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
| 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
|
40 */
41
42#include "opt_ktrace.h"
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/proc.h>
47#include <sys/kernel.h>
48#include <sys/signalvar.h>
49#include <sys/resourcevar.h>
50#include <sys/vmmeter.h>
51#include <sys/sysctl.h>
52#include <vm/vm.h>
53#include <vm/vm_extern.h>
54#ifdef KTRACE
55#include <sys/uio.h>
56#include <sys/ktrace.h>
57#endif
58
59#include <machine/cpu.h>
60#include <machine/limits.h> /* for UCHAR_MAX = typeof(p_priority)_MAX */
61
62static void rqinit __P((void *));
63SYSINIT(runqueue, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, rqinit, NULL)
64
65u_char curpriority; /* usrpri of curproc */
66int lbolt; /* once a second sleep address */
67
68static void endtsleep __P((void *));
69static void roundrobin __P((void *arg));
70static void schedcpu __P((void *arg));
71static void updatepri __P((struct proc *p));
72
73#define MAXIMUM_SCHEDULE_QUANTUM (1000000) /* arbitrary limit */
74#ifndef DEFAULT_SCHEDULE_QUANTUM
75#define DEFAULT_SCHEDULE_QUANTUM 10
76#endif
77static int quantum = DEFAULT_SCHEDULE_QUANTUM; /* default value */
78
79static int
80sysctl_kern_quantum SYSCTL_HANDLER_ARGS
81{
82 int error;
83 int new_val = quantum;
84
85 new_val = quantum;
86 error = sysctl_handle_int(oidp, &new_val, 0, req);
87 if (error == 0) {
88 if ((new_val > 0) && (new_val < MAXIMUM_SCHEDULE_QUANTUM)) {
89 quantum = new_val;
90 } else {
91 error = EINVAL;
92 }
93 }
94 return (error);
95}
96
97SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
98 0, sizeof quantum, sysctl_kern_quantum, "I", "");
99
100/* maybe_resched: Decide if you need to reschedule or not
101 * taking the priorities and schedulers into account.
102 */
103static void maybe_resched(struct proc *chk)
104{
105 struct proc *p = curproc; /* XXX */
106
107 /* If the current scheduler is the idle scheduler or
108 * the priority of the new one is higher then reschedule.
109 */
110 if (p == 0 ||
111 RTP_PRIO_BASE(p->p_rtprio.type) == RTP_PRIO_IDLE ||
112 (chk->p_priority < curpriority &&
113 RTP_PRIO_BASE(p->p_rtprio.type) == RTP_PRIO_BASE(chk->p_rtprio.type)) )
114 need_resched();
115}
116
117#define ROUNDROBIN_INTERVAL (hz / quantum)
118int roundrobin_interval(void)
119{
120 return ROUNDROBIN_INTERVAL;
121}
122
123/*
124 * Force switch among equal priority processes every 100ms.
125 */
126/* ARGSUSED */
127static void
128roundrobin(arg)
129 void *arg;
130{
131 struct proc *p = curproc; /* XXX */
132
133 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type))
134 need_resched();
135
136 timeout(roundrobin, NULL, ROUNDROBIN_INTERVAL);
137}
138
139/*
140 * Constants for digital decay and forget:
141 * 90% of (p_estcpu) usage in 5 * loadav time
142 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
143 * Note that, as ps(1) mentions, this can let percentages
144 * total over 100% (I've seen 137.9% for 3 processes).
145 *
146 * Note that statclock() updates p_estcpu and p_cpticks asynchronously.
147 *
148 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
149 * That is, the system wants to compute a value of decay such
150 * that the following for loop:
151 * for (i = 0; i < (5 * loadavg); i++)
152 * p_estcpu *= decay;
153 * will compute
154 * p_estcpu *= 0.1;
155 * for all values of loadavg:
156 *
157 * Mathematically this loop can be expressed by saying:
158 * decay ** (5 * loadavg) ~= .1
159 *
160 * The system computes decay as:
161 * decay = (2 * loadavg) / (2 * loadavg + 1)
162 *
163 * We wish to prove that the system's computation of decay
164 * will always fulfill the equation:
165 * decay ** (5 * loadavg) ~= .1
166 *
167 * If we compute b as:
168 * b = 2 * loadavg
169 * then
170 * decay = b / (b + 1)
171 *
172 * We now need to prove two things:
173 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
174 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
175 *
176 * Facts:
177 * For x close to zero, exp(x) =~ 1 + x, since
178 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
179 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
180 * For x close to zero, ln(1+x) =~ x, since
181 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
182 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
183 * ln(.1) =~ -2.30
184 *
185 * Proof of (1):
186 * Solve (factor)**(power) =~ .1 given power (5*loadav):
187 * solving for factor,
188 * ln(factor) =~ (-2.30/5*loadav), or
189 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
190 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
191 *
192 * Proof of (2):
193 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
194 * solving for power,
195 * power*ln(b/(b+1)) =~ -2.30, or
196 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
197 *
198 * Actual power values for the implemented algorithm are as follows:
199 * loadav: 1 2 3 4
200 * power: 5.68 10.32 14.94 19.55
201 */
202
203/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
204#define loadfactor(loadav) (2 * (loadav))
205#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
206
207/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
208static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
209
210/*
211 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
212 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
213 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
214 *
215 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
216 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
217 *
218 * If you don't want to bother with the faster/more-accurate formula, you
219 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
220 * (more general) method of calculating the %age of CPU used by a process.
221 */
222#define CCPU_SHIFT 11
223
224/*
225 * Recompute process priorities, every hz ticks.
226 */
227/* ARGSUSED */
228static void
229schedcpu(arg)
230 void *arg;
231{
232 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
233 register struct proc *p;
234 register int s;
235 register unsigned int newcpu;
236
237 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
238 /*
239 * Increment time in/out of memory and sleep time
240 * (if sleeping). We ignore overflow; with 16-bit int's
241 * (remember them?) overflow takes 45 days.
242 */
243 p->p_swtime++;
244 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
245 p->p_slptime++;
246 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
247 /*
248 * If the process has slept the entire second,
249 * stop recalculating its priority until it wakes up.
250 */
251 if (p->p_slptime > 1)
252 continue;
253 s = splhigh(); /* prevent state changes and protect run queue */
254 /*
255 * p_pctcpu is only for ps.
256 */
257#if (FSHIFT >= CCPU_SHIFT)
258 p->p_pctcpu += (hz == 100)?
259 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
260 100 * (((fixpt_t) p->p_cpticks)
261 << (FSHIFT - CCPU_SHIFT)) / hz;
262#else
263 p->p_pctcpu += ((FSCALE - ccpu) *
264 (p->p_cpticks * FSCALE / hz)) >> FSHIFT;
265#endif
266 p->p_cpticks = 0;
267 newcpu = (u_int) decay_cpu(loadfac, p->p_estcpu) + p->p_nice;
268 p->p_estcpu = min(newcpu, UCHAR_MAX);
269 resetpriority(p);
270 if (p->p_priority >= PUSER) {
271#define PPQ (128 / NQS) /* priorities per queue */
272 if ((p != curproc) &&
273#ifdef SMP
274 (u_char)p->p_oncpu == 0xff && /* idle */
275#endif
276 p->p_stat == SRUN &&
277 (p->p_flag & P_INMEM) &&
278 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
279 remrq(p);
280 p->p_priority = p->p_usrpri;
281 setrunqueue(p);
282 } else
283 p->p_priority = p->p_usrpri;
284 }
285 splx(s);
286 }
287 vmmeter();
288 wakeup((caddr_t)&lbolt);
289 timeout(schedcpu, (void *)0, hz);
290}
291
292/*
293 * Recalculate the priority of a process after it has slept for a while.
294 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
295 * least six times the loadfactor will decay p_estcpu to zero.
296 */
297static void
298updatepri(p)
299 register struct proc *p;
300{
301 register unsigned int newcpu = p->p_estcpu;
302 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
303
304 if (p->p_slptime > 5 * loadfac)
305 p->p_estcpu = 0;
306 else {
307 p->p_slptime--; /* the first time was done in schedcpu */
308 while (newcpu && --p->p_slptime)
309 newcpu = (int) decay_cpu(loadfac, newcpu);
310 p->p_estcpu = min(newcpu, UCHAR_MAX);
311 }
312 resetpriority(p);
313}
314
315/*
316 * We're only looking at 7 bits of the address; everything is
317 * aligned to 4, lots of things are aligned to greater powers
318 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
319 */
320#define TABLESIZE 128
321static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE];
322#define LOOKUP(x) (((long)(x) >> 8) & (TABLESIZE - 1))
323
324/*
325 * During autoconfiguration or after a panic, a sleep will simply
326 * lower the priority briefly to allow interrupts, then return.
327 * The priority to be used (safepri) is machine-dependent, thus this
328 * value is initialized and maintained in the machine-dependent layers.
329 * This priority will typically be 0, or the lowest priority
330 * that is safe for use on the interrupt stack; it can be made
331 * higher to block network software interrupts after panics.
332 */
333int safepri;
334
335void
336sleepinit()
337{
338 int i;
339
340 for (i = 0; i < TABLESIZE; i++)
341 TAILQ_INIT(&slpque[i]);
342}
343
344/*
345 * General sleep call. Suspends the current process until a wakeup is
346 * performed on the specified identifier. The process will then be made
347 * runnable with the specified priority. Sleeps at most timo/hz seconds
348 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
349 * before and after sleeping, else signals are not checked. Returns 0 if
350 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
351 * signal needs to be delivered, ERESTART is returned if the current system
352 * call should be restarted if possible, and EINTR is returned if the system
353 * call should be interrupted by the signal (return EINTR).
354 */
355int
356tsleep(ident, priority, wmesg, timo)
357 void *ident;
358 int priority, timo;
359 const char *wmesg;
360{
361 struct proc *p = curproc;
362 int s, sig, catch = priority & PCATCH;
363 struct callout_handle thandle;
364
365#ifdef KTRACE
366 if (KTRPOINT(p, KTR_CSW))
367 ktrcsw(p->p_tracep, 1, 0);
368#endif
369 s = splhigh();
370 if (cold || panicstr) {
371 /*
372 * After a panic, or during autoconfiguration,
373 * just give interrupts a chance, then just return;
374 * don't run any other procs or panic below,
375 * in case this is the idle process and already asleep.
376 */
377 splx(safepri);
378 splx(s);
379 return (0);
380 }
381#ifdef DIAGNOSTIC
382 if(p == NULL)
383 panic("tsleep1");
384 if (ident == NULL || p->p_stat != SRUN)
385 panic("tsleep");
386 /* XXX This is not exhaustive, just the most common case */
387 if ((p->p_procq.tqe_prev != NULL) && (*p->p_procq.tqe_prev == p))
388 panic("sleeping process already on another queue");
389#endif
390 p->p_wchan = ident;
391 p->p_wmesg = wmesg;
392 p->p_slptime = 0;
393 p->p_priority = priority & PRIMASK;
394 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq);
395 if (timo)
396 thandle = timeout(endtsleep, (void *)p, timo);
397 /*
398 * We put ourselves on the sleep queue and start our timeout
399 * before calling CURSIG, as we could stop there, and a wakeup
400 * or a SIGCONT (or both) could occur while we were stopped.
401 * A SIGCONT would cause us to be marked as SSLEEP
402 * without resuming us, thus we must be ready for sleep
403 * when CURSIG is called. If the wakeup happens while we're
404 * stopped, p->p_wchan will be 0 upon return from CURSIG.
405 */
406 if (catch) {
407 p->p_flag |= P_SINTR;
408 if ((sig = CURSIG(p))) {
409 if (p->p_wchan)
410 unsleep(p);
411 p->p_stat = SRUN;
412 goto resume;
413 }
414 if (p->p_wchan == 0) {
415 catch = 0;
416 goto resume;
417 }
418 } else
419 sig = 0;
420 p->p_stat = SSLEEP;
421 p->p_stats->p_ru.ru_nvcsw++;
422 mi_switch();
423resume:
424 curpriority = p->p_usrpri;
425 splx(s);
426 p->p_flag &= ~P_SINTR;
427 if (p->p_flag & P_TIMEOUT) {
428 p->p_flag &= ~P_TIMEOUT;
429 if (sig == 0) {
430#ifdef KTRACE
431 if (KTRPOINT(p, KTR_CSW))
432 ktrcsw(p->p_tracep, 0, 0);
433#endif
434 return (EWOULDBLOCK);
435 }
436 } else if (timo)
437 untimeout(endtsleep, (void *)p, thandle);
438 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
439#ifdef KTRACE
440 if (KTRPOINT(p, KTR_CSW))
441 ktrcsw(p->p_tracep, 0, 0);
442#endif
443 if (p->p_sigacts->ps_sigintr & sigmask(sig))
444 return (EINTR);
445 return (ERESTART);
446 }
447#ifdef KTRACE
448 if (KTRPOINT(p, KTR_CSW))
449 ktrcsw(p->p_tracep, 0, 0);
450#endif
451 return (0);
452}
453
454/*
455 * Implement timeout for tsleep.
456 * If process hasn't been awakened (wchan non-zero),
457 * set timeout flag and undo the sleep. If proc
458 * is stopped, just unsleep so it will remain stopped.
459 */
460static void
461endtsleep(arg)
462 void *arg;
463{
464 register struct proc *p;
465 int s;
466
467 p = (struct proc *)arg;
468 s = splhigh();
469 if (p->p_wchan) {
470 if (p->p_stat == SSLEEP)
471 setrunnable(p);
472 else
473 unsleep(p);
474 p->p_flag |= P_TIMEOUT;
475 }
476 splx(s);
477}
478
479/*
480 * Remove a process from its wait queue
481 */
482void
483unsleep(p)
484 register struct proc *p;
485{
486 int s;
487
488 s = splhigh();
489 if (p->p_wchan) {
490 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq);
491 p->p_wchan = 0;
492 }
493 splx(s);
494}
495
496/*
497 * Make all processes sleeping on the specified identifier runnable.
498 */
499void
500wakeup(ident)
501 register void *ident;
502{
503 register struct slpquehead *qp;
504 register struct proc *p;
505 int s;
506
507 s = splhigh();
508 qp = &slpque[LOOKUP(ident)];
509restart:
510 for (p = qp->tqh_first; p != NULL; p = p->p_procq.tqe_next) {
511#ifdef DIAGNOSTIC
512 if (p->p_stat != SSLEEP && p->p_stat != SSTOP)
513 panic("wakeup");
514#endif
515 if (p->p_wchan == ident) {
516 TAILQ_REMOVE(qp, p, p_procq);
517 p->p_wchan = 0;
518 if (p->p_stat == SSLEEP) {
519 /* OPTIMIZED EXPANSION OF setrunnable(p); */
520 if (p->p_slptime > 1)
521 updatepri(p);
522 p->p_slptime = 0;
523 p->p_stat = SRUN;
524 if (p->p_flag & P_INMEM) {
525 setrunqueue(p);
526 maybe_resched(p);
527 } else {
528 p->p_flag |= P_SWAPINREQ;
529 wakeup((caddr_t)&proc0);
530 }
531 /* END INLINE EXPANSION */
532 goto restart;
533 }
534 }
535 }
536 splx(s);
537}
538
539/*
540 * Make a process sleeping on the specified identifier runnable.
541 * May wake more than one process if a target prcoess is currently
542 * swapped out.
543 */
544void
545wakeup_one(ident)
546 register void *ident;
547{
548 register struct slpquehead *qp;
549 register struct proc *p;
550 int s;
551
552 s = splhigh();
553 qp = &slpque[LOOKUP(ident)];
554
555 for (p = qp->tqh_first; p != NULL; p = p->p_procq.tqe_next) {
556#ifdef DIAGNOSTIC
557 if (p->p_stat != SSLEEP && p->p_stat != SSTOP)
558 panic("wakeup_one");
559#endif
560 if (p->p_wchan == ident) {
561 TAILQ_REMOVE(qp, p, p_procq);
562 p->p_wchan = 0;
563 if (p->p_stat == SSLEEP) {
564 /* OPTIMIZED EXPANSION OF setrunnable(p); */
565 if (p->p_slptime > 1)
566 updatepri(p);
567 p->p_slptime = 0;
568 p->p_stat = SRUN;
569 if (p->p_flag & P_INMEM) {
570 setrunqueue(p);
571 maybe_resched(p);
572 break;
573 } else {
574 p->p_flag |= P_SWAPINREQ;
575 wakeup((caddr_t)&proc0);
576 }
577 /* END INLINE EXPANSION */
578 }
579 }
580 }
581 splx(s);
582}
583
584/*
585 * The machine independent parts of mi_switch().
586 * Must be called at splstatclock() or higher.
587 */
588void
589mi_switch()
590{
591 register struct proc *p = curproc; /* XXX */
592 register struct rlimit *rlim;
593 register long s, u;
594 int x;
595 struct timeval tv;
596
597 /*
598 * XXX this spl is almost unnecessary. It is partly to allow for
599 * sloppy callers that don't do it (issignal() via CURSIG() is the
600 * main offender). It is partly to work around a bug in the i386
601 * cpu_switch() (the ipl is not preserved). We ran for years
602 * without it. I think there was only a interrupt latency problem.
603 * The main caller, tsleep(), does an splx() a couple of instructions
604 * after calling here. The buggy caller, issignal(), usually calls
605 * here at spl0() and sometimes returns at splhigh(). The process
606 * then runs for a little too long at splhigh(). The ipl gets fixed
607 * when the process returns to user mode (or earlier).
608 *
609 * It would probably be better to always call here at spl0(). Callers
610 * are prepared to give up control to another process, so they must
611 * be prepared to be interrupted. The clock stuff here may not
612 * actually need splstatclock().
613 */
614 x = splstatclock();
615
616#ifdef SIMPLELOCK_DEBUG
617 if (p->p_simple_locks)
618 printf("sleep: holding simple lock\n");
619#endif
620 /*
621 * Compute the amount of time during which the current
622 * process was running, and add that to its total so far.
623 */
| 40 */
41
42#include "opt_ktrace.h"
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/proc.h>
47#include <sys/kernel.h>
48#include <sys/signalvar.h>
49#include <sys/resourcevar.h>
50#include <sys/vmmeter.h>
51#include <sys/sysctl.h>
52#include <vm/vm.h>
53#include <vm/vm_extern.h>
54#ifdef KTRACE
55#include <sys/uio.h>
56#include <sys/ktrace.h>
57#endif
58
59#include <machine/cpu.h>
60#include <machine/limits.h> /* for UCHAR_MAX = typeof(p_priority)_MAX */
61
62static void rqinit __P((void *));
63SYSINIT(runqueue, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, rqinit, NULL)
64
65u_char curpriority; /* usrpri of curproc */
66int lbolt; /* once a second sleep address */
67
68static void endtsleep __P((void *));
69static void roundrobin __P((void *arg));
70static void schedcpu __P((void *arg));
71static void updatepri __P((struct proc *p));
72
73#define MAXIMUM_SCHEDULE_QUANTUM (1000000) /* arbitrary limit */
74#ifndef DEFAULT_SCHEDULE_QUANTUM
75#define DEFAULT_SCHEDULE_QUANTUM 10
76#endif
77static int quantum = DEFAULT_SCHEDULE_QUANTUM; /* default value */
78
79static int
80sysctl_kern_quantum SYSCTL_HANDLER_ARGS
81{
82 int error;
83 int new_val = quantum;
84
85 new_val = quantum;
86 error = sysctl_handle_int(oidp, &new_val, 0, req);
87 if (error == 0) {
88 if ((new_val > 0) && (new_val < MAXIMUM_SCHEDULE_QUANTUM)) {
89 quantum = new_val;
90 } else {
91 error = EINVAL;
92 }
93 }
94 return (error);
95}
96
97SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
98 0, sizeof quantum, sysctl_kern_quantum, "I", "");
99
100/* maybe_resched: Decide if you need to reschedule or not
101 * taking the priorities and schedulers into account.
102 */
103static void maybe_resched(struct proc *chk)
104{
105 struct proc *p = curproc; /* XXX */
106
107 /* If the current scheduler is the idle scheduler or
108 * the priority of the new one is higher then reschedule.
109 */
110 if (p == 0 ||
111 RTP_PRIO_BASE(p->p_rtprio.type) == RTP_PRIO_IDLE ||
112 (chk->p_priority < curpriority &&
113 RTP_PRIO_BASE(p->p_rtprio.type) == RTP_PRIO_BASE(chk->p_rtprio.type)) )
114 need_resched();
115}
116
117#define ROUNDROBIN_INTERVAL (hz / quantum)
118int roundrobin_interval(void)
119{
120 return ROUNDROBIN_INTERVAL;
121}
122
123/*
124 * Force switch among equal priority processes every 100ms.
125 */
126/* ARGSUSED */
127static void
128roundrobin(arg)
129 void *arg;
130{
131 struct proc *p = curproc; /* XXX */
132
133 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type))
134 need_resched();
135
136 timeout(roundrobin, NULL, ROUNDROBIN_INTERVAL);
137}
138
139/*
140 * Constants for digital decay and forget:
141 * 90% of (p_estcpu) usage in 5 * loadav time
142 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
143 * Note that, as ps(1) mentions, this can let percentages
144 * total over 100% (I've seen 137.9% for 3 processes).
145 *
146 * Note that statclock() updates p_estcpu and p_cpticks asynchronously.
147 *
148 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
149 * That is, the system wants to compute a value of decay such
150 * that the following for loop:
151 * for (i = 0; i < (5 * loadavg); i++)
152 * p_estcpu *= decay;
153 * will compute
154 * p_estcpu *= 0.1;
155 * for all values of loadavg:
156 *
157 * Mathematically this loop can be expressed by saying:
158 * decay ** (5 * loadavg) ~= .1
159 *
160 * The system computes decay as:
161 * decay = (2 * loadavg) / (2 * loadavg + 1)
162 *
163 * We wish to prove that the system's computation of decay
164 * will always fulfill the equation:
165 * decay ** (5 * loadavg) ~= .1
166 *
167 * If we compute b as:
168 * b = 2 * loadavg
169 * then
170 * decay = b / (b + 1)
171 *
172 * We now need to prove two things:
173 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
174 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
175 *
176 * Facts:
177 * For x close to zero, exp(x) =~ 1 + x, since
178 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
179 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
180 * For x close to zero, ln(1+x) =~ x, since
181 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
182 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
183 * ln(.1) =~ -2.30
184 *
185 * Proof of (1):
186 * Solve (factor)**(power) =~ .1 given power (5*loadav):
187 * solving for factor,
188 * ln(factor) =~ (-2.30/5*loadav), or
189 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
190 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
191 *
192 * Proof of (2):
193 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
194 * solving for power,
195 * power*ln(b/(b+1)) =~ -2.30, or
196 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
197 *
198 * Actual power values for the implemented algorithm are as follows:
199 * loadav: 1 2 3 4
200 * power: 5.68 10.32 14.94 19.55
201 */
202
203/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
204#define loadfactor(loadav) (2 * (loadav))
205#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
206
207/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
208static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
209
210/*
211 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
212 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
213 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
214 *
215 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
216 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
217 *
218 * If you don't want to bother with the faster/more-accurate formula, you
219 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
220 * (more general) method of calculating the %age of CPU used by a process.
221 */
222#define CCPU_SHIFT 11
223
224/*
225 * Recompute process priorities, every hz ticks.
226 */
227/* ARGSUSED */
228static void
229schedcpu(arg)
230 void *arg;
231{
232 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
233 register struct proc *p;
234 register int s;
235 register unsigned int newcpu;
236
237 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
238 /*
239 * Increment time in/out of memory and sleep time
240 * (if sleeping). We ignore overflow; with 16-bit int's
241 * (remember them?) overflow takes 45 days.
242 */
243 p->p_swtime++;
244 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
245 p->p_slptime++;
246 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
247 /*
248 * If the process has slept the entire second,
249 * stop recalculating its priority until it wakes up.
250 */
251 if (p->p_slptime > 1)
252 continue;
253 s = splhigh(); /* prevent state changes and protect run queue */
254 /*
255 * p_pctcpu is only for ps.
256 */
257#if (FSHIFT >= CCPU_SHIFT)
258 p->p_pctcpu += (hz == 100)?
259 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
260 100 * (((fixpt_t) p->p_cpticks)
261 << (FSHIFT - CCPU_SHIFT)) / hz;
262#else
263 p->p_pctcpu += ((FSCALE - ccpu) *
264 (p->p_cpticks * FSCALE / hz)) >> FSHIFT;
265#endif
266 p->p_cpticks = 0;
267 newcpu = (u_int) decay_cpu(loadfac, p->p_estcpu) + p->p_nice;
268 p->p_estcpu = min(newcpu, UCHAR_MAX);
269 resetpriority(p);
270 if (p->p_priority >= PUSER) {
271#define PPQ (128 / NQS) /* priorities per queue */
272 if ((p != curproc) &&
273#ifdef SMP
274 (u_char)p->p_oncpu == 0xff && /* idle */
275#endif
276 p->p_stat == SRUN &&
277 (p->p_flag & P_INMEM) &&
278 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
279 remrq(p);
280 p->p_priority = p->p_usrpri;
281 setrunqueue(p);
282 } else
283 p->p_priority = p->p_usrpri;
284 }
285 splx(s);
286 }
287 vmmeter();
288 wakeup((caddr_t)&lbolt);
289 timeout(schedcpu, (void *)0, hz);
290}
291
292/*
293 * Recalculate the priority of a process after it has slept for a while.
294 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
295 * least six times the loadfactor will decay p_estcpu to zero.
296 */
297static void
298updatepri(p)
299 register struct proc *p;
300{
301 register unsigned int newcpu = p->p_estcpu;
302 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
303
304 if (p->p_slptime > 5 * loadfac)
305 p->p_estcpu = 0;
306 else {
307 p->p_slptime--; /* the first time was done in schedcpu */
308 while (newcpu && --p->p_slptime)
309 newcpu = (int) decay_cpu(loadfac, newcpu);
310 p->p_estcpu = min(newcpu, UCHAR_MAX);
311 }
312 resetpriority(p);
313}
314
315/*
316 * We're only looking at 7 bits of the address; everything is
317 * aligned to 4, lots of things are aligned to greater powers
318 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
319 */
320#define TABLESIZE 128
321static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE];
322#define LOOKUP(x) (((long)(x) >> 8) & (TABLESIZE - 1))
323
324/*
325 * During autoconfiguration or after a panic, a sleep will simply
326 * lower the priority briefly to allow interrupts, then return.
327 * The priority to be used (safepri) is machine-dependent, thus this
328 * value is initialized and maintained in the machine-dependent layers.
329 * This priority will typically be 0, or the lowest priority
330 * that is safe for use on the interrupt stack; it can be made
331 * higher to block network software interrupts after panics.
332 */
333int safepri;
334
335void
336sleepinit()
337{
338 int i;
339
340 for (i = 0; i < TABLESIZE; i++)
341 TAILQ_INIT(&slpque[i]);
342}
343
344/*
345 * General sleep call. Suspends the current process until a wakeup is
346 * performed on the specified identifier. The process will then be made
347 * runnable with the specified priority. Sleeps at most timo/hz seconds
348 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
349 * before and after sleeping, else signals are not checked. Returns 0 if
350 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
351 * signal needs to be delivered, ERESTART is returned if the current system
352 * call should be restarted if possible, and EINTR is returned if the system
353 * call should be interrupted by the signal (return EINTR).
354 */
355int
356tsleep(ident, priority, wmesg, timo)
357 void *ident;
358 int priority, timo;
359 const char *wmesg;
360{
361 struct proc *p = curproc;
362 int s, sig, catch = priority & PCATCH;
363 struct callout_handle thandle;
364
365#ifdef KTRACE
366 if (KTRPOINT(p, KTR_CSW))
367 ktrcsw(p->p_tracep, 1, 0);
368#endif
369 s = splhigh();
370 if (cold || panicstr) {
371 /*
372 * After a panic, or during autoconfiguration,
373 * just give interrupts a chance, then just return;
374 * don't run any other procs or panic below,
375 * in case this is the idle process and already asleep.
376 */
377 splx(safepri);
378 splx(s);
379 return (0);
380 }
381#ifdef DIAGNOSTIC
382 if(p == NULL)
383 panic("tsleep1");
384 if (ident == NULL || p->p_stat != SRUN)
385 panic("tsleep");
386 /* XXX This is not exhaustive, just the most common case */
387 if ((p->p_procq.tqe_prev != NULL) && (*p->p_procq.tqe_prev == p))
388 panic("sleeping process already on another queue");
389#endif
390 p->p_wchan = ident;
391 p->p_wmesg = wmesg;
392 p->p_slptime = 0;
393 p->p_priority = priority & PRIMASK;
394 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq);
395 if (timo)
396 thandle = timeout(endtsleep, (void *)p, timo);
397 /*
398 * We put ourselves on the sleep queue and start our timeout
399 * before calling CURSIG, as we could stop there, and a wakeup
400 * or a SIGCONT (or both) could occur while we were stopped.
401 * A SIGCONT would cause us to be marked as SSLEEP
402 * without resuming us, thus we must be ready for sleep
403 * when CURSIG is called. If the wakeup happens while we're
404 * stopped, p->p_wchan will be 0 upon return from CURSIG.
405 */
406 if (catch) {
407 p->p_flag |= P_SINTR;
408 if ((sig = CURSIG(p))) {
409 if (p->p_wchan)
410 unsleep(p);
411 p->p_stat = SRUN;
412 goto resume;
413 }
414 if (p->p_wchan == 0) {
415 catch = 0;
416 goto resume;
417 }
418 } else
419 sig = 0;
420 p->p_stat = SSLEEP;
421 p->p_stats->p_ru.ru_nvcsw++;
422 mi_switch();
423resume:
424 curpriority = p->p_usrpri;
425 splx(s);
426 p->p_flag &= ~P_SINTR;
427 if (p->p_flag & P_TIMEOUT) {
428 p->p_flag &= ~P_TIMEOUT;
429 if (sig == 0) {
430#ifdef KTRACE
431 if (KTRPOINT(p, KTR_CSW))
432 ktrcsw(p->p_tracep, 0, 0);
433#endif
434 return (EWOULDBLOCK);
435 }
436 } else if (timo)
437 untimeout(endtsleep, (void *)p, thandle);
438 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
439#ifdef KTRACE
440 if (KTRPOINT(p, KTR_CSW))
441 ktrcsw(p->p_tracep, 0, 0);
442#endif
443 if (p->p_sigacts->ps_sigintr & sigmask(sig))
444 return (EINTR);
445 return (ERESTART);
446 }
447#ifdef KTRACE
448 if (KTRPOINT(p, KTR_CSW))
449 ktrcsw(p->p_tracep, 0, 0);
450#endif
451 return (0);
452}
453
454/*
455 * Implement timeout for tsleep.
456 * If process hasn't been awakened (wchan non-zero),
457 * set timeout flag and undo the sleep. If proc
458 * is stopped, just unsleep so it will remain stopped.
459 */
460static void
461endtsleep(arg)
462 void *arg;
463{
464 register struct proc *p;
465 int s;
466
467 p = (struct proc *)arg;
468 s = splhigh();
469 if (p->p_wchan) {
470 if (p->p_stat == SSLEEP)
471 setrunnable(p);
472 else
473 unsleep(p);
474 p->p_flag |= P_TIMEOUT;
475 }
476 splx(s);
477}
478
479/*
480 * Remove a process from its wait queue
481 */
482void
483unsleep(p)
484 register struct proc *p;
485{
486 int s;
487
488 s = splhigh();
489 if (p->p_wchan) {
490 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq);
491 p->p_wchan = 0;
492 }
493 splx(s);
494}
495
496/*
497 * Make all processes sleeping on the specified identifier runnable.
498 */
499void
500wakeup(ident)
501 register void *ident;
502{
503 register struct slpquehead *qp;
504 register struct proc *p;
505 int s;
506
507 s = splhigh();
508 qp = &slpque[LOOKUP(ident)];
509restart:
510 for (p = qp->tqh_first; p != NULL; p = p->p_procq.tqe_next) {
511#ifdef DIAGNOSTIC
512 if (p->p_stat != SSLEEP && p->p_stat != SSTOP)
513 panic("wakeup");
514#endif
515 if (p->p_wchan == ident) {
516 TAILQ_REMOVE(qp, p, p_procq);
517 p->p_wchan = 0;
518 if (p->p_stat == SSLEEP) {
519 /* OPTIMIZED EXPANSION OF setrunnable(p); */
520 if (p->p_slptime > 1)
521 updatepri(p);
522 p->p_slptime = 0;
523 p->p_stat = SRUN;
524 if (p->p_flag & P_INMEM) {
525 setrunqueue(p);
526 maybe_resched(p);
527 } else {
528 p->p_flag |= P_SWAPINREQ;
529 wakeup((caddr_t)&proc0);
530 }
531 /* END INLINE EXPANSION */
532 goto restart;
533 }
534 }
535 }
536 splx(s);
537}
538
539/*
540 * Make a process sleeping on the specified identifier runnable.
541 * May wake more than one process if a target prcoess is currently
542 * swapped out.
543 */
544void
545wakeup_one(ident)
546 register void *ident;
547{
548 register struct slpquehead *qp;
549 register struct proc *p;
550 int s;
551
552 s = splhigh();
553 qp = &slpque[LOOKUP(ident)];
554
555 for (p = qp->tqh_first; p != NULL; p = p->p_procq.tqe_next) {
556#ifdef DIAGNOSTIC
557 if (p->p_stat != SSLEEP && p->p_stat != SSTOP)
558 panic("wakeup_one");
559#endif
560 if (p->p_wchan == ident) {
561 TAILQ_REMOVE(qp, p, p_procq);
562 p->p_wchan = 0;
563 if (p->p_stat == SSLEEP) {
564 /* OPTIMIZED EXPANSION OF setrunnable(p); */
565 if (p->p_slptime > 1)
566 updatepri(p);
567 p->p_slptime = 0;
568 p->p_stat = SRUN;
569 if (p->p_flag & P_INMEM) {
570 setrunqueue(p);
571 maybe_resched(p);
572 break;
573 } else {
574 p->p_flag |= P_SWAPINREQ;
575 wakeup((caddr_t)&proc0);
576 }
577 /* END INLINE EXPANSION */
578 }
579 }
580 }
581 splx(s);
582}
583
584/*
585 * The machine independent parts of mi_switch().
586 * Must be called at splstatclock() or higher.
587 */
588void
589mi_switch()
590{
591 register struct proc *p = curproc; /* XXX */
592 register struct rlimit *rlim;
593 register long s, u;
594 int x;
595 struct timeval tv;
596
597 /*
598 * XXX this spl is almost unnecessary. It is partly to allow for
599 * sloppy callers that don't do it (issignal() via CURSIG() is the
600 * main offender). It is partly to work around a bug in the i386
601 * cpu_switch() (the ipl is not preserved). We ran for years
602 * without it. I think there was only a interrupt latency problem.
603 * The main caller, tsleep(), does an splx() a couple of instructions
604 * after calling here. The buggy caller, issignal(), usually calls
605 * here at spl0() and sometimes returns at splhigh(). The process
606 * then runs for a little too long at splhigh(). The ipl gets fixed
607 * when the process returns to user mode (or earlier).
608 *
609 * It would probably be better to always call here at spl0(). Callers
610 * are prepared to give up control to another process, so they must
611 * be prepared to be interrupted. The clock stuff here may not
612 * actually need splstatclock().
613 */
614 x = splstatclock();
615
616#ifdef SIMPLELOCK_DEBUG
617 if (p->p_simple_locks)
618 printf("sleep: holding simple lock\n");
619#endif
620 /*
621 * Compute the amount of time during which the current
622 * process was running, and add that to its total so far.
623 */
|