kernel - All lwkt thread now start out mpsafe part 1/2
[dragonfly.git] / sys / kern / kern_timeout.c
CommitLineData
92b561b7
MD
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/*
984263bc
MD
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 $
bb1a3ed4 73 * $DragonFly: src/sys/kern/kern_timeout.c,v 1.27 2007/11/14 18:27:52 swildner Exp $
92b561b7
MD
74 */
75/*
76 * DRAGONFLY BGL STATUS
77 *
78 * All the API functions should be MP safe.
79 *
80 * The callback functions will be flagged as being MP safe if the
81 * timeout structure is initialized with callout_init_mp() instead of
82 * callout_init().
83 *
84 * The helper threads cannot be made preempt-capable until after we
85 * clean up all the uses of splsoftclock() and related interlocks (which
86 * require the related functions to be MP safe as well).
87 */
88/*
89 * The callout mechanism is based on the work of Adam M. Costello and
90 * George Varghese, published in a technical report entitled "Redesigning
91 * the BSD Callout and Timer Facilities" and modified slightly for inclusion
92 * in FreeBSD by Justin T. Gibbs. The original work on the data structures
93 * used in this implementation was published by G. Varghese and T. Lauck in
94 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
95 * the Efficient Implementation of a Timer Facility" in the Proceedings of
96 * the 11th ACM Annual Symposium on Operating Systems Principles,
97 * Austin, Texas Nov 1987.
98 *
99 * The per-cpu augmentation was done by Matthew Dillon.
984263bc
MD
100 */
101
102#include <sys/param.h>
103#include <sys/systm.h>
104#include <sys/callout.h>
105#include <sys/kernel.h>
ef0fdad1 106#include <sys/interrupt.h>
4b5f931b 107#include <sys/thread.h>
684a93c4 108
88c4d2f6 109#include <sys/thread2.h>
684a93c4 110#include <sys/mplock2.h>
92b561b7
MD
111
112#ifndef MAX_SOFTCLOCK_STEPS
113#define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */
114#endif
115
116
117struct softclock_pcpu {
92b561b7
MD
118 struct callout_tailq *callwheel;
119 struct callout * volatile next;
120 int softticks; /* softticks index */
121 int curticks; /* per-cpu ticks counter */
122 int isrunning;
123 struct thread thread;
124
125};
126
127typedef struct softclock_pcpu *softclock_pcpu_t;
984263bc
MD
128
129/*
130 * TODO:
131 * allocate more timeout table slots when table overflows.
132 */
92b561b7
MD
133static MALLOC_DEFINE(M_CALLOUT, "callout", "callout structures");
134static int callwheelsize;
135static int callwheelbits;
136static int callwheelmask;
137static struct softclock_pcpu softclock_pcpu_ary[MAXCPU];
138
139static void softclock_handler(void *arg);
140
141static void
142swi_softclock_setup(void *arg)
143{
144 int cpu;
145 int i;
984263bc 146
92b561b7
MD
147 /*
148 * Figure out how large a callwheel we need. It must be a power of 2.
149 */
150 callwheelsize = 1;
151 callwheelbits = 0;
152 while (callwheelsize < ncallout) {
153 callwheelsize <<= 1;
154 ++callwheelbits;
155 }
156 callwheelmask = callwheelsize - 1;
984263bc 157
92b561b7
MD
158 /*
159 * Initialize per-cpu data structures.
160 */
161 for (cpu = 0; cpu < ncpus; ++cpu) {
162 softclock_pcpu_t sc;
92b561b7
MD
163
164 sc = &softclock_pcpu_ary[cpu];
165
77652cad 166 sc->callwheel = kmalloc(sizeof(*sc->callwheel) * callwheelsize,
92b561b7
MD
167 M_CALLOUT, M_WAITOK|M_ZERO);
168 for (i = 0; i < callwheelsize; ++i)
169 TAILQ_INIT(&sc->callwheel[i]);
170
92b561b7 171 /*
c1d0893d
MD
172 * Mark the softclock handler as being an interrupt thread
173 * even though it really isn't, but do not allow it to
174 * preempt other threads (do not assign td_preemptable).
175 *
176 * Kernel code now assumes that callouts do not preempt
177 * the cpu they were scheduled on.
92b561b7
MD
178 */
179 lwkt_create(softclock_handler, sc, NULL,
c9e9fb21
MD
180 &sc->thread, TDF_STOPREQ|TDF_INTTHREAD|TDF_MPSAFE,
181 cpu, "softclock %d", cpu);
92b561b7
MD
182 }
183}
184
ed99ec35
MD
185/*
186 * Must occur after ncpus has been initialized.
187 */
ba39e2e0
MD
188SYSINIT(softclock_setup, SI_BOOT2_SOFTCLOCK, SI_ORDER_SECOND,
189 swi_softclock_setup, NULL);
984263bc
MD
190
191/*
92b561b7
MD
192 * This routine is called from the hardclock() (basically a FASTint/IPI) on
193 * each cpu in the system. sc->curticks is this cpu's notion of the timebase.
194 * It IS NOT NECESSARILY SYNCHRONIZED WITH 'ticks'! sc->softticks is where
195 * the callwheel is currently indexed.
196 *
197 * WARNING! The MP lock is not necessarily held on call, nor can it be
198 * safely obtained.
199 *
200 * sc->softticks is adjusted by either this routine or our helper thread
201 * depending on whether the helper thread is running or not.
984263bc 202 */
92b561b7
MD
203void
204hardclock_softtick(globaldata_t gd)
205{
206 softclock_pcpu_t sc;
207
208 sc = &softclock_pcpu_ary[gd->gd_cpuid];
209 ++sc->curticks;
210 if (sc->isrunning)
211 return;
212 if (sc->softticks == sc->curticks) {
213 /*
214 * in sync, only wakeup the thread if there is something to
215 * do.
216 */
217 if (TAILQ_FIRST(&sc->callwheel[sc->softticks & callwheelmask]))
218 {
219 sc->isrunning = 1;
220 lwkt_schedule(&sc->thread);
221 } else {
222 ++sc->softticks;
223 }
224 } else {
225 /*
226 * out of sync, wakeup the thread unconditionally so it can
227 * catch up.
228 */
229 sc->isrunning = 1;
230 lwkt_schedule(&sc->thread);
231 }
232}
984263bc
MD
233
234/*
92b561b7
MD
235 * This procedure is the main loop of our per-cpu helper thread. The
236 * sc->isrunning flag prevents us from racing hardclock_softtick() and
4ac1a4cd
MD
237 * a critical section is sufficient to interlock sc->curticks and protect
238 * us from remote IPI's / list removal.
92b561b7 239 *
c9e9fb21
MD
240 * The thread starts with the MP lock released and not in a critical
241 * section. The loop itself is MP safe while individual callbacks
242 * may or may not be, so we obtain or release the MP lock as appropriate.
984263bc 243 */
ef0fdad1 244static void
92b561b7 245softclock_handler(void *arg)
984263bc 246{
92b561b7 247 softclock_pcpu_t sc;
1fd87d54
RG
248 struct callout *c;
249 struct callout_tailq *bucket;
92b561b7
MD
250 void (*c_func)(void *);
251 void *c_arg;
92b561b7 252#ifdef SMP
c9e9fb21 253 int mpsafe = 1;
92b561b7
MD
254#endif
255
abb6f10a
MD
256 lwkt_setpri_self(TDPRI_SOFT_NORM);
257
92b561b7 258 sc = arg;
88c4d2f6 259 crit_enter();
92b561b7
MD
260loop:
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) {
4ac1a4cd
MD
265 if (c->c_time != sc->softticks) {
266 sc->next = TAILQ_NEXT(c, c_links.tqe);
92b561b7 267 continue;
4ac1a4cd
MD
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);
92b561b7 292 TAILQ_REMOVE(bucket, c, c_links.tqe);
4ac1a4cd 293
92b561b7
MD
294 c_func = c->c_func;
295 c_arg = c->c_arg;
92b561b7
MD
296 c->c_func = NULL;
297 KKASSERT(c->c_flags & CALLOUT_DID_INIT);
7eb3b2ae 298 c->c_flags &= ~CALLOUT_PENDING;
92b561b7 299 crit_exit();
92b561b7
MD
300 c_func(c_arg);
301 crit_enter();
302 /* NOTE: list may have changed */
984263bc 303 }
92b561b7 304 ++sc->softticks;
984263bc 305 }
92b561b7
MD
306 sc->isrunning = 0;
307 lwkt_deschedule_self(&sc->thread); /* == curthread */
308 lwkt_switch();
309 goto loop;
310 /* NOT REACHED */
984263bc
MD
311}
312
984263bc
MD
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
92b561b7
MD
319 * safely be passed to callout_reset() and callout_stop()
320 * callout_init_mp() - same but any installed functions must be MP safe.
984263bc
MD
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 */
92b561b7
MD
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.
7ac978db
MD
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.
92b561b7 341 */
984263bc 342void
92b561b7
MD
343callout_reset(struct callout *c, int to_ticks, void (*ftn)(void *),
344 void *arg)
984263bc 345{
92b561b7
MD
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);
6ea70f76 352 kprintf(
92b561b7
MD
353 "callout_reset(%p) from %p: callout was not initialized\n",
354 c, ((int **)&c)[-1]);
7ce2998e 355 print_backtrace(-1);
92b561b7
MD
356 }
357#endif
358 gd = mycpu;
359 sc = &softclock_pcpu_ary[gd->gd_cpuid];
360 crit_enter_gd(gd);
984263bc 361
984263bc
MD
362 if (c->c_flags & CALLOUT_PENDING)
363 callout_stop(c);
364
984263bc
MD
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;
92b561b7
MD
371 c->c_time = sc->curticks + to_ticks;
372#ifdef SMP
373 c->c_gd = gd;
374#endif
375
376 TAILQ_INSERT_TAIL(&sc->callwheel[c->c_time & callwheelmask],
984263bc 377 c, c_links.tqe);
92b561b7 378 crit_exit_gd(gd);
984263bc
MD
379}
380
92b561b7
MD
381/*
382 * Stop a running timer. WARNING! If called on a cpu other then the one
383 * the callout was started on this function will liveloop on its IPI to
384 * the target cpu to process the request. It is possible for the callout
385 * to execute in that case.
386 *
7ac978db
MD
387 * WARNING! This function may be called from any cpu but the caller must
388 * serialize callout_stop() and callout_reset() calls on the passed
389 * structure regardless of cpu.
390 *
92b561b7
MD
391 * WARNING! This routine may be called from an IPI
392 */
984263bc 393int
216ffe3f 394callout_stop(struct callout *c)
984263bc 395{
92b561b7
MD
396 globaldata_t gd = mycpu;
397#ifdef SMP
398 globaldata_t tgd;
399#endif
400 softclock_pcpu_t sc;
401
402#ifdef INVARIANTS
403 if ((c->c_flags & CALLOUT_DID_INIT) == 0) {
404 callout_init(c);
6ea70f76 405 kprintf(
634056cf 406 "callout_stop(%p) from %p: callout was not initialized\n",
92b561b7 407 c, ((int **)&c)[-1]);
7ce2998e 408 print_backtrace(-1);
92b561b7
MD
409 }
410#endif
411 crit_enter_gd(gd);
984263bc 412
984263bc 413 /*
7ac978db
MD
414 * Don't attempt to delete a callout that's not on the queue. The
415 * callout may not have a cpu assigned to it. Callers do not have
416 * to be on the issuing cpu but must still serialize access to the
417 * callout structure.
418 *
419 * We are not cpu-localized here and cannot safely modify the
420 * flags field in the callout structure. Note that most of the
421 * time CALLOUT_ACTIVE will be 0 if CALLOUT_PENDING is also 0.
422 *
423 * If we race another cpu's dispatch of this callout it is possible
424 * for CALLOUT_ACTIVE to be set with CALLOUT_PENDING unset. This
425 * will cause us to fall through and synchronize with the other
426 * cpu.
984263bc 427 */
92b561b7 428 if ((c->c_flags & CALLOUT_PENDING) == 0) {
7ac978db
MD
429#ifdef SMP
430 if ((c->c_flags & CALLOUT_ACTIVE) == 0) {
431 crit_exit_gd(gd);
432 return (0);
433 }
434 if (c->c_gd == NULL || c->c_gd == gd) {
435 c->c_flags &= ~CALLOUT_ACTIVE;
436 crit_exit_gd(gd);
437 return (0);
438 }
439 /* fall-through to the cpu-localization code. */
440#else
984263bc 441 c->c_flags &= ~CALLOUT_ACTIVE;
92b561b7 442 crit_exit_gd(gd);
984263bc 443 return (0);
7ac978db 444#endif
984263bc 445 }
92b561b7
MD
446#ifdef SMP
447 if ((tgd = c->c_gd) != gd) {
448 /*
449 * If the callout is owned by a different CPU we have to
450 * execute the function synchronously on the target cpu.
451 */
452 int seq;
453
35238fa5 454 cpu_ccfence(); /* don't let tgd alias c_gd */
92b561b7
MD
455 seq = lwkt_send_ipiq(tgd, (void *)callout_stop, c);
456 lwkt_wait_ipiq(tgd, seq);
457 } else
458#endif
459 {
460 /*
461 * If the callout is owned by the same CPU we can
462 * process it directly, but if we are racing our helper
463 * thread (sc->next), we have to adjust sc->next. The
464 * race is interlocked by a critical section.
465 */
466 sc = &softclock_pcpu_ary[gd->gd_cpuid];
984263bc 467
92b561b7
MD
468 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
469 if (sc->next == c)
470 sc->next = TAILQ_NEXT(c, c_links.tqe);
471
472 TAILQ_REMOVE(&sc->callwheel[c->c_time & callwheelmask],
473 c, c_links.tqe);
474 c->c_func = NULL;
984263bc 475 }
92b561b7 476 crit_exit_gd(gd);
984263bc
MD
477 return (1);
478}
479
92b561b7
MD
480/*
481 * Prepare a callout structure for use by callout_reset() and/or
482 * callout_stop(). The MP version of this routine requires that the callback
986d8a07 483 * function installed by callout_reset() be MP safe.
48e7b118
MD
484 *
485 * The init functions can be called from any cpu and do not have to be
486 * called from the cpu that the timer will eventually run on.
92b561b7 487 */
984263bc 488void
216ffe3f 489callout_init(struct callout *c)
984263bc
MD
490{
491 bzero(c, sizeof *c);
92b561b7 492 c->c_flags = CALLOUT_DID_INIT;
984263bc
MD
493}
494
92b561b7
MD
495void
496callout_init_mp(struct callout *c)
ef0fdad1 497{
92b561b7
MD
498 callout_init(c);
499 c->c_flags |= CALLOUT_MPSAFE;
ef0fdad1
MD
500}
501
92b561b7
MD
502/* What, are you joking? This is nuts! -Matt */
503#if 0
984263bc
MD
504#ifdef APM_FIXUP_CALLTODO
505/*
506 * Adjust the kernel calltodo timeout list. This routine is used after
507 * an APM resume to recalculate the calltodo timer list values with the
508 * number of hz's we have been sleeping. The next hardclock() will detect
509 * that there are fired timers and run softclock() to execute them.
510 *
511 * Please note, I have not done an exhaustive analysis of what code this
512 * might break. I am motivated to have my select()'s and alarm()'s that
513 * have expired during suspend firing upon resume so that the applications
514 * which set the timer can do the maintanence the timer was for as close
515 * as possible to the originally intended time. Testing this code for a
516 * week showed that resuming from a suspend resulted in 22 to 25 timers
517 * firing, which seemed independant on whether the suspend was 2 hours or
518 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu>
519 */
520void
216ffe3f 521adjust_timeout_calltodo(struct timeval *time_change)
984263bc 522{
1fd87d54 523 struct callout *p;
984263bc 524 unsigned long delta_ticks;
984263bc
MD
525
526 /*
527 * How many ticks were we asleep?
528 * (stolen from tvtohz()).
529 */
530
531 /* Don't do anything */
532 if (time_change->tv_sec < 0)
533 return;
534 else if (time_change->tv_sec <= LONG_MAX / 1000000)
535 delta_ticks = (time_change->tv_sec * 1000000 +
536 time_change->tv_usec + (tick - 1)) / tick + 1;
537 else if (time_change->tv_sec <= LONG_MAX / hz)
538 delta_ticks = time_change->tv_sec * hz +
539 (time_change->tv_usec + (tick - 1)) / tick + 1;
540 else
541 delta_ticks = LONG_MAX;
542
543 if (delta_ticks > INT_MAX)
544 delta_ticks = INT_MAX;
545
546 /*
547 * Now rip through the timer calltodo list looking for timers
548 * to expire.
549 */
550
551 /* don't collide with softclock() */
e43a034f 552 crit_enter();
984263bc
MD
553 for (p = calltodo.c_next; p != NULL; p = p->c_next) {
554 p->c_time -= delta_ticks;
555
556 /* Break if the timer had more time on it than delta_ticks */
557 if (p->c_time > 0)
558 break;
559
560 /* take back the ticks the timer didn't use (p->c_time <= 0) */
561 delta_ticks = -p->c_time;
562 }
e43a034f 563 crit_exit();
984263bc
MD
564
565 return;
566}
567#endif /* APM_FIXUP_CALLTODO */
92b561b7
MD
568#endif
569