Misc interrupts/LWKT 1/2: interlock the idle thread. Put execution of
[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 $
4b5f931b 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.11 2003/06/29 07:37:06 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>
53#ifdef KTRACE
54#include <sys/uio.h>
55#include <sys/ktrace.h>
56#endif
f1d1c3fa 57#include <sys/xwait.h>
984263bc
MD
58
59#include <machine/cpu.h>
60#include <machine/ipl.h>
61#include <machine/smp.h>
62
63static void sched_setup __P((void *dummy));
64SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
65
66u_char curpriority;
67int hogticks;
68int lbolt;
69int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
17a9f566 70int ncpus;
984263bc
MD
71
72static struct callout loadav_callout;
73
74struct loadavg averunnable =
75 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
76/*
77 * Constants for averages over 1, 5, and 15 minutes
78 * when sampling at 5 second intervals.
79 */
80static fixpt_t cexp[3] = {
81 0.9200444146293232 * FSCALE, /* exp(-1/12) */
82 0.9834714538216174 * FSCALE, /* exp(-1/60) */
83 0.9944598480048967 * FSCALE, /* exp(-1/180) */
84};
85
86static int curpriority_cmp __P((struct proc *p));
87static void endtsleep __P((void *));
88static void loadav __P((void *arg));
89static void maybe_resched __P((struct proc *chk));
90static void roundrobin __P((void *arg));
91static void schedcpu __P((void *arg));
92static void updatepri __P((struct proc *p));
93
94static int
95sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
96{
97 int error, new_val;
98
99 new_val = sched_quantum * tick;
100 error = sysctl_handle_int(oidp, &new_val, 0, req);
101 if (error != 0 || req->newptr == NULL)
102 return (error);
103 if (new_val < tick)
104 return (EINVAL);
105 sched_quantum = new_val / tick;
106 hogticks = 2 * sched_quantum;
107 return (0);
108}
109
110SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
111 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
112
113/*-
114 * Compare priorities. Return:
115 * <0: priority of p < current priority
116 * 0: priority of p == current priority
117 * >0: priority of p > current priority
118 * The priorities are the normal priorities or the normal realtime priorities
119 * if p is on the same scheduler as curproc. Otherwise the process on the
120 * more realtimeish scheduler has lowest priority. As usual, a higher
121 * priority really means a lower priority.
122 */
123static int
124curpriority_cmp(p)
125 struct proc *p;
126{
127 int c_class, p_class;
128
129 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type);
130 p_class = RTP_PRIO_BASE(p->p_rtprio.type);
131 if (p_class != c_class)
132 return (p_class - c_class);
133 if (p_class == RTP_PRIO_NORMAL)
134 return (((int)p->p_priority - (int)curpriority) / PPQ);
135 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio);
136}
137
138/*
139 * Arrange to reschedule if necessary, taking the priorities and
140 * schedulers into account.
141 */
142static void
143maybe_resched(chk)
144 struct proc *chk;
145{
146 struct proc *p = curproc; /* XXX */
147
148 /*
149 * XXX idle scheduler still broken because proccess stays on idle
150 * scheduler during waits (such as when getting FS locks). If a
151 * standard process becomes runaway cpu-bound, the system can lockup
152 * due to idle-scheduler processes in wakeup never getting any cpu.
153 */
154 if (p == NULL) {
155#if 0
156 need_resched();
157#endif
158 } else if (chk == p) {
159 /* We may need to yield if our priority has been raised. */
160 if (curpriority_cmp(chk) > 0)
161 need_resched();
162 } else if (curpriority_cmp(chk) < 0)
163 need_resched();
164}
165
166int
167roundrobin_interval(void)
168{
169 return (sched_quantum);
170}
171
172/*
173 * Force switch among equal priority processes every 100ms.
174 */
175/* ARGSUSED */
176static void
177roundrobin(arg)
178 void *arg;
179{
180#ifndef SMP
181 struct proc *p = curproc; /* XXX */
182#endif
183
184#ifdef SMP
185 need_resched();
186 forward_roundrobin();
187#else
188 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type))
189 need_resched();
190#endif
191
192 timeout(roundrobin, NULL, sched_quantum);
193}
194
195/*
196 * Constants for digital decay and forget:
197 * 90% of (p_estcpu) usage in 5 * loadav time
198 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
199 * Note that, as ps(1) mentions, this can let percentages
200 * total over 100% (I've seen 137.9% for 3 processes).
201 *
202 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously.
203 *
204 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
205 * That is, the system wants to compute a value of decay such
206 * that the following for loop:
207 * for (i = 0; i < (5 * loadavg); i++)
208 * p_estcpu *= decay;
209 * will compute
210 * p_estcpu *= 0.1;
211 * for all values of loadavg:
212 *
213 * Mathematically this loop can be expressed by saying:
214 * decay ** (5 * loadavg) ~= .1
215 *
216 * The system computes decay as:
217 * decay = (2 * loadavg) / (2 * loadavg + 1)
218 *
219 * We wish to prove that the system's computation of decay
220 * will always fulfill the equation:
221 * decay ** (5 * loadavg) ~= .1
222 *
223 * If we compute b as:
224 * b = 2 * loadavg
225 * then
226 * decay = b / (b + 1)
227 *
228 * We now need to prove two things:
229 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
230 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
231 *
232 * Facts:
233 * For x close to zero, exp(x) =~ 1 + x, since
234 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
235 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
236 * For x close to zero, ln(1+x) =~ x, since
237 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
238 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
239 * ln(.1) =~ -2.30
240 *
241 * Proof of (1):
242 * Solve (factor)**(power) =~ .1 given power (5*loadav):
243 * solving for factor,
244 * ln(factor) =~ (-2.30/5*loadav), or
245 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
246 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
247 *
248 * Proof of (2):
249 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
250 * solving for power,
251 * power*ln(b/(b+1)) =~ -2.30, or
252 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
253 *
254 * Actual power values for the implemented algorithm are as follows:
255 * loadav: 1 2 3 4
256 * power: 5.68 10.32 14.94 19.55
257 */
258
259/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
260#define loadfactor(loadav) (2 * (loadav))
261#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
262
263/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
264static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
265SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
266
267/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
268static int fscale __unused = FSCALE;
269SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
270
271/*
272 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
273 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
274 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
275 *
276 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
277 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
278 *
279 * If you don't want to bother with the faster/more-accurate formula, you
280 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
281 * (more general) method of calculating the %age of CPU used by a process.
282 */
283#define CCPU_SHIFT 11
284
285/*
286 * Recompute process priorities, every hz ticks.
287 */
288/* ARGSUSED */
289static void
290schedcpu(arg)
291 void *arg;
292{
4b5f931b
MD
293 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
294 struct proc *p;
295 struct proc *curp;
296 int realstathz, s;
297
298 curp = lwkt_preempted_proc(); /* YYY temporary hack */
984263bc
MD
299
300 realstathz = stathz ? stathz : hz;
301 LIST_FOREACH(p, &allproc, p_list) {
302 /*
303 * Increment time in/out of memory and sleep time
304 * (if sleeping). We ignore overflow; with 16-bit int's
305 * (remember them?) overflow takes 45 days.
306 */
307 p->p_swtime++;
308 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
309 p->p_slptime++;
310 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
311 /*
312 * If the process has slept the entire second,
313 * stop recalculating its priority until it wakes up.
314 */
315 if (p->p_slptime > 1)
316 continue;
317 s = splhigh(); /* prevent state changes and protect run queue */
318 /*
319 * p_pctcpu is only for ps.
320 */
321#if (FSHIFT >= CCPU_SHIFT)
322 p->p_pctcpu += (realstathz == 100)?
323 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
324 100 * (((fixpt_t) p->p_cpticks)
325 << (FSHIFT - CCPU_SHIFT)) / realstathz;
326#else
327 p->p_pctcpu += ((FSCALE - ccpu) *
328 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT;
329#endif
330 p->p_cpticks = 0;
331 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
332 resetpriority(p);
333 if (p->p_priority >= PUSER) {
4b5f931b 334 if ((p != curp) &&
984263bc
MD
335#ifdef SMP
336 p->p_oncpu == 0xff && /* idle */
337#endif
338 p->p_stat == SRUN &&
339 (p->p_flag & P_INMEM) &&
340 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
341 remrunqueue(p);
342 p->p_priority = p->p_usrpri;
343 setrunqueue(p);
8ad65e08 344 } else {
984263bc 345 p->p_priority = p->p_usrpri;
8ad65e08 346 }
984263bc
MD
347 }
348 splx(s);
349 }
350 wakeup((caddr_t)&lbolt);
351 timeout(schedcpu, (void *)0, hz);
352}
353
354/*
355 * Recalculate the priority of a process after it has slept for a while.
356 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
357 * least six times the loadfactor will decay p_estcpu to zero.
358 */
359static void
360updatepri(p)
361 register struct proc *p;
362{
363 register unsigned int newcpu = p->p_estcpu;
364 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
365
366 if (p->p_slptime > 5 * loadfac)
367 p->p_estcpu = 0;
368 else {
369 p->p_slptime--; /* the first time was done in schedcpu */
370 while (newcpu && --p->p_slptime)
371 newcpu = decay_cpu(loadfac, newcpu);
372 p->p_estcpu = newcpu;
373 }
374 resetpriority(p);
375}
376
377/*
378 * We're only looking at 7 bits of the address; everything is
379 * aligned to 4, lots of things are aligned to greater powers
380 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
381 */
382#define TABLESIZE 128
0cfcada1 383static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
984263bc
MD
384#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
385
386/*
387 * During autoconfiguration or after a panic, a sleep will simply
388 * lower the priority briefly to allow interrupts, then return.
389 * The priority to be used (safepri) is machine-dependent, thus this
390 * value is initialized and maintained in the machine-dependent layers.
391 * This priority will typically be 0, or the lowest priority
392 * that is safe for use on the interrupt stack; it can be made
393 * higher to block network software interrupts after panics.
394 */
395int safepri;
396
397void
398sleepinit(void)
399{
400 int i;
401
402 sched_quantum = hz/10;
403 hogticks = 2 * sched_quantum;
404 for (i = 0; i < TABLESIZE; i++)
405 TAILQ_INIT(&slpque[i]);
406}
407
408/*
409 * General sleep call. Suspends the current process until a wakeup is
410 * performed on the specified identifier. The process will then be made
411 * runnable with the specified priority. Sleeps at most timo/hz seconds
412 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
413 * before and after sleeping, else signals are not checked. Returns 0 if
414 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
415 * signal needs to be delivered, ERESTART is returned if the current system
416 * call should be restarted if possible, and EINTR is returned if the system
417 * call should be interrupted by the signal (return EINTR).
418 */
419int
420tsleep(ident, priority, wmesg, timo)
421 void *ident;
422 int priority, timo;
423 const char *wmesg;
424{
dadab5e9 425 struct thread *td = curthread;
0cfcada1
MD
426 struct proc *p = td->td_proc; /* may be NULL */
427 int s, sig = 0, catch = priority & PCATCH;
f1d1c3fa 428 int id = LOOKUP(ident);
984263bc
MD
429 struct callout_handle thandle;
430
0cfcada1
MD
431 /*
432 * NOTE: removed KTRPOINT, it could cause races due to blocking
433 * even in stable. Just scrap it for now.
434 */
984263bc 435 s = splhigh();
f1d1c3fa 436
984263bc
MD
437 if (cold || panicstr) {
438 /*
439 * After a panic, or during autoconfiguration,
440 * just give interrupts a chance, then just return;
441 * don't run any other procs or panic below,
442 * in case this is the idle process and already asleep.
443 */
444 splx(safepri);
445 splx(s);
446 return (0);
447 }
0cfcada1
MD
448 KASSERT(ident != NULL, ("tsleep: no ident"));
449 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
450 ident, wmesg, p->p_stat));
451
452 td->td_wchan = ident;
453 td->td_wmesg = wmesg;
454 if (p) {
455 p->p_slptime = 0;
456 p->p_priority = priority & PRIMASK;
457 }
458 lwkt_deschedule_self();
459 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
984263bc 460 if (timo)
0cfcada1 461 thandle = timeout(endtsleep, (void *)td, timo);
984263bc
MD
462 /*
463 * We put ourselves on the sleep queue and start our timeout
464 * before calling CURSIG, as we could stop there, and a wakeup
465 * or a SIGCONT (or both) could occur while we were stopped.
466 * A SIGCONT would cause us to be marked as SSLEEP
467 * without resuming us, thus we must be ready for sleep
468 * when CURSIG is called. If the wakeup happens while we're
469 * stopped, p->p_wchan will be 0 upon return from CURSIG.
470 */
0cfcada1
MD
471 if (p) {
472 if (catch) {
473 p->p_flag |= P_SINTR;
474 if ((sig = CURSIG(p))) {
475 if (td->td_wchan)
476 unsleep(td);
477 p->p_stat = SRUN;
478 goto resume;
479 }
480 if (p->p_wchan == 0) {
481 catch = 0;
482 goto resume;
483 }
484 } else {
485 sig = 0;
984263bc 486 }
0cfcada1
MD
487 p->p_stat = SSLEEP;
488 p->p_stats->p_ru.ru_nvcsw++;
489 mi_switch();
490 } else {
491 lwkt_switch();
492 }
984263bc 493resume:
0cfcada1
MD
494 if (p) {
495 curpriority = p->p_usrpri;
496 p->p_flag &= ~P_SINTR;
497 }
984263bc 498 splx(s);
0cfcada1
MD
499 if (td->td_flags & TDF_TIMEOUT) {
500 td->td_flags &= ~TDF_TIMEOUT;
501 if (sig == 0)
984263bc 502 return (EWOULDBLOCK);
0cfcada1
MD
503 } else if (timo) {
504 untimeout(endtsleep, (void *)td, thandle);
505 }
506 if (p) {
507 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
508 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
509 return (EINTR);
510 return (ERESTART);
984263bc 511 }
984263bc 512 }
984263bc
MD
513 return (0);
514}
515
0cfcada1
MD
516#if 0
517
984263bc 518/*
f1d1c3fa
MD
519 * General sleep call. Suspends the current process until a wakeup is
520 * performed on the specified xwait structure. The process will then be made
521 * runnable with the specified priority. Sleeps at most timo/hz seconds
522 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
523 * before and after sleeping, else signals are not checked. Returns 0 if
524 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
525 * signal needs to be delivered, ERESTART is returned if the current system
526 * call should be restarted if possible, and EINTR is returned if the system
527 * call should be interrupted by the signal (return EINTR).
984263bc 528 *
f1d1c3fa
MD
529 * If the passed generation number is different from the generation number
530 * in the xwait, return immediately.
984263bc 531 */
984263bc 532int
f1d1c3fa 533xsleep(struct xwait *w, int priority, const char *wmesg, int timo, int *gen)
984263bc 534{
dadab5e9
MD
535 struct thread *td = curthread;
536 struct proc *p = td->td_proc;
f1d1c3fa
MD
537 int s, sig, catch = priority & PCATCH;
538 struct callout_handle thandle;
984263bc 539
f1d1c3fa 540#ifdef KTRACE
dadab5e9 541 if (KTRPOINT(td, KTR_CSW))
f1d1c3fa
MD
542 ktrcsw(p->p_tracep, 1, 0);
543#endif
984263bc
MD
544 s = splhigh();
545
f1d1c3fa 546 if (cold || panicstr) {
984263bc 547 /*
f1d1c3fa
MD
548 * After a panic, or during autoconfiguration,
549 * just give interrupts a chance, then just return;
550 * don't run any other procs or panic below,
551 * in case this is the idle process and already asleep.
984263bc 552 */
f1d1c3fa
MD
553 splx(safepri);
554 splx(s);
555 return (0);
556 }
7d0bac62
MD
557 KASSERT(p != NULL, ("xsleep1"));
558 KASSERT(w != NULL && p->p_stat == SRUN, ("xsleep"));
984263bc 559
f1d1c3fa
MD
560 /*
561 * If the generation number does not match we return immediately.
562 */
563 if (*gen != w->gen) {
564 *gen = w->gen;
984263bc 565 splx(s);
984263bc 566#ifdef KTRACE
dadab5e9 567 if (KTRPOINT(td, KTR_CSW))
f1d1c3fa 568 ktrcsw(p->p_tracep, 0, 0);
984263bc 569#endif
f1d1c3fa
MD
570 return(0);
571 }
572
573 p->p_wchan = w;
574 p->p_wmesg = wmesg;
575 p->p_slptime = 0;
576 p->p_priority = priority & PRIMASK;
577 p->p_flag |= P_XSLEEP;
578 TAILQ_INSERT_TAIL(&w->waitq, p, p_procq);
579 if (timo)
580 thandle = timeout(endtsleep, (void *)p, timo);
581 /*
582 * We put ourselves on the sleep queue and start our timeout
583 * before calling CURSIG, as we could stop there, and a wakeup
584 * or a SIGCONT (or both) could occur while we were stopped.
585 * A SIGCONT would cause us to be marked as SSLEEP
586 * without resuming us, thus we must be ready for sleep
587 * when CURSIG is called. If the wakeup happens while we're
588 * stopped, p->p_wchan will be 0 upon return from CURSIG.
589 */
590 if (catch) {
591 p->p_flag |= P_SINTR;
592 if ((sig = CURSIG(p))) {
593 if (p->p_wchan)
594 unsleep(p);
595 p->p_stat = SRUN;
596 goto resume;
597 }
598 if (p->p_wchan == NULL) {
599 catch = 0;
600 goto resume;
601 }
602 } else
603 sig = 0;
604 p->p_stat = SSLEEP;
605 p->p_stats->p_ru.ru_nvcsw++;
606 mi_switch();
607resume:
608 curpriority = p->p_usrpri;
609 *gen = w->gen; /* update generation number */
610 splx(s);
611 p->p_flag &= ~P_SINTR;
612 if (p->p_flag & P_TIMEOUT) {
613 p->p_flag &= ~P_TIMEOUT;
614 if (sig == 0) {
984263bc 615#ifdef KTRACE
dadab5e9 616 if (KTRPOINT(td, KTR_CSW))
984263bc
MD
617 ktrcsw(p->p_tracep, 0, 0);
618#endif
f1d1c3fa 619 return (EWOULDBLOCK);
984263bc 620 }
f1d1c3fa
MD
621 } else if (timo)
622 untimeout(endtsleep, (void *)p, thandle);
623 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
984263bc 624#ifdef KTRACE
dadab5e9 625 if (KTRPOINT(td, KTR_CSW))
984263bc
MD
626 ktrcsw(p->p_tracep, 0, 0);
627#endif
f1d1c3fa
MD
628 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
629 return (EINTR);
630 return (ERESTART);
984263bc 631 }
f1d1c3fa 632#ifdef KTRACE
dadab5e9 633 if (KTRPOINT(td, KTR_CSW))
f1d1c3fa
MD
634 ktrcsw(p->p_tracep, 0, 0);
635#endif
984263bc
MD
636 return (0);
637}
638
0cfcada1
MD
639#endif
640
984263bc 641/*
0cfcada1
MD
642 * Implement the timeout for tsleep. We interlock against
643 * wchan when setting TDF_TIMEOUT. For processes we remove
644 * the sleep if the process is stopped rather then sleeping,
645 * so it remains stopped.
984263bc
MD
646 */
647static void
0cfcada1 648endtsleep(void *arg)
984263bc 649{
0cfcada1
MD
650 thread_t td = arg;
651 struct proc *p;
984263bc
MD
652 int s;
653
984263bc 654 s = splhigh();
0cfcada1
MD
655 if (td->td_wchan) {
656 td->td_flags |= TDF_TIMEOUT;
657 if ((p = td->td_proc) != NULL) {
658 if (p->p_stat == SSLEEP)
659 setrunnable(p);
660 else
661 unsleep(td);
662 } else {
663 unsleep(td);
664 lwkt_schedule(td);
665 }
984263bc
MD
666 }
667 splx(s);
668}
669
670/*
671 * Remove a process from its wait queue
672 */
673void
0cfcada1 674unsleep(struct thread *td)
984263bc
MD
675{
676 int s;
677
678 s = splhigh();
0cfcada1
MD
679 if (td->td_wchan) {
680#if 0
f1d1c3fa
MD
681 if (p->p_flag & P_XSLEEP) {
682 struct xwait *w = p->p_wchan;
683 TAILQ_REMOVE(&w->waitq, p, p_procq);
684 p->p_flag &= ~P_XSLEEP;
0cfcada1
MD
685 } else
686#endif
687 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
688 td->td_wchan = NULL;
f1d1c3fa
MD
689 }
690 splx(s);
691}
692
0cfcada1 693#if 0
f1d1c3fa
MD
694/*
695 * Make all processes sleeping on the explicit lock structure runnable.
696 */
697void
698xwakeup(struct xwait *w)
699{
700 struct proc *p;
701 int s;
702
703 s = splhigh();
704 ++w->gen;
705 while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
706 TAILQ_REMOVE(&w->waitq, p, p_procq);
707 KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
708 ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
709 p->p_wchan = NULL;
710 p->p_flag &= ~P_XSLEEP;
711 if (p->p_stat == SSLEEP) {
712 /* OPTIMIZED EXPANSION OF setrunnable(p); */
713 if (p->p_slptime > 1)
714 updatepri(p);
715 p->p_slptime = 0;
716 p->p_stat = SRUN;
717 if (p->p_flag & P_INMEM) {
718 setrunqueue(p);
719 maybe_resched(p);
720 } else {
721 p->p_flag |= P_SWAPINREQ;
722 wakeup((caddr_t)&proc0);
723 }
724 }
984263bc
MD
725 }
726 splx(s);
727}
0cfcada1 728#endif
984263bc
MD
729
730/*
731 * Make all processes sleeping on the specified identifier runnable.
732 */
0cfcada1
MD
733static void
734_wakeup(void *ident, int count)
984263bc 735{
0cfcada1
MD
736 struct slpquehead *qp;
737 struct thread *td;
738 struct thread *ntd;
739 struct proc *p;
984263bc 740 int s;
f1d1c3fa 741 int id = LOOKUP(ident);
984263bc
MD
742
743 s = splhigh();
f1d1c3fa 744 qp = &slpque[id];
984263bc 745restart:
0cfcada1
MD
746 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
747 ntd = TAILQ_NEXT(td, td_threadq);
748 if (td->td_wchan == ident) {
749 TAILQ_REMOVE(qp, td, td_threadq);
750 td->td_wchan = NULL;
751 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
984263bc
MD
752 /* OPTIMIZED EXPANSION OF setrunnable(p); */
753 if (p->p_slptime > 1)
754 updatepri(p);
755 p->p_slptime = 0;
756 p->p_stat = SRUN;
757 if (p->p_flag & P_INMEM) {
758 setrunqueue(p);
759 maybe_resched(p);
760 } else {
761 p->p_flag |= P_SWAPINREQ;
762 wakeup((caddr_t)&proc0);
763 }
764 /* END INLINE EXPANSION */
0cfcada1
MD
765 } else if (p == NULL) {
766 lwkt_schedule(td);
984263bc 767 }
0cfcada1
MD
768 if (--count == 0)
769 break;
770 goto restart;
984263bc
MD
771 }
772 }
773 splx(s);
774}
775
984263bc 776void
0cfcada1 777wakeup(void *ident)
984263bc 778{
0cfcada1
MD
779 _wakeup(ident, 0);
780}
984263bc 781
0cfcada1
MD
782void
783wakeup_one(void *ident)
784{
785 _wakeup(ident, 1);
984263bc
MD
786}
787
788/*
789 * The machine independent parts of mi_switch().
790 * Must be called at splstatclock() or higher.
791 */
792void
793mi_switch()
794{
d16a8831
MD
795 struct thread *td = curthread;
796 struct proc *p = td->td_proc; /* XXX */
797 struct rlimit *rlim;
984263bc 798 int x;
d16a8831 799 u_int64_t ttime;
984263bc
MD
800
801 /*
802 * XXX this spl is almost unnecessary. It is partly to allow for
803 * sloppy callers that don't do it (issignal() via CURSIG() is the
804 * main offender). It is partly to work around a bug in the i386
805 * cpu_switch() (the ipl is not preserved). We ran for years
806 * without it. I think there was only a interrupt latency problem.
807 * The main caller, tsleep(), does an splx() a couple of instructions
808 * after calling here. The buggy caller, issignal(), usually calls
809 * here at spl0() and sometimes returns at splhigh(). The process
810 * then runs for a little too long at splhigh(). The ipl gets fixed
811 * when the process returns to user mode (or earlier).
812 *
813 * It would probably be better to always call here at spl0(). Callers
814 * are prepared to give up control to another process, so they must
815 * be prepared to be interrupted. The clock stuff here may not
816 * actually need splstatclock().
817 */
818 x = splstatclock();
8ad65e08 819 clear_resched();
984263bc
MD
820
821#ifdef SIMPLELOCK_DEBUG
822 if (p->p_simple_locks)
823 printf("sleep: holding simple lock\n");
824#endif
984263bc
MD
825
826 /*
827 * Check if the process exceeds its cpu resource allocation.
d16a8831
MD
828 * If over max, kill it. Time spent in interrupts is not
829 * included. YYY 64 bit match is expensive. Ick.
984263bc 830 */
d16a8831 831 ttime = td->td_sticks + td->td_uticks;
984263bc 832 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
d16a8831 833 ttime > p->p_limit->p_cpulimit) {
984263bc 834 rlim = &p->p_rlimit[RLIMIT_CPU];
d16a8831 835 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
984263bc
MD
836 killproc(p, "exceeded maximum CPU limit");
837 } else {
838 psignal(p, SIGXCPU);
839 if (rlim->rlim_cur < rlim->rlim_max) {
840 /* XXX: we should make a private copy */
841 rlim->rlim_cur += 5;
842 }
843 }
844 }
845
846 /*
847 * Pick a new current process and record its start time.
8ad65e08
MD
848 * YYY lwkt_switch() will run the heavy weight process restoration
849 * code, which removes the target thread and process from their
850 * respective run queues to temporarily mimic 5.x behavior.
851 * YYY the userland scheduler should pick only one user process
852 * at a time to run per cpu.
984263bc
MD
853 */
854 cnt.v_swtch++;
8ad65e08
MD
855 lwkt_switch();
856 remrunqueue(p);
984263bc
MD
857
858 splx(x);
859}
860
861/*
862 * Change process state to be runnable,
863 * placing it on the run queue if it is in memory,
864 * and awakening the swapper if it isn't in memory.
865 */
866void
0cfcada1 867setrunnable(struct proc *p)
984263bc 868{
0cfcada1 869 int s;
984263bc
MD
870
871 s = splhigh();
872 switch (p->p_stat) {
873 case 0:
874 case SRUN:
875 case SZOMB:
876 default:
877 panic("setrunnable");
878 case SSTOP:
879 case SSLEEP:
0cfcada1 880 unsleep(p->p_thread); /* e.g. when sending signals */
984263bc
MD
881 break;
882
883 case SIDL:
884 break;
885 }
886 p->p_stat = SRUN;
887 if (p->p_flag & P_INMEM)
888 setrunqueue(p);
889 splx(s);
890 if (p->p_slptime > 1)
891 updatepri(p);
892 p->p_slptime = 0;
893 if ((p->p_flag & P_INMEM) == 0) {
894 p->p_flag |= P_SWAPINREQ;
895 wakeup((caddr_t)&proc0);
896 }
897 else
898 maybe_resched(p);
899}
900
901/*
902 * Compute the priority of a process when running in user mode.
903 * Arrange to reschedule if the resulting priority is better
904 * than that of the current process.
905 */
906void
907resetpriority(p)
908 register struct proc *p;
909{
910 register unsigned int newpriority;
911
912 if (p->p_rtprio.type == RTP_PRIO_NORMAL) {
913 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT +
914 NICE_WEIGHT * p->p_nice;
915 newpriority = min(newpriority, MAXPRI);
916 p->p_usrpri = newpriority;
917 }
918 maybe_resched(p);
919}
920
921/*
922 * Compute a tenex style load average of a quantity on
923 * 1, 5 and 15 minute intervals.
924 */
925static void
926loadav(void *arg)
927{
928 int i, nrun;
929 struct loadavg *avg;
930 struct proc *p;
931
932 avg = &averunnable;
933 nrun = 0;
934 LIST_FOREACH(p, &allproc, p_list) {
935 switch (p->p_stat) {
936 case SRUN:
937 case SIDL:
938 nrun++;
939 }
940 }
941 for (i = 0; i < 3; i++)
942 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
943 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
944
945 /*
946 * Schedule the next update to occur after 5 seconds, but add a
947 * random variation to avoid synchronisation with processes that
948 * run at regular intervals.
949 */
950 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
951 loadav, NULL);
952}
953
954/* ARGSUSED */
955static void
956sched_setup(dummy)
957 void *dummy;
958{
959
960 callout_init(&loadav_callout);
961
962 /* Kick off timeout driven events by calling first time. */
963 roundrobin(NULL);
964 schedcpu(NULL);
965 loadav(NULL);
966}
967
968/*
969 * We adjust the priority of the current process. The priority of
970 * a process gets worse as it accumulates CPU time. The cpu usage
971 * estimator (p_estcpu) is increased here. resetpriority() will
972 * compute a different priority each time p_estcpu increases by
973 * INVERSE_ESTCPU_WEIGHT
974 * (until MAXPRI is reached). The cpu usage estimator ramps up
975 * quite quickly when the process is running (linearly), and decays
976 * away exponentially, at a rate which is proportionally slower when
977 * the system is busy. The basic principle is that the system will
978 * 90% forget that the process used a lot of CPU time in 5 * loadav
979 * seconds. This causes the system to favor processes which haven't
980 * run much recently, and to round-robin among other processes.
981 */
982void
983schedclock(p)
984 struct proc *p;
985{
986
987 p->p_cpticks++;
988 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
989 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
990 resetpriority(p);
991 if (p->p_priority >= PUSER)
992 p->p_priority = p->p_usrpri;
993 }
994}