kernel - Major signal path adjustments to fix races, tsleep race fixes, +more
[dragonfly.git] / sys / kern / kern_synch.c
CommitLineData
984263bc
MD
1/*-
2 * Copyright (c) 1982, 1986, 1990, 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 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * SUCH DAMAGE.
37 *
38 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
39 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $
c730be20 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.91 2008/09/09 04:06:13 dillon Exp $
984263bc
MD
41 */
42
43#include "opt_ktrace.h"
44
45#include <sys/param.h>
46#include <sys/systm.h>
47#include <sys/proc.h>
48#include <sys/kernel.h>
49#include <sys/signalvar.h>
50#include <sys/resourcevar.h>
51#include <sys/vmmeter.h>
52#include <sys/sysctl.h>
344ad853 53#include <sys/lock.h>
984263bc 54#include <sys/uio.h>
fc9ae81d 55#ifdef KTRACE
984263bc
MD
56#include <sys/ktrace.h>
57#endif
f1d1c3fa 58#include <sys/xwait.h>
9afb0ffd 59#include <sys/ktr.h>
684a93c4 60#include <sys/serialize.h>
984263bc 61
684a93c4 62#include <sys/signal2.h>
bf765287
MD
63#include <sys/thread2.h>
64#include <sys/spinlock2.h>
7f6220a9 65#include <sys/mutex2.h>
bf765287 66
984263bc 67#include <machine/cpu.h>
984263bc
MD
68#include <machine/smp.h>
69
fc17ad60
MD
70TAILQ_HEAD(tslpque, thread);
71
402ed7e1 72static void sched_setup (void *dummy);
984263bc
MD
73SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
74
984263bc
MD
75int hogticks;
76int lbolt;
344ad853 77int lbolt_syncer;
984263bc 78int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
17a9f566 79int ncpus;
da23a592
MD
80int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */
81int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */
e43a034f 82int safepri;
dbcd0c9b 83int tsleep_now_works;
5ea440eb 84int tsleep_crypto_dump = 0;
984263bc
MD
85
86static struct callout loadav_callout;
35f9d051 87static struct callout schedcpu_callout;
fc17ad60 88MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
984263bc 89
5decebc7
MD
90#define __DEALL(ident) __DEQUALIFY(void *, ident)
91
9afb0ffd
MD
92#if !defined(KTR_TSLEEP)
93#define KTR_TSLEEP KTR_ALL
94#endif
95KTR_INFO_MASTER(tsleep);
8aa3430c
MD
96KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *));
97KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0);
98KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *));
99KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0);
d9345d3a 100KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *));
8aa3430c
MD
101
102#define logtsleep1(name) KTR_LOG(tsleep_ ## name)
103#define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
9afb0ffd 104
984263bc
MD
105struct loadavg averunnable =
106 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
107/*
108 * Constants for averages over 1, 5, and 15 minutes
109 * when sampling at 5 second intervals.
110 */
111static fixpt_t cexp[3] = {
112 0.9200444146293232 * FSCALE, /* exp(-1/12) */
113 0.9834714538216174 * FSCALE, /* exp(-1/60) */
114 0.9944598480048967 * FSCALE, /* exp(-1/180) */
115};
116
402ed7e1
RG
117static void endtsleep (void *);
118static void loadav (void *arg);
402ed7e1 119static void schedcpu (void *arg);
984263bc 120
a46fac56
MD
121/*
122 * Adjust the scheduler quantum. The quantum is specified in microseconds.
123 * Note that 'tick' is in microseconds per tick.
124 */
984263bc
MD
125static int
126sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
127{
128 int error, new_val;
129
a591f597 130 new_val = sched_quantum * ustick;
984263bc
MD
131 error = sysctl_handle_int(oidp, &new_val, 0, req);
132 if (error != 0 || req->newptr == NULL)
133 return (error);
a591f597 134 if (new_val < ustick)
984263bc 135 return (EINVAL);
a591f597 136 sched_quantum = new_val / ustick;
984263bc
MD
137 hogticks = 2 * sched_quantum;
138 return (0);
139}
140
141SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
142 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
143
984263bc
MD
144/*
145 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
146 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
147 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
148 *
149 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
dcc99b62 150 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
984263bc
MD
151 *
152 * If you don't want to bother with the faster/more-accurate formula, you
153 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
154 * (more general) method of calculating the %age of CPU used by a process.
dcc99b62 155 *
08f2f1bb 156 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
dcc99b62
MD
157 */
158#define CCPU_SHIFT 11
159
160static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
161SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
162
163/*
164 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
984263bc 165 */
460426e6 166int fscale __unused = FSCALE; /* exported to systat */
dcc99b62 167SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
984263bc
MD
168
169/*
0a3f9b47 170 * Recompute process priorities, once a second.
dcc99b62
MD
171 *
172 * Since the userland schedulers are typically event oriented, if the
173 * estcpu calculation at wakeup() time is not sufficient to make a
174 * process runnable relative to other processes in the system we have
175 * a 1-second recalc to help out.
176 *
177 * This code also allows us to store sysclock_t data in the process structure
178 * without fear of an overrun, since sysclock_t are guarenteed to hold
179 * several seconds worth of count.
8fa76237
MD
180 *
181 * WARNING! callouts can preempt normal threads. However, they will not
182 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
984263bc 183 */
8fa76237
MD
184static int schedcpu_stats(struct proc *p, void *data __unused);
185static int schedcpu_resource(struct proc *p, void *data __unused);
186
984263bc 187static void
26a0694b 188schedcpu(void *arg)
984263bc 189{
8fa76237
MD
190 allproc_scan(schedcpu_stats, NULL);
191 allproc_scan(schedcpu_resource, NULL);
192 wakeup((caddr_t)&lbolt);
193 wakeup((caddr_t)&lbolt_syncer);
194 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
195}
196
197/*
198 * General process statistics once a second
199 */
200static int
201schedcpu_stats(struct proc *p, void *data __unused)
202{
08f2f1bb
SS
203 struct lwp *lp;
204
7bea4e64
MD
205 /*
206 * Threads may not be completely set up if process in SIDL state.
207 */
208 if (p->p_stat == SIDL)
209 return(0);
210
0d78b86e 211 PHOLD(p);
85946b6c
MD
212 if (lwkt_trytoken(&p->p_token) == FALSE) {
213 PRELE(p);
214 return(0);
215 }
0d78b86e 216
8fa76237 217 p->p_swtime++;
c7e98b2f
SS
218 FOREACH_LWP_IN_PROC(lp, p) {
219 if (lp->lwp_stat == LSSLEEP)
220 lp->lwp_slptime++;
4b5f931b 221
c7e98b2f
SS
222 /*
223 * Only recalculate processes that are active or have slept
224 * less then 2 seconds. The schedulers understand this.
225 */
226 if (lp->lwp_slptime <= 1) {
227 p->p_usched->recalculate(lp);
228 } else {
229 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
230 }
8fa76237 231 }
0d78b86e 232 lwkt_reltoken(&p->p_token);
d2d8515b 233 lwkt_yield();
0d78b86e 234 PRELE(p);
8fa76237
MD
235 return(0);
236}
a46fac56 237
8fa76237 238/*
84204577 239 * Resource checks. XXX break out since ksignal/killproc can block,
8fa76237
MD
240 * limiting us to one process killed per second. There is probably
241 * a better way.
242 */
243static int
244schedcpu_resource(struct proc *p, void *data __unused)
245{
246 u_int64_t ttime;
08f2f1bb 247 struct lwp *lp;
8fa76237 248
0d78b86e
MD
249 if (p->p_stat == SIDL)
250 return(0);
251
252 PHOLD(p);
85946b6c
MD
253 if (lwkt_trytoken(&p->p_token) == FALSE) {
254 PRELE(p);
255 return(0);
256 }
0d78b86e
MD
257
258 if (p->p_stat == SZOMB || p->p_limit == NULL) {
259 lwkt_reltoken(&p->p_token);
260 PRELE(p);
8fa76237 261 return(0);
984263bc 262 }
344ad853 263
c7e98b2f
SS
264 ttime = 0;
265 FOREACH_LWP_IN_PROC(lp, p) {
e595c6cd
MD
266 /*
267 * We may have caught an lp in the middle of being
268 * created, lwp_thread can be NULL.
269 */
270 if (lp->lwp_thread) {
271 ttime += lp->lwp_thread->td_sticks;
272 ttime += lp->lwp_thread->td_uticks;
273 }
c7e98b2f 274 }
8fa76237
MD
275
276 switch(plimit_testcpulimit(p->p_limit, ttime)) {
277 case PLIMIT_TESTCPU_KILL:
278 killproc(p, "exceeded maximum CPU limit");
279 break;
280 case PLIMIT_TESTCPU_XCPU:
4643740a
MD
281 if ((p->p_flags & P_XCPU) == 0) {
282 p->p_flags |= P_XCPU;
84204577 283 ksignal(p, SIGXCPU);
344ad853 284 }
8fa76237
MD
285 break;
286 default:
c0b8a06d 287 break;
344ad853 288 }
0d78b86e 289 lwkt_reltoken(&p->p_token);
d2d8515b 290 lwkt_yield();
0d78b86e 291 PRELE(p);
8fa76237 292 return(0);
984263bc
MD
293}
294
295/*
dcc99b62
MD
296 * This is only used by ps. Generate a cpu percentage use over
297 * a period of one second.
52eedfb5
MD
298 *
299 * MPSAFE
984263bc 300 */
dcc99b62 301void
553ea3c8 302updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
984263bc 303{
dcc99b62
MD
304 fixpt_t acc;
305 int remticks;
306
307 acc = (cpticks << FSHIFT) / ttlticks;
308 if (ttlticks >= ESTCPUFREQ) {
553ea3c8 309 lp->lwp_pctcpu = acc;
dcc99b62
MD
310 } else {
311 remticks = ESTCPUFREQ - ttlticks;
553ea3c8 312 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
dcc99b62 313 ESTCPUFREQ;
a46fac56 314 }
984263bc
MD
315}
316
317/*
8aa3430c
MD
318 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
319 * like addresses being slept on.
984263bc 320 */
b12defdc
MD
321#define TABLESIZE 4001
322#define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE)
984263bc 323
fc17ad60
MD
324static cpumask_t slpque_cpumasks[TABLESIZE];
325
984263bc 326/*
a46fac56 327 * General scheduler initialization. We force a reschedule 25 times
fc17ad60
MD
328 * a second by default. Note that cpu0 is initialized in early boot and
329 * cannot make any high level calls.
330 *
331 * Each cpu has its own sleep queue.
984263bc 332 */
984263bc 333void
fc17ad60 334sleep_gdinit(globaldata_t gd)
984263bc 335{
fc17ad60 336 static struct tslpque slpque_cpu0[TABLESIZE];
9c1fad94 337 int i;
984263bc 338
fc17ad60
MD
339 if (gd->gd_cpuid == 0) {
340 sched_quantum = (hz + 24) / 25;
341 hogticks = 2 * sched_quantum;
342
343 gd->gd_tsleep_hash = slpque_cpu0;
344 } else {
77652cad 345 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
fc17ad60
MD
346 M_TSLEEP, M_WAITOK | M_ZERO);
347 }
348 for (i = 0; i < TABLESIZE; ++i)
349 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
984263bc
MD
350}
351
352/*
ae8e83e6
MD
353 * This is a dandy function that allows us to interlock tsleep/wakeup
354 * operations with unspecified upper level locks, such as lockmgr locks,
355 * simply by holding a critical section. The sequence is:
356 *
357 * (acquire upper level lock)
358 * tsleep_interlock(blah)
359 * (release upper level lock)
360 * tsleep(blah, ...)
361 *
362 * Basically this functions queues us on the tsleep queue without actually
363 * descheduling us. When tsleep() is later called with PINTERLOCK it
364 * assumes the thread was already queued, otherwise it queues it there.
365 *
366 * Thus it is possible to receive the wakeup prior to going to sleep and
367 * the race conditions are covered.
368 */
369static __inline void
5decebc7 370_tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags)
ae8e83e6
MD
371{
372 thread_t td = gd->gd_curthread;
373 int id;
374
375 crit_enter_quick(td);
376 if (td->td_flags & TDF_TSLEEPQ) {
377 id = LOOKUP(td->td_wchan);
378 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
b12defdc
MD
379 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) {
380 atomic_clear_cpumask(&slpque_cpumasks[id],
381 gd->gd_cpumask);
382 }
ae8e83e6
MD
383 } else {
384 td->td_flags |= TDF_TSLEEPQ;
385 }
386 id = LOOKUP(ident);
387 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq);
da23a592 388 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
ae8e83e6
MD
389 td->td_wchan = ident;
390 td->td_wdomain = flags & PDOMAIN_MASK;
ae8e83e6
MD
391 crit_exit_quick(td);
392}
393
394void
5decebc7 395tsleep_interlock(const volatile void *ident, int flags)
ae8e83e6
MD
396{
397 _tsleep_interlock(mycpu, ident, flags);
398}
399
400/*
401 * Remove thread from sleepq. Must be called with a critical section held.
4643740a 402 * The thread must not be migrating.
ae8e83e6
MD
403 */
404static __inline void
405_tsleep_remove(thread_t td)
406{
407 globaldata_t gd = mycpu;
408 int id;
409
957625b2 410 KKASSERT(td->td_gd == gd && IN_CRITICAL_SECT(td));
4643740a 411 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
ae8e83e6
MD
412 if (td->td_flags & TDF_TSLEEPQ) {
413 td->td_flags &= ~TDF_TSLEEPQ;
414 id = LOOKUP(td->td_wchan);
415 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq);
416 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
da23a592 417 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask);
ae8e83e6
MD
418 td->td_wchan = NULL;
419 td->td_wdomain = 0;
420 }
421}
422
423void
424tsleep_remove(thread_t td)
425{
426 _tsleep_remove(td);
427}
428
429/*
984263bc
MD
430 * General sleep call. Suspends the current process until a wakeup is
431 * performed on the specified identifier. The process will then be made
432 * runnable with the specified priority. Sleeps at most timo/hz seconds
377d4740 433 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
984263bc
MD
434 * before and after sleeping, else signals are not checked. Returns 0 if
435 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
436 * signal needs to be delivered, ERESTART is returned if the current system
437 * call should be restarted if possible, and EINTR is returned if the system
438 * call should be interrupted by the signal (return EINTR).
26a0694b 439 *
0a3f9b47
MD
440 * Note that if we are a process, we release_curproc() before messing with
441 * the LWKT scheduler.
a46fac56
MD
442 *
443 * During autoconfiguration or after a panic, a sleep will simply
444 * lower the priority briefly to allow interrupts, then return.
94f98873
MD
445 *
446 * WARNING! This code can't block (short of switching away), or bad things
447 * will happen. No getting tokens, no blocking locks, etc.
984263bc
MD
448 */
449int
5decebc7 450tsleep(const volatile void *ident, int flags, const char *wmesg, int timo)
984263bc 451{
dadab5e9 452 struct thread *td = curthread;
08f2f1bb 453 struct lwp *lp = td->td_lwp;
0cfcada1 454 struct proc *p = td->td_proc; /* may be NULL */
fc17ad60 455 globaldata_t gd;
344ad853
MD
456 int sig;
457 int catch;
344ad853 458 int error;
e43a034f 459 int oldpri;
076fecef 460 struct callout thandle;
984263bc 461
0cfcada1
MD
462 /*
463 * NOTE: removed KTRPOINT, it could cause races due to blocking
464 * even in stable. Just scrap it for now.
465 */
5ea440eb 466 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) {
984263bc 467 /*
dbcd0c9b
MD
468 * After a panic, or before we actually have an operational
469 * softclock, just give interrupts a chance, then just return;
470 *
984263bc
MD
471 * don't run any other procs or panic below,
472 * in case this is the idle process and already asleep.
473 */
e43a034f 474 splz();
f9235b6d 475 oldpri = td->td_pri;
e43a034f
MD
476 lwkt_setpri_self(safepri);
477 lwkt_switch();
478 lwkt_setpri_self(oldpri);
984263bc
MD
479 return (0);
480 }
8aa3430c 481 logtsleep2(tsleep_beg, ident);
fc17ad60
MD
482 gd = td->td_gd;
483 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
344ad853
MD
484
485 /*
486 * NOTE: all of this occurs on the current cpu, including any
487 * callout-based wakeups, so a critical section is a sufficient
488 * interlock.
489 *
490 * The entire sequence through to where we actually sleep must
491 * run without breaking the critical section.
492 */
344ad853
MD
493 catch = flags & PCATCH;
494 error = 0;
495 sig = 0;
496
37af14fe 497 crit_enter_quick(td);
344ad853 498
0cfcada1 499 KASSERT(ident != NULL, ("tsleep: no ident"));
7278a846
SS
500 KASSERT(lp == NULL ||
501 lp->lwp_stat == LSRUN || /* Obvious */
502 lp->lwp_stat == LSSTOP, /* Set in tstop */
503 ("tsleep %p %s %d",
504 ident, wmesg, lp->lwp_stat));
0cfcada1 505
344ad853 506 /*
5686ec5a
MD
507 * We interlock the sleep queue if the caller has not already done
508 * it for us. This must be done before we potentially acquire any
509 * tokens or we can loose the wakeup.
510 */
511 if ((flags & PINTERLOCKED) == 0) {
5686ec5a
MD
512 _tsleep_interlock(gd, ident, flags);
513 }
514
515 /*
4643740a
MD
516 * Setup for the current process (if this is a process). We must
517 * interlock with lwp_token to avoid remote wakeup races via
518 * setrunnable()
344ad853 519 */
08f2f1bb 520 if (lp) {
4643740a 521 lwkt_gettoken(&lp->lwp_token);
344ad853
MD
522 if (catch) {
523 /*
524 * Early termination if PCATCH was set and a
525 * signal is pending, interlocked with the
526 * critical section.
527 *
528 * Early termination only occurs when tsleep() is
164b8401 529 * entered while in a normal LSRUN state.
344ad853 530 */
08f2f1bb 531 if ((sig = CURSIG(lp)) != 0)
344ad853
MD
532 goto resume;
533
534 /*
5686ec5a
MD
535 * Causes ksignal to wake us up if a signal is
536 * received (interlocked with p->p_token).
344ad853 537 */
4643740a 538 lp->lwp_flags |= LWP_SINTR;
344ad853 539 }
5686ec5a
MD
540 } else {
541 KKASSERT(p == NULL);
4ecd8190 542 }
344ad853 543
4ecd8190 544 /*
4ecd8190
MD
545 * Make sure the current process has been untangled from
546 * the userland scheduler and initialize slptime to start
5686ec5a 547 * counting.
4ecd8190
MD
548 */
549 if (lp) {
08f2f1bb
SS
550 p->p_usched->release_curproc(lp);
551 lp->lwp_slptime = 0;
0a3f9b47 552 }
fc17ad60
MD
553
554 /*
d9345d3a
MD
555 * If the interlocked flag is set but our cpu bit in the slpqueue
556 * is no longer set, then a wakeup was processed inbetween the
4ecd8190
MD
557 * tsleep_interlock() (ours or the callers), and here. This can
558 * occur under numerous circumstances including when we release the
559 * current process.
d9345d3a 560 *
4ecd8190
MD
561 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s)
562 * to process incoming IPIs, thus draining incoming wakeups.
d9345d3a 563 */
4ecd8190
MD
564 if ((td->td_flags & TDF_TSLEEPQ) == 0) {
565 logtsleep2(ilockfail, ident);
566 goto resume;
d9345d3a 567 }
4ecd8190
MD
568
569 /*
570 * scheduling is blocked while in a critical section. Coincide
571 * the descheduled-by-tsleep flag with the descheduling of the
572 * lwkt.
8d446850
MD
573 *
574 * The timer callout is localized on our cpu and interlocked by
575 * our critical section.
4ecd8190 576 */
37af14fe 577 lwkt_deschedule_self(td);
ae8e83e6 578 td->td_flags |= TDF_TSLEEP_DESCHEDULED;
344ad853 579 td->td_wmesg = wmesg;
344ad853
MD
580
581 /*
8d446850
MD
582 * Setup the timeout, if any. The timeout is only operable while
583 * the thread is flagged descheduled.
344ad853 584 */
8d446850 585 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0);
076fecef 586 if (timo) {
8d446850 587 callout_init_mp(&thandle);
076fecef
MD
588 callout_reset(&thandle, timo, endtsleep, td);
589 }
344ad853 590
984263bc 591 /*
344ad853 592 * Beddy bye bye.
984263bc 593 */
08f2f1bb 594 if (lp) {
26a0694b 595 /*
52eedfb5 596 * Ok, we are sleeping. Place us in the SSLEEP state.
26a0694b 597 */
4643740a 598 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
7278a846
SS
599 /*
600 * tstop() sets LSSTOP, so don't fiddle with that.
601 */
602 if (lp->lwp_stat != LSSTOP)
603 lp->lwp_stat = LSSLEEP;
08f2f1bb 604 lp->lwp_ru.ru_nvcsw++;
344ad853 605 lwkt_switch();
ab554892
MD
606
607 /*
164b8401 608 * And when we are woken up, put us back in LSRUN. If we
ab554892
MD
609 * slept for over a second, recalculate our estcpu.
610 */
164b8401 611 lp->lwp_stat = LSRUN;
08f2f1bb
SS
612 if (lp->lwp_slptime)
613 p->p_usched->recalculate(lp);
614 lp->lwp_slptime = 0;
0cfcada1
MD
615 } else {
616 lwkt_switch();
617 }
344ad853 618
fc17ad60
MD
619 /*
620 * Make sure we haven't switched cpus while we were asleep. It's
344ad853 621 * not supposed to happen. Cleanup our temporary flags.
fc17ad60
MD
622 */
623 KKASSERT(gd == td->td_gd);
344ad853
MD
624
625 /*
8d446850 626 * Cleanup the timeout. If the timeout has already occured thandle
4643740a
MD
627 * has already been stopped, otherwise stop thandle. If the timeout
628 * is running (the callout thread must be blocked trying to get
629 * lwp_token) then wait for us to get scheduled.
344ad853
MD
630 */
631 if (timo) {
4643740a
MD
632 while (td->td_flags & TDF_TIMEOUT_RUNNING) {
633 lwkt_deschedule_self(td);
634 td->td_wmesg = "tsrace";
635 lwkt_switch();
636 kprintf("td %p %s: timeout race\n", td, td->td_comm);
637 }
344ad853
MD
638 if (td->td_flags & TDF_TIMEOUT) {
639 td->td_flags &= ~TDF_TIMEOUT;
a40da8f0 640 error = EWOULDBLOCK;
344ad853 641 } else {
8d446850 642 /* does not block when on same cpu */
344ad853
MD
643 callout_stop(&thandle);
644 }
0cfcada1 645 }
4643740a 646 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED;
344ad853
MD
647
648 /*
8d446850
MD
649 * Make sure we have been removed from the sleepq. In most
650 * cases this will have been done for us already but it is
651 * possible for a scheduling IPI to be in-flight from a
652 * previous tsleep/tsleep_interlock() or due to a straight-out
653 * call to lwkt_schedule() (in the case of an interrupt thread),
654 * causing a spurious wakeup.
344ad853 655 */
ae8e83e6 656 _tsleep_remove(td);
344ad853 657 td->td_wmesg = NULL;
344ad853
MD
658
659 /*
7c1212ec
MD
660 * Figure out the correct error return. If interrupted by a
661 * signal we want to return EINTR or ERESTART.
344ad853
MD
662 */
663resume:
4643740a 664 if (lp) {
7c1212ec 665 if (catch && error == 0) {
94f98873 666 if (sig != 0 || (sig = CURSIG(lp))) {
7c1212ec
MD
667 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
668 error = EINTR;
669 else
670 error = ERESTART;
671 }
984263bc 672 }
4643740a
MD
673 lp->lwp_flags &= ~LWP_SINTR;
674 lwkt_reltoken(&lp->lwp_token);
984263bc 675 }
8aa3430c 676 logtsleep1(tsleep_end);
344ad853
MD
677 crit_exit_quick(td);
678 return (error);
984263bc
MD
679}
680
984263bc 681/*
bf765287 682 * Interlocked spinlock sleep. An exclusively held spinlock must
e590ee86 683 * be passed to ssleep(). The function will atomically release the
bf765287
MD
684 * spinlock and tsleep on the ident, then reacquire the spinlock and
685 * return.
686 *
687 * This routine is fairly important along the critical path, so optimize it
688 * heavily.
689 */
690int
5decebc7 691ssleep(const volatile void *ident, struct spinlock *spin, int flags,
bf765287
MD
692 const char *wmesg, int timo)
693{
694 globaldata_t gd = mycpu;
695 int error;
16523a43 696
ae8e83e6 697 _tsleep_interlock(gd, ident, flags);
7cfe2b28 698 spin_unlock_quick(gd, spin);
ef48be0d 699 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
7cfe2b28 700 spin_lock_quick(gd, spin);
bf765287
MD
701
702 return (error);
16523a43
MD
703}
704
bed060de 705int
5decebc7
MD
706lksleep(const volatile void *ident, struct lock *lock, int flags,
707 const char *wmesg, int timo)
bed060de
AH
708{
709 globaldata_t gd = mycpu;
710 int error;
711
712 _tsleep_interlock(gd, ident, flags);
713 lockmgr(lock, LK_RELEASE);
714 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
715 lockmgr(lock, LK_EXCLUSIVE);
716
717 return (error);
718}
719
16523a43 720/*
7f6220a9
MD
721 * Interlocked mutex sleep. An exclusively held mutex must be passed
722 * to mtxsleep(). The function will atomically release the mutex
723 * and tsleep on the ident, then reacquire the mutex and return.
724 */
725int
5decebc7 726mtxsleep(const volatile void *ident, struct mtx *mtx, int flags,
7f6220a9
MD
727 const char *wmesg, int timo)
728{
729 globaldata_t gd = mycpu;
730 int error;
731
732 _tsleep_interlock(gd, ident, flags);
733 mtx_unlock(mtx);
734 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
735 mtx_lock_ex_quick(mtx, wmesg);
736
737 return (error);
738}
739
740/*
362e59be 741 * Interlocked serializer sleep. An exclusively held serializer must
ed3f6624 742 * be passed to zsleep(). The function will atomically release
362e59be
SZ
743 * the serializer and tsleep on the ident, then reacquire the serializer
744 * and return.
745 */
746int
5decebc7 747zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags,
ed3f6624 748 const char *wmesg, int timo)
362e59be 749{
ae8e83e6 750 globaldata_t gd = mycpu;
362e59be
SZ
751 int ret;
752
753 ASSERT_SERIALIZED(slz);
754
ae8e83e6 755 _tsleep_interlock(gd, ident, flags);
362e59be 756 lwkt_serialize_exit(slz);
ef48be0d 757 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo);
362e59be 758 lwkt_serialize_enter(slz);
362e59be
SZ
759
760 return ret;
761}
762
763/*
a22c590e
MD
764 * Directly block on the LWKT thread by descheduling it. This
765 * is much faster then tsleep(), but the only legal way to wake
766 * us up is to directly schedule the thread.
767 *
768 * Setting TDF_SINTR will cause new signals to directly schedule us.
769 *
ae8e83e6 770 * This routine must be called while in a critical section.
a22c590e
MD
771 */
772int
773lwkt_sleep(const char *wmesg, int flags)
774{
775 thread_t td = curthread;
776 int sig;
777
778 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
779 td->td_flags |= TDF_BLOCKED;
780 td->td_wmesg = wmesg;
781 lwkt_deschedule_self(td);
782 lwkt_switch();
783 td->td_wmesg = NULL;
784 td->td_flags &= ~TDF_BLOCKED;
785 return(0);
786 }
787 if ((sig = CURSIG(td->td_lwp)) != 0) {
788 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
789 return(EINTR);
790 else
791 return(ERESTART);
792
793 }
794 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
795 td->td_wmesg = wmesg;
796 lwkt_deschedule_self(td);
797 lwkt_switch();
798 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
799 td->td_wmesg = NULL;
800 return(0);
801}
802
803/*
344ad853 804 * Implement the timeout for tsleep.
fc17ad60 805 *
344ad853
MD
806 * This type of callout timeout is scheduled on the same cpu the process
807 * is sleeping on. Also, at the moment, the MP lock is held.
984263bc
MD
808 */
809static void
0cfcada1 810endtsleep(void *arg)
984263bc 811{
0cfcada1 812 thread_t td = arg;
9a379a4a 813 struct lwp *lp;
984263bc 814
8d446850 815 /*
4643740a
MD
816 * We are going to have to get the lwp_token, which means we might
817 * block. This can race a tsleep getting woken up by other means
818 * so set TDF_TIMEOUT_RUNNING to force the tsleep to wait for our
819 * processing to complete (sorry tsleep!).
820 *
821 * We can safely set td_flags because td MUST be on the same cpu
822 * as we are.
8d446850 823 */
4643740a
MD
824 KKASSERT(td->td_gd == mycpu);
825 crit_enter();
826 td->td_flags |= TDF_TIMEOUT_RUNNING | TDF_TIMEOUT;
344ad853
MD
827
828 /*
4643740a
MD
829 * This can block but TDF_TIMEOUT_RUNNING will prevent the thread
830 * from exiting the tsleep on us. The flag is interlocked by virtue
831 * of lp being on the same cpu as we are.
344ad853 832 */
8d446850 833 if ((lp = td->td_lwp) != NULL)
e2b148c6 834 lwkt_gettoken(&lp->lwp_token);
344ad853 835
4643740a
MD
836 KKASSERT(td->td_flags & TDF_TSLEEP_DESCHEDULED);
837
838 if (lp) {
839 if (lp->lwp_proc->p_stat != SSTOP)
840 setrunnable(lp);
e2b148c6 841 lwkt_reltoken(&lp->lwp_token);
4643740a
MD
842 } else {
843 _tsleep_remove(td);
844 lwkt_schedule(td);
845 }
846 KKASSERT(td->td_gd == mycpu);
847 td->td_flags &= ~TDF_TIMEOUT_RUNNING;
37af14fe 848 crit_exit();
984263bc
MD
849}
850
984263bc 851/*
8fb8bca6 852 * Make all processes sleeping on the specified identifier runnable.
fc17ad60
MD
853 * count may be zero or one only.
854 *
855 * The domain encodes the sleep/wakeup domain AND the first cpu to check
856 * (which is always the current cpu). As we iterate across cpus
344ad853
MD
857 *
858 * This call may run without the MP lock held. We can only manipulate thread
859 * state on the cpu owning the thread. We CANNOT manipulate process state
860 * at all.
5decebc7
MD
861 *
862 * _wakeup() can be passed to an IPI so we can't use (const volatile
863 * void *ident).
8fb8bca6
EN
864 */
865static void
fc17ad60 866_wakeup(void *ident, int domain)
8fb8bca6 867{
fc17ad60 868 struct tslpque *qp;
0cfcada1
MD
869 struct thread *td;
870 struct thread *ntd;
fc17ad60 871 globaldata_t gd;
fc17ad60
MD
872#ifdef SMP
873 cpumask_t mask;
fc17ad60
MD
874#endif
875 int id;
984263bc 876
37af14fe 877 crit_enter();
8aa3430c 878 logtsleep2(wakeup_beg, ident);
fc17ad60
MD
879 gd = mycpu;
880 id = LOOKUP(ident);
881 qp = &gd->gd_tsleep_hash[id];
984263bc 882restart:
0cfcada1 883 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
ae8e83e6 884 ntd = TAILQ_NEXT(td, td_sleepq);
fc17ad60
MD
885 if (td->td_wchan == ident &&
886 td->td_wdomain == (domain & PDOMAIN_MASK)
887 ) {
ae8e83e6
MD
888 KKASSERT(td->td_gd == gd);
889 _tsleep_remove(td);
890 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) {
ae8e83e6
MD
891 lwkt_schedule(td);
892 if (domain & PWAKEUP_ONE)
893 goto done;
fc17ad60 894 }
0cfcada1 895 goto restart;
984263bc
MD
896 }
897 }
fc17ad60
MD
898
899#ifdef SMP
900 /*
901 * We finished checking the current cpu but there still may be
902 * more work to do. Either wakeup_one was requested and no matching
903 * thread was found, or a normal wakeup was requested and we have
904 * to continue checking cpus.
905 *
fc17ad60
MD
906 * It should be noted that this scheme is actually less expensive then
907 * the old scheme when waking up multiple threads, since we send
908 * only one IPI message per target candidate which may then schedule
909 * multiple threads. Before we could have wound up sending an IPI
910 * message for each thread on the target cpu (!= current cpu) that
911 * needed to be woken up.
912 *
913 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
914 * should be ok since we are passing idents in the IPI rather then
915 * thread pointers.
916 */
1f4f6e0b
MD
917 if ((domain & PWAKEUP_MYCPU) == 0 &&
918 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
919 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
920 domain | PWAKEUP_MYCPU);
fc17ad60
MD
921 }
922#endif
923done:
8aa3430c 924 logtsleep1(wakeup_end);
37af14fe 925 crit_exit();
984263bc
MD
926}
927
b336a9b1
MD
928/*
929 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
930 */
984263bc 931void
5decebc7 932wakeup(const volatile void *ident)
984263bc 933{
5decebc7 934 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
0cfcada1 935}
984263bc 936
b336a9b1
MD
937/*
938 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
939 */
0cfcada1 940void
5decebc7 941wakeup_one(const volatile void *ident)
0cfcada1 942{
fc17ad60 943 /* XXX potentially round-robin the first responding cpu */
5decebc7 944 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
da5fb9ef
MD
945}
946
b336a9b1
MD
947/*
948 * Wakeup threads tsleep()ing on the specified ident on the current cpu
949 * only.
950 */
951void
5decebc7 952wakeup_mycpu(const volatile void *ident)
b336a9b1 953{
5decebc7 954 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
b336a9b1
MD
955}
956
957/*
958 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
959 * only.
960 */
961void
5decebc7 962wakeup_mycpu_one(const volatile void *ident)
b336a9b1
MD
963{
964 /* XXX potentially round-robin the first responding cpu */
5decebc7 965 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE);
b336a9b1
MD
966}
967
968/*
969 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
970 * only.
971 */
972void
5decebc7 973wakeup_oncpu(globaldata_t gd, const volatile void *ident)
b336a9b1 974{
1699d292 975#ifdef SMP
b336a9b1 976 if (gd == mycpu) {
5decebc7 977 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
b336a9b1 978 } else {
5decebc7 979 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU);
b336a9b1 980 }
1699d292 981#else
8e44b950 982 _wakeup(__DEALL(ident), PWAKEUP_MYCPU);
1699d292 983#endif
b336a9b1
MD
984}
985
986/*
987 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
988 * only.
989 */
990void
5decebc7 991wakeup_oncpu_one(globaldata_t gd, const volatile void *ident)
b336a9b1 992{
1699d292 993#ifdef SMP
b336a9b1 994 if (gd == mycpu) {
5decebc7 995 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
b336a9b1 996 } else {
5decebc7
MD
997 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident),
998 PWAKEUP_MYCPU | PWAKEUP_ONE);
b336a9b1 999 }
1699d292 1000#else
8e44b950 1001 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE);
1699d292 1002#endif
b336a9b1
MD
1003}
1004
1005/*
1006 * Wakeup all threads waiting on the specified ident that slept using
1007 * the specified domain, on all cpus.
1008 */
da5fb9ef 1009void
5decebc7 1010wakeup_domain(const volatile void *ident, int domain)
da5fb9ef 1011{
5decebc7 1012 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
da5fb9ef
MD
1013}
1014
b336a9b1
MD
1015/*
1016 * Wakeup one thread waiting on the specified ident that slept using
1017 * the specified domain, on any cpu.
1018 */
da5fb9ef 1019void
5decebc7 1020wakeup_domain_one(const volatile void *ident, int domain)
da5fb9ef 1021{
fc17ad60 1022 /* XXX potentially round-robin the first responding cpu */
5decebc7
MD
1023 _wakeup(__DEALL(ident),
1024 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
984263bc
MD
1025}
1026
1027/*
344ad853
MD
1028 * setrunnable()
1029 *
4643740a
MD
1030 * Make a process runnable. lp->lwp_token must be held on call and this
1031 * function must be called from the cpu owning lp.
37af14fe 1032 *
4643740a 1033 * This only has an effect if we are in LSSTOP or LSSLEEP.
984263bc
MD
1034 */
1035void
9a379a4a 1036setrunnable(struct lwp *lp)
984263bc 1037{
4643740a
MD
1038 thread_t td = lp->lwp_thread;
1039
e2b148c6 1040 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_token);
4643740a 1041 KKASSERT(td->td_gd == mycpu);
344ad853 1042 crit_enter();
2daf83b0
SS
1043 if (lp->lwp_stat == LSSTOP)
1044 lp->lwp_stat = LSSLEEP;
4643740a
MD
1045 if (lp->lwp_stat == LSSLEEP) {
1046 _tsleep_remove(td);
1047 lwkt_schedule(td);
1048 } else if (td->td_flags & TDF_SINTR) {
1049 lwkt_schedule(td);
1050 }
344ad853 1051 crit_exit();
984263bc
MD
1052}
1053
1054/*
164b8401
SS
1055 * The process is stopped due to some condition, usually because p_stat is
1056 * set to SSTOP, but also possibly due to being traced.
fc17ad60 1057 *
4643740a
MD
1058 * Caller must hold p->p_token
1059 *
164b8401 1060 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
344ad853
MD
1061 * because the parent may check the child's status before the child actually
1062 * gets to this routine.
1063 *
9a379a4a 1064 * This routine is called with the current lwp only, typically just
4643740a
MD
1065 * before returning to userland if the process state is detected as
1066 * possibly being in a stopped state.
984263bc
MD
1067 */
1068void
9a379a4a 1069tstop(void)
984263bc 1070{
9a379a4a 1071 struct lwp *lp = curthread->td_lwp;
7278a846 1072 struct proc *p = lp->lwp_proc;
8c986a82 1073 struct proc *q;
9a379a4a 1074
4643740a 1075 lwkt_gettoken(&lp->lwp_token);
7278a846 1076 crit_enter();
4643740a 1077
f33e8653 1078 /*
4643740a 1079 * If LWP_MP_WSTOP is set, we were sleeping
f33e8653
SS
1080 * while our process was stopped. At this point
1081 * we were already counted as stopped.
1082 */
4643740a 1083 if ((lp->lwp_mpflags & LWP_MP_WSTOP) == 0) {
f33e8653
SS
1084 /*
1085 * If we're the last thread to stop, signal
1086 * our parent.
1087 */
1088 p->p_nstopped++;
4643740a 1089 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
ea59a697 1090 wakeup(&p->p_nstopped);
f33e8653 1091 if (p->p_nstopped == p->p_nthreads) {
8c986a82
MD
1092 /*
1093 * Token required to interlock kern_wait()
1094 */
1095 q = p->p_pptr;
1096 PHOLD(q);
1097 lwkt_gettoken(&q->p_token);
4643740a 1098 p->p_flags &= ~P_WAITED;
f33e8653 1099 wakeup(p->p_pptr);
8c986a82
MD
1100 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
1101 ksignal(q, SIGCHLD);
1102 lwkt_reltoken(&q->p_token);
1103 PRELE(q);
f33e8653
SS
1104 }
1105 }
ea59a697 1106 while (p->p_stat == SSTOP) {
ea59a697
SS
1107 lp->lwp_stat = LSSTOP;
1108 tsleep(p, 0, "stop", 0);
1109 }
7278a846 1110 p->p_nstopped--;
4643740a 1111 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP);
7278a846 1112 crit_exit();
4643740a 1113 lwkt_reltoken(&lp->lwp_token);
26a0694b
MD
1114}
1115
1116/*
984263bc
MD
1117 * Compute a tenex style load average of a quantity on
1118 * 1, 5 and 15 minute intervals.
1119 */
c7e98b2f 1120static int loadav_count_runnable(struct lwp *p, void *data);
8fa76237 1121
984263bc
MD
1122static void
1123loadav(void *arg)
1124{
984263bc 1125 struct loadavg *avg;
8fa76237 1126 int i, nrun;
984263bc 1127
984263bc 1128 nrun = 0;
c7e98b2f 1129 alllwp_scan(loadav_count_runnable, &nrun);
8fa76237
MD
1130 avg = &averunnable;
1131 for (i = 0; i < 3; i++) {
984263bc
MD
1132 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1133 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
8fa76237 1134 }
984263bc
MD
1135
1136 /*
1137 * Schedule the next update to occur after 5 seconds, but add a
1138 * random variation to avoid synchronisation with processes that
1139 * run at regular intervals.
1140 */
cddfb7bb 1141 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
8fa76237
MD
1142 loadav, NULL);
1143}
1144
1145static int
c7e98b2f 1146loadav_count_runnable(struct lwp *lp, void *data)
8fa76237
MD
1147{
1148 int *nrunp = data;
1149 thread_t td;
1150
164b8401
SS
1151 switch (lp->lwp_stat) {
1152 case LSRUN:
08f2f1bb 1153 if ((td = lp->lwp_thread) == NULL)
8fa76237
MD
1154 break;
1155 if (td->td_flags & TDF_BLOCKED)
1156 break;
8fa76237
MD
1157 ++*nrunp;
1158 break;
1159 default:
1160 break;
1161 }
d2d8515b 1162 lwkt_yield();
8fa76237 1163 return(0);
984263bc
MD
1164}
1165
1166/* ARGSUSED */
1167static void
6656cd91 1168sched_setup(void *dummy)
984263bc 1169{
8d446850
MD
1170 callout_init_mp(&loadav_callout);
1171 callout_init_mp(&schedcpu_callout);
984263bc
MD
1172
1173 /* Kick off timeout driven events by calling first time. */
984263bc
MD
1174 schedcpu(NULL);
1175 loadav(NULL);
1176}
1177