Rename functions to avoid conflicts with libc.
[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 $
ab554892 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.64 2006/07/11 01:01:50 dillon Exp $
984263bc
MD
41 */
42
43#include "opt_ktrace.h"
44
45#include <sys/param.h>
46#include <sys/systm.h>
47#include <sys/proc.h>
48#include <sys/kernel.h>
49#include <sys/signalvar.h>
50#include <sys/resourcevar.h>
51#include <sys/vmmeter.h>
52#include <sys/sysctl.h>
344ad853 53#include <sys/lock.h>
984263bc
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
MD
64#include <machine/cpu.h>
65#include <machine/ipl.h>
66#include <machine/smp.h>
67
fc17ad60
MD
68TAILQ_HEAD(tslpque, thread);
69
402ed7e1 70static void sched_setup (void *dummy);
984263bc
MD
71SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
72
984263bc
MD
73int hogticks;
74int lbolt;
344ad853 75int lbolt_syncer;
984263bc 76int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
17a9f566 77int ncpus;
90100055 78int ncpus2, ncpus2_shift, ncpus2_mask;
e43a034f 79int safepri;
984263bc
MD
80
81static struct callout loadav_callout;
35f9d051 82static struct callout schedcpu_callout;
fc17ad60 83MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues");
984263bc 84
9afb0ffd
MD
85#if !defined(KTR_TSLEEP)
86#define KTR_TSLEEP KTR_ALL
87#endif
88KTR_INFO_MASTER(tsleep);
89KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter", 0);
90KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 0, "tsleep exit", 0);
91KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 0, "wakeup enter", 0);
92KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 0, "wakeup exit", 0);
93#define logtsleep(name) KTR_LOG(tsleep_ ## name)
94
984263bc
MD
95struct loadavg averunnable =
96 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
97/*
98 * Constants for averages over 1, 5, and 15 minutes
99 * when sampling at 5 second intervals.
100 */
101static fixpt_t cexp[3] = {
102 0.9200444146293232 * FSCALE, /* exp(-1/12) */
103 0.9834714538216174 * FSCALE, /* exp(-1/60) */
104 0.9944598480048967 * FSCALE, /* exp(-1/180) */
105};
106
402ed7e1 107static void endtsleep (void *);
344ad853 108static void unsleep_and_wakeup_thread(struct thread *td);
402ed7e1 109static void loadav (void *arg);
402ed7e1 110static void schedcpu (void *arg);
984263bc 111
a46fac56
MD
112/*
113 * Adjust the scheduler quantum. The quantum is specified in microseconds.
114 * Note that 'tick' is in microseconds per tick.
115 */
984263bc
MD
116static int
117sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
118{
119 int error, new_val;
120
121 new_val = sched_quantum * tick;
122 error = sysctl_handle_int(oidp, &new_val, 0, req);
123 if (error != 0 || req->newptr == NULL)
124 return (error);
125 if (new_val < tick)
126 return (EINVAL);
127 sched_quantum = new_val / tick;
128 hogticks = 2 * sched_quantum;
129 return (0);
130}
131
132SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
133 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
134
984263bc
MD
135/*
136 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
137 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
138 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
139 *
140 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
dcc99b62 141 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
984263bc
MD
142 *
143 * If you don't want to bother with the faster/more-accurate formula, you
144 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
145 * (more general) method of calculating the %age of CPU used by a process.
dcc99b62
MD
146 *
147 * decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing
148 */
149#define CCPU_SHIFT 11
150
151static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
152SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
153
154/*
155 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale
984263bc 156 */
dcc99b62
MD
157static int fscale __unused = FSCALE;
158SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
984263bc
MD
159
160/*
0a3f9b47 161 * Recompute process priorities, once a second.
dcc99b62
MD
162 *
163 * Since the userland schedulers are typically event oriented, if the
164 * estcpu calculation at wakeup() time is not sufficient to make a
165 * process runnable relative to other processes in the system we have
166 * a 1-second recalc to help out.
167 *
168 * This code also allows us to store sysclock_t data in the process structure
169 * without fear of an overrun, since sysclock_t are guarenteed to hold
170 * several seconds worth of count.
8fa76237
MD
171 *
172 * WARNING! callouts can preempt normal threads. However, they will not
173 * preempt a thread holding a spinlock so we *can* safely use spinlocks.
984263bc 174 */
8fa76237
MD
175static int schedcpu_stats(struct proc *p, void *data __unused);
176static int schedcpu_resource(struct proc *p, void *data __unused);
177
984263bc 178static void
26a0694b 179schedcpu(void *arg)
984263bc 180{
8fa76237
MD
181 allproc_scan(schedcpu_stats, NULL);
182 allproc_scan(schedcpu_resource, NULL);
183 wakeup((caddr_t)&lbolt);
184 wakeup((caddr_t)&lbolt_syncer);
185 callout_reset(&schedcpu_callout, hz, schedcpu, NULL);
186}
187
188/*
189 * General process statistics once a second
190 */
191static int
192schedcpu_stats(struct proc *p, void *data __unused)
193{
194 crit_enter();
195 p->p_swtime++;
196 if (p->p_stat == SSLEEP)
197 p->p_slptime++;
4b5f931b 198
344ad853 199 /*
8fa76237
MD
200 * Only recalculate processes that are active or have slept
201 * less then 2 seconds. The schedulers understand this.
344ad853 202 */
8fa76237
MD
203 if (p->p_slptime <= 1) {
204 p->p_usched->recalculate(&p->p_lwp);
205 } else {
206 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
207 }
208 crit_exit();
209 return(0);
210}
a46fac56 211
8fa76237
MD
212/*
213 * Resource checks. XXX break out since psignal/killproc can block,
214 * limiting us to one process killed per second. There is probably
215 * a better way.
216 */
217static int
218schedcpu_resource(struct proc *p, void *data __unused)
219{
220 u_int64_t ttime;
221
222 crit_enter();
223 if (p->p_stat == SIDL ||
224 (p->p_flag & P_ZOMBIE) ||
225 p->p_limit == NULL ||
226 p->p_thread == NULL
227 ) {
e43a034f 228 crit_exit();
8fa76237 229 return(0);
984263bc 230 }
344ad853 231
8fa76237
MD
232 ttime = p->p_thread->td_sticks + p->p_thread->td_uticks;
233
234 switch(plimit_testcpulimit(p->p_limit, ttime)) {
235 case PLIMIT_TESTCPU_KILL:
236 killproc(p, "exceeded maximum CPU limit");
237 break;
238 case PLIMIT_TESTCPU_XCPU:
239 if ((p->p_flag & P_XCPU) == 0) {
240 p->p_flag |= P_XCPU;
241 psignal(p, SIGXCPU);
344ad853 242 }
8fa76237
MD
243 break;
244 default:
c0b8a06d 245 break;
344ad853 246 }
8fa76237
MD
247 crit_exit();
248 return(0);
984263bc
MD
249}
250
251/*
dcc99b62
MD
252 * This is only used by ps. Generate a cpu percentage use over
253 * a period of one second.
52eedfb5
MD
254 *
255 * MPSAFE
984263bc 256 */
dcc99b62 257void
553ea3c8 258updatepcpu(struct lwp *lp, int cpticks, int ttlticks)
984263bc 259{
dcc99b62
MD
260 fixpt_t acc;
261 int remticks;
262
263 acc = (cpticks << FSHIFT) / ttlticks;
264 if (ttlticks >= ESTCPUFREQ) {
553ea3c8 265 lp->lwp_pctcpu = acc;
dcc99b62
MD
266 } else {
267 remticks = ESTCPUFREQ - ttlticks;
553ea3c8 268 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) /
dcc99b62 269 ESTCPUFREQ;
a46fac56 270 }
984263bc
MD
271}
272
273/*
274 * We're only looking at 7 bits of the address; everything is
275 * aligned to 4, lots of things are aligned to greater powers
276 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
277 */
278#define TABLESIZE 128
984263bc
MD
279#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
280
fc17ad60
MD
281static cpumask_t slpque_cpumasks[TABLESIZE];
282
984263bc 283/*
a46fac56 284 * General scheduler initialization. We force a reschedule 25 times
fc17ad60
MD
285 * a second by default. Note that cpu0 is initialized in early boot and
286 * cannot make any high level calls.
287 *
288 * Each cpu has its own sleep queue.
984263bc 289 */
984263bc 290void
fc17ad60 291sleep_gdinit(globaldata_t gd)
984263bc 292{
fc17ad60 293 static struct tslpque slpque_cpu0[TABLESIZE];
9c1fad94 294 int i;
984263bc 295
fc17ad60
MD
296 if (gd->gd_cpuid == 0) {
297 sched_quantum = (hz + 24) / 25;
298 hogticks = 2 * sched_quantum;
299
300 gd->gd_tsleep_hash = slpque_cpu0;
301 } else {
302 gd->gd_tsleep_hash = malloc(sizeof(slpque_cpu0),
303 M_TSLEEP, M_WAITOK | M_ZERO);
304 }
305 for (i = 0; i < TABLESIZE; ++i)
306 TAILQ_INIT(&gd->gd_tsleep_hash[i]);
984263bc
MD
307}
308
309/*
310 * General sleep call. Suspends the current process until a wakeup is
311 * performed on the specified identifier. The process will then be made
312 * runnable with the specified priority. Sleeps at most timo/hz seconds
377d4740 313 * (0 means no timeout). If flags includes PCATCH flag, signals are checked
984263bc
MD
314 * before and after sleeping, else signals are not checked. Returns 0 if
315 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
316 * signal needs to be delivered, ERESTART is returned if the current system
317 * call should be restarted if possible, and EINTR is returned if the system
318 * call should be interrupted by the signal (return EINTR).
26a0694b 319 *
0a3f9b47
MD
320 * Note that if we are a process, we release_curproc() before messing with
321 * the LWKT scheduler.
a46fac56
MD
322 *
323 * During autoconfiguration or after a panic, a sleep will simply
324 * lower the priority briefly to allow interrupts, then return.
984263bc
MD
325 */
326int
6656cd91 327tsleep(void *ident, int flags, const char *wmesg, int timo)
984263bc 328{
dadab5e9 329 struct thread *td = curthread;
0cfcada1 330 struct proc *p = td->td_proc; /* may be NULL */
fc17ad60 331 globaldata_t gd;
344ad853
MD
332 int sig;
333 int catch;
334 int id;
335 int error;
e43a034f 336 int oldpri;
076fecef 337 struct callout thandle;
984263bc 338
0cfcada1
MD
339 /*
340 * NOTE: removed KTRPOINT, it could cause races due to blocking
341 * even in stable. Just scrap it for now.
342 */
984263bc
MD
343 if (cold || panicstr) {
344 /*
345 * After a panic, or during autoconfiguration,
346 * just give interrupts a chance, then just return;
347 * don't run any other procs or panic below,
348 * in case this is the idle process and already asleep.
349 */
e43a034f
MD
350 splz();
351 oldpri = td->td_pri & TDPRI_MASK;
352 lwkt_setpri_self(safepri);
353 lwkt_switch();
354 lwkt_setpri_self(oldpri);
984263bc
MD
355 return (0);
356 }
9afb0ffd 357 logtsleep(tsleep_beg);
fc17ad60
MD
358 gd = td->td_gd;
359 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */
344ad853
MD
360
361 /*
362 * NOTE: all of this occurs on the current cpu, including any
363 * callout-based wakeups, so a critical section is a sufficient
364 * interlock.
365 *
366 * The entire sequence through to where we actually sleep must
367 * run without breaking the critical section.
368 */
369 id = LOOKUP(ident);
370 catch = flags & PCATCH;
371 error = 0;
372 sig = 0;
373
37af14fe 374 crit_enter_quick(td);
344ad853 375
0cfcada1
MD
376 KASSERT(ident != NULL, ("tsleep: no ident"));
377 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
378 ident, wmesg, p->p_stat));
379
344ad853
MD
380 /*
381 * Setup for the current process (if this is a process).
382 */
0a3f9b47 383 if (p) {
344ad853
MD
384 if (catch) {
385 /*
386 * Early termination if PCATCH was set and a
387 * signal is pending, interlocked with the
388 * critical section.
389 *
390 * Early termination only occurs when tsleep() is
391 * entered while in a normal SRUN state.
392 */
393 if ((sig = CURSIG(p)) != 0)
394 goto resume;
395
396 /*
397 * Causes psignal to wake us up when.
398 */
399 p->p_flag |= P_SINTR;
400 }
401
402 /*
403 * Make sure the current process has been untangled from
404 * the userland scheduler and initialize slptime to start
405 * counting.
406 */
0a3f9b47
MD
407 if (flags & PNORESCHED)
408 td->td_flags |= TDF_NORESCHED;
553ea3c8 409 p->p_usched->release_curproc(&p->p_lwp);
0cfcada1 410 p->p_slptime = 0;
0a3f9b47 411 }
fc17ad60
MD
412
413 /*
344ad853 414 * Move our thread to the correct queue and setup our wchan, etc.
fc17ad60 415 */
37af14fe 416 lwkt_deschedule_self(td);
344ad853 417 td->td_flags |= TDF_TSLEEPQ;
fc17ad60
MD
418 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq);
419 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
344ad853
MD
420
421 td->td_wchan = ident;
422 td->td_wmesg = wmesg;
423 td->td_wdomain = flags & PDOMAIN_MASK;
424
425 /*
426 * Setup the timeout, if any
427 */
076fecef
MD
428 if (timo) {
429 callout_init(&thandle);
430 callout_reset(&thandle, timo, endtsleep, td);
431 }
344ad853 432
984263bc 433 /*
344ad853 434 * Beddy bye bye.
984263bc 435 */
0cfcada1 436 if (p) {
26a0694b 437 /*
52eedfb5 438 * Ok, we are sleeping. Place us in the SSLEEP state.
26a0694b 439 */
52eedfb5 440 KKASSERT((p->p_flag & P_ONRUNQ) == 0);
344ad853 441 p->p_stat = SSLEEP;
0cfcada1 442 p->p_stats->p_ru.ru_nvcsw++;
344ad853 443 lwkt_switch();
ab554892
MD
444
445 /*
446 * And when we are woken up, put us back in SRUN. If we
447 * slept for over a second, recalculate our estcpu.
448 */
344ad853 449 p->p_stat = SRUN;
ab554892
MD
450 if (p->p_slptime)
451 p->p_usched->recalculate(&p->p_lwp);
452 p->p_slptime = 0;
0cfcada1
MD
453 } else {
454 lwkt_switch();
455 }
344ad853 456
fc17ad60
MD
457 /*
458 * Make sure we haven't switched cpus while we were asleep. It's
344ad853 459 * not supposed to happen. Cleanup our temporary flags.
fc17ad60
MD
460 */
461 KKASSERT(gd == td->td_gd);
0a3f9b47 462 td->td_flags &= ~TDF_NORESCHED;
344ad853
MD
463
464 /*
465 * Cleanup the timeout.
466 */
467 if (timo) {
468 if (td->td_flags & TDF_TIMEOUT) {
469 td->td_flags &= ~TDF_TIMEOUT;
470 if (sig == 0)
471 error = EWOULDBLOCK;
472 } else {
473 callout_stop(&thandle);
474 }
0cfcada1 475 }
344ad853
MD
476
477 /*
478 * Since td_threadq is used both for our run queue AND for the
479 * tsleep hash queue, we can't still be on it at this point because
480 * we've gotten cpu back.
481 */
afbfc034 482 KASSERT((td->td_flags & TDF_TSLEEPQ) == 0, ("tsleep: impossible thread flags %08x", td->td_flags));
344ad853
MD
483 td->td_wchan = NULL;
484 td->td_wmesg = NULL;
485 td->td_wdomain = 0;
486
487 /*
488 * Figure out the correct error return
489 */
490resume:
0cfcada1 491 if (p) {
344ad853
MD
492 p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR);
493 if (catch && error == 0 && (sig != 0 || (sig = CURSIG(p)))) {
0cfcada1 494 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
344ad853
MD
495 error = EINTR;
496 else
497 error = ERESTART;
984263bc 498 }
984263bc 499 }
9afb0ffd 500 logtsleep(tsleep_end);
344ad853
MD
501 crit_exit_quick(td);
502 return (error);
984263bc
MD
503}
504
16523a43
MD
505/*
506 * This is a dandy function that allows us to interlock tsleep/wakeup
507 * operations with unspecified upper level locks, such as lockmgr locks,
508 * simply by holding a critical section. The sequence is:
509 *
510 * (enter critical section)
511 * (acquire upper level lock)
512 * tsleep_interlock(blah)
513 * (release upper level lock)
514 * tsleep(blah, ...)
515 * (exit critical section)
516 *
517 * Basically this function sets our cpumask for the ident which informs
518 * other cpus that our cpu 'might' be waiting (or about to wait on) the
519 * hash index related to the ident. The critical section prevents another
520 * cpu's wakeup() from being processed on our cpu until we are actually
521 * able to enter the tsleep(). Thus, no race occurs between our attempt
522 * to release a resource and sleep, and another cpu's attempt to acquire
523 * a resource and call wakeup.
524 *
525 * There isn't much of a point to this function unless you call it while
526 * holding a critical section.
527 */
bf765287
MD
528static __inline void
529_tsleep_interlock(globaldata_t gd, void *ident)
530{
531 int id = LOOKUP(ident);
532
533 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask);
534}
535
16523a43
MD
536void
537tsleep_interlock(void *ident)
538{
bf765287
MD
539 _tsleep_interlock(mycpu, ident);
540}
541
542/*
543 * Interlocked spinlock sleep. An exclusively held spinlock must
544 * be passed to msleep(). The function will atomically release the
545 * spinlock and tsleep on the ident, then reacquire the spinlock and
546 * return.
547 *
548 * This routine is fairly important along the critical path, so optimize it
549 * heavily.
550 */
551int
552msleep(void *ident, struct spinlock *spin, int flags,
553 const char *wmesg, int timo)
554{
555 globaldata_t gd = mycpu;
556 int error;
16523a43 557
bf765287
MD
558 crit_enter_gd(gd);
559 _tsleep_interlock(gd, ident);
560 spin_unlock_wr_quick(gd, spin);
561 error = tsleep(ident, flags, wmesg, timo);
562 spin_lock_wr_quick(gd, spin);
563 crit_exit_gd(gd);
564
565 return (error);
16523a43
MD
566}
567
984263bc 568/*
344ad853 569 * Implement the timeout for tsleep.
fc17ad60 570 *
344ad853
MD
571 * We set P_BREAKTSLEEP to indicate that an event has occured, but
572 * we only call setrunnable if the process is not stopped.
573 *
574 * This type of callout timeout is scheduled on the same cpu the process
575 * is sleeping on. Also, at the moment, the MP lock is held.
984263bc
MD
576 */
577static void
0cfcada1 578endtsleep(void *arg)
984263bc 579{
0cfcada1
MD
580 thread_t td = arg;
581 struct proc *p;
984263bc 582
344ad853 583 ASSERT_MP_LOCK_HELD(curthread);
37af14fe 584 crit_enter();
344ad853
MD
585
586 /*
587 * cpu interlock. Thread flags are only manipulated on
588 * the cpu owning the thread. proc flags are only manipulated
589 * by the older of the MP lock. We have both.
590 */
591 if (td->td_flags & TDF_TSLEEPQ) {
0cfcada1 592 td->td_flags |= TDF_TIMEOUT;
344ad853 593
0cfcada1 594 if ((p = td->td_proc) != NULL) {
344ad853
MD
595 p->p_flag |= P_BREAKTSLEEP;
596 if ((p->p_flag & P_STOPPED) == 0)
0cfcada1 597 setrunnable(p);
0cfcada1 598 } else {
344ad853 599 unsleep_and_wakeup_thread(td);
0cfcada1 600 }
984263bc 601 }
37af14fe 602 crit_exit();
984263bc
MD
603}
604
984263bc 605/*
344ad853
MD
606 * Unsleep and wakeup a thread. This function runs without the MP lock
607 * which means that it can only manipulate thread state on the owning cpu,
608 * and cannot touch the process state at all.
984263bc 609 */
344ad853 610static
8fb8bca6 611void
344ad853 612unsleep_and_wakeup_thread(struct thread *td)
8fb8bca6 613{
344ad853 614 globaldata_t gd = mycpu;
fc17ad60
MD
615 int id;
616
344ad853
MD
617#ifdef SMP
618 if (td->td_gd != gd) {
619 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td);
620 return;
621 }
622#endif
9c1fad94 623 crit_enter();
344ad853
MD
624 if (td->td_flags & TDF_TSLEEPQ) {
625 td->td_flags &= ~TDF_TSLEEPQ;
626 id = LOOKUP(td->td_wchan);
627 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq);
628 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL)
629 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask);
630 lwkt_schedule(td);
8fb8bca6 631 }
9c1fad94 632 crit_exit();
8fb8bca6 633}
8fb8bca6
EN
634
635/*
636 * Make all processes sleeping on the specified identifier runnable.
fc17ad60
MD
637 * count may be zero or one only.
638 *
639 * The domain encodes the sleep/wakeup domain AND the first cpu to check
640 * (which is always the current cpu). As we iterate across cpus
344ad853
MD
641 *
642 * This call may run without the MP lock held. We can only manipulate thread
643 * state on the cpu owning the thread. We CANNOT manipulate process state
644 * at all.
8fb8bca6
EN
645 */
646static void
fc17ad60 647_wakeup(void *ident, int domain)
984263bc 648{
fc17ad60 649 struct tslpque *qp;
0cfcada1
MD
650 struct thread *td;
651 struct thread *ntd;
fc17ad60 652 globaldata_t gd;
fc17ad60
MD
653#ifdef SMP
654 cpumask_t mask;
655 cpumask_t tmask;
656 int startcpu;
657 int nextcpu;
658#endif
659 int id;
984263bc 660
37af14fe 661 crit_enter();
9afb0ffd 662 logtsleep(wakeup_beg);
fc17ad60
MD
663 gd = mycpu;
664 id = LOOKUP(ident);
665 qp = &gd->gd_tsleep_hash[id];
984263bc 666restart:
0cfcada1
MD
667 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
668 ntd = TAILQ_NEXT(td, td_threadq);
fc17ad60
MD
669 if (td->td_wchan == ident &&
670 td->td_wdomain == (domain & PDOMAIN_MASK)
671 ) {
344ad853
MD
672 KKASSERT(td->td_flags & TDF_TSLEEPQ);
673 td->td_flags &= ~TDF_TSLEEPQ;
0cfcada1 674 TAILQ_REMOVE(qp, td, td_threadq);
fc17ad60
MD
675 if (TAILQ_FIRST(qp) == NULL) {
676 atomic_clear_int(&slpque_cpumasks[id],
677 gd->gd_cpumask);
678 }
344ad853 679 lwkt_schedule(td);
fc17ad60
MD
680 if (domain & PWAKEUP_ONE)
681 goto done;
0cfcada1 682 goto restart;
984263bc
MD
683 }
684 }
fc17ad60
MD
685
686#ifdef SMP
687 /*
688 * We finished checking the current cpu but there still may be
689 * more work to do. Either wakeup_one was requested and no matching
690 * thread was found, or a normal wakeup was requested and we have
691 * to continue checking cpus.
692 *
693 * The cpu that started the wakeup sequence is encoded in the domain.
694 * We use this information to determine which cpus still need to be
695 * checked, locate a candidate cpu, and chain the wakeup
696 * asynchronously with an IPI message.
697 *
698 * It should be noted that this scheme is actually less expensive then
699 * the old scheme when waking up multiple threads, since we send
700 * only one IPI message per target candidate which may then schedule
701 * multiple threads. Before we could have wound up sending an IPI
702 * message for each thread on the target cpu (!= current cpu) that
703 * needed to be woken up.
704 *
705 * NOTE: Wakeups occuring on remote cpus are asynchronous. This
706 * should be ok since we are passing idents in the IPI rather then
707 * thread pointers.
708 */
b336a9b1
MD
709 if ((domain & PWAKEUP_MYCPU) == 0 &&
710 (mask = slpque_cpumasks[id]) != 0
711 ) {
fc17ad60
MD
712 /*
713 * Look for a cpu that might have work to do. Mask out cpus
714 * which have already been processed.
715 *
716 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0
717 * ^ ^ ^
718 * start currentcpu start
719 * case2 case1
720 * * * *
721 * 11111111111111110000000000000111 case1
722 * 00000000111111110000000000000000 case2
723 *
724 * case1: We started at start_case1 and processed through
725 * to the current cpu. We have to check any bits
726 * after the current cpu, then check bits before
727 * the starting cpu.
728 *
729 * case2: We have already checked all the bits from
730 * start_case2 to the end, and from 0 to the current
731 * cpu. We just have the bits from the current cpu
732 * to start_case2 left to check.
733 */
734 startcpu = PWAKEUP_DECODE(domain);
735 if (gd->gd_cpuid >= startcpu) {
736 /*
737 * CASE1
738 */
739 tmask = mask & ~((gd->gd_cpumask << 1) - 1);
740 if (mask & tmask) {
741 nextcpu = bsfl(mask & tmask);
742 lwkt_send_ipiq2(globaldata_find(nextcpu),
743 _wakeup, ident, domain);
744 } else {
745 tmask = (1 << startcpu) - 1;
746 if (mask & tmask) {
747 nextcpu = bsfl(mask & tmask);
748 lwkt_send_ipiq2(
749 globaldata_find(nextcpu),
750 _wakeup, ident, domain);
751 }
752 }
753 } else {
754 /*
755 * CASE2
756 */
757 tmask = ~((gd->gd_cpumask << 1) - 1) &
758 ((1 << startcpu) - 1);
759 if (mask & tmask) {
760 nextcpu = bsfl(mask & tmask);
761 lwkt_send_ipiq2(globaldata_find(nextcpu),
762 _wakeup, ident, domain);
763 }
764 }
765 }
766#endif
767done:
9afb0ffd 768 logtsleep(wakeup_end);
37af14fe 769 crit_exit();
984263bc
MD
770}
771
b336a9b1
MD
772/*
773 * Wakeup all threads tsleep()ing on the specified ident, on all cpus
774 */
984263bc 775void
0cfcada1 776wakeup(void *ident)
984263bc 777{
fc17ad60 778 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid));
0cfcada1 779}
984263bc 780
b336a9b1
MD
781/*
782 * Wakeup one thread tsleep()ing on the specified ident, on any cpu.
783 */
0cfcada1
MD
784void
785wakeup_one(void *ident)
786{
fc17ad60
MD
787 /* XXX potentially round-robin the first responding cpu */
788 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE);
da5fb9ef
MD
789}
790
b336a9b1
MD
791/*
792 * Wakeup threads tsleep()ing on the specified ident on the current cpu
793 * only.
794 */
795void
796wakeup_mycpu(void *ident)
797{
798 _wakeup(ident, PWAKEUP_MYCPU);
799}
800
801/*
802 * Wakeup one thread tsleep()ing on the specified ident on the current cpu
803 * only.
804 */
805void
806wakeup_mycpu_one(void *ident)
807{
808 /* XXX potentially round-robin the first responding cpu */
809 _wakeup(ident, PWAKEUP_MYCPU|PWAKEUP_ONE);
810}
811
812/*
813 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu
814 * only.
815 */
816void
817wakeup_oncpu(globaldata_t gd, void *ident)
818{
1699d292 819#ifdef SMP
b336a9b1
MD
820 if (gd == mycpu) {
821 _wakeup(ident, PWAKEUP_MYCPU);
822 } else {
823 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU);
824 }
1699d292
MD
825#else
826 _wakeup(ident, PWAKEUP_MYCPU);
827#endif
b336a9b1
MD
828}
829
830/*
831 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu
832 * only.
833 */
834void
835wakeup_oncpu_one(globaldata_t gd, void *ident)
836{
1699d292 837#ifdef SMP
b336a9b1
MD
838 if (gd == mycpu) {
839 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
840 } else {
841 lwkt_send_ipiq2(gd, _wakeup, ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
842 }
1699d292
MD
843#else
844 _wakeup(ident, PWAKEUP_MYCPU | PWAKEUP_ONE);
845#endif
b336a9b1
MD
846}
847
848/*
849 * Wakeup all threads waiting on the specified ident that slept using
850 * the specified domain, on all cpus.
851 */
da5fb9ef
MD
852void
853wakeup_domain(void *ident, int domain)
854{
fc17ad60 855 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid));
da5fb9ef
MD
856}
857
b336a9b1
MD
858/*
859 * Wakeup one thread waiting on the specified ident that slept using
860 * the specified domain, on any cpu.
861 */
da5fb9ef
MD
862void
863wakeup_domain_one(void *ident, int domain)
864{
fc17ad60
MD
865 /* XXX potentially round-robin the first responding cpu */
866 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE);
984263bc
MD
867}
868
869/*
344ad853
MD
870 * setrunnable()
871 *
872 * Make a process runnable. The MP lock must be held on call. This only
873 * has an effect if we are in SSLEEP. We only break out of the
874 * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state.
37af14fe 875 *
344ad853
MD
876 * NOTE: With the MP lock held we can only safely manipulate the process
877 * structure. We cannot safely manipulate the thread structure.
984263bc
MD
878 */
879void
344ad853 880setrunnable(struct proc *p)
984263bc 881{
344ad853
MD
882 crit_enter();
883 ASSERT_MP_LOCK_HELD(curthread);
884 p->p_flag &= ~P_STOPPED;
885 if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) {
886 unsleep_and_wakeup_thread(p->p_thread);
984263bc 887 }
344ad853 888 crit_exit();
984263bc
MD
889}
890
891/*
344ad853
MD
892 * The process is stopped due to some condition, usually because P_STOPPED
893 * is set but also possibly due to being traced.
fc17ad60 894 *
344ad853
MD
895 * NOTE! If the caller sets P_STOPPED, the caller must also clear P_WAITED
896 * because the parent may check the child's status before the child actually
897 * gets to this routine.
898 *
899 * This routine is called with the current process only, typically just
900 * before returning to userland.
901 *
902 * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive
903 * SIGCONT to break out of the tsleep.
984263bc
MD
904 */
905void
344ad853 906tstop(struct proc *p)
984263bc 907{
344ad853
MD
908 wakeup((caddr_t)p->p_pptr);
909 p->p_flag |= P_BREAKTSLEEP;
910 tsleep(p, 0, "stop", 0);
26a0694b
MD
911}
912
a77ac49d
MD
913/*
914 * Yield / synchronous reschedule. This is a bit tricky because the trap
915 * code might have set a lazy release on the switch function. Setting
916 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
917 * switch, and that we are given a greater chance of affinity with our
918 * current cpu.
919 *
920 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt
921 * run queue. lwkt_switch() will also execute any assigned passive release
922 * (which usually calls release_curproc()), allowing a same/higher priority
923 * process to be designated as the current process.
924 *
925 * While it is possible for a lower priority process to be designated,
926 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely
927 * round-robin back to us and we will be able to re-acquire the current
928 * process designation.
929 */
930void
931uio_yield(void)
932{
933 struct thread *td = curthread;
934 struct proc *p = td->td_proc;
935
936 lwkt_setpri_self(td->td_pri & TDPRI_MASK);
937 if (p) {
938 p->p_flag |= P_PASSIVE_ACQ;
939 lwkt_switch();
940 p->p_flag &= ~P_PASSIVE_ACQ;
941 } else {
942 lwkt_switch();
943 }
944}
945
984263bc
MD
946/*
947 * Compute a tenex style load average of a quantity on
948 * 1, 5 and 15 minute intervals.
949 */
8fa76237
MD
950static int loadav_count_runnable(struct proc *p, void *data);
951
984263bc
MD
952static void
953loadav(void *arg)
954{
984263bc 955 struct loadavg *avg;
8fa76237 956 int i, nrun;
984263bc 957
984263bc 958 nrun = 0;
8fa76237
MD
959 allproc_scan(loadav_count_runnable, &nrun);
960 avg = &averunnable;
961 for (i = 0; i < 3; i++) {
984263bc
MD
962 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
963 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
8fa76237 964 }
984263bc
MD
965
966 /*
967 * Schedule the next update to occur after 5 seconds, but add a
968 * random variation to avoid synchronisation with processes that
969 * run at regular intervals.
970 */
971 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
8fa76237
MD
972 loadav, NULL);
973}
974
975static int
976loadav_count_runnable(struct proc *p, void *data)
977{
978 int *nrunp = data;
979 thread_t td;
980
981 switch (p->p_stat) {
982 case SRUN:
983 if ((td = p->p_thread) == NULL)
984 break;
985 if (td->td_flags & TDF_BLOCKED)
986 break;
987 /* fall through */
988 case SIDL:
989 ++*nrunp;
990 break;
991 default:
992 break;
993 }
994 return(0);
984263bc
MD
995}
996
997/* ARGSUSED */
998static void
6656cd91 999sched_setup(void *dummy)
984263bc 1000{
984263bc 1001 callout_init(&loadav_callout);
35f9d051 1002 callout_init(&schedcpu_callout);
984263bc
MD
1003
1004 /* Kick off timeout driven events by calling first time. */
984263bc
MD
1005 schedcpu(NULL);
1006 loadav(NULL);
1007}
1008