If a process forks while being scanned, a non-zero p_lock will be inherited
[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 $
9afb0ffd 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.57 2005/12/10 18:27:24 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>
26a0694b 53#include <sys/thread2.h>
344ad853 54#include <sys/lock.h>
984263bc
MD
55#ifdef KTRACE
56#include <sys/uio.h>
57#include <sys/ktrace.h>
58#endif
f1d1c3fa 59#include <sys/xwait.h>
9afb0ffd 60#include <sys/ktr.h>
984263bc
MD
61
62#include <machine/cpu.h>
63#include <machine/ipl.h>
64#include <machine/smp.h>
65
fc17ad60
MD
66TAILQ_HEAD(tslpque, thread);
67
402ed7e1 68static void sched_setup (void *dummy);
984263bc
MD
69SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
70
984263bc
MD
71int hogticks;
72int lbolt;
344ad853 73int lbolt_syncer;
984263bc 74int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
17a9f566 75int ncpus;
90100055 76int ncpus2, ncpus2_shift, ncpus2_mask;
e43a034f 77int safepri;
984263bc
MD
78
79static struct callout loadav_callout;
35f9d051 80static struct callout schedcpu_callout;
fc17ad60 81MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
984263bc 82
9afb0ffd
MD
83#if !defined(KTR_TSLEEP)
84#define KTR_TSLEEP KTR_ALL
85#endif
86KTR_INFO_MASTER(tsleep);
87KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter", 0);
88KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0);
89KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0);
90KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0);
91#define logtsleep(name) KTR_LOG(tsleep_ ## name)
92
984263bc
MD
93struct loadavg averunnable =
94 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
95/*
96 * Constants for averages over 1, 5, and 15 minutes
97 * when sampling at 5 second intervals.
98 */
99static fixpt_t cexp[3] = {
100 0.9200444146293232 * FSCALE, /* exp(-1/12) */
101 0.9834714538216174 * FSCALE, /* exp(-1/60) */
102 0.9944598480048967 * FSCALE, /* exp(-1/180) */
103};
104
402ed7e1 105static void endtsleep (void *);
344ad853 106static void unsleep_and_wakeup_thread(struct thread *td);
402ed7e1 107static void loadav (void *arg);
402ed7e1 108static void schedcpu (void *arg);
984263bc 109
a46fac56
MD
110/*
111 * Adjust the scheduler quantum. The quantum is specified in microseconds.
112 * Note that 'tick' is in microseconds per tick.
113 */
984263bc
MD
114static int
115sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
116{
117 int error, new_val;
118
119 new_val = sched_quantum * tick;
120 error = sysctl_handle_int(oidp, &new_val, 0, req);
121 if (error != 0 || req->newptr == NULL)
122 return (error);
123 if (new_val < tick)
124 return (EINVAL);
125 sched_quantum = new_val / tick;
126 hogticks = 2 * sched_quantum;
127 return (0);
128}
129
130SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
131 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
132
984263bc
MD
133/*
134 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
135 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
136 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
137 *
138 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
dcc99b62 139 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
984263bc
MD
140 *
141 * If you don't want to bother with the faster/more-accurate formula, you
142 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
143 * (more general) method of calculating the %age of CPU used by a process.
dcc99b62
MD
144 *
145 * decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
146 */
147#define CCPU_SHIFT 11
148
149static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
150SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
151
152/*
153 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
984263bc 154 */
dcc99b62
MD
155static int fscale __unused = FSCALE;
156SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
984263bc
MD
157
158/*
0a3f9b47 159 * Recompute process priorities, once a second.
dcc99b62
MD
160 *
161 * Since the userland schedulers are typically event oriented, if the
162 * estcpu calculation at wakeup() time is not sufficient to make a
163 * process runnable relative to other processes in the system we have
164 * a 1-second recalc to help out.
165 *
166 * This code also allows us to store sysclock_t data in the process structure
167 * without fear of an overrun, since sysclock_t are guarenteed to hold
168 * several seconds worth of count.
984263bc
MD
169 */
170/* ARGSUSED */
171static void
26a0694b 172schedcpu(void *arg)
984263bc 173{
344ad853 174 struct rlimit *rlim;
4b5f931b 175 struct proc *p;
344ad853 176 u_int64_t ttime;
4b5f931b 177
344ad853
MD
178 /*
179 * General process statistics once a second
180 */
f62004ad 181 FOREACH_PROC_IN_SYSTEM(p) {
dcc99b62 182 crit_enter();
984263bc 183 p->p_swtime++;
344ad853 184 if (p->p_stat == SSLEEP)
984263bc 185 p->p_slptime++;
a46fac56
MD
186
187 /*
dcc99b62
MD
188 * Only recalculate processes that are active or have slept
189 * less then 2 seconds. The schedulers understand this.
a46fac56 190 */
dcc99b62 191 if (p->p_slptime <= 1) {
553ea3c8 192 p->p_usched->recalculate(&p->p_lwp);
dcc99b62
MD
193 } else {
194 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
a46fac56 195 }
e43a034f 196 crit_exit();
984263bc 197 }
344ad853
MD
198
199 /*
200 * Resource checks. XXX break out since psignal/killproc can block,
201 * limiting us to one process killed per second. There is probably
202 * a better way.
203 */
204 FOREACH_PROC_IN_SYSTEM(p) {
205 crit_enter();
f026f3b0
MD
206 if (p->p_stat == SIDL ||
207 (p->p_flag & P_ZOMBIE) ||
344ad853
MD
208 p->p_limit == NULL ||
209 p->p_thread == NULL
210 ) {
211 crit_exit();
212 continue;
213 }
214 ttime = p->p_thread->td_sticks + p->p_thread->td_uticks;
215 if (p->p_limit->p_cpulimit != RLIM_INFINITY &&
216 ttime > p->p_limit->p_cpulimit
217 ) {
218 rlim = &p->p_rlimit[RLIMIT_CPU];
219 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
220 killproc(p, "exceeded maximum CPU limit");
221 } else {
222 psignal(p, SIGXCPU);
223 if (rlim->rlim_cur < rlim->rlim_max) {
224 /* XXX: we should make a private copy */
225 rlim->rlim_cur += 5;
226 }
227 }
228 crit_exit();
229 break;
230 }
231 crit_exit();
232 }
233
984263bc 234 wakeup((caddr_t)&lbolt);
344ad853 235 wakeup((caddr_t)&lbolt_syncer);
35f9d051 236 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
984263bc
MD
237}
238
239/*
dcc99b62
MD
240 * This is only used by ps. Generate a cpu percentage use over
241 * a period of one second.
984263bc 242 */
dcc99b62 243void
553ea3c8 244updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
984263bc 245{
dcc99b62
MD
246 fixpt_t acc;
247 int remticks;
248
249 acc = (cpticks << FSHIFT) / ttlticks;
250 if (ttlticks >= ESTCPUFREQ) {
553ea3c8 251 lp->lwp_pctcpu = acc;
dcc99b62
MD
252 } else {
253 remticks = ESTCPUFREQ - ttlticks;
553ea3c8 254 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
dcc99b62 255 ESTCPUFREQ;
a46fac56 256 }
984263bc
MD
257}
258
259/*
260 * We're only looking at 7 bits of the address; everything is
261 * aligned to 4, lots of things are aligned to greater powers
262 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
263 */
264#define TABLESIZE 128
984263bc
MD
265#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
266
fc17ad60
MD
267static cpumask_t slpque_cpumasks[TABLESIZE];
268
984263bc 269/*
a46fac56 270 * General scheduler initialization. We force a reschedule 25 times
fc17ad60
MD
271 * a second by default. Note that cpu0 is initialized in early boot and
272 * cannot make any high level calls.
273 *
274 * Each cpu has its own sleep queue.
984263bc 275 */
984263bc 276void
fc17ad60 277sleep_gdinit(globaldata_t gd)
984263bc 278{
fc17ad60 279 static struct tslpque slpque_cpu0[TABLESIZE];
9c1fad94 280 int i;
984263bc 281
fc17ad60
MD
282 if (gd->gd_cpuid == 0) {
283 sched_quantum = (hz + 24) / 25;
284 hogticks = 2 * sched_quantum;
285
286 gd->gd_tsleep_hash = slpque_cpu0;
287 } else {
288 gd->gd_tsleep_hash = malloc(sizeof(slpque_cpu0),
289 M_TSLEEP, M_WAITOK | M_ZERO);
290 }
291 for (i = 0; i < TABLESIZE; ++i)
292 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
984263bc
MD
293}
294
295/*
296 * General sleep call. Suspends the current process until a wakeup is
297 * performed on the specified identifier. The process will then be made
298 * runnable with the specified priority. Sleeps at most timo/hz seconds
377d4740 299 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
984263bc
MD
300 * before and after sleeping, else signals are not checked. Returns 0 if
301 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
302 * signal needs to be delivered, ERESTART is returned if the current system
303 * call should be restarted if possible, and EINTR is returned if the system
304 * call should be interrupted by the signal (return EINTR).
26a0694b 305 *
0a3f9b47
MD
306 * Note that if we are a process, we release_curproc() before messing with
307 * the LWKT scheduler.
a46fac56
MD
308 *
309 * During autoconfiguration or after a panic, a sleep will simply
310 * lower the priority briefly to allow interrupts, then return.
984263bc
MD
311 */
312int
6656cd91 313tsleep(void *ident, int flags, const char *wmesg, int timo)
984263bc 314{
dadab5e9 315 struct thread *td = curthread;
0cfcada1 316 struct proc *p = td->td_proc; /* may be NULL */
fc17ad60 317 globaldata_t gd;
344ad853
MD
318 int sig;
319 int catch;
320 int id;
321 int error;
e43a034f 322 int oldpri;
076fecef 323 struct callout thandle;
984263bc 324
0cfcada1
MD
325 /*
326 * NOTE: removed KTRPOINT, it could cause races due to blocking
327 * even in stable. Just scrap it for now.
328 */
984263bc
MD
329 if (cold || panicstr) {
330 /*
331 * After a panic, or during autoconfiguration,
332 * just give interrupts a chance, then just return;
333 * don't run any other procs or panic below,
334 * in case this is the idle process and already asleep.
335 */
e43a034f
MD
336 splz();
337 oldpri = td->td_pri & TDPRI_MASK;
338 lwkt_setpri_self(safepri);
339 lwkt_switch();
340 lwkt_setpri_self(oldpri);
984263bc
MD
341 return (0);
342 }
9afb0ffd 343 logtsleep(tsleep_beg);
fc17ad60
MD
344 gd = td->td_gd;
345 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
344ad853
MD
346
347 /*
348 * NOTE: all of this occurs on the current cpu, including any
349 * callout-based wakeups, so a critical section is a sufficient
350 * interlock.
351 *
352 * The entire sequence through to where we actually sleep must
353 * run without breaking the critical section.
354 */
355 id = LOOKUP(ident);
356 catch = flags & PCATCH;
357 error = 0;
358 sig = 0;
359
37af14fe 360 crit_enter_quick(td);
344ad853 361
0cfcada1
MD
362 KASSERT(ident != NULL, ("tsleep: no ident"));
363 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
364 ident, wmesg, p->p_stat));
365
344ad853
MD
366 /*
367 * Setup for the current process (if this is a process).
368 */
0a3f9b47 369 if (p) {
344ad853
MD
370 if (catch) {
371 /*
372 * Early termination if PCATCH was set and a
373 * signal is pending, interlocked with the
374 * critical section.
375 *
376 * Early termination only occurs when tsleep() is
377 * entered while in a normal SRUN state.
378 */
379 if ((sig = CURSIG(p)) != 0)
380 goto resume;
381
382 /*
383 * Causes psignal to wake us up when.
384 */
385 p->p_flag |= P_SINTR;
386 }
387
388 /*
389 * Make sure the current process has been untangled from
390 * the userland scheduler and initialize slptime to start
391 * counting.
392 */
0a3f9b47
MD
393 if (flags & PNORESCHED)
394 td->td_flags |= TDF_NORESCHED;
553ea3c8 395 p->p_usched->release_curproc(&p->p_lwp);
0cfcada1 396 p->p_slptime = 0;
0a3f9b47 397 }
fc17ad60
MD
398
399 /*
344ad853 400 * Move our thread to the correct queue and setup our wchan, etc.
fc17ad60 401 */
37af14fe 402 lwkt_deschedule_self(td);
344ad853 403 td->td_flags |= TDF_TSLEEPQ;
fc17ad60
MD
404 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
405 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
344ad853
MD
406
407 td->td_wchan = ident;
408 td->td_wmesg = wmesg;
409 td->td_wdomain = flags & PDOMAIN_MASK;
410
411 /*
412 * Setup the timeout, if any
413 */
076fecef
MD
414 if (timo) {
415 callout_init(&thandle);
416 callout_reset(&thandle, timo, endtsleep, td);
417 }
344ad853 418
984263bc 419 /*
344ad853 420 * Beddy bye bye.
984263bc 421 */
0cfcada1 422 if (p) {
26a0694b 423 /*
344ad853
MD
424 * Ok, we are sleeping. Remove us from the userland runq
425 * and place us in the SSLEEP state.
26a0694b 426 */
344ad853
MD
427 if (p->p_flag & P_ONRUNQ)
428 p->p_usched->remrunqueue(&p->p_lwp);
429 p->p_stat = SSLEEP;
0cfcada1 430 p->p_stats->p_ru.ru_nvcsw++;
344ad853
MD
431 lwkt_switch();
432 p->p_stat = SRUN;
0cfcada1
MD
433 } else {
434 lwkt_switch();
435 }
344ad853 436
fc17ad60
MD
437 /*
438 * Make sure we haven't switched cpus while we were asleep. It's
344ad853 439 * not supposed to happen. Cleanup our temporary flags.
fc17ad60
MD
440 */
441 KKASSERT(gd == td->td_gd);
0a3f9b47 442 td->td_flags &= ~TDF_NORESCHED;
344ad853
MD
443
444 /*
445 * Cleanup the timeout.
446 */
447 if (timo) {
448 if (td->td_flags & TDF_TIMEOUT) {
449 td->td_flags &= ~TDF_TIMEOUT;
450 if (sig == 0)
451 error = EWOULDBLOCK;
452 } else {
453 callout_stop(&thandle);
454 }
0cfcada1 455 }
344ad853
MD
456
457 /*
458 * Since td_threadq is used both for our run queue AND for the
459 * tsleep hash queue, we can't still be on it at this point because
460 * we've gotten cpu back.
461 */
afbfc034 462 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
344ad853
MD
463 td->td_wchan = NULL;
464 td->td_wmesg = NULL;
465 td->td_wdomain = 0;
466
467 /*
468 * Figure out the correct error return
469 */
470resume:
0cfcada1 471 if (p) {
344ad853
MD
472 p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR);
473 if (catch && error == 0 && (sig != 0 || (sig = CURSIG(p)))) {
0cfcada1 474 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
344ad853
MD
475 error = EINTR;
476 else
477 error = ERESTART;
984263bc 478 }
984263bc 479 }
9afb0ffd 480 logtsleep(tsleep_end);
344ad853
MD
481 crit_exit_quick(td);
482 return (error);
984263bc
MD
483}
484
16523a43
MD
485/*
486 * This is a dandy function that allows us to interlock tsleep/wakeup
487 * operations with unspecified upper level locks, such as lockmgr locks,
488 * simply by holding a critical section. The sequence is:
489 *
490 * (enter critical section)
491 * (acquire upper level lock)
492 * tsleep_interlock(blah)
493 * (release upper level lock)
494 * tsleep(blah, ...)
495 * (exit critical section)
496 *
497 * Basically this function sets our cpumask for the ident which informs
498 * other cpus that our cpu 'might' be waiting (or about to wait on) the
499 * hash index related to the ident. The critical section prevents another
500 * cpu's wakeup() from being processed on our cpu until we are actually
501 * able to enter the tsleep(). Thus, no race occurs between our attempt
502 * to release a resource and sleep, and another cpu's attempt to acquire
503 * a resource and call wakeup.
504 *
505 * There isn't much of a point to this function unless you call it while
506 * holding a critical section.
507 */
508void
509tsleep_interlock(void *ident)
510{
511 int id = LOOKUP(ident);
512
513 atomic_set_int(&slpque_cpumasks[id], mycpu->gd_cpumask);
514}
515
984263bc 516/*
344ad853 517 * Implement the timeout for tsleep.
fc17ad60 518 *
344ad853
MD
519 * We set P_BREAKTSLEEP to indicate that an event has occured, but
520 * we only call setrunnable if the process is not stopped.
521 *
522 * This type of callout timeout is scheduled on the same cpu the process
523 * is sleeping on. Also, at the moment, the MP lock is held.
984263bc
MD
524 */
525static void
0cfcada1 526endtsleep(void *arg)
984263bc 527{
0cfcada1
MD
528 thread_t td = arg;
529 struct proc *p;
984263bc 530
344ad853 531 ASSERT_MP_LOCK_HELD(curthread);
37af14fe 532 crit_enter();
344ad853
MD
533
534 /*
535 * cpu interlock. Thread flags are only manipulated on
536 * the cpu owning the thread. proc flags are only manipulated
537 * by the older of the MP lock. We have both.
538 */
539 if (td->td_flags & TDF_TSLEEPQ) {
0cfcada1 540 td->td_flags |= TDF_TIMEOUT;
344ad853 541
0cfcada1 542 if ((p = td->td_proc) != NULL) {
344ad853
MD
543 p->p_flag |= P_BREAKTSLEEP;
544 if ((p->p_flag & P_STOPPED) == 0)
0cfcada1 545 setrunnable(p);
0cfcada1 546 } else {
344ad853 547 unsleep_and_wakeup_thread(td);
0cfcada1 548 }
984263bc 549 }
37af14fe 550 crit_exit();
984263bc
MD
551}
552
984263bc 553/*
344ad853
MD
554 * Unsleep and wakeup a thread. This function runs without the MP lock
555 * which means that it can only manipulate thread state on the owning cpu,
556 * and cannot touch the process state at all.
984263bc 557 */
344ad853 558static
8fb8bca6 559void
344ad853 560unsleep_and_wakeup_thread(struct thread *td)
8fb8bca6 561{
344ad853 562 globaldata_t gd = mycpu;
fc17ad60
MD
563 int id;
564
344ad853
MD
565#ifdef SMP
566 if (td->td_gd != gd) {
567 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
568 return;
569 }
570#endif
9c1fad94 571 crit_enter();
344ad853
MD
572 if (td->td_flags & TDF_TSLEEPQ) {
573 td->td_flags &= ~TDF_TSLEEPQ;
574 id = LOOKUP(td->td_wchan);
575 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
576 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
577 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
578 lwkt_schedule(td);
8fb8bca6 579 }
9c1fad94 580 crit_exit();
8fb8bca6 581}
8fb8bca6
EN
582
583/*
584 * Make all processes sleeping on the specified identifier runnable.
fc17ad60
MD
585 * count may be zero or one only.
586 *
587 * The domain encodes the sleep/wakeup domain AND the first cpu to check
588 * (which is always the current cpu). As we iterate across cpus
344ad853
MD
589 *
590 * This call may run without the MP lock held. We can only manipulate thread
591 * state on the cpu owning the thread. We CANNOT manipulate process state
592 * at all.
8fb8bca6
EN
593 */
594static void
fc17ad60 595_wakeup(void *ident, int domain)
984263bc 596{
fc17ad60 597 struct tslpque *qp;
0cfcada1
MD
598 struct thread *td;
599 struct thread *ntd;
fc17ad60 600 globaldata_t gd;
fc17ad60
MD
601#ifdef SMP
602 cpumask_t mask;
603 cpumask_t tmask;
604 int startcpu;
605 int nextcpu;
606#endif
607 int id;
984263bc 608
37af14fe 609 crit_enter();
9afb0ffd 610 logtsleep(wakeup_beg);
fc17ad60
MD
611 gd = mycpu;
612 id = LOOKUP(ident);
613 qp = &gd->gd_tsleep_hash[id];
984263bc 614restart:
0cfcada1
MD
615 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
616 ntd = TAILQ_NEXT(td, td_threadq);
fc17ad60
MD
617 if (td->td_wchan == ident &&
618 td->td_wdomain == (domain & PDOMAIN_MASK)
619 ) {
344ad853
MD
620 KKASSERT(td->td_flags & TDF_TSLEEPQ);
621 td->td_flags &= ~TDF_TSLEEPQ;
0cfcada1 622 TAILQ_REMOVE(qp, td, td_threadq);
fc17ad60
MD
623 if (TAILQ_FIRST(qp) == NULL) {
624 atomic_clear_int(&slpque_cpumasks[id],
625 gd->gd_cpumask);
626 }
344ad853 627 lwkt_schedule(td);
fc17ad60
MD
628 if (domain & PWAKEUP_ONE)
629 goto done;
0cfcada1 630 goto restart;
984263bc
MD
631 }
632 }
fc17ad60
MD
633
634#ifdef SMP
635 /*
636 * We finished checking the current cpu but there still may be
637 * more work to do. Either wakeup_one was requested and no matching
638 * thread was found, or a normal wakeup was requested and we have
639 * to continue checking cpus.
640 *
641 * The cpu that started the wakeup sequence is encoded in the domain.
642 * We use this information to determine which cpus still need to be
643 * checked, locate a candidate cpu, and chain the wakeup
644 * asynchronously with an IPI message.
645 *
646 * It should be noted that this scheme is actually less expensive then
647 * the old scheme when waking up multiple threads, since we send
648 * only one IPI message per target candidate which may then schedule
649 * multiple threads. Before we could have wound up sending an IPI
650 * message for each thread on the target cpu (!= current cpu) that
651 * needed to be woken up.
652 *
653 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
654 * should be ok since we are passing idents in the IPI rather then
655 * thread pointers.
656 */
657 if ((mask = slpque_cpumasks[id]) != 0) {
658 /*
659 * Look for a cpu that might have work to do. Mask out cpus
660 * which have already been processed.
661 *
662 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
663 * ^ ^ ^
664 * start currentcpu start
665 * case2 case1
666 * * * *
667 * 11111111111111110000000000000111 case1
668 * 00000000111111110000000000000000 case2
669 *
670 * case1: We started at start_case1 and processed through
671 * to the current cpu. We have to check any bits
672 * after the current cpu, then check bits before
673 * the starting cpu.
674 *
675 * case2: We have already checked all the bits from
676 * start_case2 to the end, and from 0 to the current
677 * cpu. We just have the bits from the current cpu
678 * to start_case2 left to check.
679 */
680 startcpu = PWAKEUP_DECODE(domain);
681 if (gd->gd_cpuid >= startcpu) {
682 /*
683 * CASE1
684 */
685 tmask = mask & ~((gd->gd_cpumask << 1) - 1);
686 if (mask & tmask) {
687 nextcpu = bsfl(mask & tmask);
688 lwkt_send_ipiq2(globaldata_find(nextcpu),
689 _wakeup, ident, domain);
690 } else {
691 tmask = (1 << startcpu) - 1;
692 if (mask & tmask) {
693 nextcpu = bsfl(mask & tmask);
694 lwkt_send_ipiq2(
695 globaldata_find(nextcpu),
696 _wakeup, ident, domain);
697 }
698 }
699 } else {
700 /*
701 * CASE2
702 */
703 tmask = ~((gd->gd_cpumask << 1) - 1) &
704 ((1 << startcpu) - 1);
705 if (mask & tmask) {
706 nextcpu = bsfl(mask & tmask);
707 lwkt_send_ipiq2(globaldata_find(nextcpu),
708 _wakeup, ident, domain);
709 }
710 }
711 }
712#endif
713done:
9afb0ffd 714 logtsleep(wakeup_end);
37af14fe 715 crit_exit();
984263bc
MD
716}
717
984263bc 718void
0cfcada1 719wakeup(void *ident)
984263bc 720{
fc17ad60 721 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
0cfcada1 722}
984263bc 723
0cfcada1
MD
724void
725wakeup_one(void *ident)
726{
fc17ad60
MD
727 /* XXX potentially round-robin the first responding cpu */
728 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
da5fb9ef
MD
729}
730
731void
732wakeup_domain(void *ident, int domain)
733{
fc17ad60 734 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
da5fb9ef
MD
735}
736
737void
738wakeup_domain_one(void *ident, int domain)
739{
fc17ad60
MD
740 /* XXX potentially round-robin the first responding cpu */
741 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
984263bc
MD
742}
743
744/*
344ad853
MD
745 * setrunnable()
746 *
747 * Make a process runnable. The MP lock must be held on call. This only
748 * has an effect if we are in SSLEEP. We only break out of the
749 * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state.
37af14fe 750 *
344ad853
MD
751 * NOTE: With the MP lock held we can only safely manipulate the process
752 * structure. We cannot safely manipulate the thread structure.
984263bc
MD
753 */
754void
344ad853 755setrunnable(struct proc *p)
984263bc 756{
344ad853
MD
757 crit_enter();
758 ASSERT_MP_LOCK_HELD(curthread);
759 p->p_flag &= ~P_STOPPED;
760 if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) {
761 unsleep_and_wakeup_thread(p->p_thread);
984263bc 762 }
344ad853 763 crit_exit();
984263bc
MD
764}
765
766/*
344ad853
MD
767 * The process is stopped due to some condition, usually because P_STOPPED
768 * is set but also possibly due to being traced.
fc17ad60 769 *
344ad853
MD
770 * NOTE! If the caller sets P_STOPPED, the caller must also clear P_WAITED
771 * because the parent may check the child's status before the child actually
772 * gets to this routine.
773 *
774 * This routine is called with the current process only, typically just
775 * before returning to userland.
776 *
777 * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive
778 * SIGCONT to break out of the tsleep.
984263bc
MD
779 */
780void
344ad853 781tstop(struct proc *p)
984263bc 782{
344ad853
MD
783 wakeup((caddr_t)p->p_pptr);
784 p->p_flag |= P_BREAKTSLEEP;
785 tsleep(p, 0, "stop", 0);
26a0694b
MD
786}
787
a77ac49d
MD
788/*
789 * Yield / synchronous reschedule. This is a bit tricky because the trap
790 * code might have set a lazy release on the switch function. Setting
791 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
792 * switch, and that we are given a greater chance of affinity with our
793 * current cpu.
794 *
795 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
796 * run queue. lwkt_switch() will also execute any assigned passive release
797 * (which usually calls release_curproc()), allowing a same/higher priority
798 * process to be designated as the current process.
799 *
800 * While it is possible for a lower priority process to be designated,
801 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
802 * round-robin back to us and we will be able to re-acquire the current
803 * process designation.
804 */
805void
806uio_yield(void)
807{
808 struct thread *td = curthread;
809 struct proc *p = td->td_proc;
810
811 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
812 if (p) {
813 p->p_flag |= P_PASSIVE_ACQ;
814 lwkt_switch();
815 p->p_flag &= ~P_PASSIVE_ACQ;
816 } else {
817 lwkt_switch();
818 }
819}
820
984263bc
MD
821/*
822 * Compute a tenex style load average of a quantity on
823 * 1, 5 and 15 minute intervals.
824 */
825static void
826loadav(void *arg)
827{
828 int i, nrun;
829 struct loadavg *avg;
830 struct proc *p;
8ec60c3f 831 thread_t td;
984263bc
MD
832
833 avg = &averunnable;
834 nrun = 0;
f62004ad 835 FOREACH_PROC_IN_SYSTEM(p) {
8ec60c3f
MD
836 switch (p->p_stat) {
837 case SRUN:
838 if ((td = p->p_thread) == NULL)
839 break;
840 if (td->td_flags & TDF_BLOCKED)
841 break;
842 /* fall through */
843 case SIDL:
844 nrun++;
845 break;
846 default:
847 break;
984263bc
MD
848 }
849 }
850 for (i = 0; i < 3; i++)
851 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
852 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
853
854 /*
855 * Schedule the next update to occur after 5 seconds, but add a
856 * random variation to avoid synchronisation with processes that
857 * run at regular intervals.
858 */
859 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
860 loadav, NULL);
861}
862
863/* ARGSUSED */
864static void
6656cd91 865sched_setup(void *dummy)
984263bc 866{
984263bc 867 callout_init(&loadav_callout);
35f9d051 868 callout_init(&schedcpu_callout);
984263bc
MD
869
870 /* Kick off timeout driven events by calling first time. */
984263bc
MD
871 schedcpu(NULL);
872 loadav(NULL);
873}
874