thread stage 7: Implement basic LWKTs, use a straight round-robin model for
[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 $
8ad65e08 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.4 2003/06/20 02:09:56 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
57
58#include <machine/cpu.h>
59#include <machine/ipl.h>
60#include <machine/smp.h>
61
62static void sched_setup __P((void *dummy));
63SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
64
65u_char curpriority;
66int hogticks;
67int lbolt;
68int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
69
70static struct callout loadav_callout;
71
72struct loadavg averunnable =
73 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */
74/*
75 * Constants for averages over 1, 5, and 15 minutes
76 * when sampling at 5 second intervals.
77 */
78static fixpt_t cexp[3] = {
79 0.9200444146293232 * FSCALE, /* exp(-1/12) */
80 0.9834714538216174 * FSCALE, /* exp(-1/60) */
81 0.9944598480048967 * FSCALE, /* exp(-1/180) */
82};
83
84static int curpriority_cmp __P((struct proc *p));
85static void endtsleep __P((void *));
86static void loadav __P((void *arg));
87static void maybe_resched __P((struct proc *chk));
88static void roundrobin __P((void *arg));
89static void schedcpu __P((void *arg));
90static void updatepri __P((struct proc *p));
91
92static int
93sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
94{
95 int error, new_val;
96
97 new_val = sched_quantum * tick;
98 error = sysctl_handle_int(oidp, &new_val, 0, req);
99 if (error != 0 || req->newptr == NULL)
100 return (error);
101 if (new_val < tick)
102 return (EINVAL);
103 sched_quantum = new_val / tick;
104 hogticks = 2 * sched_quantum;
105 return (0);
106}
107
108SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
109 0, sizeof sched_quantum, sysctl_kern_quantum, "I", "");
110
111/*-
112 * Compare priorities. Return:
113 * <0: priority of p < current priority
114 * 0: priority of p == current priority
115 * >0: priority of p > current priority
116 * The priorities are the normal priorities or the normal realtime priorities
117 * if p is on the same scheduler as curproc. Otherwise the process on the
118 * more realtimeish scheduler has lowest priority. As usual, a higher
119 * priority really means a lower priority.
120 */
121static int
122curpriority_cmp(p)
123 struct proc *p;
124{
125 int c_class, p_class;
126
127 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type);
128 p_class = RTP_PRIO_BASE(p->p_rtprio.type);
129 if (p_class != c_class)
130 return (p_class - c_class);
131 if (p_class == RTP_PRIO_NORMAL)
132 return (((int)p->p_priority - (int)curpriority) / PPQ);
133 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio);
134}
135
136/*
137 * Arrange to reschedule if necessary, taking the priorities and
138 * schedulers into account.
139 */
140static void
141maybe_resched(chk)
142 struct proc *chk;
143{
144 struct proc *p = curproc; /* XXX */
145
146 /*
147 * XXX idle scheduler still broken because proccess stays on idle
148 * scheduler during waits (such as when getting FS locks). If a
149 * standard process becomes runaway cpu-bound, the system can lockup
150 * due to idle-scheduler processes in wakeup never getting any cpu.
151 */
152 if (p == NULL) {
153#if 0
154 need_resched();
155#endif
156 } else if (chk == p) {
157 /* We may need to yield if our priority has been raised. */
158 if (curpriority_cmp(chk) > 0)
159 need_resched();
160 } else if (curpriority_cmp(chk) < 0)
161 need_resched();
162}
163
164int
165roundrobin_interval(void)
166{
167 return (sched_quantum);
168}
169
170/*
171 * Force switch among equal priority processes every 100ms.
172 */
173/* ARGSUSED */
174static void
175roundrobin(arg)
176 void *arg;
177{
178#ifndef SMP
179 struct proc *p = curproc; /* XXX */
180#endif
181
182#ifdef SMP
183 need_resched();
184 forward_roundrobin();
185#else
186 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type))
187 need_resched();
188#endif
189
190 timeout(roundrobin, NULL, sched_quantum);
191}
192
193/*
194 * Constants for digital decay and forget:
195 * 90% of (p_estcpu) usage in 5 * loadav time
196 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
197 * Note that, as ps(1) mentions, this can let percentages
198 * total over 100% (I've seen 137.9% for 3 processes).
199 *
200 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously.
201 *
202 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
203 * That is, the system wants to compute a value of decay such
204 * that the following for loop:
205 * for (i = 0; i < (5 * loadavg); i++)
206 * p_estcpu *= decay;
207 * will compute
208 * p_estcpu *= 0.1;
209 * for all values of loadavg:
210 *
211 * Mathematically this loop can be expressed by saying:
212 * decay ** (5 * loadavg) ~= .1
213 *
214 * The system computes decay as:
215 * decay = (2 * loadavg) / (2 * loadavg + 1)
216 *
217 * We wish to prove that the system's computation of decay
218 * will always fulfill the equation:
219 * decay ** (5 * loadavg) ~= .1
220 *
221 * If we compute b as:
222 * b = 2 * loadavg
223 * then
224 * decay = b / (b + 1)
225 *
226 * We now need to prove two things:
227 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
228 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
229 *
230 * Facts:
231 * For x close to zero, exp(x) =~ 1 + x, since
232 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
233 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
234 * For x close to zero, ln(1+x) =~ x, since
235 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
236 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
237 * ln(.1) =~ -2.30
238 *
239 * Proof of (1):
240 * Solve (factor)**(power) =~ .1 given power (5*loadav):
241 * solving for factor,
242 * ln(factor) =~ (-2.30/5*loadav), or
243 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
244 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
245 *
246 * Proof of (2):
247 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
248 * solving for power,
249 * power*ln(b/(b+1)) =~ -2.30, or
250 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
251 *
252 * Actual power values for the implemented algorithm are as follows:
253 * loadav: 1 2 3 4
254 * power: 5.68 10.32 14.94 19.55
255 */
256
257/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
258#define loadfactor(loadav) (2 * (loadav))
259#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
260
261/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
262static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
263SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
264
265/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
266static int fscale __unused = FSCALE;
267SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
268
269/*
270 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
271 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
272 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
273 *
274 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
275 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
276 *
277 * If you don't want to bother with the faster/more-accurate formula, you
278 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
279 * (more general) method of calculating the %age of CPU used by a process.
280 */
281#define CCPU_SHIFT 11
282
283/*
284 * Recompute process priorities, every hz ticks.
285 */
286/* ARGSUSED */
287static void
288schedcpu(arg)
289 void *arg;
290{
291 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
292 register struct proc *p;
293 register int realstathz, s;
294
295 realstathz = stathz ? stathz : hz;
296 LIST_FOREACH(p, &allproc, p_list) {
297 /*
298 * Increment time in/out of memory and sleep time
299 * (if sleeping). We ignore overflow; with 16-bit int's
300 * (remember them?) overflow takes 45 days.
301 */
302 p->p_swtime++;
303 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
304 p->p_slptime++;
305 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
306 /*
307 * If the process has slept the entire second,
308 * stop recalculating its priority until it wakes up.
309 */
310 if (p->p_slptime > 1)
311 continue;
312 s = splhigh(); /* prevent state changes and protect run queue */
313 /*
314 * p_pctcpu is only for ps.
315 */
316#if (FSHIFT >= CCPU_SHIFT)
317 p->p_pctcpu += (realstathz == 100)?
318 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
319 100 * (((fixpt_t) p->p_cpticks)
320 << (FSHIFT - CCPU_SHIFT)) / realstathz;
321#else
322 p->p_pctcpu += ((FSCALE - ccpu) *
323 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT;
324#endif
325 p->p_cpticks = 0;
326 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
327 resetpriority(p);
328 if (p->p_priority >= PUSER) {
329 if ((p != curproc) &&
330#ifdef SMP
331 p->p_oncpu == 0xff && /* idle */
332#endif
333 p->p_stat == SRUN &&
334 (p->p_flag & P_INMEM) &&
335 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
336 remrunqueue(p);
337 p->p_priority = p->p_usrpri;
338 setrunqueue(p);
8ad65e08 339 } else {
984263bc 340 p->p_priority = p->p_usrpri;
8ad65e08 341 }
984263bc
MD
342 }
343 splx(s);
344 }
345 wakeup((caddr_t)&lbolt);
346 timeout(schedcpu, (void *)0, hz);
347}
348
349/*
350 * Recalculate the priority of a process after it has slept for a while.
351 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
352 * least six times the loadfactor will decay p_estcpu to zero.
353 */
354static void
355updatepri(p)
356 register struct proc *p;
357{
358 register unsigned int newcpu = p->p_estcpu;
359 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
360
361 if (p->p_slptime > 5 * loadfac)
362 p->p_estcpu = 0;
363 else {
364 p->p_slptime--; /* the first time was done in schedcpu */
365 while (newcpu && --p->p_slptime)
366 newcpu = decay_cpu(loadfac, newcpu);
367 p->p_estcpu = newcpu;
368 }
369 resetpriority(p);
370}
371
372/*
373 * We're only looking at 7 bits of the address; everything is
374 * aligned to 4, lots of things are aligned to greater powers
375 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
376 */
377#define TABLESIZE 128
378static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE];
379#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
380
381/*
382 * During autoconfiguration or after a panic, a sleep will simply
383 * lower the priority briefly to allow interrupts, then return.
384 * The priority to be used (safepri) is machine-dependent, thus this
385 * value is initialized and maintained in the machine-dependent layers.
386 * This priority will typically be 0, or the lowest priority
387 * that is safe for use on the interrupt stack; it can be made
388 * higher to block network software interrupts after panics.
389 */
390int safepri;
391
392void
393sleepinit(void)
394{
395 int i;
396
397 sched_quantum = hz/10;
398 hogticks = 2 * sched_quantum;
399 for (i = 0; i < TABLESIZE; i++)
400 TAILQ_INIT(&slpque[i]);
401}
402
403/*
404 * General sleep call. Suspends the current process until a wakeup is
405 * performed on the specified identifier. The process will then be made
406 * runnable with the specified priority. Sleeps at most timo/hz seconds
407 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
408 * before and after sleeping, else signals are not checked. Returns 0 if
409 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
410 * signal needs to be delivered, ERESTART is returned if the current system
411 * call should be restarted if possible, and EINTR is returned if the system
412 * call should be interrupted by the signal (return EINTR).
413 */
414int
415tsleep(ident, priority, wmesg, timo)
416 void *ident;
417 int priority, timo;
418 const char *wmesg;
419{
420 struct proc *p = curproc;
421 int s, sig, catch = priority & PCATCH;
422 struct callout_handle thandle;
423
424#ifdef KTRACE
425 if (p && KTRPOINT(p, KTR_CSW))
426 ktrcsw(p->p_tracep, 1, 0);
427#endif
428 s = splhigh();
429 if (cold || panicstr) {
430 /*
431 * After a panic, or during autoconfiguration,
432 * just give interrupts a chance, then just return;
433 * don't run any other procs or panic below,
434 * in case this is the idle process and already asleep.
435 */
436 splx(safepri);
437 splx(s);
438 return (0);
439 }
440 KASSERT(p != NULL, ("tsleep1"));
441 KASSERT(ident != NULL && p->p_stat == SRUN, ("tsleep"));
442 /*
443 * Process may be sitting on a slpque if asleep() was called, remove
444 * it before re-adding.
445 */
446 if (p->p_wchan != NULL)
447 unsleep(p);
448
449 p->p_wchan = ident;
450 p->p_wmesg = wmesg;
451 p->p_slptime = 0;
452 p->p_priority = priority & PRIMASK;
453 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq);
454 if (timo)
455 thandle = timeout(endtsleep, (void *)p, timo);
456 /*
457 * We put ourselves on the sleep queue and start our timeout
458 * before calling CURSIG, as we could stop there, and a wakeup
459 * or a SIGCONT (or both) could occur while we were stopped.
460 * A SIGCONT would cause us to be marked as SSLEEP
461 * without resuming us, thus we must be ready for sleep
462 * when CURSIG is called. If the wakeup happens while we're
463 * stopped, p->p_wchan will be 0 upon return from CURSIG.
464 */
465 if (catch) {
466 p->p_flag |= P_SINTR;
467 if ((sig = CURSIG(p))) {
468 if (p->p_wchan)
469 unsleep(p);
470 p->p_stat = SRUN;
471 goto resume;
472 }
473 if (p->p_wchan == 0) {
474 catch = 0;
475 goto resume;
476 }
477 } else
478 sig = 0;
479 p->p_stat = SSLEEP;
480 p->p_stats->p_ru.ru_nvcsw++;
481 mi_switch();
482resume:
483 curpriority = p->p_usrpri;
484 splx(s);
485 p->p_flag &= ~P_SINTR;
486 if (p->p_flag & P_TIMEOUT) {
487 p->p_flag &= ~P_TIMEOUT;
488 if (sig == 0) {
489#ifdef KTRACE
490 if (KTRPOINT(p, KTR_CSW))
491 ktrcsw(p->p_tracep, 0, 0);
492#endif
493 return (EWOULDBLOCK);
494 }
495 } else if (timo)
496 untimeout(endtsleep, (void *)p, thandle);
497 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
498#ifdef KTRACE
499 if (KTRPOINT(p, KTR_CSW))
500 ktrcsw(p->p_tracep, 0, 0);
501#endif
502 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
503 return (EINTR);
504 return (ERESTART);
505 }
506#ifdef KTRACE
507 if (KTRPOINT(p, KTR_CSW))
508 ktrcsw(p->p_tracep, 0, 0);
509#endif
510 return (0);
511}
512
513/*
514 * asleep() - async sleep call. Place process on wait queue and return
515 * immediately without blocking. The process stays runnable until await()
516 * is called. If ident is NULL, remove process from wait queue if it is still
517 * on one.
518 *
519 * Only the most recent sleep condition is effective when making successive
520 * calls to asleep() or when calling tsleep().
521 *
522 * The timeout, if any, is not initiated until await() is called. The sleep
523 * priority, signal, and timeout is specified in the asleep() call but may be
524 * overriden in the await() call.
525 *
526 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>>
527 */
528
529int
530asleep(void *ident, int priority, const char *wmesg, int timo)
531{
532 struct proc *p = curproc;
533 int s;
534
535 /*
536 * splhigh() while manipulating sleep structures and slpque.
537 *
538 * Remove preexisting wait condition (if any) and place process
539 * on appropriate slpque, but do not put process to sleep.
540 */
541
542 s = splhigh();
543
544 if (p->p_wchan != NULL)
545 unsleep(p);
546
547 if (ident) {
548 p->p_wchan = ident;
549 p->p_wmesg = wmesg;
550 p->p_slptime = 0;
551 p->p_asleep.as_priority = priority;
552 p->p_asleep.as_timo = timo;
553 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq);
554 }
555
556 splx(s);
557
558 return(0);
559}
560
561/*
562 * await() - wait for async condition to occur. The process blocks until
563 * wakeup() is called on the most recent asleep() address. If wakeup is called
564 * priority to await(), await() winds up being a NOP.
565 *
566 * If await() is called more then once (without an intervening asleep() call),
567 * await() is still effectively a NOP but it calls mi_switch() to give other
568 * processes some cpu before returning. The process is left runnable.
569 *
570 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>>
571 */
572
573int
574await(int priority, int timo)
575{
576 struct proc *p = curproc;
577 int s;
578
579 s = splhigh();
580
581 if (p->p_wchan != NULL) {
582 struct callout_handle thandle;
583 int sig;
584 int catch;
585
586 /*
587 * The call to await() can override defaults specified in
588 * the original asleep().
589 */
590 if (priority < 0)
591 priority = p->p_asleep.as_priority;
592 if (timo < 0)
593 timo = p->p_asleep.as_timo;
594
595 /*
596 * Install timeout
597 */
598
599 if (timo)
600 thandle = timeout(endtsleep, (void *)p, timo);
601
602 sig = 0;
603 catch = priority & PCATCH;
604
605 if (catch) {
606 p->p_flag |= P_SINTR;
607 if ((sig = CURSIG(p))) {
608 if (p->p_wchan)
609 unsleep(p);
610 p->p_stat = SRUN;
611 goto resume;
612 }
613 if (p->p_wchan == NULL) {
614 catch = 0;
615 goto resume;
616 }
617 }
618 p->p_stat = SSLEEP;
619 p->p_stats->p_ru.ru_nvcsw++;
620 mi_switch();
621resume:
622 curpriority = p->p_usrpri;
623
624 splx(s);
625 p->p_flag &= ~P_SINTR;
626 if (p->p_flag & P_TIMEOUT) {
627 p->p_flag &= ~P_TIMEOUT;
628 if (sig == 0) {
629#ifdef KTRACE
630 if (KTRPOINT(p, KTR_CSW))
631 ktrcsw(p->p_tracep, 0, 0);
632#endif
633 return (EWOULDBLOCK);
634 }
635 } else if (timo)
636 untimeout(endtsleep, (void *)p, thandle);
637 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
638#ifdef KTRACE
639 if (KTRPOINT(p, KTR_CSW))
640 ktrcsw(p->p_tracep, 0, 0);
641#endif
642 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
643 return (EINTR);
644 return (ERESTART);
645 }
646#ifdef KTRACE
647 if (KTRPOINT(p, KTR_CSW))
648 ktrcsw(p->p_tracep, 0, 0);
649#endif
650 } else {
651 /*
652 * If as_priority is 0, await() has been called without an
653 * intervening asleep(). We are still effectively a NOP,
654 * but we call mi_switch() for safety.
655 */
656
657 if (p->p_asleep.as_priority == 0) {
658 p->p_stats->p_ru.ru_nvcsw++;
659 mi_switch();
660 }
661 splx(s);
662 }
663
664 /*
665 * clear p_asleep.as_priority as an indication that await() has been
666 * called. If await() is called again without an intervening asleep(),
667 * await() is still effectively a NOP but the above mi_switch() code
668 * is triggered as a safety.
669 */
670 p->p_asleep.as_priority = 0;
671
672 return (0);
673}
674
675/*
676 * Implement timeout for tsleep or asleep()/await()
677 *
678 * If process hasn't been awakened (wchan non-zero),
679 * set timeout flag and undo the sleep. If proc
680 * is stopped, just unsleep so it will remain stopped.
681 */
682static void
683endtsleep(arg)
684 void *arg;
685{
686 register struct proc *p;
687 int s;
688
689 p = (struct proc *)arg;
690 s = splhigh();
691 if (p->p_wchan) {
692 if (p->p_stat == SSLEEP)
693 setrunnable(p);
694 else
695 unsleep(p);
696 p->p_flag |= P_TIMEOUT;
697 }
698 splx(s);
699}
700
701/*
702 * Remove a process from its wait queue
703 */
704void
705unsleep(p)
706 register struct proc *p;
707{
708 int s;
709
710 s = splhigh();
711 if (p->p_wchan) {
712 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq);
713 p->p_wchan = 0;
714 }
715 splx(s);
716}
717
718/*
719 * Make all processes sleeping on the specified identifier runnable.
720 */
721void
722wakeup(ident)
723 register void *ident;
724{
725 register struct slpquehead *qp;
726 register struct proc *p;
727 struct proc *np;
728 int s;
729
730 s = splhigh();
731 qp = &slpque[LOOKUP(ident)];
732restart:
733 for (p = TAILQ_FIRST(qp); p != NULL; p = np) {
734 np = TAILQ_NEXT(p, p_procq);
735 if (p->p_wchan == ident) {
736 TAILQ_REMOVE(qp, p, p_procq);
737 p->p_wchan = 0;
738 if (p->p_stat == SSLEEP) {
739 /* OPTIMIZED EXPANSION OF setrunnable(p); */
740 if (p->p_slptime > 1)
741 updatepri(p);
742 p->p_slptime = 0;
743 p->p_stat = SRUN;
744 if (p->p_flag & P_INMEM) {
745 setrunqueue(p);
746 maybe_resched(p);
747 } else {
748 p->p_flag |= P_SWAPINREQ;
749 wakeup((caddr_t)&proc0);
750 }
751 /* END INLINE EXPANSION */
752 goto restart;
753 }
754 }
755 }
756 splx(s);
757}
758
759/*
760 * Make a process sleeping on the specified identifier runnable.
761 * May wake more than one process if a target process is currently
762 * swapped out.
763 */
764void
765wakeup_one(ident)
766 register void *ident;
767{
768 register struct slpquehead *qp;
769 register struct proc *p;
770 struct proc *np;
771 int s;
772
773 s = splhigh();
774 qp = &slpque[LOOKUP(ident)];
775
776restart:
777 for (p = TAILQ_FIRST(qp); p != NULL; p = np) {
778 np = TAILQ_NEXT(p, p_procq);
779 if (p->p_wchan == ident) {
780 TAILQ_REMOVE(qp, p, p_procq);
781 p->p_wchan = 0;
782 if (p->p_stat == SSLEEP) {
783 /* OPTIMIZED EXPANSION OF setrunnable(p); */
784 if (p->p_slptime > 1)
785 updatepri(p);
786 p->p_slptime = 0;
787 p->p_stat = SRUN;
788 if (p->p_flag & P_INMEM) {
789 setrunqueue(p);
790 maybe_resched(p);
791 break;
792 } else {
793 p->p_flag |= P_SWAPINREQ;
794 wakeup((caddr_t)&proc0);
795 }
796 /* END INLINE EXPANSION */
797 goto restart;
798 }
799 }
800 }
801 splx(s);
802}
803
804/*
805 * The machine independent parts of mi_switch().
806 * Must be called at splstatclock() or higher.
807 */
808void
809mi_switch()
810{
811 struct timeval new_switchtime;
812 register struct proc *p = curproc; /* XXX */
813 register struct rlimit *rlim;
814 int x;
815
816 /*
817 * XXX this spl is almost unnecessary. It is partly to allow for
818 * sloppy callers that don't do it (issignal() via CURSIG() is the
819 * main offender). It is partly to work around a bug in the i386
820 * cpu_switch() (the ipl is not preserved). We ran for years
821 * without it. I think there was only a interrupt latency problem.
822 * The main caller, tsleep(), does an splx() a couple of instructions
823 * after calling here. The buggy caller, issignal(), usually calls
824 * here at spl0() and sometimes returns at splhigh(). The process
825 * then runs for a little too long at splhigh(). The ipl gets fixed
826 * when the process returns to user mode (or earlier).
827 *
828 * It would probably be better to always call here at spl0(). Callers
829 * are prepared to give up control to another process, so they must
830 * be prepared to be interrupted. The clock stuff here may not
831 * actually need splstatclock().
832 */
833 x = splstatclock();
8ad65e08 834 clear_resched();
984263bc
MD
835
836#ifdef SIMPLELOCK_DEBUG
837 if (p->p_simple_locks)
838 printf("sleep: holding simple lock\n");
839#endif
840 /*
841 * Compute the amount of time during which the current
842 * process was running, and add that to its total so far.
843 */
844 microuptime(&new_switchtime);
3020e3be 845 if (timevalcmp(&new_switchtime, &mycpu->gd_switchtime, <)) {
984263bc 846 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n",
3020e3be 847 mycpu->gd_switchtime.tv_sec, mycpu->gd_switchtime.tv_usec,
984263bc 848 new_switchtime.tv_sec, new_switchtime.tv_usec);
3020e3be 849 new_switchtime = mycpu->gd_switchtime;
984263bc 850 } else {
3020e3be
MD
851 p->p_runtime +=
852 (new_switchtime.tv_usec - mycpu->gd_switchtime.tv_usec) +
853 (new_switchtime.tv_sec - mycpu->gd_switchtime.tv_sec) *
854 (int64_t)1000000;
984263bc
MD
855 }
856
857 /*
858 * Check if the process exceeds its cpu resource allocation.
859 * If over max, kill it.
860 */
861 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
862 p->p_runtime > p->p_limit->p_cpulimit) {
863 rlim = &p->p_rlimit[RLIMIT_CPU];
864 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) {
865 killproc(p, "exceeded maximum CPU limit");
866 } else {
867 psignal(p, SIGXCPU);
868 if (rlim->rlim_cur < rlim->rlim_max) {
869 /* XXX: we should make a private copy */
870 rlim->rlim_cur += 5;
871 }
872 }
873 }
874
875 /*
876 * Pick a new current process and record its start time.
8ad65e08
MD
877 * YYY lwkt_switch() will run the heavy weight process restoration
878 * code, which removes the target thread and process from their
879 * respective run queues to temporarily mimic 5.x behavior.
880 * YYY the userland scheduler should pick only one user process
881 * at a time to run per cpu.
984263bc
MD
882 */
883 cnt.v_swtch++;
3020e3be 884 mycpu->gd_switchtime = new_switchtime;
8ad65e08
MD
885 lwkt_switch();
886 remrunqueue(p);
3020e3be
MD
887 if (mycpu->gd_switchtime.tv_sec == 0)
888 microuptime(&mycpu->gd_switchtime);
889 mycpu->gd_switchticks = ticks;
984263bc
MD
890
891 splx(x);
892}
893
894/*
895 * Change process state to be runnable,
896 * placing it on the run queue if it is in memory,
897 * and awakening the swapper if it isn't in memory.
898 */
899void
900setrunnable(p)
901 register struct proc *p;
902{
903 register int s;
904
905 s = splhigh();
906 switch (p->p_stat) {
907 case 0:
908 case SRUN:
909 case SZOMB:
910 default:
911 panic("setrunnable");
912 case SSTOP:
913 case SSLEEP:
914 unsleep(p); /* e.g. when sending signals */
915 break;
916
917 case SIDL:
918 break;
919 }
920 p->p_stat = SRUN;
921 if (p->p_flag & P_INMEM)
922 setrunqueue(p);
923 splx(s);
924 if (p->p_slptime > 1)
925 updatepri(p);
926 p->p_slptime = 0;
927 if ((p->p_flag & P_INMEM) == 0) {
928 p->p_flag |= P_SWAPINREQ;
929 wakeup((caddr_t)&proc0);
930 }
931 else
932 maybe_resched(p);
933}
934
935/*
936 * Compute the priority of a process when running in user mode.
937 * Arrange to reschedule if the resulting priority is better
938 * than that of the current process.
939 */
940void
941resetpriority(p)
942 register struct proc *p;
943{
944 register unsigned int newpriority;
945
946 if (p->p_rtprio.type == RTP_PRIO_NORMAL) {
947 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT +
948 NICE_WEIGHT * p->p_nice;
949 newpriority = min(newpriority, MAXPRI);
950 p->p_usrpri = newpriority;
951 }
952 maybe_resched(p);
953}
954
955/*
956 * Compute a tenex style load average of a quantity on
957 * 1, 5 and 15 minute intervals.
958 */
959static void
960loadav(void *arg)
961{
962 int i, nrun;
963 struct loadavg *avg;
964 struct proc *p;
965
966 avg = &averunnable;
967 nrun = 0;
968 LIST_FOREACH(p, &allproc, p_list) {
969 switch (p->p_stat) {
970 case SRUN:
971 case SIDL:
972 nrun++;
973 }
974 }
975 for (i = 0; i < 3; i++)
976 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
977 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
978
979 /*
980 * Schedule the next update to occur after 5 seconds, but add a
981 * random variation to avoid synchronisation with processes that
982 * run at regular intervals.
983 */
984 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
985 loadav, NULL);
986}
987
988/* ARGSUSED */
989static void
990sched_setup(dummy)
991 void *dummy;
992{
993
994 callout_init(&loadav_callout);
995
996 /* Kick off timeout driven events by calling first time. */
997 roundrobin(NULL);
998 schedcpu(NULL);
999 loadav(NULL);
1000}
1001
1002/*
1003 * We adjust the priority of the current process. The priority of
1004 * a process gets worse as it accumulates CPU time. The cpu usage
1005 * estimator (p_estcpu) is increased here. resetpriority() will
1006 * compute a different priority each time p_estcpu increases by
1007 * INVERSE_ESTCPU_WEIGHT
1008 * (until MAXPRI is reached). The cpu usage estimator ramps up
1009 * quite quickly when the process is running (linearly), and decays
1010 * away exponentially, at a rate which is proportionally slower when
1011 * the system is busy. The basic principle is that the system will
1012 * 90% forget that the process used a lot of CPU time in 5 * loadav
1013 * seconds. This causes the system to favor processes which haven't
1014 * run much recently, and to round-robin among other processes.
1015 */
1016void
1017schedclock(p)
1018 struct proc *p;
1019{
1020
1021 p->p_cpticks++;
1022 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
1023 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
1024 resetpriority(p);
1025 if (p->p_priority >= PUSER)
1026 p->p_priority = p->p_usrpri;
1027 }
1028}