HAMMER - new hammer_lock code, fix mplock bug in last commit, mpsafe getattr.
[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>
b1b4e5a6 50#include <sys/signal2.h>
984263bc
MD
51#include <sys/resourcevar.h>
52#include <sys/vmmeter.h>
53#include <sys/sysctl.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 61
bf765287
MD
62#include <sys/thread2.h>
63#include <sys/spinlock2.h>
362e59be 64#include <sys/serialize.h>
bf765287 65
984263bc 66#include <machine/cpu.h>
984263bc
MD
67#include <machine/smp.h>
68
fc17ad60
MD
69TAILQ_HEAD(tslpque, thread);
70
402ed7e1 71static void sched_setup (void *dummy);
984263bc
MD
72SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
73
984263bc
MD
74int hogticks;
75int lbolt;
344ad853 76int lbolt_syncer;
984263bc 77int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
17a9f566 78int ncpus;
90100055 79int ncpus2, ncpus2_shift, ncpus2_mask;
b45759e1 80int ncpus_fit, ncpus_fit_mask;
e43a034f 81int safepri;
dbcd0c9b 82int tsleep_now_works;
984263bc
MD
83
84static struct callout loadav_callout;
35f9d051 85static struct callout schedcpu_callout;
fc17ad60 86MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
984263bc 87
9afb0ffd
MD
88#if !defined(KTR_TSLEEP)
89#define KTR_TSLEEP KTR_ALL
90#endif
91KTR_INFO_MASTER(tsleep);
8aa3430c
MD
92KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *));
93KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0);
94KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *));
95KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0);
96
97#define logtsleep1(name) KTR_LOG(tsleep_ ## name)
98#define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val)
9afb0ffd 99
984263bc
MD
100struct loadavg averunnable =
101 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
102/*
103 * Constants for averages over 1, 5, and 15 minutes
104 * when sampling at 5 second intervals.
105 */
106static fixpt_t cexp[3] = {
107 0.9200444146293232 * FSCALE, /* exp(-1/12) */
108 0.9834714538216174 * FSCALE, /* exp(-1/60) */
109 0.9944598480048967 * FSCALE, /* exp(-1/180) */
110};
111
402ed7e1 112static void endtsleep (void *);
344ad853 113static void unsleep_and_wakeup_thread(struct thread *td);
402ed7e1 114static void loadav (void *arg);
402ed7e1 115static void schedcpu (void *arg);
984263bc 116
a46fac56
MD
117/*
118 * Adjust the scheduler quantum. The quantum is specified in microseconds.
119 * Note that 'tick' is in microseconds per tick.
120 */
984263bc
MD
121static int
122sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
123{
124 int error, new_val;
125
126 new_val = sched_quantum * tick;
127 error = sysctl_handle_int(oidp, &new_val, 0, req);
128 if (error != 0 || req->newptr == NULL)
129 return (error);
130 if (new_val < tick)
131 return (EINVAL);
132 sched_quantum = new_val / tick;
133 hogticks = 2 * sched_quantum;
134 return (0);
135}
136
137SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
138 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
139
984263bc
MD
140/*
141 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
142 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
143 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
144 *
145 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
dcc99b62 146 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
984263bc
MD
147 *
148 * If you don't want to bother with the faster/more-accurate formula, you
149 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
150 * (more general) method of calculating the %age of CPU used by a process.
dcc99b62 151 *
08f2f1bb 152 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
dcc99b62
MD
153 */
154#define CCPU_SHIFT 11
155
156static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
157SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
158
159/*
160 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
984263bc 161 */
460426e6 162int fscale __unused = FSCALE; /* exported to systat */
dcc99b62 163SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
984263bc
MD
164
165/*
0a3f9b47 166 * Recompute process priorities, once a second.
dcc99b62
MD
167 *
168 * Since the userland schedulers are typically event oriented, if the
169 * estcpu calculation at wakeup() time is not sufficient to make a
170 * process runnable relative to other processes in the system we have
171 * a 1-second recalc to help out.
172 *
173 * This code also allows us to store sysclock_t data in the process structure
174 * without fear of an overrun, since sysclock_t are guarenteed to hold
175 * several seconds worth of count.
8fa76237
MD
176 *
177 * WARNING! callouts can preempt normal threads. However, they will not
178 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
984263bc 179 */
8fa76237
MD
180static int schedcpu_stats(struct proc *p, void *data __unused);
181static int schedcpu_resource(struct proc *p, void *data __unused);
182
984263bc 183static void
26a0694b 184schedcpu(void *arg)
984263bc 185{
8fa76237
MD
186 allproc_scan(schedcpu_stats, NULL);
187 allproc_scan(schedcpu_resource, NULL);
188 wakeup((caddr_t)&lbolt);
189 wakeup((caddr_t)&lbolt_syncer);
190 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
191}
192
193/*
194 * General process statistics once a second
195 */
196static int
197schedcpu_stats(struct proc *p, void *data __unused)
198{
08f2f1bb
SS
199 struct lwp *lp;
200
8fa76237
MD
201 crit_enter();
202 p->p_swtime++;
c7e98b2f
SS
203 FOREACH_LWP_IN_PROC(lp, p) {
204 if (lp->lwp_stat == LSSLEEP)
205 lp->lwp_slptime++;
4b5f931b 206
c7e98b2f
SS
207 /*
208 * Only recalculate processes that are active or have slept
209 * less then 2 seconds. The schedulers understand this.
210 */
211 if (lp->lwp_slptime <= 1) {
212 p->p_usched->recalculate(lp);
213 } else {
214 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
215 }
8fa76237
MD
216 }
217 crit_exit();
218 return(0);
219}
a46fac56 220
8fa76237 221/*
84204577 222 * Resource checks. XXX break out since ksignal/killproc can block,
8fa76237
MD
223 * limiting us to one process killed per second. There is probably
224 * a better way.
225 */
226static int
227schedcpu_resource(struct proc *p, void *data __unused)
228{
229 u_int64_t ttime;
08f2f1bb 230 struct lwp *lp;
8fa76237
MD
231
232 crit_enter();
233 if (p->p_stat == SIDL ||
416d05d7 234 p->p_stat == SZOMB ||
c7e98b2f 235 p->p_limit == NULL
8fa76237 236 ) {
e43a034f 237 crit_exit();
8fa76237 238 return(0);
984263bc 239 }
344ad853 240
c7e98b2f
SS
241 ttime = 0;
242 FOREACH_LWP_IN_PROC(lp, p) {
e595c6cd
MD
243 /*
244 * We may have caught an lp in the middle of being
245 * created, lwp_thread can be NULL.
246 */
247 if (lp->lwp_thread) {
248 ttime += lp->lwp_thread->td_sticks;
249 ttime += lp->lwp_thread->td_uticks;
250 }
c7e98b2f 251 }
8fa76237
MD
252
253 switch(plimit_testcpulimit(p->p_limit, ttime)) {
254 case PLIMIT_TESTCPU_KILL:
255 killproc(p, "exceeded maximum CPU limit");
256 break;
257 case PLIMIT_TESTCPU_XCPU:
258 if ((p->p_flag & P_XCPU) == 0) {
259 p->p_flag |= P_XCPU;
84204577 260 ksignal(p, SIGXCPU);
344ad853 261 }
8fa76237
MD
262 break;
263 default:
c0b8a06d 264 break;
344ad853 265 }
8fa76237
MD
266 crit_exit();
267 return(0);
984263bc
MD
268}
269
270/*
dcc99b62
MD
271 * This is only used by ps. Generate a cpu percentage use over
272 * a period of one second.
52eedfb5
MD
273 *
274 * MPSAFE
984263bc 275 */
dcc99b62 276void
553ea3c8 277updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
984263bc 278{
dcc99b62
MD
279 fixpt_t acc;
280 int remticks;
281
282 acc = (cpticks << FSHIFT) / ttlticks;
283 if (ttlticks >= ESTCPUFREQ) {
553ea3c8 284 lp->lwp_pctcpu = acc;
dcc99b62
MD
285 } else {
286 remticks = ESTCPUFREQ - ttlticks;
553ea3c8 287 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
dcc99b62 288 ESTCPUFREQ;
a46fac56 289 }
984263bc
MD
290}
291
292/*
8aa3430c
MD
293 * tsleep/wakeup hash table parameters. Try to find the sweet spot for
294 * like addresses being slept on.
984263bc 295 */
8aa3430c
MD
296#define TABLESIZE 1024
297#define LOOKUP(x) (((intptr_t)(x) >> 6) & (TABLESIZE - 1))
984263bc 298
fc17ad60
MD
299static cpumask_t slpque_cpumasks[TABLESIZE];
300
984263bc 301/*
a46fac56 302 * General scheduler initialization. We force a reschedule 25 times
fc17ad60
MD
303 * a second by default. Note that cpu0 is initialized in early boot and
304 * cannot make any high level calls.
305 *
306 * Each cpu has its own sleep queue.
984263bc 307 */
984263bc 308void
fc17ad60 309sleep_gdinit(globaldata_t gd)
984263bc 310{
fc17ad60 311 static struct tslpque slpque_cpu0[TABLESIZE];
9c1fad94 312 int i;
984263bc 313
fc17ad60
MD
314 if (gd->gd_cpuid == 0) {
315 sched_quantum = (hz + 24) / 25;
316 hogticks = 2 * sched_quantum;
317
318 gd->gd_tsleep_hash = slpque_cpu0;
319 } else {
77652cad 320 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
fc17ad60
MD
321 M_TSLEEP, M_WAITOK | M_ZERO);
322 }
323 for (i = 0; i < TABLESIZE; ++i)
324 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
984263bc
MD
325}
326
327/*
328 * General sleep call. Suspends the current process until a wakeup is
329 * performed on the specified identifier. The process will then be made
330 * runnable with the specified priority. Sleeps at most timo/hz seconds
377d4740 331 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
984263bc
MD
332 * before and after sleeping, else signals are not checked. Returns 0 if
333 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
334 * signal needs to be delivered, ERESTART is returned if the current system
335 * call should be restarted if possible, and EINTR is returned if the system
336 * call should be interrupted by the signal (return EINTR).
26a0694b 337 *
0a3f9b47
MD
338 * Note that if we are a process, we release_curproc() before messing with
339 * the LWKT scheduler.
a46fac56
MD
340 *
341 * During autoconfiguration or after a panic, a sleep will simply
342 * lower the priority briefly to allow interrupts, then return.
984263bc
MD
343 */
344int
6656cd91 345tsleep(void *ident, int flags, const char *wmesg, int timo)
984263bc 346{
dadab5e9 347 struct thread *td = curthread;
08f2f1bb 348 struct lwp *lp = td->td_lwp;
0cfcada1 349 struct proc *p = td->td_proc; /* may be NULL */
fc17ad60 350 globaldata_t gd;
344ad853
MD
351 int sig;
352 int catch;
353 int id;
354 int error;
e43a034f 355 int oldpri;
076fecef 356 struct callout thandle;
984263bc 357
0cfcada1
MD
358 /*
359 * NOTE: removed KTRPOINT, it could cause races due to blocking
360 * even in stable. Just scrap it for now.
361 */
dbcd0c9b 362 if (tsleep_now_works == 0 || panicstr) {
984263bc 363 /*
dbcd0c9b
MD
364 * After a panic, or before we actually have an operational
365 * softclock, just give interrupts a chance, then just return;
366 *
984263bc
MD
367 * don't run any other procs or panic below,
368 * in case this is the idle process and already asleep.
369 */
e43a034f
MD
370 splz();
371 oldpri = td->td_pri & TDPRI_MASK;
372 lwkt_setpri_self(safepri);
373 lwkt_switch();
374 lwkt_setpri_self(oldpri);
984263bc
MD
375 return (0);
376 }
8aa3430c 377 logtsleep2(tsleep_beg, ident);
fc17ad60
MD
378 gd = td->td_gd;
379 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
344ad853
MD
380
381 /*
382 * NOTE: all of this occurs on the current cpu, including any
383 * callout-based wakeups, so a critical section is a sufficient
384 * interlock.
385 *
386 * The entire sequence through to where we actually sleep must
387 * run without breaking the critical section.
388 */
389 id = LOOKUP(ident);
390 catch = flags & PCATCH;
391 error = 0;
392 sig = 0;
393
37af14fe 394 crit_enter_quick(td);
344ad853 395
0cfcada1 396 KASSERT(ident != NULL, ("tsleep: no ident"));
7278a846
SS
397 KASSERT(lp == NULL ||
398 lp->lwp_stat == LSRUN || /* Obvious */
399 lp->lwp_stat == LSSTOP, /* Set in tstop */
400 ("tsleep %p %s %d",
401 ident, wmesg, lp->lwp_stat));
0cfcada1 402
344ad853
MD
403 /*
404 * Setup for the current process (if this is a process).
405 */
08f2f1bb 406 if (lp) {
344ad853
MD
407 if (catch) {
408 /*
409 * Early termination if PCATCH was set and a
410 * signal is pending, interlocked with the
411 * critical section.
412 *
413 * Early termination only occurs when tsleep() is
164b8401 414 * entered while in a normal LSRUN state.
344ad853 415 */
08f2f1bb 416 if ((sig = CURSIG(lp)) != 0)
344ad853
MD
417 goto resume;
418
7c1212ec
MD
419 /*
420 * Early termination if PCATCH was set and a
421 * mailbox signal was possibly delivered prior to
422 * the system call even being made, in order to
423 * allow the user to interlock without having to
424 * make additional system calls.
425 */
426 if (p->p_flag & P_MAILBOX)
427 goto resume;
428
344ad853 429 /*
84204577 430 * Causes ksignal to wake us up when.
344ad853 431 */
9a379a4a 432 lp->lwp_flag |= LWP_SINTR;
344ad853
MD
433 }
434
435 /*
436 * Make sure the current process has been untangled from
437 * the userland scheduler and initialize slptime to start
438 * counting.
439 */
08f2f1bb
SS
440 p->p_usched->release_curproc(lp);
441 lp->lwp_slptime = 0;
0a3f9b47 442 }
fc17ad60
MD
443
444 /*
344ad853 445 * Move our thread to the correct queue and setup our wchan, etc.
fc17ad60 446 */
37af14fe 447 lwkt_deschedule_self(td);
344ad853 448 td->td_flags |= TDF_TSLEEPQ;
fc17ad60
MD
449 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
450 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
344ad853
MD
451
452 td->td_wchan = ident;
453 td->td_wmesg = wmesg;
454 td->td_wdomain = flags & PDOMAIN_MASK;
455
456 /*
457 * Setup the timeout, if any
458 */
076fecef
MD
459 if (timo) {
460 callout_init(&thandle);
461 callout_reset(&thandle, timo, endtsleep, td);
462 }
344ad853 463
984263bc 464 /*
344ad853 465 * Beddy bye bye.
984263bc 466 */
08f2f1bb 467 if (lp) {
26a0694b 468 /*
52eedfb5 469 * Ok, we are sleeping. Place us in the SSLEEP state.
26a0694b 470 */
9388413d 471 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
7278a846
SS
472 /*
473 * tstop() sets LSSTOP, so don't fiddle with that.
474 */
475 if (lp->lwp_stat != LSSTOP)
476 lp->lwp_stat = LSSLEEP;
08f2f1bb 477 lp->lwp_ru.ru_nvcsw++;
344ad853 478 lwkt_switch();
ab554892
MD
479
480 /*
164b8401 481 * And when we are woken up, put us back in LSRUN. If we
ab554892
MD
482 * slept for over a second, recalculate our estcpu.
483 */
164b8401 484 lp->lwp_stat = LSRUN;
08f2f1bb
SS
485 if (lp->lwp_slptime)
486 p->p_usched->recalculate(lp);
487 lp->lwp_slptime = 0;
0cfcada1
MD
488 } else {
489 lwkt_switch();
490 }
344ad853 491
fc17ad60
MD
492 /*
493 * Make sure we haven't switched cpus while we were asleep. It's
344ad853 494 * not supposed to happen. Cleanup our temporary flags.
fc17ad60
MD
495 */
496 KKASSERT(gd == td->td_gd);
344ad853
MD
497
498 /*
499 * Cleanup the timeout.
500 */
501 if (timo) {
502 if (td->td_flags & TDF_TIMEOUT) {
503 td->td_flags &= ~TDF_TIMEOUT;
a40da8f0 504 error = EWOULDBLOCK;
344ad853
MD
505 } else {
506 callout_stop(&thandle);
507 }
0cfcada1 508 }
344ad853
MD
509
510 /*
511 * Since td_threadq is used both for our run queue AND for the
512 * tsleep hash queue, we can't still be on it at this point because
513 * we've gotten cpu back.
514 */
afbfc034 515 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
344ad853
MD
516 td->td_wchan = NULL;
517 td->td_wmesg = NULL;
518 td->td_wdomain = 0;
519
520 /*
7c1212ec
MD
521 * Figure out the correct error return. If interrupted by a
522 * signal we want to return EINTR or ERESTART.
523 *
524 * If P_MAILBOX is set no automatic system call restart occurs
525 * and we return EINTR. P_MAILBOX is meant to be used as an
526 * interlock, the user must poll it prior to any system call
527 * that it wishes to interlock a mailbox signal against since
528 * the flag is cleared on *any* system call that sleeps.
344ad853
MD
529 */
530resume:
0cfcada1 531 if (p) {
7c1212ec
MD
532 if (catch && error == 0) {
533 if ((p->p_flag & P_MAILBOX) && sig == 0) {
344ad853 534 error = EINTR;
08f2f1bb 535 } else if (sig != 0 || (sig = CURSIG(lp))) {
7c1212ec
MD
536 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
537 error = EINTR;
538 else
539 error = ERESTART;
540 }
984263bc 541 }
9a379a4a
SS
542 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
543 p->p_flag &= ~P_MAILBOX;
984263bc 544 }
8aa3430c 545 logtsleep1(tsleep_end);
344ad853
MD
546 crit_exit_quick(td);
547 return (error);
984263bc
MD
548}
549
16523a43
MD
550/*
551 * This is a dandy function that allows us to interlock tsleep/wakeup
552 * operations with unspecified upper level locks, such as lockmgr locks,
553 * simply by holding a critical section. The sequence is:
554 *
555 * (enter critical section)
556 * (acquire upper level lock)
557 * tsleep_interlock(blah)
558 * (release upper level lock)
559 * tsleep(blah, ...)
560 * (exit critical section)
561 *
562 * Basically this function sets our cpumask for the ident which informs
563 * other cpus that our cpu 'might' be waiting (or about to wait on) the
564 * hash index related to the ident. The critical section prevents another
565 * cpu's wakeup() from being processed on our cpu until we are actually
566 * able to enter the tsleep(). Thus, no race occurs between our attempt
567 * to release a resource and sleep, and another cpu's attempt to acquire
568 * a resource and call wakeup.
569 *
570 * There isn't much of a point to this function unless you call it while
571 * holding a critical section.
572 */
bf765287
MD
573static __inline void
574_tsleep_interlock(globaldata_t gd, void *ident)
575{
576 int id = LOOKUP(ident);
577
578 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
579}
580
16523a43
MD
581void
582tsleep_interlock(void *ident)
583{
bf765287
MD
584 _tsleep_interlock(mycpu, ident);
585}
586
587/*
588 * Interlocked spinlock sleep. An exclusively held spinlock must
589 * be passed to msleep(). The function will atomically release the
590 * spinlock and tsleep on the ident, then reacquire the spinlock and
591 * return.
592 *
593 * This routine is fairly important along the critical path, so optimize it
594 * heavily.
595 */
596int
597msleep(void *ident, struct spinlock *spin, int flags,
598 const char *wmesg, int timo)
599{
600 globaldata_t gd = mycpu;
601 int error;
16523a43 602
bf765287
MD
603 crit_enter_gd(gd);
604 _tsleep_interlock(gd, ident);
605 spin_unlock_wr_quick(gd, spin);
606 error = tsleep(ident, flags, wmesg, timo);
607 spin_lock_wr_quick(gd, spin);
608 crit_exit_gd(gd);
609
610 return (error);
16523a43
MD
611}
612
362e59be
SZ
613/*
614 * Interlocked serializer sleep. An exclusively held serializer must
615 * be passed to serialize_sleep(). The function will atomically release
616 * the serializer and tsleep on the ident, then reacquire the serializer
617 * and return.
618 */
619int
620serialize_sleep(void *ident, struct lwkt_serialize *slz, int flags,
621 const char *wmesg, int timo)
622{
623 int ret;
624
625 ASSERT_SERIALIZED(slz);
626
627 crit_enter();
628 tsleep_interlock(ident);
629 lwkt_serialize_exit(slz);
630 ret = tsleep(ident, flags, wmesg, timo);
631 lwkt_serialize_enter(slz);
632 crit_exit();
633
634 return ret;
635}
636
a22c590e
MD
637/*
638 * Directly block on the LWKT thread by descheduling it. This
639 * is much faster then tsleep(), but the only legal way to wake
640 * us up is to directly schedule the thread.
641 *
642 * Setting TDF_SINTR will cause new signals to directly schedule us.
643 *
644 * This routine is typically called while in a critical section.
645 */
646int
647lwkt_sleep(const char *wmesg, int flags)
648{
649 thread_t td = curthread;
650 int sig;
651
652 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) {
653 td->td_flags |= TDF_BLOCKED;
654 td->td_wmesg = wmesg;
655 lwkt_deschedule_self(td);
656 lwkt_switch();
657 td->td_wmesg = NULL;
658 td->td_flags &= ~TDF_BLOCKED;
659 return(0);
660 }
661 if ((sig = CURSIG(td->td_lwp)) != 0) {
662 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig))
663 return(EINTR);
664 else
665 return(ERESTART);
666
667 }
668 td->td_flags |= TDF_BLOCKED | TDF_SINTR;
669 td->td_wmesg = wmesg;
670 lwkt_deschedule_self(td);
671 lwkt_switch();
672 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR);
673 td->td_wmesg = NULL;
674 return(0);
675}
676
984263bc 677/*
344ad853 678 * Implement the timeout for tsleep.
fc17ad60 679 *
9a379a4a 680 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
344ad853
MD
681 * we only call setrunnable if the process is not stopped.
682 *
683 * This type of callout timeout is scheduled on the same cpu the process
684 * is sleeping on. Also, at the moment, the MP lock is held.
984263bc
MD
685 */
686static void
0cfcada1 687endtsleep(void *arg)
984263bc 688{
0cfcada1 689 thread_t td = arg;
9a379a4a 690 struct lwp *lp;
984263bc 691
344ad853 692 ASSERT_MP_LOCK_HELD(curthread);
37af14fe 693 crit_enter();
344ad853
MD
694
695 /*
696 * cpu interlock. Thread flags are only manipulated on
697 * the cpu owning the thread. proc flags are only manipulated
698 * by the older of the MP lock. We have both.
699 */
700 if (td->td_flags & TDF_TSLEEPQ) {
0cfcada1 701 td->td_flags |= TDF_TIMEOUT;
344ad853 702
9a379a4a
SS
703 if ((lp = td->td_lwp) != NULL) {
704 lp->lwp_flag |= LWP_BREAKTSLEEP;
705 if (lp->lwp_proc->p_stat != SSTOP)
706 setrunnable(lp);
0cfcada1 707 } else {
344ad853 708 unsleep_and_wakeup_thread(td);
0cfcada1 709 }
984263bc 710 }
37af14fe 711 crit_exit();
984263bc
MD
712}
713
984263bc 714/*
344ad853
MD
715 * Unsleep and wakeup a thread. This function runs without the MP lock
716 * which means that it can only manipulate thread state on the owning cpu,
717 * and cannot touch the process state at all.
984263bc 718 */
344ad853 719static
8fb8bca6 720void
344ad853 721unsleep_and_wakeup_thread(struct thread *td)
8fb8bca6 722{
344ad853 723 globaldata_t gd = mycpu;
fc17ad60
MD
724 int id;
725
344ad853
MD
726#ifdef SMP
727 if (td->td_gd != gd) {
728 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
729 return;
730 }
731#endif
9c1fad94 732 crit_enter();
344ad853
MD
733 if (td->td_flags & TDF_TSLEEPQ) {
734 td->td_flags &= ~TDF_TSLEEPQ;
735 id = LOOKUP(td->td_wchan);
736 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
737 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
738 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
739 lwkt_schedule(td);
8fb8bca6 740 }
9c1fad94 741 crit_exit();
8fb8bca6 742}
8fb8bca6
EN
743
744/*
745 * Make all processes sleeping on the specified identifier runnable.
fc17ad60
MD
746 * count may be zero or one only.
747 *
748 * The domain encodes the sleep/wakeup domain AND the first cpu to check
749 * (which is always the current cpu). As we iterate across cpus
344ad853
MD
750 *
751 * This call may run without the MP lock held. We can only manipulate thread
752 * state on the cpu owning the thread. We CANNOT manipulate process state
753 * at all.
8fb8bca6
EN
754 */
755static void
fc17ad60 756_wakeup(void *ident, int domain)
984263bc 757{
fc17ad60 758 struct tslpque *qp;
0cfcada1
MD
759 struct thread *td;
760 struct thread *ntd;
fc17ad60 761 globaldata_t gd;
fc17ad60
MD
762#ifdef SMP
763 cpumask_t mask;
fc17ad60
MD
764#endif
765 int id;
984263bc 766
37af14fe 767 crit_enter();
8aa3430c 768 logtsleep2(wakeup_beg, ident);
fc17ad60
MD
769 gd = mycpu;
770 id = LOOKUP(ident);
771 qp = &gd->gd_tsleep_hash[id];
984263bc 772restart:
0cfcada1
MD
773 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
774 ntd = TAILQ_NEXT(td, td_threadq);
fc17ad60
MD
775 if (td->td_wchan == ident &&
776 td->td_wdomain == (domain & PDOMAIN_MASK)
777 ) {
344ad853
MD
778 KKASSERT(td->td_flags & TDF_TSLEEPQ);
779 td->td_flags &= ~TDF_TSLEEPQ;
0cfcada1 780 TAILQ_REMOVE(qp, td, td_threadq);
fc17ad60
MD
781 if (TAILQ_FIRST(qp) == NULL) {
782 atomic_clear_int(&slpque_cpumasks[id],
783 gd->gd_cpumask);
784 }
344ad853 785 lwkt_schedule(td);
fc17ad60
MD
786 if (domain & PWAKEUP_ONE)
787 goto done;
0cfcada1 788 goto restart;
984263bc
MD
789 }
790 }
fc17ad60
MD
791
792#ifdef SMP
793 /*
794 * We finished checking the current cpu but there still may be
795 * more work to do. Either wakeup_one was requested and no matching
796 * thread was found, or a normal wakeup was requested and we have
797 * to continue checking cpus.
798 *
fc17ad60
MD
799 * It should be noted that this scheme is actually less expensive then
800 * the old scheme when waking up multiple threads, since we send
801 * only one IPI message per target candidate which may then schedule
802 * multiple threads. Before we could have wound up sending an IPI
803 * message for each thread on the target cpu (!= current cpu) that
804 * needed to be woken up.
805 *
806 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
807 * should be ok since we are passing idents in the IPI rather then
808 * thread pointers.
809 */
1f4f6e0b
MD
810 if ((domain & PWAKEUP_MYCPU) == 0 &&
811 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) {
812 lwkt_send_ipiq2_mask(mask, _wakeup, ident,
813 domain | PWAKEUP_MYCPU);
fc17ad60
MD
814 }
815#endif
816done:
8aa3430c 817 logtsleep1(wakeup_end);
37af14fe 818 crit_exit();
984263bc
MD
819}
820
b336a9b1
MD
821/*
822 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
823 */
984263bc 824void
0cfcada1 825wakeup(void *ident)
984263bc 826{
fc17ad60 827 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
0cfcada1 828}
984263bc 829
b336a9b1
MD
830/*
831 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
832 */
0cfcada1
MD
833void
834wakeup_one(void *ident)
835{
fc17ad60
MD
836 /* XXX potentially round-robin the first responding cpu */
837 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
da5fb9ef
MD
838}
839
b336a9b1
MD
840/*
841 * Wakeup threads tsleep()ing on the specified ident on the current cpu
842 * only.
843 */
844void
845wakeup_mycpu(void *ident)
846{
847 _wakeup(ident, PWAKEUP_MYCPU);
848}
849
850/*
851 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
852 * only.
853 */
854void
855wakeup_mycpu_one(void *ident)
856{
857 /* XXX potentially round-robin the first responding cpu */
858 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
859}
860
861/*
862 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
863 * only.
864 */
865void
866wakeup_oncpu(globaldata_t gd, void *ident)
867{
1699d292 868#ifdef SMP
b336a9b1
MD
869 if (gd == mycpu) {
870 _wakeup(ident, PWAKEUP_MYCPU);
871 } else {
872 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
873 }
1699d292
MD
874#else
875 _wakeup(ident, PWAKEUP_MYCPU);
876#endif
b336a9b1
MD
877}
878
879/*
880 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
881 * only.
882 */
883void
884wakeup_oncpu_one(globaldata_t gd, void *ident)
885{
1699d292 886#ifdef SMP
b336a9b1
MD
887 if (gd == mycpu) {
888 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
889 } else {
890 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
891 }
1699d292
MD
892#else
893 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
894#endif
b336a9b1
MD
895}
896
897/*
898 * Wakeup all threads waiting on the specified ident that slept using
899 * the specified domain, on all cpus.
900 */
da5fb9ef
MD
901void
902wakeup_domain(void *ident, int domain)
903{
fc17ad60 904 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
da5fb9ef
MD
905}
906
b336a9b1
MD
907/*
908 * Wakeup one thread waiting on the specified ident that slept using
909 * the specified domain, on any cpu.
910 */
da5fb9ef
MD
911void
912wakeup_domain_one(void *ident, int domain)
913{
fc17ad60
MD
914 /* XXX potentially round-robin the first responding cpu */
915 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
984263bc
MD
916}
917
918/*
344ad853
MD
919 * setrunnable()
920 *
921 * Make a process runnable. The MP lock must be held on call. This only
922 * has an effect if we are in SSLEEP. We only break out of the
9a379a4a 923 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
37af14fe 924 *
344ad853
MD
925 * NOTE: With the MP lock held we can only safely manipulate the process
926 * structure. We cannot safely manipulate the thread structure.
984263bc
MD
927 */
928void
9a379a4a 929setrunnable(struct lwp *lp)
984263bc 930{
344ad853
MD
931 crit_enter();
932 ASSERT_MP_LOCK_HELD(curthread);
2daf83b0
SS
933 if (lp->lwp_stat == LSSTOP)
934 lp->lwp_stat = LSSLEEP;
935 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
08f2f1bb 936 unsleep_and_wakeup_thread(lp->lwp_thread);
344ad853 937 crit_exit();
984263bc
MD
938}
939
940/*
164b8401
SS
941 * The process is stopped due to some condition, usually because p_stat is
942 * set to SSTOP, but also possibly due to being traced.
fc17ad60 943 *
164b8401 944 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
344ad853
MD
945 * because the parent may check the child's status before the child actually
946 * gets to this routine.
947 *
9a379a4a 948 * This routine is called with the current lwp only, typically just
344ad853
MD
949 * before returning to userland.
950 *
9a379a4a 951 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
344ad853 952 * SIGCONT to break out of the tsleep.
984263bc
MD
953 */
954void
9a379a4a 955tstop(void)
984263bc 956{
9a379a4a 957 struct lwp *lp = curthread->td_lwp;
7278a846 958 struct proc *p = lp->lwp_proc;
9a379a4a 959
7278a846 960 crit_enter();
f33e8653
SS
961 /*
962 * If LWP_WSTOP is set, we were sleeping
963 * while our process was stopped. At this point
964 * we were already counted as stopped.
965 */
966 if ((lp->lwp_flag & LWP_WSTOP) == 0) {
967 /*
968 * If we're the last thread to stop, signal
969 * our parent.
970 */
971 p->p_nstopped++;
972 lp->lwp_flag |= LWP_WSTOP;
ea59a697 973 wakeup(&p->p_nstopped);
f33e8653
SS
974 if (p->p_nstopped == p->p_nthreads) {
975 p->p_flag &= ~P_WAITED;
976 wakeup(p->p_pptr);
977 if ((p->p_pptr->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0)
978 ksignal(p->p_pptr, SIGCHLD);
979 }
980 }
ea59a697
SS
981 while (p->p_stat == SSTOP) {
982 lp->lwp_flag |= LWP_BREAKTSLEEP;
983 lp->lwp_stat = LSSTOP;
984 tsleep(p, 0, "stop", 0);
985 }
7278a846 986 p->p_nstopped--;
a5ff9d37 987 lp->lwp_flag &= ~LWP_WSTOP;
7278a846 988 crit_exit();
26a0694b
MD
989}
990
a77ac49d
MD
991/*
992 * Yield / synchronous reschedule. This is a bit tricky because the trap
993 * code might have set a lazy release on the switch function. Setting
994 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
995 * switch, and that we are given a greater chance of affinity with our
996 * current cpu.
997 *
998 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
999 * run queue. lwkt_switch() will also execute any assigned passive release
1000 * (which usually calls release_curproc()), allowing a same/higher priority
1001 * process to be designated as the current process.
1002 *
1003 * While it is possible for a lower priority process to be designated,
1004 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
1005 * round-robin back to us and we will be able to re-acquire the current
1006 * process designation.
1007 */
1008void
1009uio_yield(void)
1010{
1011 struct thread *td = curthread;
1012 struct proc *p = td->td_proc;
1013
1014 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
1015 if (p) {
1016 p->p_flag |= P_PASSIVE_ACQ;
1017 lwkt_switch();
1018 p->p_flag &= ~P_PASSIVE_ACQ;
1019 } else {
1020 lwkt_switch();
1021 }
1022}
1023
984263bc
MD
1024/*
1025 * Compute a tenex style load average of a quantity on
1026 * 1, 5 and 15 minute intervals.
1027 */
c7e98b2f 1028static int loadav_count_runnable(struct lwp *p, void *data);
8fa76237 1029
984263bc
MD
1030static void
1031loadav(void *arg)
1032{
984263bc 1033 struct loadavg *avg;
8fa76237 1034 int i, nrun;
984263bc 1035
984263bc 1036 nrun = 0;
c7e98b2f 1037 alllwp_scan(loadav_count_runnable, &nrun);
8fa76237
MD
1038 avg = &averunnable;
1039 for (i = 0; i < 3; i++) {
984263bc
MD
1040 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
1041 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
8fa76237 1042 }
984263bc
MD
1043
1044 /*
1045 * Schedule the next update to occur after 5 seconds, but add a
1046 * random variation to avoid synchronisation with processes that
1047 * run at regular intervals.
1048 */
cddfb7bb 1049 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
8fa76237
MD
1050 loadav, NULL);
1051}
1052
1053static int
c7e98b2f 1054loadav_count_runnable(struct lwp *lp, void *data)
8fa76237
MD
1055{
1056 int *nrunp = data;
1057 thread_t td;
1058
164b8401
SS
1059 switch (lp->lwp_stat) {
1060 case LSRUN:
08f2f1bb 1061 if ((td = lp->lwp_thread) == NULL)
8fa76237
MD
1062 break;
1063 if (td->td_flags & TDF_BLOCKED)
1064 break;
8fa76237
MD
1065 ++*nrunp;
1066 break;
1067 default:
1068 break;
1069 }
1070 return(0);
984263bc
MD
1071}
1072
1073/* ARGSUSED */
1074static void
6656cd91 1075sched_setup(void *dummy)
984263bc 1076{
984263bc 1077 callout_init(&loadav_callout);
35f9d051 1078 callout_init(&schedcpu_callout);
984263bc
MD
1079
1080 /* Kick off timeout driven events by calling first time. */
984263bc
MD
1081 schedcpu(NULL);
1082 loadav(NULL);
1083}
1084