MP Implmentation 3A/4: Fix stupid bug introduced in last commit.
[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 $
a2a5ad0d 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.16 2003/07/10 04:47:54 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>
26a0694b 53#include <sys/thread2.h>
984263bc
MD
54#ifdef KTRACE
55#include <sys/uio.h>
56#include <sys/ktrace.h>
57#endif
f1d1c3fa 58#include <sys/xwait.h>
984263bc
MD
59
60#include <machine/cpu.h>
61#include <machine/ipl.h>
62#include <machine/smp.h>
63
64static void sched_setup __P((void *dummy));
65SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL)
66
984263bc
MD
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
984263bc
MD
86static void endtsleep __P((void *));
87static void loadav __P((void *arg));
88static void maybe_resched __P((struct proc *chk));
89static void roundrobin __P((void *arg));
90static void schedcpu __P((void *arg));
91static void updatepri __P((struct proc *p));
8a8d5d85 92static void crit_panicints(void);
984263bc
MD
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
984263bc 113/*
26a0694b
MD
114 * Arrange to reschedule if necessary by checking to see if the current
115 * process is on the highest priority user scheduling queue. This may
116 * be run from an interrupt so we have to follow any preemption chains
117 * back to the original process.
984263bc
MD
118 */
119static void
26a0694b 120maybe_resched(struct proc *chk)
984263bc 121{
26a0694b
MD
122 struct proc *cur = lwkt_preempted_proc();
123
124 if (cur == NULL)
125 return;
984263bc
MD
126
127 /*
26a0694b
MD
128 * Check the user queue (realtime, normal, idle). Lower numbers
129 * indicate higher priority queues. Lower numbers are also better
130 * for p_priority.
984263bc 131 */
26a0694b 132 if (chk->p_rtprio.type < cur->p_rtprio.type) {
984263bc 133 need_resched();
26a0694b
MD
134 } else if (chk->p_rtprio.type == cur->p_rtprio.type) {
135 if (chk->p_rtprio.type == RTP_PRIO_NORMAL) {
136 if (chk->p_priority / PPQ < cur->p_priority / PPQ)
137 need_resched();
138 } else {
139 if (chk->p_rtprio.prio < cur->p_rtprio.prio)
140 need_resched();
141 }
142 }
984263bc
MD
143}
144
145int
146roundrobin_interval(void)
147{
148 return (sched_quantum);
149}
150
151/*
152 * Force switch among equal priority processes every 100ms.
153 */
154/* ARGSUSED */
155static void
156roundrobin(arg)
157 void *arg;
158{
8a8d5d85 159 struct proc *p = lwkt_preempted_proc();
984263bc 160#ifdef SMP
8a8d5d85
MD
161 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
162 need_resched();
984263bc
MD
163 forward_roundrobin();
164#else
8a8d5d85 165 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type))
984263bc
MD
166 need_resched();
167#endif
984263bc
MD
168 timeout(roundrobin, NULL, sched_quantum);
169}
170
171/*
172 * Constants for digital decay and forget:
173 * 90% of (p_estcpu) usage in 5 * loadav time
174 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
175 * Note that, as ps(1) mentions, this can let percentages
176 * total over 100% (I've seen 137.9% for 3 processes).
177 *
178 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously.
179 *
180 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
181 * That is, the system wants to compute a value of decay such
182 * that the following for loop:
183 * for (i = 0; i < (5 * loadavg); i++)
184 * p_estcpu *= decay;
185 * will compute
186 * p_estcpu *= 0.1;
187 * for all values of loadavg:
188 *
189 * Mathematically this loop can be expressed by saying:
190 * decay ** (5 * loadavg) ~= .1
191 *
192 * The system computes decay as:
193 * decay = (2 * loadavg) / (2 * loadavg + 1)
194 *
195 * We wish to prove that the system's computation of decay
196 * will always fulfill the equation:
197 * decay ** (5 * loadavg) ~= .1
198 *
199 * If we compute b as:
200 * b = 2 * loadavg
201 * then
202 * decay = b / (b + 1)
203 *
204 * We now need to prove two things:
205 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
206 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
207 *
208 * Facts:
209 * For x close to zero, exp(x) =~ 1 + x, since
210 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
211 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
212 * For x close to zero, ln(1+x) =~ x, since
213 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
214 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
215 * ln(.1) =~ -2.30
216 *
217 * Proof of (1):
218 * Solve (factor)**(power) =~ .1 given power (5*loadav):
219 * solving for factor,
220 * ln(factor) =~ (-2.30/5*loadav), or
221 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
222 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
223 *
224 * Proof of (2):
225 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
226 * solving for power,
227 * power*ln(b/(b+1)) =~ -2.30, or
228 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
229 *
230 * Actual power values for the implemented algorithm are as follows:
231 * loadav: 1 2 3 4
232 * power: 5.68 10.32 14.94 19.55
233 */
234
235/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
236#define loadfactor(loadav) (2 * (loadav))
237#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
238
239/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
240static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
241SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
242
243/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */
244static int fscale __unused = FSCALE;
245SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, "");
246
247/*
248 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
249 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
250 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
251 *
252 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
253 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
254 *
255 * If you don't want to bother with the faster/more-accurate formula, you
256 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
257 * (more general) method of calculating the %age of CPU used by a process.
258 */
259#define CCPU_SHIFT 11
260
261/*
262 * Recompute process priorities, every hz ticks.
263 */
264/* ARGSUSED */
265static void
26a0694b 266schedcpu(void *arg)
984263bc 267{
4b5f931b
MD
268 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
269 struct proc *p;
270 struct proc *curp;
271 int realstathz, s;
272
273 curp = lwkt_preempted_proc(); /* YYY temporary hack */
984263bc
MD
274
275 realstathz = stathz ? stathz : hz;
276 LIST_FOREACH(p, &allproc, p_list) {
277 /*
278 * Increment time in/out of memory and sleep time
279 * (if sleeping). We ignore overflow; with 16-bit int's
280 * (remember them?) overflow takes 45 days.
281 */
282 p->p_swtime++;
283 if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
284 p->p_slptime++;
285 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
286 /*
287 * If the process has slept the entire second,
288 * stop recalculating its priority until it wakes up.
289 */
290 if (p->p_slptime > 1)
291 continue;
292 s = splhigh(); /* prevent state changes and protect run queue */
293 /*
294 * p_pctcpu is only for ps.
295 */
296#if (FSHIFT >= CCPU_SHIFT)
297 p->p_pctcpu += (realstathz == 100)?
298 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
299 100 * (((fixpt_t) p->p_cpticks)
300 << (FSHIFT - CCPU_SHIFT)) / realstathz;
301#else
302 p->p_pctcpu += ((FSCALE - ccpu) *
303 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT;
304#endif
305 p->p_cpticks = 0;
306 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu);
307 resetpriority(p);
984263bc
MD
308 splx(s);
309 }
310 wakeup((caddr_t)&lbolt);
311 timeout(schedcpu, (void *)0, hz);
312}
313
314/*
315 * Recalculate the priority of a process after it has slept for a while.
316 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at
317 * least six times the loadfactor will decay p_estcpu to zero.
318 */
319static void
26a0694b 320updatepri(struct proc *p)
984263bc 321{
26a0694b
MD
322 unsigned int newcpu = p->p_estcpu;
323 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
984263bc 324
26a0694b 325 if (p->p_slptime > 5 * loadfac) {
984263bc 326 p->p_estcpu = 0;
26a0694b 327 } else {
984263bc
MD
328 p->p_slptime--; /* the first time was done in schedcpu */
329 while (newcpu && --p->p_slptime)
330 newcpu = decay_cpu(loadfac, newcpu);
331 p->p_estcpu = newcpu;
332 }
333 resetpriority(p);
334}
335
336/*
337 * We're only looking at 7 bits of the address; everything is
338 * aligned to 4, lots of things are aligned to greater powers
339 * of 2. Shift right by 8, i.e. drop the bottom 256 worth.
340 */
341#define TABLESIZE 128
0cfcada1 342static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE];
984263bc
MD
343#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1))
344
345/*
346 * During autoconfiguration or after a panic, a sleep will simply
347 * lower the priority briefly to allow interrupts, then return.
348 * The priority to be used (safepri) is machine-dependent, thus this
349 * value is initialized and maintained in the machine-dependent layers.
350 * This priority will typically be 0, or the lowest priority
351 * that is safe for use on the interrupt stack; it can be made
352 * higher to block network software interrupts after panics.
353 */
354int safepri;
355
356void
357sleepinit(void)
358{
359 int i;
360
361 sched_quantum = hz/10;
362 hogticks = 2 * sched_quantum;
363 for (i = 0; i < TABLESIZE; i++)
364 TAILQ_INIT(&slpque[i]);
365}
366
367/*
368 * General sleep call. Suspends the current process until a wakeup is
369 * performed on the specified identifier. The process will then be made
370 * runnable with the specified priority. Sleeps at most timo/hz seconds
371 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
372 * before and after sleeping, else signals are not checked. Returns 0 if
373 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
374 * signal needs to be delivered, ERESTART is returned if the current system
375 * call should be restarted if possible, and EINTR is returned if the system
376 * call should be interrupted by the signal (return EINTR).
26a0694b
MD
377 *
378 * If the process has P_CURPROC set mi_switch() will not re-queue it to
379 * the userland scheduler queues because we are in a SSLEEP state. If
380 * we are not the current process then we have to remove ourselves from
381 * the scheduler queues.
382 *
383 * YYY priority now unused
984263bc
MD
384 */
385int
386tsleep(ident, priority, wmesg, timo)
387 void *ident;
388 int priority, timo;
389 const char *wmesg;
390{
dadab5e9 391 struct thread *td = curthread;
0cfcada1
MD
392 struct proc *p = td->td_proc; /* may be NULL */
393 int s, sig = 0, catch = priority & PCATCH;
f1d1c3fa 394 int id = LOOKUP(ident);
984263bc
MD
395 struct callout_handle thandle;
396
0cfcada1
MD
397 /*
398 * NOTE: removed KTRPOINT, it could cause races due to blocking
399 * even in stable. Just scrap it for now.
400 */
984263bc
MD
401 if (cold || panicstr) {
402 /*
403 * After a panic, or during autoconfiguration,
404 * just give interrupts a chance, then just return;
405 * don't run any other procs or panic below,
406 * in case this is the idle process and already asleep.
407 */
8a8d5d85 408 crit_panicints();
984263bc
MD
409 return (0);
410 }
a2a5ad0d 411 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */
8a8d5d85 412 s = splhigh();
0cfcada1
MD
413 KASSERT(ident != NULL, ("tsleep: no ident"));
414 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d",
415 ident, wmesg, p->p_stat));
416
26a0694b 417 crit_enter();
0cfcada1
MD
418 td->td_wchan = ident;
419 td->td_wmesg = wmesg;
26a0694b 420 if (p)
0cfcada1 421 p->p_slptime = 0;
0cfcada1
MD
422 lwkt_deschedule_self();
423 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq);
984263bc 424 if (timo)
0cfcada1 425 thandle = timeout(endtsleep, (void *)td, timo);
984263bc
MD
426 /*
427 * We put ourselves on the sleep queue and start our timeout
428 * before calling CURSIG, as we could stop there, and a wakeup
429 * or a SIGCONT (or both) could occur while we were stopped.
430 * A SIGCONT would cause us to be marked as SSLEEP
431 * without resuming us, thus we must be ready for sleep
432 * when CURSIG is called. If the wakeup happens while we're
a2a5ad0d 433 * stopped, td->td_wchan will be 0 upon return from CURSIG.
984263bc 434 */
0cfcada1
MD
435 if (p) {
436 if (catch) {
437 p->p_flag |= P_SINTR;
438 if ((sig = CURSIG(p))) {
26a0694b 439 if (td->td_wchan) {
0cfcada1 440 unsleep(td);
26a0694b
MD
441 lwkt_schedule_self();
442 }
0cfcada1
MD
443 p->p_stat = SRUN;
444 goto resume;
445 }
a2a5ad0d 446 if (td->td_wchan == NULL) {
0cfcada1
MD
447 catch = 0;
448 goto resume;
449 }
450 } else {
451 sig = 0;
984263bc 452 }
26a0694b
MD
453
454 /*
455 * If we are not the current process we have to remove ourself
456 * from the run queue.
457 */
458 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat));
459 /*
460 * If this is the current 'user' process schedule another one.
461 */
462 clrrunnable(p, SSLEEP);
0cfcada1 463 p->p_stats->p_ru.ru_nvcsw++;
a2a5ad0d 464 KKASSERT(td->td_release || (p->p_flag & P_CURPROC) == 0);
0cfcada1 465 mi_switch();
26a0694b 466 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun"));
0cfcada1
MD
467 } else {
468 lwkt_switch();
469 }
984263bc 470resume:
26a0694b
MD
471 crit_exit();
472 if (p)
0cfcada1 473 p->p_flag &= ~P_SINTR;
984263bc 474 splx(s);
0cfcada1
MD
475 if (td->td_flags & TDF_TIMEOUT) {
476 td->td_flags &= ~TDF_TIMEOUT;
477 if (sig == 0)
984263bc 478 return (EWOULDBLOCK);
0cfcada1
MD
479 } else if (timo) {
480 untimeout(endtsleep, (void *)td, thandle);
481 }
482 if (p) {
483 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
484 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
485 return (EINTR);
486 return (ERESTART);
984263bc 487 }
984263bc 488 }
984263bc
MD
489 return (0);
490}
491
0cfcada1
MD
492#if 0
493
984263bc 494/*
f1d1c3fa
MD
495 * General sleep call. Suspends the current process until a wakeup is
496 * performed on the specified xwait structure. The process will then be made
497 * runnable with the specified priority. Sleeps at most timo/hz seconds
498 * (0 means no timeout). If pri includes PCATCH flag, signals are checked
499 * before and after sleeping, else signals are not checked. Returns 0 if
500 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
501 * signal needs to be delivered, ERESTART is returned if the current system
502 * call should be restarted if possible, and EINTR is returned if the system
503 * call should be interrupted by the signal (return EINTR).
984263bc 504 *
f1d1c3fa
MD
505 * If the passed generation number is different from the generation number
506 * in the xwait, return immediately.
984263bc 507 */
984263bc 508int
f1d1c3fa 509xsleep(struct xwait *w, int priority, const char *wmesg, int timo, int *gen)
984263bc 510{
dadab5e9
MD
511 struct thread *td = curthread;
512 struct proc *p = td->td_proc;
f1d1c3fa
MD
513 int s, sig, catch = priority & PCATCH;
514 struct callout_handle thandle;
984263bc 515
f1d1c3fa 516#ifdef KTRACE
dadab5e9 517 if (KTRPOINT(td, KTR_CSW))
f1d1c3fa
MD
518 ktrcsw(p->p_tracep, 1, 0);
519#endif
f1d1c3fa 520 if (cold || panicstr) {
984263bc 521 /*
f1d1c3fa
MD
522 * After a panic, or during autoconfiguration,
523 * just give interrupts a chance, then just return;
524 * don't run any other procs or panic below,
525 * in case this is the idle process and already asleep.
984263bc 526 */
8a8d5d85 527 crit_panicints();
f1d1c3fa
MD
528 return (0);
529 }
8a8d5d85 530 s = splhigh();
7d0bac62
MD
531 KASSERT(p != NULL, ("xsleep1"));
532 KASSERT(w != NULL && p->p_stat == SRUN, ("xsleep"));
984263bc 533
f1d1c3fa
MD
534 /*
535 * If the generation number does not match we return immediately.
536 */
537 if (*gen != w->gen) {
538 *gen = w->gen;
984263bc 539 splx(s);
984263bc 540#ifdef KTRACE
dadab5e9 541 if (KTRPOINT(td, KTR_CSW))
f1d1c3fa 542 ktrcsw(p->p_tracep, 0, 0);
984263bc 543#endif
f1d1c3fa
MD
544 return(0);
545 }
546
547 p->p_wchan = w;
548 p->p_wmesg = wmesg;
549 p->p_slptime = 0;
f1d1c3fa
MD
550 p->p_flag |= P_XSLEEP;
551 TAILQ_INSERT_TAIL(&w->waitq, p, p_procq);
552 if (timo)
553 thandle = timeout(endtsleep, (void *)p, timo);
554 /*
555 * We put ourselves on the sleep queue and start our timeout
556 * before calling CURSIG, as we could stop there, and a wakeup
557 * or a SIGCONT (or both) could occur while we were stopped.
558 * A SIGCONT would cause us to be marked as SSLEEP
559 * without resuming us, thus we must be ready for sleep
560 * when CURSIG is called. If the wakeup happens while we're
561 * stopped, p->p_wchan will be 0 upon return from CURSIG.
562 */
563 if (catch) {
564 p->p_flag |= P_SINTR;
565 if ((sig = CURSIG(p))) {
26a0694b 566 if (p->p_wchan) {
f1d1c3fa 567 unsleep(p);
26a0694b
MD
568 lwkt_schedule_self();
569 }
f1d1c3fa
MD
570 p->p_stat = SRUN;
571 goto resume;
572 }
573 if (p->p_wchan == NULL) {
574 catch = 0;
575 goto resume;
576 }
26a0694b 577 } else {
f1d1c3fa 578 sig = 0;
26a0694b
MD
579 }
580 clrrunnable(p, SSLEEP);
f1d1c3fa
MD
581 p->p_stats->p_ru.ru_nvcsw++;
582 mi_switch();
583resume:
f1d1c3fa
MD
584 *gen = w->gen; /* update generation number */
585 splx(s);
586 p->p_flag &= ~P_SINTR;
587 if (p->p_flag & P_TIMEOUT) {
588 p->p_flag &= ~P_TIMEOUT;
589 if (sig == 0) {
984263bc 590#ifdef KTRACE
dadab5e9 591 if (KTRPOINT(td, KTR_CSW))
984263bc
MD
592 ktrcsw(p->p_tracep, 0, 0);
593#endif
f1d1c3fa 594 return (EWOULDBLOCK);
984263bc 595 }
f1d1c3fa
MD
596 } else if (timo)
597 untimeout(endtsleep, (void *)p, thandle);
598 if (catch && (sig != 0 || (sig = CURSIG(p)))) {
984263bc 599#ifdef KTRACE
dadab5e9 600 if (KTRPOINT(td, KTR_CSW))
984263bc
MD
601 ktrcsw(p->p_tracep, 0, 0);
602#endif
f1d1c3fa
MD
603 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig))
604 return (EINTR);
605 return (ERESTART);
984263bc 606 }
f1d1c3fa 607#ifdef KTRACE
dadab5e9 608 if (KTRPOINT(td, KTR_CSW))
f1d1c3fa
MD
609 ktrcsw(p->p_tracep, 0, 0);
610#endif
984263bc
MD
611 return (0);
612}
613
0cfcada1
MD
614#endif
615
984263bc 616/*
0cfcada1
MD
617 * Implement the timeout for tsleep. We interlock against
618 * wchan when setting TDF_TIMEOUT. For processes we remove
619 * the sleep if the process is stopped rather then sleeping,
620 * so it remains stopped.
984263bc
MD
621 */
622static void
0cfcada1 623endtsleep(void *arg)
984263bc 624{
0cfcada1
MD
625 thread_t td = arg;
626 struct proc *p;
984263bc
MD
627 int s;
628
984263bc 629 s = splhigh();
0cfcada1
MD
630 if (td->td_wchan) {
631 td->td_flags |= TDF_TIMEOUT;
632 if ((p = td->td_proc) != NULL) {
633 if (p->p_stat == SSLEEP)
634 setrunnable(p);
635 else
636 unsleep(td);
637 } else {
638 unsleep(td);
639 lwkt_schedule(td);
640 }
984263bc
MD
641 }
642 splx(s);
643}
644
645/*
646 * Remove a process from its wait queue
647 */
648void
0cfcada1 649unsleep(struct thread *td)
984263bc
MD
650{
651 int s;
652
653 s = splhigh();
0cfcada1
MD
654 if (td->td_wchan) {
655#if 0
f1d1c3fa
MD
656 if (p->p_flag & P_XSLEEP) {
657 struct xwait *w = p->p_wchan;
658 TAILQ_REMOVE(&w->waitq, p, p_procq);
659 p->p_flag &= ~P_XSLEEP;
0cfcada1
MD
660 } else
661#endif
662 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq);
663 td->td_wchan = NULL;
f1d1c3fa
MD
664 }
665 splx(s);
666}
667
0cfcada1 668#if 0
f1d1c3fa
MD
669/*
670 * Make all processes sleeping on the explicit lock structure runnable.
671 */
672void
673xwakeup(struct xwait *w)
674{
675 struct proc *p;
676 int s;
677
678 s = splhigh();
679 ++w->gen;
680 while ((p = TAILQ_FIRST(&w->waitq)) != NULL) {
681 TAILQ_REMOVE(&w->waitq, p, p_procq);
682 KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP),
683 ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP));
684 p->p_wchan = NULL;
685 p->p_flag &= ~P_XSLEEP;
686 if (p->p_stat == SSLEEP) {
687 /* OPTIMIZED EXPANSION OF setrunnable(p); */
688 if (p->p_slptime > 1)
689 updatepri(p);
690 p->p_slptime = 0;
691 p->p_stat = SRUN;
692 if (p->p_flag & P_INMEM) {
693 setrunqueue(p);
694 maybe_resched(p);
695 } else {
696 p->p_flag |= P_SWAPINREQ;
697 wakeup((caddr_t)&proc0);
698 }
699 }
984263bc
MD
700 }
701 splx(s);
702}
0cfcada1 703#endif
984263bc
MD
704
705/*
706 * Make all processes sleeping on the specified identifier runnable.
707 */
0cfcada1
MD
708static void
709_wakeup(void *ident, int count)
984263bc 710{
0cfcada1
MD
711 struct slpquehead *qp;
712 struct thread *td;
713 struct thread *ntd;
714 struct proc *p;
984263bc 715 int s;
f1d1c3fa 716 int id = LOOKUP(ident);
984263bc
MD
717
718 s = splhigh();
f1d1c3fa 719 qp = &slpque[id];
984263bc 720restart:
0cfcada1
MD
721 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) {
722 ntd = TAILQ_NEXT(td, td_threadq);
723 if (td->td_wchan == ident) {
724 TAILQ_REMOVE(qp, td, td_threadq);
725 td->td_wchan = NULL;
726 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) {
984263bc
MD
727 /* OPTIMIZED EXPANSION OF setrunnable(p); */
728 if (p->p_slptime > 1)
729 updatepri(p);
730 p->p_slptime = 0;
731 p->p_stat = SRUN;
732 if (p->p_flag & P_INMEM) {
733 setrunqueue(p);
a2a5ad0d
MD
734 if (p->p_flag & P_CURPROC)
735 maybe_resched(p);
984263bc
MD
736 } else {
737 p->p_flag |= P_SWAPINREQ;
738 wakeup((caddr_t)&proc0);
739 }
740 /* END INLINE EXPANSION */
0cfcada1
MD
741 } else if (p == NULL) {
742 lwkt_schedule(td);
984263bc 743 }
0cfcada1
MD
744 if (--count == 0)
745 break;
746 goto restart;
984263bc
MD
747 }
748 }
749 splx(s);
750}
751
984263bc 752void
0cfcada1 753wakeup(void *ident)
984263bc 754{
0cfcada1
MD
755 _wakeup(ident, 0);
756}
984263bc 757
0cfcada1
MD
758void
759wakeup_one(void *ident)
760{
761 _wakeup(ident, 1);
984263bc
MD
762}
763
764/*
765 * The machine independent parts of mi_switch().
766 * Must be called at splstatclock() or higher.
767 */
768void
769mi_switch()
770{
d16a8831
MD
771 struct thread *td = curthread;
772 struct proc *p = td->td_proc; /* XXX */
773 struct rlimit *rlim;
984263bc 774 int x;
d16a8831 775 u_int64_t ttime;
984263bc
MD
776
777 /*
778 * XXX this spl is almost unnecessary. It is partly to allow for
779 * sloppy callers that don't do it (issignal() via CURSIG() is the
780 * main offender). It is partly to work around a bug in the i386
781 * cpu_switch() (the ipl is not preserved). We ran for years
782 * without it. I think there was only a interrupt latency problem.
783 * The main caller, tsleep(), does an splx() a couple of instructions
784 * after calling here. The buggy caller, issignal(), usually calls
785 * here at spl0() and sometimes returns at splhigh(). The process
786 * then runs for a little too long at splhigh(). The ipl gets fixed
787 * when the process returns to user mode (or earlier).
788 *
789 * It would probably be better to always call here at spl0(). Callers
790 * are prepared to give up control to another process, so they must
791 * be prepared to be interrupted. The clock stuff here may not
792 * actually need splstatclock().
793 */
794 x = splstatclock();
8ad65e08 795 clear_resched();
984263bc 796
984263bc
MD
797 /*
798 * Check if the process exceeds its cpu resource allocation.
d16a8831
MD
799 * If over max, kill it. Time spent in interrupts is not
800 * included. YYY 64 bit match is expensive. Ick.
984263bc 801 */
d16a8831 802 ttime = td->td_sticks + td->td_uticks;
984263bc 803 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY &&
d16a8831 804 ttime > p->p_limit->p_cpulimit) {
984263bc 805 rlim = &p->p_rlimit[RLIMIT_CPU];
d16a8831 806 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) {
984263bc
MD
807 killproc(p, "exceeded maximum CPU limit");
808 } else {
809 psignal(p, SIGXCPU);
810 if (rlim->rlim_cur < rlim->rlim_max) {
811 /* XXX: we should make a private copy */
812 rlim->rlim_cur += 5;
813 }
814 }
815 }
816
817 /*
a2a5ad0d
MD
818 * Pick a new current process and record its start time. If we
819 * are in a SSTOPped state we deschedule ourselves. YYY this needs
820 * to be cleaned up, remember that LWKTs stay on their run queue
821 * which works differently then the user scheduler which removes
822 * the process from the runq when it runs it.
984263bc 823 */
12e4aaff 824 mycpu->gd_cnt.v_swtch++;
a2a5ad0d
MD
825 if (p->p_stat == SSTOP)
826 lwkt_deschedule_self();
8ad65e08 827 lwkt_switch();
984263bc
MD
828
829 splx(x);
830}
831
832/*
833 * Change process state to be runnable,
834 * placing it on the run queue if it is in memory,
835 * and awakening the swapper if it isn't in memory.
836 */
837void
0cfcada1 838setrunnable(struct proc *p)
984263bc 839{
0cfcada1 840 int s;
984263bc
MD
841
842 s = splhigh();
843 switch (p->p_stat) {
844 case 0:
845 case SRUN:
846 case SZOMB:
847 default:
848 panic("setrunnable");
849 case SSTOP:
850 case SSLEEP:
0cfcada1 851 unsleep(p->p_thread); /* e.g. when sending signals */
984263bc
MD
852 break;
853
854 case SIDL:
855 break;
856 }
857 p->p_stat = SRUN;
858 if (p->p_flag & P_INMEM)
859 setrunqueue(p);
860 splx(s);
861 if (p->p_slptime > 1)
862 updatepri(p);
863 p->p_slptime = 0;
864 if ((p->p_flag & P_INMEM) == 0) {
865 p->p_flag |= P_SWAPINREQ;
866 wakeup((caddr_t)&proc0);
26a0694b 867 } else {
984263bc 868 maybe_resched(p);
26a0694b
MD
869 }
870}
871
872/*
873 * Change the process state to NOT be runnable, removing it from the run
874 * queue. If P_CURPROC is not set and we are in SRUN the process is on the
875 * run queue (If P_INMEM is not set then it isn't because it is swapped).
876 */
877void
878clrrunnable(struct proc *p, int stat)
879{
880 int s;
881
882 s = splhigh();
883 switch(p->p_stat) {
884 case SRUN:
a2a5ad0d 885 if (p->p_flag & P_ONRUNQ)
26a0694b
MD
886 remrunqueue(p);
887 break;
888 default:
889 break;
890 }
891 p->p_stat = stat;
892 splx(s);
893}
894
984263bc
MD
895/*
896 * Compute the priority of a process when running in user mode.
897 * Arrange to reschedule if the resulting priority is better
898 * than that of the current process.
26a0694b
MD
899 *
900 * YYY real time / idle procs do not use p_priority XXX
984263bc
MD
901 */
902void
26a0694b 903resetpriority(struct proc *p)
984263bc 904{
26a0694b
MD
905 unsigned int newpriority;
906 int opq;
907 int npq;
908
909 if (p->p_rtprio.type != RTP_PRIO_NORMAL)
910 return;
911 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT +
912 NICE_WEIGHT * p->p_nice;
913 newpriority = min(newpriority, MAXPRI);
26a0694b 914 npq = newpriority / PPQ;
d6dd2af9
MD
915 crit_enter();
916 opq = p->p_priority / PPQ;
a2a5ad0d 917 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) {
26a0694b
MD
918 /*
919 * We have to move the process to another queue
920 */
921 remrunqueue(p);
922 p->p_priority = newpriority;
923 setrunqueue(p);
924 } else {
925 /*
a2a5ad0d
MD
926 * We can just adjust the priority and it will be picked
927 * up later.
26a0694b 928 */
a2a5ad0d 929 KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0);
26a0694b 930 p->p_priority = newpriority;
984263bc 931 }
d6dd2af9 932 crit_exit();
984263bc
MD
933 maybe_resched(p);
934}
935
936/*
937 * Compute a tenex style load average of a quantity on
938 * 1, 5 and 15 minute intervals.
939 */
940static void
941loadav(void *arg)
942{
943 int i, nrun;
944 struct loadavg *avg;
945 struct proc *p;
946
947 avg = &averunnable;
948 nrun = 0;
949 LIST_FOREACH(p, &allproc, p_list) {
950 switch (p->p_stat) {
951 case SRUN:
952 case SIDL:
953 nrun++;
954 }
955 }
956 for (i = 0; i < 3; i++)
957 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] +
958 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT;
959
960 /*
961 * Schedule the next update to occur after 5 seconds, but add a
962 * random variation to avoid synchronisation with processes that
963 * run at regular intervals.
964 */
965 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)),
966 loadav, NULL);
967}
968
969/* ARGSUSED */
970static void
971sched_setup(dummy)
972 void *dummy;
973{
974
975 callout_init(&loadav_callout);
976
977 /* Kick off timeout driven events by calling first time. */
978 roundrobin(NULL);
979 schedcpu(NULL);
980 loadav(NULL);
981}
982
983/*
984 * We adjust the priority of the current process. The priority of
985 * a process gets worse as it accumulates CPU time. The cpu usage
986 * estimator (p_estcpu) is increased here. resetpriority() will
987 * compute a different priority each time p_estcpu increases by
988 * INVERSE_ESTCPU_WEIGHT
989 * (until MAXPRI is reached). The cpu usage estimator ramps up
990 * quite quickly when the process is running (linearly), and decays
991 * away exponentially, at a rate which is proportionally slower when
992 * the system is busy. The basic principle is that the system will
993 * 90% forget that the process used a lot of CPU time in 5 * loadav
994 * seconds. This causes the system to favor processes which haven't
995 * run much recently, and to round-robin among other processes.
996 */
997void
998schedclock(p)
999 struct proc *p;
1000{
1001
1002 p->p_cpticks++;
1003 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
26a0694b 1004 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0)
984263bc 1005 resetpriority(p);
984263bc 1006}
8a8d5d85
MD
1007
1008static
1009void
1010crit_panicints(void)
1011{
1012 int s;
1013 int cpri;
1014
1015 s = splhigh();
1016 cpri = crit_panic_save();
1017 splx(safepri);
1018 crit_panic_restore(cpri);
1019 splx(s);
1020}
1021