Make setrunnable() behave more like it is intended.
[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 $
2daf83b0 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.78 2007/02/22 15:49:08 corecode 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
MD
54#ifdef KTRACE
55#include <sys/uio.h>
56#include <sys/ktrace.h>
57#endif
f1d1c3fa 58#include <sys/xwait.h>
9afb0ffd 59#include <sys/ktr.h>
984263bc 60
bf765287
MD
61#include <sys/thread2.h>
62#include <sys/spinlock2.h>
63
984263bc 64#include <machine/cpu.h>
984263bc
MD
65#include <machine/smp.h>
66
fc17ad60
MD
67TAILQ_HEAD(tslpque, thread);
68
402ed7e1 69static void sched_setup (void *dummy);
984263bc
MD
70SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
71
984263bc
MD
72int hogticks;
73int lbolt;
344ad853 74int lbolt_syncer;
984263bc 75int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
17a9f566 76int ncpus;
90100055 77int ncpus2, ncpus2_shift, ncpus2_mask;
e43a034f 78int safepri;
984263bc
MD
79
80static struct callout loadav_callout;
35f9d051 81static struct callout schedcpu_callout;
fc17ad60 82MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
984263bc 83
9afb0ffd
MD
84#if !defined(KTR_TSLEEP)
85#define KTR_TSLEEP KTR_ALL
86#endif
87KTR_INFO_MASTER(tsleep);
88KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter", 0);
89KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0);
90KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0);
91KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0);
92#define logtsleep(name) KTR_LOG(tsleep_ ## name)
93
984263bc
MD
94struct loadavg averunnable =
95 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
96/*
97 * Constants for averages over 1, 5, and 15 minutes
98 * when sampling at 5 second intervals.
99 */
100static fixpt_t cexp[3] = {
101 0.9200444146293232 * FSCALE, /* exp(-1/12) */
102 0.9834714538216174 * FSCALE, /* exp(-1/60) */
103 0.9944598480048967 * FSCALE, /* exp(-1/180) */
104};
105
402ed7e1 106static void endtsleep (void *);
344ad853 107static void unsleep_and_wakeup_thread(struct thread *td);
402ed7e1 108static void loadav (void *arg);
402ed7e1 109static void schedcpu (void *arg);
984263bc 110
a46fac56
MD
111/*
112 * Adjust the scheduler quantum. The quantum is specified in microseconds.
113 * Note that 'tick' is in microseconds per tick.
114 */
984263bc
MD
115static int
116sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
117{
118 int error, new_val;
119
120 new_val = sched_quantum * tick;
121 error = sysctl_handle_int(oidp, &new_val, 0, req);
122 if (error != 0 || req->newptr == NULL)
123 return (error);
124 if (new_val < tick)
125 return (EINVAL);
126 sched_quantum = new_val / tick;
127 hogticks = 2 * sched_quantum;
128 return (0);
129}
130
131SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
132 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
133
984263bc
MD
134/*
135 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
136 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
137 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
138 *
139 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
dcc99b62 140 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
984263bc
MD
141 *
142 * If you don't want to bother with the faster/more-accurate formula, you
143 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
144 * (more general) method of calculating the %age of CPU used by a process.
dcc99b62 145 *
08f2f1bb 146 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
dcc99b62
MD
147 */
148#define CCPU_SHIFT 11
149
150static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
151SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
152
153/*
154 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
984263bc 155 */
460426e6 156int fscale __unused = FSCALE; /* exported to systat */
dcc99b62 157SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
984263bc
MD
158
159/*
0a3f9b47 160 * Recompute process priorities, once a second.
dcc99b62
MD
161 *
162 * Since the userland schedulers are typically event oriented, if the
163 * estcpu calculation at wakeup() time is not sufficient to make a
164 * process runnable relative to other processes in the system we have
165 * a 1-second recalc to help out.
166 *
167 * This code also allows us to store sysclock_t data in the process structure
168 * without fear of an overrun, since sysclock_t are guarenteed to hold
169 * several seconds worth of count.
8fa76237
MD
170 *
171 * WARNING! callouts can preempt normal threads. However, they will not
172 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
984263bc 173 */
8fa76237
MD
174static int schedcpu_stats(struct proc *p, void *data __unused);
175static int schedcpu_resource(struct proc *p, void *data __unused);
176
984263bc 177static void
26a0694b 178schedcpu(void *arg)
984263bc 179{
8fa76237
MD
180 allproc_scan(schedcpu_stats, NULL);
181 allproc_scan(schedcpu_resource, NULL);
182 wakeup((caddr_t)&lbolt);
183 wakeup((caddr_t)&lbolt_syncer);
184 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
185}
186
187/*
188 * General process statistics once a second
189 */
190static int
191schedcpu_stats(struct proc *p, void *data __unused)
192{
08f2f1bb
SS
193 struct lwp *lp;
194
8fa76237
MD
195 crit_enter();
196 p->p_swtime++;
c7e98b2f
SS
197 FOREACH_LWP_IN_PROC(lp, p) {
198 if (lp->lwp_stat == LSSLEEP)
199 lp->lwp_slptime++;
4b5f931b 200
c7e98b2f
SS
201 /*
202 * Only recalculate processes that are active or have slept
203 * less then 2 seconds. The schedulers understand this.
204 */
205 if (lp->lwp_slptime <= 1) {
206 p->p_usched->recalculate(lp);
207 } else {
208 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT;
209 }
8fa76237
MD
210 }
211 crit_exit();
212 return(0);
213}
a46fac56 214
8fa76237 215/*
84204577 216 * Resource checks. XXX break out since ksignal/killproc can block,
8fa76237
MD
217 * limiting us to one process killed per second. There is probably
218 * a better way.
219 */
220static int
221schedcpu_resource(struct proc *p, void *data __unused)
222{
223 u_int64_t ttime;
08f2f1bb 224 struct lwp *lp;
8fa76237
MD
225
226 crit_enter();
227 if (p->p_stat == SIDL ||
416d05d7 228 p->p_stat == SZOMB ||
c7e98b2f 229 p->p_limit == NULL
8fa76237 230 ) {
e43a034f 231 crit_exit();
8fa76237 232 return(0);
984263bc 233 }
344ad853 234
c7e98b2f
SS
235 ttime = 0;
236 FOREACH_LWP_IN_PROC(lp, p) {
237 ttime += lp->lwp_thread->td_sticks;
238 ttime += lp->lwp_thread->td_uticks;
239 }
8fa76237
MD
240
241 switch(plimit_testcpulimit(p->p_limit, ttime)) {
242 case PLIMIT_TESTCPU_KILL:
243 killproc(p, "exceeded maximum CPU limit");
244 break;
245 case PLIMIT_TESTCPU_XCPU:
246 if ((p->p_flag & P_XCPU) == 0) {
247 p->p_flag |= P_XCPU;
84204577 248 ksignal(p, SIGXCPU);
344ad853 249 }
8fa76237
MD
250 break;
251 default:
c0b8a06d 252 break;
344ad853 253 }
8fa76237
MD
254 crit_exit();
255 return(0);
984263bc
MD
256}
257
258/*
dcc99b62
MD
259 * This is only used by ps. Generate a cpu percentage use over
260 * a period of one second.
52eedfb5
MD
261 *
262 * MPSAFE
984263bc 263 */
dcc99b62 264void
553ea3c8 265updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
984263bc 266{
dcc99b62
MD
267 fixpt_t acc;
268 int remticks;
269
270 acc = (cpticks << FSHIFT) / ttlticks;
271 if (ttlticks >= ESTCPUFREQ) {
553ea3c8 272 lp->lwp_pctcpu = acc;
dcc99b62
MD
273 } else {
274 remticks = ESTCPUFREQ - ttlticks;
553ea3c8 275 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
dcc99b62 276 ESTCPUFREQ;
a46fac56 277 }
984263bc
MD
278}
279
280/*
281 * We're only looking at 7 bits of the address; everything is
282 * aligned to 4, lots of things are aligned to greater powers
283 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
284 */
285#define TABLESIZE 128
984263bc
MD
286#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
287
fc17ad60
MD
288static cpumask_t slpque_cpumasks[TABLESIZE];
289
984263bc 290/*
a46fac56 291 * General scheduler initialization. We force a reschedule 25 times
fc17ad60
MD
292 * a second by default. Note that cpu0 is initialized in early boot and
293 * cannot make any high level calls.
294 *
295 * Each cpu has its own sleep queue.
984263bc 296 */
984263bc 297void
fc17ad60 298sleep_gdinit(globaldata_t gd)
984263bc 299{
fc17ad60 300 static struct tslpque slpque_cpu0[TABLESIZE];
9c1fad94 301 int i;
984263bc 302
fc17ad60
MD
303 if (gd->gd_cpuid == 0) {
304 sched_quantum = (hz + 24) / 25;
305 hogticks = 2 * sched_quantum;
306
307 gd->gd_tsleep_hash = slpque_cpu0;
308 } else {
77652cad 309 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0),
fc17ad60
MD
310 M_TSLEEP, M_WAITOK | M_ZERO);
311 }
312 for (i = 0; i < TABLESIZE; ++i)
313 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
984263bc
MD
314}
315
316/*
317 * General sleep call. Suspends the current process until a wakeup is
318 * performed on the specified identifier. The process will then be made
319 * runnable with the specified priority. Sleeps at most timo/hz seconds
377d4740 320 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
984263bc
MD
321 * before and after sleeping, else signals are not checked. Returns 0 if
322 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
323 * signal needs to be delivered, ERESTART is returned if the current system
324 * call should be restarted if possible, and EINTR is returned if the system
325 * call should be interrupted by the signal (return EINTR).
26a0694b 326 *
0a3f9b47
MD
327 * Note that if we are a process, we release_curproc() before messing with
328 * the LWKT scheduler.
a46fac56
MD
329 *
330 * During autoconfiguration or after a panic, a sleep will simply
331 * lower the priority briefly to allow interrupts, then return.
984263bc
MD
332 */
333int
6656cd91 334tsleep(void *ident, int flags, const char *wmesg, int timo)
984263bc 335{
dadab5e9 336 struct thread *td = curthread;
08f2f1bb 337 struct lwp *lp = td->td_lwp;
0cfcada1 338 struct proc *p = td->td_proc; /* may be NULL */
fc17ad60 339 globaldata_t gd;
344ad853
MD
340 int sig;
341 int catch;
342 int id;
343 int error;
e43a034f 344 int oldpri;
076fecef 345 struct callout thandle;
984263bc 346
0cfcada1
MD
347 /*
348 * NOTE: removed KTRPOINT, it could cause races due to blocking
349 * even in stable. Just scrap it for now.
350 */
984263bc
MD
351 if (cold || panicstr) {
352 /*
353 * After a panic, or during autoconfiguration,
354 * just give interrupts a chance, then just return;
355 * don't run any other procs or panic below,
356 * in case this is the idle process and already asleep.
357 */
e43a034f
MD
358 splz();
359 oldpri = td->td_pri & TDPRI_MASK;
360 lwkt_setpri_self(safepri);
361 lwkt_switch();
362 lwkt_setpri_self(oldpri);
984263bc
MD
363 return (0);
364 }
9afb0ffd 365 logtsleep(tsleep_beg);
fc17ad60
MD
366 gd = td->td_gd;
367 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
344ad853
MD
368
369 /*
370 * NOTE: all of this occurs on the current cpu, including any
371 * callout-based wakeups, so a critical section is a sufficient
372 * interlock.
373 *
374 * The entire sequence through to where we actually sleep must
375 * run without breaking the critical section.
376 */
377 id = LOOKUP(ident);
378 catch = flags & PCATCH;
379 error = 0;
380 sig = 0;
381
37af14fe 382 crit_enter_quick(td);
344ad853 383
0cfcada1 384 KASSERT(ident != NULL, ("tsleep: no ident"));
164b8401
SS
385 KASSERT(lp == NULL || lp->lwp_stat == LSRUN, ("tsleep %p %s %d",
386 ident, wmesg, lp->lwp_stat));
0cfcada1 387
344ad853
MD
388 /*
389 * Setup for the current process (if this is a process).
390 */
08f2f1bb 391 if (lp) {
344ad853
MD
392 if (catch) {
393 /*
394 * Early termination if PCATCH was set and a
395 * signal is pending, interlocked with the
396 * critical section.
397 *
398 * Early termination only occurs when tsleep() is
164b8401 399 * entered while in a normal LSRUN state.
344ad853 400 */
08f2f1bb 401 if ((sig = CURSIG(lp)) != 0)
344ad853
MD
402 goto resume;
403
7c1212ec
MD
404 /*
405 * Early termination if PCATCH was set and a
406 * mailbox signal was possibly delivered prior to
407 * the system call even being made, in order to
408 * allow the user to interlock without having to
409 * make additional system calls.
410 */
411 if (p->p_flag & P_MAILBOX)
412 goto resume;
413
344ad853 414 /*
84204577 415 * Causes ksignal to wake us up when.
344ad853 416 */
9a379a4a 417 lp->lwp_flag |= LWP_SINTR;
344ad853
MD
418 }
419
420 /*
421 * Make sure the current process has been untangled from
422 * the userland scheduler and initialize slptime to start
423 * counting.
424 */
0a3f9b47
MD
425 if (flags & PNORESCHED)
426 td->td_flags |= TDF_NORESCHED;
08f2f1bb
SS
427 p->p_usched->release_curproc(lp);
428 lp->lwp_slptime = 0;
0a3f9b47 429 }
fc17ad60
MD
430
431 /*
344ad853 432 * Move our thread to the correct queue and setup our wchan, etc.
fc17ad60 433 */
37af14fe 434 lwkt_deschedule_self(td);
344ad853 435 td->td_flags |= TDF_TSLEEPQ;
fc17ad60
MD
436 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
437 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
344ad853
MD
438
439 td->td_wchan = ident;
440 td->td_wmesg = wmesg;
441 td->td_wdomain = flags & PDOMAIN_MASK;
442
443 /*
444 * Setup the timeout, if any
445 */
076fecef
MD
446 if (timo) {
447 callout_init(&thandle);
448 callout_reset(&thandle, timo, endtsleep, td);
449 }
344ad853 450
984263bc 451 /*
344ad853 452 * Beddy bye bye.
984263bc 453 */
08f2f1bb 454 if (lp) {
26a0694b 455 /*
52eedfb5 456 * Ok, we are sleeping. Place us in the SSLEEP state.
26a0694b 457 */
9388413d 458 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
164b8401 459 lp->lwp_stat = LSSLEEP;
08f2f1bb 460 lp->lwp_ru.ru_nvcsw++;
344ad853 461 lwkt_switch();
ab554892
MD
462
463 /*
164b8401 464 * And when we are woken up, put us back in LSRUN. If we
ab554892
MD
465 * slept for over a second, recalculate our estcpu.
466 */
164b8401 467 lp->lwp_stat = LSRUN;
08f2f1bb
SS
468 if (lp->lwp_slptime)
469 p->p_usched->recalculate(lp);
470 lp->lwp_slptime = 0;
0cfcada1
MD
471 } else {
472 lwkt_switch();
473 }
344ad853 474
fc17ad60
MD
475 /*
476 * Make sure we haven't switched cpus while we were asleep. It's
344ad853 477 * not supposed to happen. Cleanup our temporary flags.
fc17ad60
MD
478 */
479 KKASSERT(gd == td->td_gd);
0a3f9b47 480 td->td_flags &= ~TDF_NORESCHED;
344ad853
MD
481
482 /*
483 * Cleanup the timeout.
484 */
485 if (timo) {
486 if (td->td_flags & TDF_TIMEOUT) {
487 td->td_flags &= ~TDF_TIMEOUT;
488 if (sig == 0)
489 error = EWOULDBLOCK;
490 } else {
491 callout_stop(&thandle);
492 }
0cfcada1 493 }
344ad853
MD
494
495 /*
496 * Since td_threadq is used both for our run queue AND for the
497 * tsleep hash queue, we can't still be on it at this point because
498 * we've gotten cpu back.
499 */
afbfc034 500 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
344ad853
MD
501 td->td_wchan = NULL;
502 td->td_wmesg = NULL;
503 td->td_wdomain = 0;
504
505 /*
7c1212ec
MD
506 * Figure out the correct error return. If interrupted by a
507 * signal we want to return EINTR or ERESTART.
508 *
509 * If P_MAILBOX is set no automatic system call restart occurs
510 * and we return EINTR. P_MAILBOX is meant to be used as an
511 * interlock, the user must poll it prior to any system call
512 * that it wishes to interlock a mailbox signal against since
513 * the flag is cleared on *any* system call that sleeps.
344ad853
MD
514 */
515resume:
0cfcada1 516 if (p) {
7c1212ec
MD
517 if (catch && error == 0) {
518 if ((p->p_flag & P_MAILBOX) && sig == 0) {
344ad853 519 error = EINTR;
08f2f1bb 520 } else if (sig != 0 || (sig = CURSIG(lp))) {
7c1212ec
MD
521 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
522 error = EINTR;
523 else
524 error = ERESTART;
525 }
984263bc 526 }
9a379a4a
SS
527 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR);
528 p->p_flag &= ~P_MAILBOX;
984263bc 529 }
9afb0ffd 530 logtsleep(tsleep_end);
344ad853
MD
531 crit_exit_quick(td);
532 return (error);
984263bc
MD
533}
534
16523a43
MD
535/*
536 * This is a dandy function that allows us to interlock tsleep/wakeup
537 * operations with unspecified upper level locks, such as lockmgr locks,
538 * simply by holding a critical section. The sequence is:
539 *
540 * (enter critical section)
541 * (acquire upper level lock)
542 * tsleep_interlock(blah)
543 * (release upper level lock)
544 * tsleep(blah, ...)
545 * (exit critical section)
546 *
547 * Basically this function sets our cpumask for the ident which informs
548 * other cpus that our cpu 'might' be waiting (or about to wait on) the
549 * hash index related to the ident. The critical section prevents another
550 * cpu's wakeup() from being processed on our cpu until we are actually
551 * able to enter the tsleep(). Thus, no race occurs between our attempt
552 * to release a resource and sleep, and another cpu's attempt to acquire
553 * a resource and call wakeup.
554 *
555 * There isn't much of a point to this function unless you call it while
556 * holding a critical section.
557 */
bf765287
MD
558static __inline void
559_tsleep_interlock(globaldata_t gd, void *ident)
560{
561 int id = LOOKUP(ident);
562
563 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
564}
565
16523a43
MD
566void
567tsleep_interlock(void *ident)
568{
bf765287
MD
569 _tsleep_interlock(mycpu, ident);
570}
571
572/*
573 * Interlocked spinlock sleep. An exclusively held spinlock must
574 * be passed to msleep(). The function will atomically release the
575 * spinlock and tsleep on the ident, then reacquire the spinlock and
576 * return.
577 *
578 * This routine is fairly important along the critical path, so optimize it
579 * heavily.
580 */
581int
582msleep(void *ident, struct spinlock *spin, int flags,
583 const char *wmesg, int timo)
584{
585 globaldata_t gd = mycpu;
586 int error;
16523a43 587
bf765287
MD
588 crit_enter_gd(gd);
589 _tsleep_interlock(gd, ident);
590 spin_unlock_wr_quick(gd, spin);
591 error = tsleep(ident, flags, wmesg, timo);
592 spin_lock_wr_quick(gd, spin);
593 crit_exit_gd(gd);
594
595 return (error);
16523a43
MD
596}
597
984263bc 598/*
344ad853 599 * Implement the timeout for tsleep.
fc17ad60 600 *
9a379a4a 601 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but
344ad853
MD
602 * we only call setrunnable if the process is not stopped.
603 *
604 * This type of callout timeout is scheduled on the same cpu the process
605 * is sleeping on. Also, at the moment, the MP lock is held.
984263bc
MD
606 */
607static void
0cfcada1 608endtsleep(void *arg)
984263bc 609{
0cfcada1 610 thread_t td = arg;
9a379a4a 611 struct lwp *lp;
984263bc 612
344ad853 613 ASSERT_MP_LOCK_HELD(curthread);
37af14fe 614 crit_enter();
344ad853
MD
615
616 /*
617 * cpu interlock. Thread flags are only manipulated on
618 * the cpu owning the thread. proc flags are only manipulated
619 * by the older of the MP lock. We have both.
620 */
621 if (td->td_flags & TDF_TSLEEPQ) {
0cfcada1 622 td->td_flags |= TDF_TIMEOUT;
344ad853 623
9a379a4a
SS
624 if ((lp = td->td_lwp) != NULL) {
625 lp->lwp_flag |= LWP_BREAKTSLEEP;
626 if (lp->lwp_proc->p_stat != SSTOP)
627 setrunnable(lp);
0cfcada1 628 } else {
344ad853 629 unsleep_and_wakeup_thread(td);
0cfcada1 630 }
984263bc 631 }
37af14fe 632 crit_exit();
984263bc
MD
633}
634
984263bc 635/*
344ad853
MD
636 * Unsleep and wakeup a thread. This function runs without the MP lock
637 * which means that it can only manipulate thread state on the owning cpu,
638 * and cannot touch the process state at all.
984263bc 639 */
344ad853 640static
8fb8bca6 641void
344ad853 642unsleep_and_wakeup_thread(struct thread *td)
8fb8bca6 643{
344ad853 644 globaldata_t gd = mycpu;
fc17ad60
MD
645 int id;
646
344ad853
MD
647#ifdef SMP
648 if (td->td_gd != gd) {
649 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
650 return;
651 }
652#endif
9c1fad94 653 crit_enter();
344ad853
MD
654 if (td->td_flags & TDF_TSLEEPQ) {
655 td->td_flags &= ~TDF_TSLEEPQ;
656 id = LOOKUP(td->td_wchan);
657 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
658 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
659 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
660 lwkt_schedule(td);
8fb8bca6 661 }
9c1fad94 662 crit_exit();
8fb8bca6 663}
8fb8bca6
EN
664
665/*
666 * Make all processes sleeping on the specified identifier runnable.
fc17ad60
MD
667 * count may be zero or one only.
668 *
669 * The domain encodes the sleep/wakeup domain AND the first cpu to check
670 * (which is always the current cpu). As we iterate across cpus
344ad853
MD
671 *
672 * This call may run without the MP lock held. We can only manipulate thread
673 * state on the cpu owning the thread. We CANNOT manipulate process state
674 * at all.
8fb8bca6
EN
675 */
676static void
fc17ad60 677_wakeup(void *ident, int domain)
984263bc 678{
fc17ad60 679 struct tslpque *qp;
0cfcada1
MD
680 struct thread *td;
681 struct thread *ntd;
fc17ad60 682 globaldata_t gd;
fc17ad60
MD
683#ifdef SMP
684 cpumask_t mask;
685 cpumask_t tmask;
686 int startcpu;
687 int nextcpu;
688#endif
689 int id;
984263bc 690
37af14fe 691 crit_enter();
9afb0ffd 692 logtsleep(wakeup_beg);
fc17ad60
MD
693 gd = mycpu;
694 id = LOOKUP(ident);
695 qp = &gd->gd_tsleep_hash[id];
984263bc 696restart:
0cfcada1
MD
697 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
698 ntd = TAILQ_NEXT(td, td_threadq);
fc17ad60
MD
699 if (td->td_wchan == ident &&
700 td->td_wdomain == (domain & PDOMAIN_MASK)
701 ) {
344ad853
MD
702 KKASSERT(td->td_flags & TDF_TSLEEPQ);
703 td->td_flags &= ~TDF_TSLEEPQ;
0cfcada1 704 TAILQ_REMOVE(qp, td, td_threadq);
fc17ad60
MD
705 if (TAILQ_FIRST(qp) == NULL) {
706 atomic_clear_int(&slpque_cpumasks[id],
707 gd->gd_cpumask);
708 }
344ad853 709 lwkt_schedule(td);
fc17ad60
MD
710 if (domain & PWAKEUP_ONE)
711 goto done;
0cfcada1 712 goto restart;
984263bc
MD
713 }
714 }
fc17ad60
MD
715
716#ifdef SMP
717 /*
718 * We finished checking the current cpu but there still may be
719 * more work to do. Either wakeup_one was requested and no matching
720 * thread was found, or a normal wakeup was requested and we have
721 * to continue checking cpus.
722 *
723 * The cpu that started the wakeup sequence is encoded in the domain.
724 * We use this information to determine which cpus still need to be
725 * checked, locate a candidate cpu, and chain the wakeup
726 * asynchronously with an IPI message.
727 *
728 * It should be noted that this scheme is actually less expensive then
729 * the old scheme when waking up multiple threads, since we send
730 * only one IPI message per target candidate which may then schedule
731 * multiple threads. Before we could have wound up sending an IPI
732 * message for each thread on the target cpu (!= current cpu) that
733 * needed to be woken up.
734 *
735 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
736 * should be ok since we are passing idents in the IPI rather then
737 * thread pointers.
738 */
b336a9b1
MD
739 if ((domain & PWAKEUP_MYCPU) == 0 &&
740 (mask = slpque_cpumasks[id]) != 0
741 ) {
fc17ad60
MD
742 /*
743 * Look for a cpu that might have work to do. Mask out cpus
744 * which have already been processed.
745 *
746 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
747 * ^ ^ ^
748 * start currentcpu start
749 * case2 case1
750 * * * *
751 * 11111111111111110000000000000111 case1
752 * 00000000111111110000000000000000 case2
753 *
754 * case1: We started at start_case1 and processed through
755 * to the current cpu. We have to check any bits
756 * after the current cpu, then check bits before
757 * the starting cpu.
758 *
759 * case2: We have already checked all the bits from
760 * start_case2 to the end, and from 0 to the current
761 * cpu. We just have the bits from the current cpu
762 * to start_case2 left to check.
763 */
764 startcpu = PWAKEUP_DECODE(domain);
765 if (gd->gd_cpuid >= startcpu) {
766 /*
767 * CASE1
768 */
769 tmask = mask & ~((gd->gd_cpumask << 1) - 1);
770 if (mask & tmask) {
771 nextcpu = bsfl(mask & tmask);
772 lwkt_send_ipiq2(globaldata_find(nextcpu),
773 _wakeup, ident, domain);
774 } else {
775 tmask = (1 << startcpu) - 1;
776 if (mask & tmask) {
777 nextcpu = bsfl(mask & tmask);
778 lwkt_send_ipiq2(
779 globaldata_find(nextcpu),
780 _wakeup, ident, domain);
781 }
782 }
783 } else {
784 /*
785 * CASE2
786 */
787 tmask = ~((gd->gd_cpumask << 1) - 1) &
788 ((1 << startcpu) - 1);
789 if (mask & tmask) {
790 nextcpu = bsfl(mask & tmask);
791 lwkt_send_ipiq2(globaldata_find(nextcpu),
792 _wakeup, ident, domain);
793 }
794 }
795 }
796#endif
797done:
9afb0ffd 798 logtsleep(wakeup_end);
37af14fe 799 crit_exit();
984263bc
MD
800}
801
b336a9b1
MD
802/*
803 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
804 */
984263bc 805void
0cfcada1 806wakeup(void *ident)
984263bc 807{
fc17ad60 808 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
0cfcada1 809}
984263bc 810
b336a9b1
MD
811/*
812 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
813 */
0cfcada1
MD
814void
815wakeup_one(void *ident)
816{
fc17ad60
MD
817 /* XXX potentially round-robin the first responding cpu */
818 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
da5fb9ef
MD
819}
820
b336a9b1
MD
821/*
822 * Wakeup threads tsleep()ing on the specified ident on the current cpu
823 * only.
824 */
825void
826wakeup_mycpu(void *ident)
827{
828 _wakeup(ident, PWAKEUP_MYCPU);
829}
830
831/*
832 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
833 * only.
834 */
835void
836wakeup_mycpu_one(void *ident)
837{
838 /* XXX potentially round-robin the first responding cpu */
839 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
840}
841
842/*
843 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
844 * only.
845 */
846void
847wakeup_oncpu(globaldata_t gd, void *ident)
848{
1699d292 849#ifdef SMP
b336a9b1
MD
850 if (gd == mycpu) {
851 _wakeup(ident, PWAKEUP_MYCPU);
852 } else {
853 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
854 }
1699d292
MD
855#else
856 _wakeup(ident, PWAKEUP_MYCPU);
857#endif
b336a9b1
MD
858}
859
860/*
861 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
862 * only.
863 */
864void
865wakeup_oncpu_one(globaldata_t gd, void *ident)
866{
1699d292 867#ifdef SMP
b336a9b1
MD
868 if (gd == mycpu) {
869 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
870 } else {
871 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
872 }
1699d292
MD
873#else
874 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
875#endif
b336a9b1
MD
876}
877
878/*
879 * Wakeup all threads waiting on the specified ident that slept using
880 * the specified domain, on all cpus.
881 */
da5fb9ef
MD
882void
883wakeup_domain(void *ident, int domain)
884{
fc17ad60 885 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
da5fb9ef
MD
886}
887
b336a9b1
MD
888/*
889 * Wakeup one thread waiting on the specified ident that slept using
890 * the specified domain, on any cpu.
891 */
da5fb9ef
MD
892void
893wakeup_domain_one(void *ident, int domain)
894{
fc17ad60
MD
895 /* XXX potentially round-robin the first responding cpu */
896 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
984263bc
MD
897}
898
899/*
344ad853
MD
900 * setrunnable()
901 *
902 * Make a process runnable. The MP lock must be held on call. This only
903 * has an effect if we are in SSLEEP. We only break out of the
9a379a4a 904 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state.
37af14fe 905 *
344ad853
MD
906 * NOTE: With the MP lock held we can only safely manipulate the process
907 * structure. We cannot safely manipulate the thread structure.
984263bc
MD
908 */
909void
9a379a4a 910setrunnable(struct lwp *lp)
984263bc 911{
344ad853
MD
912 crit_enter();
913 ASSERT_MP_LOCK_HELD(curthread);
2daf83b0
SS
914 if (lp->lwp_stat == LSSTOP)
915 lp->lwp_stat = LSSLEEP;
916 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP))
08f2f1bb 917 unsleep_and_wakeup_thread(lp->lwp_thread);
344ad853 918 crit_exit();
984263bc
MD
919}
920
921/*
164b8401
SS
922 * The process is stopped due to some condition, usually because p_stat is
923 * set to SSTOP, but also possibly due to being traced.
fc17ad60 924 *
164b8401 925 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED
344ad853
MD
926 * because the parent may check the child's status before the child actually
927 * gets to this routine.
928 *
9a379a4a 929 * This routine is called with the current lwp only, typically just
344ad853
MD
930 * before returning to userland.
931 *
9a379a4a 932 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive
344ad853 933 * SIGCONT to break out of the tsleep.
984263bc
MD
934 */
935void
9a379a4a 936tstop(void)
984263bc 937{
9a379a4a
SS
938 struct lwp *lp = curthread->td_lwp;
939
940 lp->lwp_flag |= LWP_BREAKTSLEEP;
941 tsleep(lp->lwp_proc, 0, "stop", 0);
26a0694b
MD
942}
943
a77ac49d
MD
944/*
945 * Yield / synchronous reschedule. This is a bit tricky because the trap
946 * code might have set a lazy release on the switch function. Setting
947 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
948 * switch, and that we are given a greater chance of affinity with our
949 * current cpu.
950 *
951 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
952 * run queue. lwkt_switch() will also execute any assigned passive release
953 * (which usually calls release_curproc()), allowing a same/higher priority
954 * process to be designated as the current process.
955 *
956 * While it is possible for a lower priority process to be designated,
957 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
958 * round-robin back to us and we will be able to re-acquire the current
959 * process designation.
960 */
961void
962uio_yield(void)
963{
964 struct thread *td = curthread;
965 struct proc *p = td->td_proc;
966
967 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
968 if (p) {
969 p->p_flag |= P_PASSIVE_ACQ;
970 lwkt_switch();
971 p->p_flag &= ~P_PASSIVE_ACQ;
972 } else {
973 lwkt_switch();
974 }
975}
976
984263bc
MD
977/*
978 * Compute a tenex style load average of a quantity on
979 * 1, 5 and 15 minute intervals.
980 */
c7e98b2f 981static int loadav_count_runnable(struct lwp *p, void *data);
8fa76237 982
984263bc
MD
983static void
984loadav(void *arg)
985{
984263bc 986 struct loadavg *avg;
8fa76237 987 int i, nrun;
984263bc 988
984263bc 989 nrun = 0;
c7e98b2f 990 alllwp_scan(loadav_count_runnable, &nrun);
8fa76237
MD
991 avg = &averunnable;
992 for (i = 0; i < 3; i++) {
984263bc
MD
993 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
994 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
8fa76237 995 }
984263bc
MD
996
997 /*
998 * Schedule the next update to occur after 5 seconds, but add a
999 * random variation to avoid synchronisation with processes that
1000 * run at regular intervals.
1001 */
cddfb7bb 1002 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)),
8fa76237
MD
1003 loadav, NULL);
1004}
1005
1006static int
c7e98b2f 1007loadav_count_runnable(struct lwp *lp, void *data)
8fa76237
MD
1008{
1009 int *nrunp = data;
1010 thread_t td;
1011
164b8401
SS
1012 switch (lp->lwp_stat) {
1013 case LSRUN:
08f2f1bb 1014 if ((td = lp->lwp_thread) == NULL)
8fa76237
MD
1015 break;
1016 if (td->td_flags & TDF_BLOCKED)
1017 break;
8fa76237
MD
1018 ++*nrunp;
1019 break;
1020 default:
1021 break;
1022 }
1023 return(0);
984263bc
MD
1024}
1025
1026/* ARGSUSED */
1027static void
6656cd91 1028sched_setup(void *dummy)
984263bc 1029{
984263bc 1030 callout_init(&loadav_callout);
35f9d051 1031 callout_init(&schedcpu_callout);
984263bc
MD
1032
1033 /* Kick off timeout driven events by calling first time. */
984263bc
MD
1034 schedcpu(NULL);
1035 loadav(NULL);
1036}
1037