK(cc);
 280         new_cc = CC_CPU(new_cpu);
 281         CC_LOCK(new_cc);
 282         spinlock_exit();
 283         c->c_cpu = new_cpu;
 284         return (new_cc);
 285 }
 286 #endif
 287
 288 /*
 289  * kern_timeout_callwheel_init() - initialize previously reserved callwheel
 290  *                                 space.
 291  *
 292  *      This code is called just once, after the space reserved for the
 293  *      callout wheel has been finalized.
 294  */
 295 void
 296 kern_timeout_callwheel_init(void)
 297 {
 298         callout_cpu_init(CC_CPU(timeout_cpu));
 299 }
 300
 301 /*
 302  * Start standard softclock thread.
 303  */
 304 static void
 305 start_softclock(void *dummy)
 306 {
 307         struct callout_cpu *cc;
 308 #ifdef SMP
 309         int cpu;
 310 #endif
 311
 312         cc = CC_CPU(timeout_cpu);
 313         if (swi_add(&clk_intr_event, "clock", softclock, cc, SWI_CLOCK,
 314             INTR_MPSAFE, &cc->cc_cookie))
 315                 panic("died while creating standard software ithreads");
 316 #ifdef SMP
 317         CPU_FOREACH(cpu) {
 318                 if (cpu == timeout_cpu)
 319                         continue;
 320                 cc = CC_CPU(cpu);
 321                 if (swi_add(NULL, "clock", softclock, cc, SWI_CLOCK,
 322                     INTR_MPSAFE, &cc->cc_cookie))
 323                         panic("died while creating standard software ithreads");
 324                 cc->cc_callout = NULL;  /* Only cpu0 handles timeout(). */
 325                 cc->cc_callwheel = malloc(
 326                     sizeof(struct callout_tailq) * callwheelsize, M_CALLOUT,
 327                     M_WAITOK);
 328                 callout_cpu_init(cc);
 329         }
 330 #endif
 331 }
 332
 333 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
 334
 335 void
 336 callout_tick(void)
 337 {
 338         struct callout_cpu *cc;
 339         int need_softclock;
 340         int bucket;
 341
 342         /*
 343          * Process callouts at a very low cpu priority, so we don't keep the
 344          * relatively high clock interrupt priority any longer than necessary.
 345          */
 346         need_softclock = 0;
 347         cc = CC_SELF();
 348         mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
 349         cc->cc_firsttick = cc->cc_ticks = ticks;
 350         for (; (cc->cc_softticks - cc->cc_ticks) <= 0; cc->cc_softticks++) {
 351                 bucket = cc->cc_softticks & callwheelmask;
 352                 if (!TAILQ_EMPTY(&cc->cc_callwheel[bucket])) {
 353                         need_softclock = 1;
 354                         break;
 355                 }
 356         }
 357         mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
 358         /*
 359          * swi_sched acquires the thread lock, so we don't want to call it
 360          * with cc_lock held; incorrect locking order.
 361          */
 362         if (need_softclock)
 363                 swi_sched(cc->cc_cookie, 0);
 364 }
 365
 366 int
 367 callout_tickstofirst(int limit)
 368 {
 369         struct callout_cpu *cc;
 370         struct callout *c;
 371         struct callout_tailq *sc;
 372         int curticks;
 373         int skip = 1;
 374
 375         cc = CC_SELF();
 376         mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
 377         curticks = cc->cc_ticks;
 378         while( skip < ncallout && skip < limit ) {
 379                 sc = &cc->cc_callwheel[ (curticks+skip) & callwheelmask ];
 380                 /* search scanning ticks */
 381                 TAILQ_FOREACH( c, sc, c_links.tqe ){
 382                         if (c->c_time - curticks <= ncallout)
 383                                 goto out;
 384                 }
 385                 skip++;
 386         }
 387 out:
 388         cc->cc_firsttick = curticks + skip;
 389         mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
 390         return (skip);
 391 }
 392
 393 static struct callout_cpu *
 394 callout_lock(struct callout *c)
 395 {
 396         struct callout_cpu *cc;
 397         int cpu;
 398
 399         for (;;) {
 400                 cpu = c->c_cpu;
 401 #ifdef SMP
 402                 if (cpu == CPUBLOCK) {
 403                         while (c->c_cpu == CPUBLOCK)
 404                                 cpu_spinwait();
 405                         continue;
 406                 }
 407 #endif
 408                 cc = CC_CPU(cpu);
 409                 CC_LOCK(cc);
 410                 if (cpu == c->c_cpu)
 411                         break;
 412                 CC_UNLOCK(cc);
 413         }
 414         return (cc);
 415 }
 416
 417 static void
 418 callout_cc_add(struct callout *c, struct callout_cpu *cc, int to_ticks,
 419     void (*func)(void *), void *arg, int cpu)
 420 {
 421
 422         CC_LOCK_ASSERT(cc);
 423
 424         if (to_ticks <= 0)
 425                 to_ticks = 1;
 426         c->c_arg = arg;
 427         c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
 428         c->c_func = func;
 429         c->c_time = ticks + to_ticks;
 430         TAILQ_INSERT_TAIL(&cc->cc_callwheel[c->c_time & callwheelmask],
 431             c, c_links.tqe);
 432         if ((c->c_time - cc->cc_firsttick) < 0 &&
 433             callout_new_inserted != NULL) {
 434                 cc->cc_firsttick = c->c_time;
 435                 (*callout_new_inserted)(cpu,
 436                     to_ticks + (ticks - cc->cc_ticks));
 437         }
 438 }
 439
 440 /*
 441  * The callout mechanism is based on the work of Adam M. Costello and
 442  * George Varghese, published in a technical report entitled "Redesigning
 443  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
 444  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
 445  * used in this implementation was published by G. Varghese and T. Lauck in
 446  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
 447  * the Efficient Implementation of a Timer Facility" in the Proceedings of
 448  * the 11th ACM Annual Symposium on Operating Systems Principles,
 449  * Austin, Texas Nov 1987.
 450  */
 451
 452 /*
 453  * Software (low priority) clock interrupt.
 454  * Run periodic events from timeout queue.
 455  */
 456 void
 457 softclock(void *arg)
 458 {
 459         struct callout_cpu *cc;
 460         struct callout *c;
 461         struct callout_tailq *bucket;
 462         int curticks;
 463         int steps;      /* #steps since we last allowed interrupts */
 464         int depth;
 465         int mpcalls;
 466         int lockcalls;
 467         int gcalls;
 468 #ifdef DIAGNOSTIC
 469         struct bintime bt1, bt2;
 470         struct timespec ts2;
 471         static uint64_t maxdt = 36893488147419102LL;    /* 2 msec */
 472         static timeout_t *lastfunc;
 473 #endif
 474
 475 #ifndef MAX_SOFTCLOCK_STEPS
 476 #define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */
 477 #endif /* MAX_SOFTCLOCK_STEPS */
 478
 479         mpcalls = 0;
 480         lockcalls = 0;
 481         gcalls = 0;
 482         depth = 0;
 483         steps = 0;
 484         cc = (struct callout_cpu *)arg;
 485         CC_LOCK(cc);
 486         while (cc->cc_softticks - 1 != cc->cc_ticks) {
 487                 /*
 488                  * cc_softticks may be modified by hard clock, so cache
 489                  * it while we work on a given bucket.
 490                  */
 491                 curticks = cc->cc_softticks;
 492                 cc->cc_softticks++;
 493                 bucket = &cc->cc_callwheel[curticks & callwheelmask];
 494                 c = TAILQ_FIRST(bucket);
 495                 while (c) {
 496                         depth++;
 497                         if (c->c_time != curticks) {
 498                                 c = TAILQ_NEXT(c, c_links.tqe);
 499                                 ++steps;
 500                                 if (steps >= MAX_SOFTCLOCK_STEPS) {
 501                                         cc->cc_next = c;
 502                                         /* Give interrupts a chance. */
 503                                         CC_UNLOCK(cc);
 504                                         ;       /* nothing */
 505                                         CC_LOCK(cc);
 506                                         c = cc->cc_next;
 507                                         steps = 0;
 508                                 }
 509                         } else {
 510                                 void (*c_func)(void *);
 511                                 void *c_arg;
 512                                 struct lock_class *class;
 513                                 struct lock_object *c_lock;
 514                                 int c_flags, sharedlock;
 515
 516                                 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
 517                                 TAILQ_REMOVE(bucket, c, c_links.tqe);
 518                                 class = (c->c_lock != NULL) ?
 519                                     LOCK_CLASS(c->c_lock) : NULL;
 520                                 sharedlock = (c->c_flags & CALLOUT_SHAREDLOCK) ?
 521                                     0 : 1;
 522                                 c_lock = c->c_lock;
 523                                 c_func = c->c_func;
 524                                 c_arg = c->c_arg;
 525                                 c_flags = c->c_flags;
 526                                 if (c->c_flags & CALLOUT_LOCAL_ALLOC) {
 527                                         c->c_flags = CALLOUT_LOCAL_ALLOC;
 528                                 } else {
 529                                         c->c_flags =
 530                                             (c->c_flags & ~CALLOUT_PENDING);
 531                                 }
 532                                 cc->cc_curr = c;
 533                                 cc->cc_cancel = 0;
 534                                 CC_UNLOCK(cc);
 535                                 if (c_lock != NULL) {
 536                                         class->lc_lock(c_lock, sharedlock);
 537                                         /*
 538                                          * The callout may have been cancelled
 539                                          * while we switched locks.
 540                                          */
 541                                         if (cc->cc_cancel) {
 542                                                 class->lc_unlock(c_lock);
 543                                                 goto skip;
 544                                         }
 545                                         /* The callout cannot be stopped now. */
 546                                         cc->cc_cancel = 1;
 547
 548                                         if (c_lock == &Giant.lock_object) {
 549                                                 gcalls++;
 550                                                 CTR3(KTR_CALLOUT,
 551                                                     "callout %p func %p arg %p",
 552                                                     c, c_func, c_arg);
 553                                         } else {
 554                                                 lockcalls++;
 555                                                 CTR3(KTR_CALLOUT, "callout lock"
 556                                                     " %p func %p arg %p",
 557                                                     c, c_func, c_arg);
 558                                         }
 559                                 } else {
 560                                         mpcalls++;
 561                                         CTR3(KTR_CALLOUT,
 562                                             "callout mpsafe %p func %p arg %p",
 563                                             c, c_func, c_arg);
 564                                 }
 565 #ifdef DIAGNOSTIC
 566                                 binuptime(&bt1);
 567 #endif
 568                                 THREAD_NO_SLEEPING();
 569                                 SDT_PROBE(callout_execute, kernel, ,
 570                                     callout_start, c, 0, 0, 0, 0);
 571                                 c_func(c_arg);
 572                                 SDT_PROBE(callout_execute, kernel, ,
 573                                     callout_end, c, 0, 0, 0, 0);
 574                                 THREAD_SLEEPING_OK();
 575 #ifdef DIAGNOSTIC
 576                                 binuptime(&bt2);
 577                                 bintime_sub(&bt2, &bt1);
 578                                 if (bt2.frac > maxdt) {
 579                                         if (lastfunc != c_func ||
 580                                             bt2.frac > maxdt * 2) {
 581                                                 bintime2timespec(&bt2, &ts2);
 582                                                 printf(
 583                         "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
 584                                                     c_func, c_arg,
 585                                                     (intmax_t)ts2.tv_sec,
 586                                                     ts2.tv_nsec);
 587                                         }
 588                                         maxdt = bt2.frac;
 589                                         lastfunc = c_func;
 590                                 }
 591 #endif
 592                                 CTR1(KTR_CALLOUT, "callout %p finished", c);
 593                                 if ((c_flags & CALLOUT_RETURNUNLOCKED) == 0)
 594                                         class->lc_unlock(c_lock);
 595                         skip:
 596                                 CC_LOCK(cc);
 597                                 /*
 598                                  * If the current callout is locally
 599                                  * allocated (from timeout(9))
 600                                  * then put it on the freelist.
 601                                  *
 602                                  * Note: we need to check the cached
 603                                  * copy of c_flags because if it was not
 604                                  * local, then it's not safe to deref the
 605                                  * callout pointer.
 606                                  */
 607                                 if (c_flags & CALLOUT_LOCAL_ALLOC) {
 608                                         KASSERT(c->c_flags ==
 609                                             CALLOUT_LOCAL_ALLOC,
 610                                             ("corrupted callout"));
 611                                         c->c_func = NULL;
 612                                         SLIST_INSERT_HEAD(&cc->cc_callfree, c,
 613                                             c_links.sle);
 614                                 }
 615                                 cc->cc_curr = NULL;
 616                                 if (cc->cc_waiting) {
 617
 618                                         /*
 619                                          * There is someone waiting for the
 620                                          * callout to complete.
 621                                          * If the callout was scheduled for
 622                                          * migration just cancel it.
 623                                          */
 624                                         if (cc_cme_migrating(cc))
 625                                                 cc_cme_cleanup(cc);
 626                                         cc->cc_waiting = 0;
 627                                         CC_UNLOCK(cc);
 628                                         wakeup(&cc->cc_waiting);
 629                                         CC_LOCK(cc);
 630                                 } else if (cc_cme_migrating(cc)) {
 631 #ifdef SMP
 632                                         struct callout_cpu *new_cc;
 633                                         void (*new_func)(void *);
 634                                         void *new_arg;
 635                                         int new_cpu, new_ticks;
 636
 637                                         /*
 638                                          * If the callout was scheduled for
 639                                          * migration just perform it now.
 640                                          */
 641                                         new_cpu = cc->cc_migration_cpu;
 642                                         new_ticks = cc->cc_migration_ticks;
 643                                         new_func = cc->cc_migration_func;
 644                                         new_arg = cc->cc_migration_arg;
 645                                         cc_cme_cleanup(cc);
 646
 647                                         /*
 648                                          * Handle deferred callout stops
 649                                          */
 650                                         if ((c->c_flags & CALLOUT_DFRMIGRATION)
 651                                             == 0) {
 652                                                 CTR3(KTR_CALLOUT,
 653                                         "deferred cancelled %p func %p arg %p",
 654                                                     c, new_func, new_arg);
 655                                                 if (cc->cc_next == c) {
 656                                                         cc->cc_next =
 657                                                             TAILQ_NEXT(c,
 658                                                             c_links.tqe);
 659                                                 }
 660                                                 if (c->c_flags &
 661                                                     CALLOUT_LOCAL_ALLOC) {
 662                                                         c->c_func = NULL;
 663                                                         SLIST_INSERT_HEAD(
 664                                                             &cc->cc_callfree, c,
 665                                                             c_links.sle);
 666                                                 }
 667                                                 goto nextc;
 668                                         } else {
 669                                                 c->c_flags &= ~
 670                                                     CALLOUT_DFRMIGRATION;
 671                                         }
 672
 673                                         /*
 674                                          * It should be assert here that the
 675                                          * callout is not destroyed but that
 676                                          * is not easy.
 677                                          */
 678                                         new_cc = callout_cpu_switch(c, cc,
 679                                             new_cpu);
 680                                         callout_cc_add(c, new_cc, new_ticks,
 681                                             new_func, new_arg, new_cpu);
 682                                         CC_UNLOCK(new_cc);
 683                                         CC_LOCK(cc);
 684 #else
 685                                         panic("migration should not happen");
 686 #endif
 687                                 }
 688 #ifdef SMP
 689 nextc:
 690 #endif
 691                                 steps = 0;
 692                                 c = cc->cc_next;
 693                         }
 694                 }
 695         }
 696         avg_depth += (depth * 1000 - avg_depth) >> 8;
 697         avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
 698         avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
 699         avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
 700         cc->cc_next = NULL;
 701         CC_UNLOCK(cc);
 702 }
 703
 704 /*
 705  * timeout --
 706  *      Execute a function after a specified length of time.
 707  *
 708  * untimeout --
 709  *      Cancel previous timeout function call.
 710  *
 711  * callout_handle_init --
 712  *      Initialize a handle so that using it with untimeout is benign.
 713  *
 714  *      See AT&T BCI Driver Reference Manual for specification.  This
 715  *      implementation differs from that one in that although an
 716  *      identification value is returned from timeout, the original
 717  *      arguments to timeout as well as the identifier are used to
 718  *      identify entries for untimeout.
 719  */
 720 struct callout_handle
 721 timeout(ftn, arg, to_ticks)
 722         timeout_t *ftn;
 723         void *arg;
 724         int to_ticks;
 725 {
 726         struct callout_cpu *cc;
 727         struct callout *new;
 728         struct callout_handle handle;
 729
 730         cc = CC_CPU(timeout_cpu);
 731         CC_LOCK(cc);
 732         /* Fill in the next free callout structure. */
 733         new = SLIST_FIRST(&cc->cc_callfree);
 734         if (new == NULL)
 735                 /* XXX Attempt to malloc first */
 736                 panic("timeout table full");
 737         SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
 738         callout_reset(new, to_ticks, ftn, arg);
 739         handle.callout = new;
 740         CC_UNLOCK(cc);
 741
 742         return (handle);
 743 }
 744
 745 void
 746 untimeout(ftn, arg, handle)
 747         timeout_t *ftn;
 748         void *arg;
 749         struct callout_handle handle;
 750 {
 751         struct callout_cpu *cc;
 752
 753         /*
 754          * Check for a handle that was initialized
 755          * by callout_handle_init, but never used
 756          * for a real timeout.
 757          */
 758         if (handle.callout == NULL)
 759                 return;
 760
 761         cc = callout_lock(handle.callout);
 762         if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
 763                 callout_stop(handle.callout);
 764         CC_UNLOCK(cc);
 765 }
 766
 767 void
 768 callout_handle_init(struct callout_handle *handle)
 769 {
 770         handle->callout = NULL;
 771 }
 772
 773 /*
 774  * New interface; clients allocate their own callout structures.
 775  *
 776  * callout_reset() - establish or change a timeout
 777  * callout_stop() - disestablish a timeout
 778  * callout_init() - initialize a callout structure so that it can
 779  *      safely be passed to callout_reset() and callout_stop()
 780  *
 781  * <sys/callout.h> defines three convenience macros:
 782  *
 783  * callout_active() - returns truth if callout has not been stopped,
 784  *      drained, or deactivated since the last time the callout was
 785  *      reset.
 786  * callout_pending() - returns truth if callout is still waiting for timeout
 787  * callout_deactivate() - marks the callout as having been serviced
 788  */
 789 int
 790 callout_reset_on(struct callout *c, int to_ticks, void (*ftn)(void *),
 791     void *arg, int cpu)
 792 {
 793         struct callout_cpu *cc;
 794         int cancelled = 0;
 795
 796         /*
 797          * Don't allow migration of pre-allocated callouts lest they
 798          * become unbalanced.
 799          */
 800         if (c->c_flags & CALLOUT_LOCAL_ALLOC)
 801                 cpu = c->c_cpu;
 802         cc = callout_lock(c);
 803         if (cc->cc_curr == c) {
 804                 /*
 805                  * We're being asked to reschedule a callout which is
 806                  * currently in progress.  If there is a lock then we
 807                  * can cancel the callout if it has not really started.
 808                  */
 809                 if (c->c_lock != NULL && !cc->cc_cancel)
 810                         cancelled = cc->cc_cancel = 1;
 811                 if (cc->cc_waiting) {
 812                         /*
 813                          * Someone has called callout_drain to kill this
 814                          * callout.  Don't reschedule.
 815                          */
 816                         CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
 817                             cancelled ? "cancelled" : "failed to cancel",
 818                             c, c->c_func, c->c_arg);
 819                         CC_UNLOCK(cc);
 820                         return (cancelled);
 821                 }
 822         }
 823         if (c->c_flags & CALLOUT_PENDING) {
 824                 if (cc->cc_next == c) {
 825                         cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
 826                 }
 827                 TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
 828                     c_links.tqe);
 829
 830                 cancelled = 1;
 831                 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
 832         }
 833
 834 #ifdef SMP
 835         /*
 836          * If the callout must migrate try to perform it immediately.
 837          * If the callout is currently running, just defer the migration
 838          * to a more appropriate moment.
 839          */
 840         if (c->c_cpu != cpu) {
 841                 if (cc->cc_curr == c) {
 842                         cc->cc_migration_cpu = cpu;
 843                         cc->cc_migration_ticks = to_ticks;
 844                         cc->cc_migration_func = ftn;
 845                         cc->cc_migration_arg = arg;
 846                         c->c_flags |= CALLOUT_DFRMIGRATION;
 847                         CTR5(KTR_CALLOUT,
 848                     "migration of %p func %p arg %p in %d to %u deferred",
 849                             c, c->c_func, c->c_arg, to_ticks, cpu);
 850                         CC_UNLOCK(cc);
 851                         return (cancelled);
 852                 }
 853                 cc = callout_cpu_switch(c, cc, cpu);
 854         }
 855 #endif
 856
 857         callout_cc_add(c, cc, to_ticks, ftn, arg, cpu);
 858         CTR5(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d",
 859             cancelled ? "re" : "", c, c->c_func, c->c_arg, to_ticks);
 860         CC_UNLOCK(cc);
 861
 862         return (cancelled);
 863 }
 864
 865 /*
 866  * Common idioms that can be optimized in the future.
 867  */
 868 int
 869 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
 870 {
 871         return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
 872 }
 873
 874 int
 875 callout_schedule(struct callout *c, int to_ticks)
 876 {
 877         return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
 878 }
 879
 880 int
 881 _callout_stop_safe(c, safe)
 882         struct  callout *c;
 883         int     safe;
 884 {
 885         struct callout_cpu *cc, *old_cc;
 886         struct lock_class *class;
 887         int use_lock, sq_locked;
 888
 889         /*
 890          * Some old subsystems don't hold Giant while running a callout_stop(),
 891          * so just discard this check for the moment.
 892          */
 893         if (!safe && c->c_lock != NULL) {
 894                 if (c->c_lock == &Giant.lock_object)
 895                         use_lock = mtx_owned(&Giant);
 896                 else {
 897                         use_lock = 1;
 898                         class = LOCK_CLASS(c->c_lock);
 899                         class->lc_assert(c->c_lock, LA_XLOCKED);
 900                 }
 901         } else
 902                 use_lock = 0;
 903
 904         sq_locked = 0;
 905         old_cc = NULL;
 906 again:
 907         cc = callout_lock(c);
 908
 909         /*
 910          * If the callout was migrating while the callout cpu lock was
 911          * dropped,  just drop the sleepqueue lock and check the states
 912          * again.
 913          */
 914         if (sq_locked != 0 && cc != old_cc) {
 915 #ifdef SMP
 916                 CC_UNLOCK(cc);
 917                 sleepq_release(&old_cc->cc_waiting);
 918                 sq_locked = 0;
 919                 old_cc = NULL;
 920                 goto again;
 921 #else
 922                 panic("migration should not happen");
 923 #endif
 924         }
 925
 926         /*
 927          * If the callout isn't pending, it's not on the queue, so
 928          * don't attempt to remove it from the queue.  We can try to
 929          * stop it by other means however.
 930          */
 931         if (!(c->c_flags & CALLOUT_PENDING)) {
 932                 c->c_flags &= ~CALLOUT_ACTIVE;
 933
 934                 /*
 935                  * If it wasn't on the queue and it isn't the current
 936                  * callout, then we can't stop it, so just bail.
 937                  */
 938                 if (cc->cc_curr != c) {
 939                         CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
 940                             c, c->c_func, c->c_arg);
 941                         CC_UNLOCK(cc);
 942                         if (sq_locked)
 943                                 sleepq_release(&cc->cc_waiting);
 944                         return (0);
 945                 }
 946
 947                 if (safe) {
 948                         /*
 949                          * The current callout is running (or just
 950                          * about to run) and blocking is allowed, so
 951                          * just wait for the current invocation to
 952                          * finish.
 953                          */
 954                         while (cc->cc_curr == c) {
 955
 956                                 /*
 957                                  * Use direct calls to sleepqueue interface
 958                                  * instead of cv/msleep in order to avoid
 959                                  * a LOR between cc_lock and sleepqueue
 960                                  * chain spinlocks.  This piece of code
 961                                  * emulates a msleep_spin() call actually.
 962                                  *
 963                                  * If we already have the sleepqueue chain
 964                                  * locked, then we can safely block.  If we
 965                                  * don't already have it locked, however,
 966                                  * we have to drop the cc_lock to lock
 967                                  * it.  This opens several races, so we
 968                                  * restart at the beginning once we have
 969                                  * both locks.  If nothing has changed, then
 970                                  * we will end up back here with sq_locked
 971                                  * set.
 972                                  */
 973                                 if (!sq_locked) {
 974                                         CC_UNLOCK(cc);
 975                                         sleepq_lock(&cc->cc_waiting);
 976                                         sq_locked = 1;
 977                                         old_cc = cc;
 978                                         goto again;
 979                                 }
 980
 981                                 /*
 982                                  * Migration could be cancelled here, but
 983                                  * as long as it is still not sure when it
 984                                  * will be packed up, just let softclock()
 985                                  * take care of it.
 986                                  */
 987                                 cc->cc_waiting = 1;
 988                                 DROP_GIANT();
 989                                 CC_UNLOCK(cc);
 990                                 sleepq_add(&cc->cc_waiting,
 991                                     &cc->cc_lock.lock_object, "codrain",
 992                                     SLEEPQ_SLEEP, 0);
 993                                 sleepq_wait(&cc->cc_waiting, 0);
 994                                 sq_locked = 0;
 995                                 old_cc = NULL;
 996
 997                                 /* Reacquire locks previously released. */
 998                                 PICKUP_GIANT();
 999                                 CC_LOCK(cc);
1000                         }
1001                 } else if (use_lock && !cc->cc_cancel) {
1002                         /*
1003                          * The current callout is waiting for its
1004                          * lock which we hold.  Cancel the callout
1005                          * and return.  After our caller drops the
1006                          * lock, the callout will be skipped in
1007                          * softclock().
1008                          */
1009                         cc->cc_cancel = 1;
1010                         CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1011                             c, c->c_func, c->c_arg);
1012                         KASSERT(!cc_cme_migrating(cc),
1013                             ("callout wrongly scheduled for migration"));
1014                         CC_UNLOCK(cc);
1015                         KASSERT(!sq_locked, ("sleepqueue chain locked"));
1016                         return (1);
1017                 } else if ((c->c_flags & CALLOUT_DFRMIGRATION) != 0) {
1018                         c->c_flags &= ~CALLOUT_DFRMIGRATION;
1019                         CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1020                             c, c->c_func, c->c_arg);
1021                         CC_UNLOCK(cc);
1022                         return (1);
1023                 }
1024                 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1025                     c, c->c_func, c->c_arg);
1026                 CC_UNLOCK(cc);
1027                 KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1028                 return (0);
1029         }
1030         if (sq_locked)
1031                 sleepq_release(&cc->cc_waiting);
1032
1033         c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
1034
1035         if (cc->cc_next == c) {
1036                 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
1037         }
1038         TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
1039             c_links.tqe);
1040
1041         CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1042             c, c->c_func, c->c_arg);
1043
1044         if (c->c_flags & CALLOUT_LOCAL_ALLOC) {
1045                 c->c_func = NULL;
1046                 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
1047         }
1048         CC_UNLOCK(cc);
1049         return (1);
1050 }
1051
1052 void
1053 callout_init(c, mpsafe)
1054         struct  callout *c;
1055         int mpsafe;
1056 {
1057         bzero(c, sizeof *c);
1058         if (mpsafe) {
1059                 c->c_lock = NULL;
1060                 c->c_flags = CALLOUT_RETURNUNLOCKED;
1061         } else {
1062                 c->c_lock = &Giant.lock_object;
1063                 c->c_flags = 0;
1064         }
1065         c->c_cpu = timeout_cpu;
1066 }
1067
1068 void
1069 _callout_init_lock(c, lock, flags)
1070         struct  callout *c;
1071         struct  lock_object *lock;
1072         int flags;
1073 {
1074         bzero(c, sizeof *c);
1075         c->c_lock = lock;
1076         KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1077             ("callout_init_lock: bad flags %d", flags));
1078         KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1079             ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1080         KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1081             (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1082             __func__));
1083         c->c_flags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1084         c->c_cpu = timeout_cpu;
1085 }
1086
1087 #ifdef APM_FIXUP_CALLTODO
1088 /*
1089  * Adjust the kernel calltodo timeout list.  This routine is used after
1090  * an APM resume to recalculate the calltodo timer list values with the
1091  * number of hz's we have been sleeping.  The next hardclock() will detect
1092  * that there are fired timers and run softclock() to execute them.
1093  *
1094  * Please note, I have not done an exhaustive analysis of what code this
1095  * might break.  I am motivated to have my select()'s and alarm()'s that
1096  * have expired during suspend firing upon resume so that the applications
1097  * which set the timer can do the maintanence the timer was for as close
1098  * as possible to the originally intended time.  Testing this code for a
1099  * week showed that resuming from a suspend resulted in 22 to 25 timers
1100  * firing, which seemed independant on whether the suspend was 2 hours or
1101  * 2 days.  Your milage may vary.   - Ken Key <key@cs.utk.edu>
1102  */
1103 void
1104 adjust_timeout_calltodo(time_change)
1105     struct timeval *time_change;
1106 {
1107         register struct callout *p;
1108         unsigned long delta_ticks;
1109
1110         /*
1111          * How many ticks were we asleep?
1112          * (stolen from tvtohz()).
1113          */
1114
1115         /* Don't do anything */
1116         if (time_change->tv_sec < 0)
1117                 return;
1118         else if (time_change->tv_sec <= LONG_MAX / 1000000)
1119                 delta_ticks = (time_change->tv_sec * 1000000 +
1120                                time_change->tv_usec + (tick - 1)) / tick + 1;
1121         else if (time_change->tv_sec <= LONG_MAX / hz)
1122                 delta_ticks = time_change->tv_sec * hz +
1123                               (time_change->tv_usec + (tick - 1)) / tick + 1;
1124         else
1125                 delta_ticks = LONG_MAX;
1126
1127         if (delta_ticks > INT_MAX)
1128                 delta_ticks = INT_MAX;
1129
1130         /*
1131          * Now rip through the timer calltodo list looking for timers
1132          * to expire.
1133          */
1134
1135         /* don't collide with softclock() */
1136         CC_LOCK(cc);
1137         for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1138                 p->c_time -= delta_ticks;
1139
1140                 /* Break if the timer had more time on it than delta_ticks */
1141                 if (p->c_time > 0)
1142                         break;
1143
1144                 /* take back the ticks the timer didn't use (p->c_time <= 0) */
1145                 delta_ticks = -p->c_time;
1146         }
1147         CC_UNLOCK(cc);
1148
1149         return;
1150 }
1151 #endif /* APM_FIXUP_CALLTODO */