FreeBSD/Linux Kernel Cross Reference
sys/kern/sched_4bsd.c
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 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 */
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37
38 #include "opt_hwpmc_hooks.h"
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/ktr.h>
44 #include <sys/lock.h>
45 #include <sys/kthread.h>
46 #include <sys/mutex.h>
47 #include <sys/proc.h>
48 #include <sys/resourcevar.h>
49 #include <sys/sched.h>
50 #include <sys/smp.h>
51 #include <sys/sysctl.h>
52 #include <sys/sx.h>
53 #include <sys/turnstile.h>
54 #include <sys/umtx.h>
55 #include <machine/pcb.h>
56 #include <machine/smp.h>
57
58 #ifdef HWPMC_HOOKS
59 #include <sys/pmckern.h>
60 #endif
61
62 /*
63 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
64 * the range 100-256 Hz (approximately).
65 */
66 #define ESTCPULIM(e) \
67 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
68 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
69 #ifdef SMP
70 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
71 #else
72 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
73 #endif
74 #define NICE_WEIGHT 1 /* Priorities per nice level. */
75
76 /*
77 * The schedulable entity that runs a context.
78 * This is an extension to the thread structure and is tailored to
79 * the requirements of this scheduler
80 */
81 struct td_sched {
82 TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */
83 struct thread *ts_thread; /* (*) Active associated thread. */
84 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
85 u_char ts_rqindex; /* (j) Run queue index. */
86 int ts_cpticks; /* (j) Ticks of cpu time. */
87 int ts_slptime; /* (j) Seconds !RUNNING. */
88 struct runq *ts_runq; /* runq the thread is currently on */
89 };
90
91 /* flags kept in td_flags */
92 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
93 #define TDF_EXIT TDF_SCHED1 /* thread is being killed. */
94 #define TDF_BOUND TDF_SCHED2
95
96 #define ts_flags ts_thread->td_flags
97 #define TSF_DIDRUN TDF_DIDRUN /* thread actually ran. */
98 #define TSF_EXIT TDF_EXIT /* thread is being killed. */
99 #define TSF_BOUND TDF_BOUND /* stuck to one CPU */
100
101 #define SKE_RUNQ_PCPU(ts) \
102 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
103
104 static struct td_sched td_sched0;
105 struct mtx sched_lock;
106
107 static int sched_tdcnt; /* Total runnable threads in the system. */
108 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
109 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
110
111 static struct callout roundrobin_callout;
112
113 static void setup_runqs(void);
114 static void roundrobin(void *arg);
115 static void schedcpu(void);
116 static void schedcpu_thread(void);
117 static void sched_priority(struct thread *td, u_char prio);
118 static void sched_setup(void *dummy);
119 static void maybe_resched(struct thread *td);
120 static void updatepri(struct thread *td);
121 static void resetpriority(struct thread *td);
122 static void resetpriority_thread(struct thread *td);
123 #ifdef SMP
124 static int forward_wakeup(int cpunum);
125 #endif
126
127 static struct kproc_desc sched_kp = {
128 "schedcpu",
129 schedcpu_thread,
130 NULL
131 };
132 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp)
133 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
134
135 /*
136 * Global run queue.
137 */
138 static struct runq runq;
139
140 #ifdef SMP
141 /*
142 * Per-CPU run queues
143 */
144 static struct runq runq_pcpu[MAXCPU];
145 #endif
146
147 static void
148 setup_runqs(void)
149 {
150 #ifdef SMP
151 int i;
152
153 for (i = 0; i < MAXCPU; ++i)
154 runq_init(&runq_pcpu[i]);
155 #endif
156
157 runq_init(&runq);
158 }
159
160 static int
161 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
162 {
163 int error, new_val;
164
165 new_val = sched_quantum * tick;
166 error = sysctl_handle_int(oidp, &new_val, 0, req);
167 if (error != 0 || req->newptr == NULL)
168 return (error);
169 if (new_val < tick)
170 return (EINVAL);
171 sched_quantum = new_val / tick;
172 hogticks = 2 * sched_quantum;
173 return (0);
174 }
175
176 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
177
178 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
179 "Scheduler name");
180
181 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
182 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
183 "Roundrobin scheduling quantum in microseconds");
184
185 #ifdef SMP
186 /* Enable forwarding of wakeups to all other cpus */
187 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
188
189 static int forward_wakeup_enabled = 1;
190 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
191 &forward_wakeup_enabled, 0,
192 "Forwarding of wakeup to idle CPUs");
193
194 static int forward_wakeups_requested = 0;
195 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
196 &forward_wakeups_requested, 0,
197 "Requests for Forwarding of wakeup to idle CPUs");
198
199 static int forward_wakeups_delivered = 0;
200 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
201 &forward_wakeups_delivered, 0,
202 "Completed Forwarding of wakeup to idle CPUs");
203
204 static int forward_wakeup_use_mask = 1;
205 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
206 &forward_wakeup_use_mask, 0,
207 "Use the mask of idle cpus");
208
209 static int forward_wakeup_use_loop = 0;
210 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
211 &forward_wakeup_use_loop, 0,
212 "Use a loop to find idle cpus");
213
214 static int forward_wakeup_use_single = 0;
215 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
216 &forward_wakeup_use_single, 0,
217 "Only signal one idle cpu");
218
219 static int forward_wakeup_use_htt = 0;
220 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
221 &forward_wakeup_use_htt, 0,
222 "account for htt");
223
224 #endif
225 #if 0
226 static int sched_followon = 0;
227 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
228 &sched_followon, 0,
229 "allow threads to share a quantum");
230 #endif
231
232 static __inline void
233 sched_load_add(void)
234 {
235 sched_tdcnt++;
236 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
237 }
238
239 static __inline void
240 sched_load_rem(void)
241 {
242 sched_tdcnt--;
243 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt);
244 }
245 /*
246 * Arrange to reschedule if necessary, taking the priorities and
247 * schedulers into account.
248 */
249 static void
250 maybe_resched(struct thread *td)
251 {
252
253 THREAD_LOCK_ASSERT(td, MA_OWNED);
254 if (td->td_priority < curthread->td_priority)
255 curthread->td_flags |= TDF_NEEDRESCHED;
256 }
257
258 /*
259 * Force switch among equal priority processes every 100ms.
260 * We don't actually need to force a context switch of the current process.
261 * The act of firing the event triggers a context switch to softclock() and
262 * then switching back out again which is equivalent to a preemption, thus
263 * no further work is needed on the local CPU.
264 */
265 /* ARGSUSED */
266 static void
267 roundrobin(void *arg)
268 {
269
270 #ifdef SMP
271 mtx_lock_spin(&sched_lock);
272 forward_roundrobin();
273 mtx_unlock_spin(&sched_lock);
274 #endif
275
276 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
277 }
278
279 /*
280 * Constants for digital decay and forget:
281 * 90% of (td_estcpu) usage in 5 * loadav time
282 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
283 * Note that, as ps(1) mentions, this can let percentages
284 * total over 100% (I've seen 137.9% for 3 processes).
285 *
286 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
287 *
288 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
289 * That is, the system wants to compute a value of decay such
290 * that the following for loop:
291 * for (i = 0; i < (5 * loadavg); i++)
292 * td_estcpu *= decay;
293 * will compute
294 * td_estcpu *= 0.1;
295 * for all values of loadavg:
296 *
297 * Mathematically this loop can be expressed by saying:
298 * decay ** (5 * loadavg) ~= .1
299 *
300 * The system computes decay as:
301 * decay = (2 * loadavg) / (2 * loadavg + 1)
302 *
303 * We wish to prove that the system's computation of decay
304 * will always fulfill the equation:
305 * decay ** (5 * loadavg) ~= .1
306 *
307 * If we compute b as:
308 * b = 2 * loadavg
309 * then
310 * decay = b / (b + 1)
311 *
312 * We now need to prove two things:
313 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
314 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
315 *
316 * Facts:
317 * For x close to zero, exp(x) =~ 1 + x, since
318 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
319 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
320 * For x close to zero, ln(1+x) =~ x, since
321 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
322 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
323 * ln(.1) =~ -2.30
324 *
325 * Proof of (1):
326 * Solve (factor)**(power) =~ .1 given power (5*loadav):
327 * solving for factor,
328 * ln(factor) =~ (-2.30/5*loadav), or
329 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
330 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
331 *
332 * Proof of (2):
333 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
334 * solving for power,
335 * power*ln(b/(b+1)) =~ -2.30, or
336 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
337 *
338 * Actual power values for the implemented algorithm are as follows:
339 * loadav: 1 2 3 4
340 * power: 5.68 10.32 14.94 19.55
341 */
342
343 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
344 #define loadfactor(loadav) (2 * (loadav))
345 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
346
347 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
348 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
349 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
350
351 /*
352 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
353 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
354 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
355 *
356 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
357 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
358 *
359 * If you don't want to bother with the faster/more-accurate formula, you
360 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
361 * (more general) method of calculating the %age of CPU used by a process.
362 */
363 #define CCPU_SHIFT 11
364
365 /*
366 * Recompute process priorities, every hz ticks.
367 * MP-safe, called without the Giant mutex.
368 */
369 /* ARGSUSED */
370 static void
371 schedcpu(void)
372 {
373 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
374 struct thread *td;
375 struct proc *p;
376 struct td_sched *ts;
377 int awake, realstathz;
378
379 realstathz = stathz ? stathz : hz;
380 sx_slock(&allproc_lock);
381 FOREACH_PROC_IN_SYSTEM(p) {
382 PROC_SLOCK(p);
383 FOREACH_THREAD_IN_PROC(p, td) {
384 awake = 0;
385 thread_lock(td);
386 ts = td->td_sched;
387 /*
388 * Increment sleep time (if sleeping). We
389 * ignore overflow, as above.
390 */
391 /*
392 * The td_sched slptimes are not touched in wakeup
393 * because the thread may not HAVE everything in
394 * memory? XXX I think this is out of date.
395 */
396 if (TD_ON_RUNQ(td)) {
397 awake = 1;
398 ts->ts_flags &= ~TSF_DIDRUN;
399 } else if (TD_IS_RUNNING(td)) {
400 awake = 1;
401 /* Do not clear TSF_DIDRUN */
402 } else if (ts->ts_flags & TSF_DIDRUN) {
403 awake = 1;
404 ts->ts_flags &= ~TSF_DIDRUN;
405 }
406
407 /*
408 * ts_pctcpu is only for ps and ttyinfo().
409 * Do it per td_sched, and add them up at the end?
410 * XXXKSE
411 */
412 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
413 /*
414 * If the td_sched has been idle the entire second,
415 * stop recalculating its priority until
416 * it wakes up.
417 */
418 if (ts->ts_cpticks != 0) {
419 #if (FSHIFT >= CCPU_SHIFT)
420 ts->ts_pctcpu += (realstathz == 100)
421 ? ((fixpt_t) ts->ts_cpticks) <<
422 (FSHIFT - CCPU_SHIFT) :
423 100 * (((fixpt_t) ts->ts_cpticks)
424 << (FSHIFT - CCPU_SHIFT)) / realstathz;
425 #else
426 ts->ts_pctcpu += ((FSCALE - ccpu) *
427 (ts->ts_cpticks *
428 FSCALE / realstathz)) >> FSHIFT;
429 #endif
430 ts->ts_cpticks = 0;
431 }
432 /*
433 * If there are ANY running threads in this process,
434 * then don't count it as sleeping.
435 XXX this is broken
436
437 */
438 if (awake) {
439 if (ts->ts_slptime > 1) {
440 /*
441 * In an ideal world, this should not
442 * happen, because whoever woke us
443 * up from the long sleep should have
444 * unwound the slptime and reset our
445 * priority before we run at the stale
446 * priority. Should KASSERT at some
447 * point when all the cases are fixed.
448 */
449 updatepri(td);
450 }
451 ts->ts_slptime = 0;
452 } else
453 ts->ts_slptime++;
454 if (ts->ts_slptime > 1) {
455 thread_unlock(td);
456 continue;
457 }
458 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
459 resetpriority(td);
460 resetpriority_thread(td);
461 thread_unlock(td);
462 } /* end of thread loop */
463 PROC_SUNLOCK(p);
464 } /* end of process loop */
465 sx_sunlock(&allproc_lock);
466 }
467
468 /*
469 * Main loop for a kthread that executes schedcpu once a second.
470 */
471 static void
472 schedcpu_thread(void)
473 {
474
475 for (;;) {
476 schedcpu();
477 pause("-", hz);
478 }
479 }
480
481 /*
482 * Recalculate the priority of a process after it has slept for a while.
483 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
484 * least six times the loadfactor will decay td_estcpu to zero.
485 */
486 static void
487 updatepri(struct thread *td)
488 {
489 struct td_sched *ts;
490 fixpt_t loadfac;
491 unsigned int newcpu;
492
493 ts = td->td_sched;
494 loadfac = loadfactor(averunnable.ldavg[0]);
495 if (ts->ts_slptime > 5 * loadfac)
496 td->td_estcpu = 0;
497 else {
498 newcpu = td->td_estcpu;
499 ts->ts_slptime--; /* was incremented in schedcpu() */
500 while (newcpu && --ts->ts_slptime)
501 newcpu = decay_cpu(loadfac, newcpu);
502 td->td_estcpu = newcpu;
503 }
504 }
505
506 /*
507 * Compute the priority of a process when running in user mode.
508 * Arrange to reschedule if the resulting priority is better
509 * than that of the current process.
510 */
511 static void
512 resetpriority(struct thread *td)
513 {
514 register unsigned int newpriority;
515
516 if (td->td_pri_class == PRI_TIMESHARE) {
517 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
518 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
519 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
520 PRI_MAX_TIMESHARE);
521 sched_user_prio(td, newpriority);
522 }
523 }
524
525 /*
526 * Update the thread's priority when the associated process's user
527 * priority changes.
528 */
529 static void
530 resetpriority_thread(struct thread *td)
531 {
532
533 /* Only change threads with a time sharing user priority. */
534 if (td->td_priority < PRI_MIN_TIMESHARE ||
535 td->td_priority > PRI_MAX_TIMESHARE)
536 return;
537
538 /* XXX the whole needresched thing is broken, but not silly. */
539 maybe_resched(td);
540
541 sched_prio(td, td->td_user_pri);
542 }
543
544 /* ARGSUSED */
545 static void
546 sched_setup(void *dummy)
547 {
548 setup_runqs();
549
550 if (sched_quantum == 0)
551 sched_quantum = SCHED_QUANTUM;
552 hogticks = 2 * sched_quantum;
553
554 callout_init(&roundrobin_callout, CALLOUT_MPSAFE);
555
556 /* Kick off timeout driven events by calling first time. */
557 roundrobin(NULL);
558
559 /* Account for thread0. */
560 sched_load_add();
561 }
562
563 /* External interfaces start here */
564 /*
565 * Very early in the boot some setup of scheduler-specific
566 * parts of proc0 and of some scheduler resources needs to be done.
567 * Called from:
568 * proc0_init()
569 */
570 void
571 schedinit(void)
572 {
573 /*
574 * Set up the scheduler specific parts of proc0.
575 */
576 proc0.p_sched = NULL; /* XXX */
577 thread0.td_sched = &td_sched0;
578 thread0.td_lock = &sched_lock;
579 td_sched0.ts_thread = &thread0;
580 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
581 }
582
583 int
584 sched_runnable(void)
585 {
586 #ifdef SMP
587 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
588 #else
589 return runq_check(&runq);
590 #endif
591 }
592
593 int
594 sched_rr_interval(void)
595 {
596 if (sched_quantum == 0)
597 sched_quantum = SCHED_QUANTUM;
598 return (sched_quantum);
599 }
600
601 /*
602 * We adjust the priority of the current process. The priority of
603 * a process gets worse as it accumulates CPU time. The cpu usage
604 * estimator (td_estcpu) is increased here. resetpriority() will
605 * compute a different priority each time td_estcpu increases by
606 * INVERSE_ESTCPU_WEIGHT
607 * (until MAXPRI is reached). The cpu usage estimator ramps up
608 * quite quickly when the process is running (linearly), and decays
609 * away exponentially, at a rate which is proportionally slower when
610 * the system is busy. The basic principle is that the system will
611 * 90% forget that the process used a lot of CPU time in 5 * loadav
612 * seconds. This causes the system to favor processes which haven't
613 * run much recently, and to round-robin among other processes.
614 */
615 void
616 sched_clock(struct thread *td)
617 {
618 struct td_sched *ts;
619
620 THREAD_LOCK_ASSERT(td, MA_OWNED);
621 ts = td->td_sched;
622
623 ts->ts_cpticks++;
624 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
625 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
626 resetpriority(td);
627 resetpriority_thread(td);
628 }
629 }
630
631 /*
632 * charge childs scheduling cpu usage to parent.
633 */
634 void
635 sched_exit(struct proc *p, struct thread *td)
636 {
637
638 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
639 td, td->td_proc->p_comm, td->td_priority);
640 PROC_SLOCK_ASSERT(p, MA_OWNED);
641 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
642 }
643
644 void
645 sched_exit_thread(struct thread *td, struct thread *child)
646 {
647
648 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
649 child, child->td_proc->p_comm, child->td_priority);
650 thread_lock(td);
651 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
652 thread_unlock(td);
653 mtx_lock_spin(&sched_lock);
654 if ((child->td_proc->p_flag & P_NOLOAD) == 0)
655 sched_load_rem();
656 mtx_unlock_spin(&sched_lock);
657 }
658
659 void
660 sched_fork(struct thread *td, struct thread *childtd)
661 {
662 sched_fork_thread(td, childtd);
663 }
664
665 void
666 sched_fork_thread(struct thread *td, struct thread *childtd)
667 {
668 childtd->td_estcpu = td->td_estcpu;
669 childtd->td_lock = &sched_lock;
670 sched_newthread(childtd);
671 }
672
673 void
674 sched_nice(struct proc *p, int nice)
675 {
676 struct thread *td;
677
678 PROC_LOCK_ASSERT(p, MA_OWNED);
679 PROC_SLOCK_ASSERT(p, MA_OWNED);
680 p->p_nice = nice;
681 FOREACH_THREAD_IN_PROC(p, td) {
682 thread_lock(td);
683 resetpriority(td);
684 resetpriority_thread(td);
685 thread_unlock(td);
686 }
687 }
688
689 void
690 sched_class(struct thread *td, int class)
691 {
692 THREAD_LOCK_ASSERT(td, MA_OWNED);
693 td->td_pri_class = class;
694 }
695
696 /*
697 * Adjust the priority of a thread.
698 */
699 static void
700 sched_priority(struct thread *td, u_char prio)
701 {
702 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
703 td, td->td_proc->p_comm, td->td_priority, prio, curthread,
704 curthread->td_proc->p_comm);
705
706 THREAD_LOCK_ASSERT(td, MA_OWNED);
707 if (td->td_priority == prio)
708 return;
709 td->td_priority = prio;
710 if (TD_ON_RUNQ(td) &&
711 td->td_sched->ts_rqindex != (prio / RQ_PPQ)) {
712 sched_rem(td);
713 sched_add(td, SRQ_BORING);
714 }
715 }
716
717 /*
718 * Update a thread's priority when it is lent another thread's
719 * priority.
720 */
721 void
722 sched_lend_prio(struct thread *td, u_char prio)
723 {
724
725 td->td_flags |= TDF_BORROWING;
726 sched_priority(td, prio);
727 }
728
729 /*
730 * Restore a thread's priority when priority propagation is
731 * over. The prio argument is the minimum priority the thread
732 * needs to have to satisfy other possible priority lending
733 * requests. If the thread's regulary priority is less
734 * important than prio the thread will keep a priority boost
735 * of prio.
736 */
737 void
738 sched_unlend_prio(struct thread *td, u_char prio)
739 {
740 u_char base_pri;
741
742 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
743 td->td_base_pri <= PRI_MAX_TIMESHARE)
744 base_pri = td->td_user_pri;
745 else
746 base_pri = td->td_base_pri;
747 if (prio >= base_pri) {
748 td->td_flags &= ~TDF_BORROWING;
749 sched_prio(td, base_pri);
750 } else
751 sched_lend_prio(td, prio);
752 }
753
754 void
755 sched_prio(struct thread *td, u_char prio)
756 {
757 u_char oldprio;
758
759 /* First, update the base priority. */
760 td->td_base_pri = prio;
761
762 /*
763 * If the thread is borrowing another thread's priority, don't ever
764 * lower the priority.
765 */
766 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
767 return;
768
769 /* Change the real priority. */
770 oldprio = td->td_priority;
771 sched_priority(td, prio);
772
773 /*
774 * If the thread is on a turnstile, then let the turnstile update
775 * its state.
776 */
777 if (TD_ON_LOCK(td) && oldprio != prio)
778 turnstile_adjust(td, oldprio);
779 }
780
781 void
782 sched_user_prio(struct thread *td, u_char prio)
783 {
784 u_char oldprio;
785
786 THREAD_LOCK_ASSERT(td, MA_OWNED);
787 td->td_base_user_pri = prio;
788 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
789 return;
790 oldprio = td->td_user_pri;
791 td->td_user_pri = prio;
792 }
793
794 void
795 sched_lend_user_prio(struct thread *td, u_char prio)
796 {
797 u_char oldprio;
798
799 THREAD_LOCK_ASSERT(td, MA_OWNED);
800 td->td_flags |= TDF_UBORROWING;
801
802 oldprio = td->td_user_pri;
803 td->td_user_pri = prio;
804 }
805
806 void
807 sched_unlend_user_prio(struct thread *td, u_char prio)
808 {
809 u_char base_pri;
810
811 THREAD_LOCK_ASSERT(td, MA_OWNED);
812 base_pri = td->td_base_user_pri;
813 if (prio >= base_pri) {
814 td->td_flags &= ~TDF_UBORROWING;
815 sched_user_prio(td, base_pri);
816 } else {
817 sched_lend_user_prio(td, prio);
818 }
819 }
820
821 void
822 sched_sleep(struct thread *td)
823 {
824
825 THREAD_LOCK_ASSERT(td, MA_OWNED);
826 td->td_slptick = ticks;
827 td->td_sched->ts_slptime = 0;
828 }
829
830 void
831 sched_switch(struct thread *td, struct thread *newtd, int flags)
832 {
833 struct td_sched *ts;
834 struct proc *p;
835
836 ts = td->td_sched;
837 p = td->td_proc;
838
839 THREAD_LOCK_ASSERT(td, MA_OWNED);
840 /*
841 * Switch to the sched lock to fix things up and pick
842 * a new thread.
843 */
844 if (td->td_lock != &sched_lock) {
845 mtx_lock_spin(&sched_lock);
846 thread_unlock(td);
847 }
848
849 if ((p->p_flag & P_NOLOAD) == 0)
850 sched_load_rem();
851
852 if (newtd)
853 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
854
855 td->td_lastcpu = td->td_oncpu;
856 td->td_flags &= ~TDF_NEEDRESCHED;
857 td->td_owepreempt = 0;
858 td->td_oncpu = NOCPU;
859 /*
860 * At the last moment, if this thread is still marked RUNNING,
861 * then put it back on the run queue as it has not been suspended
862 * or stopped or any thing else similar. We never put the idle
863 * threads on the run queue, however.
864 */
865 if (td->td_flags & TDF_IDLETD) {
866 TD_SET_CAN_RUN(td);
867 #ifdef SMP
868 idle_cpus_mask &= ~PCPU_GET(cpumask);
869 #endif
870 } else {
871 if (TD_IS_RUNNING(td)) {
872 /* Put us back on the run queue. */
873 sched_add(td, (flags & SW_PREEMPT) ?
874 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
875 SRQ_OURSELF|SRQ_YIELDING);
876 }
877 }
878 if (newtd) {
879 /*
880 * The thread we are about to run needs to be counted
881 * as if it had been added to the run queue and selected.
882 * It came from:
883 * * A preemption
884 * * An upcall
885 * * A followon
886 */
887 KASSERT((newtd->td_inhibitors == 0),
888 ("trying to run inhibited thread"));
889 newtd->td_sched->ts_flags |= TSF_DIDRUN;
890 TD_SET_RUNNING(newtd);
891 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0)
892 sched_load_add();
893 } else {
894 newtd = choosethread();
895 }
896 MPASS(newtd->td_lock == &sched_lock);
897
898 if (td != newtd) {
899 #ifdef HWPMC_HOOKS
900 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
901 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
902 #endif
903
904 /* I feel sleepy */
905 cpu_switch(td, newtd, td->td_lock);
906 /*
907 * Where am I? What year is it?
908 * We are in the same thread that went to sleep above,
909 * but any amount of time may have passed. All out context
910 * will still be available as will local variables.
911 * PCPU values however may have changed as we may have
912 * changed CPU so don't trust cached values of them.
913 * New threads will go to fork_exit() instead of here
914 * so if you change things here you may need to change
915 * things there too.
916 * If the thread above was exiting it will never wake
917 * up again here, so either it has saved everything it
918 * needed to, or the thread_wait() or wait() will
919 * need to reap it.
920 */
921 #ifdef HWPMC_HOOKS
922 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
923 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
924 #endif
925 }
926
927 #ifdef SMP
928 if (td->td_flags & TDF_IDLETD)
929 idle_cpus_mask |= PCPU_GET(cpumask);
930 #endif
931 sched_lock.mtx_lock = (uintptr_t)td;
932 td->td_oncpu = PCPU_GET(cpuid);
933 MPASS(td->td_lock == &sched_lock);
934 }
935
936 void
937 sched_wakeup(struct thread *td)
938 {
939 struct td_sched *ts;
940
941 THREAD_LOCK_ASSERT(td, MA_OWNED);
942 ts = td->td_sched;
943 if (ts->ts_slptime > 1) {
944 updatepri(td);
945 resetpriority(td);
946 }
947 td->td_slptick = ticks;
948 ts->ts_slptime = 0;
949 sched_add(td, SRQ_BORING);
950 }
951
952 #ifdef SMP
953 /* enable HTT_2 if you have a 2-way HTT cpu.*/
954 static int
955 forward_wakeup(int cpunum)
956 {
957 cpumask_t map, me, dontuse;
958 cpumask_t map2;
959 struct pcpu *pc;
960 cpumask_t id, map3;
961
962 mtx_assert(&sched_lock, MA_OWNED);
963
964 CTR0(KTR_RUNQ, "forward_wakeup()");
965
966 if ((!forward_wakeup_enabled) ||
967 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
968 return (0);
969 if (!smp_started || cold || panicstr)
970 return (0);
971
972 forward_wakeups_requested++;
973
974 /*
975 * check the idle mask we received against what we calculated before
976 * in the old version.
977 */
978 me = PCPU_GET(cpumask);
979 /*
980 * don't bother if we should be doing it ourself..
981 */
982 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
983 return (0);
984
985 dontuse = me | stopped_cpus | hlt_cpus_mask;
986 map3 = 0;
987 if (forward_wakeup_use_loop) {
988 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
989 id = pc->pc_cpumask;
990 if ( (id & dontuse) == 0 &&
991 pc->pc_curthread == pc->pc_idlethread) {
992 map3 |= id;
993 }
994 }
995 }
996
997 if (forward_wakeup_use_mask) {
998 map = 0;
999 map = idle_cpus_mask & ~dontuse;
1000
1001 /* If they are both on, compare and use loop if different */
1002 if (forward_wakeup_use_loop) {
1003 if (map != map3) {
1004 printf("map (%02X) != map3 (%02X)\n",
1005 map, map3);
1006 map = map3;
1007 }
1008 }
1009 } else {
1010 map = map3;
1011 }
1012 /* If we only allow a specific CPU, then mask off all the others */
1013 if (cpunum != NOCPU) {
1014 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1015 map &= (1 << cpunum);
1016 } else {
1017 /* Try choose an idle die. */
1018 if (forward_wakeup_use_htt) {
1019 map2 = (map & (map >> 1)) & 0x5555;
1020 if (map2) {
1021 map = map2;
1022 }
1023 }
1024
1025 /* set only one bit */
1026 if (forward_wakeup_use_single) {
1027 map = map & ((~map) + 1);
1028 }
1029 }
1030 if (map) {
1031 forward_wakeups_delivered++;
1032 ipi_selected(map, IPI_AST);
1033 return (1);
1034 }
1035 if (cpunum == NOCPU)
1036 printf("forward_wakeup: Idle processor not found\n");
1037 return (0);
1038 }
1039 #endif
1040
1041 #ifdef SMP
1042 static void kick_other_cpu(int pri,int cpuid);
1043
1044 static void
1045 kick_other_cpu(int pri,int cpuid)
1046 {
1047 struct pcpu * pcpu = pcpu_find(cpuid);
1048 int cpri = pcpu->pc_curthread->td_priority;
1049
1050 if (idle_cpus_mask & pcpu->pc_cpumask) {
1051 forward_wakeups_delivered++;
1052 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1053 return;
1054 }
1055
1056 if (pri >= cpri)
1057 return;
1058
1059 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1060 #if !defined(FULL_PREEMPTION)
1061 if (pri <= PRI_MAX_ITHD)
1062 #endif /* ! FULL_PREEMPTION */
1063 {
1064 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1065 return;
1066 }
1067 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1068
1069 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1070 ipi_selected( pcpu->pc_cpumask , IPI_AST);
1071 return;
1072 }
1073 #endif /* SMP */
1074
1075 void
1076 sched_add(struct thread *td, int flags)
1077 #ifdef SMP
1078 {
1079 struct td_sched *ts;
1080 int forwarded = 0;
1081 int cpu;
1082 int single_cpu = 0;
1083
1084 ts = td->td_sched;
1085 THREAD_LOCK_ASSERT(td, MA_OWNED);
1086 KASSERT((td->td_inhibitors == 0),
1087 ("sched_add: trying to run inhibited thread"));
1088 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1089 ("sched_add: bad thread state"));
1090 KASSERT(td->td_flags & TDF_INMEM,
1091 ("sched_add: thread swapped out"));
1092 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1093 td, td->td_proc->p_comm, td->td_priority, curthread,
1094 curthread->td_proc->p_comm);
1095 /*
1096 * Now that the thread is moving to the run-queue, set the lock
1097 * to the scheduler's lock.
1098 */
1099 if (td->td_lock != &sched_lock) {
1100 mtx_lock_spin(&sched_lock);
1101 thread_lock_set(td, &sched_lock);
1102 }
1103 TD_SET_RUNQ(td);
1104
1105 if (td->td_pinned != 0) {
1106 cpu = td->td_lastcpu;
1107 ts->ts_runq = &runq_pcpu[cpu];
1108 single_cpu = 1;
1109 CTR3(KTR_RUNQ,
1110 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1111 } else if ((ts)->ts_flags & TSF_BOUND) {
1112 /* Find CPU from bound runq */
1113 KASSERT(SKE_RUNQ_PCPU(ts),("sched_add: bound td_sched not on cpu runq"));
1114 cpu = ts->ts_runq - &runq_pcpu[0];
1115 single_cpu = 1;
1116 CTR3(KTR_RUNQ,
1117 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu);
1118 } else {
1119 CTR2(KTR_RUNQ,
1120 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, td);
1121 cpu = NOCPU;
1122 ts->ts_runq = &runq;
1123 }
1124
1125 if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1126 kick_other_cpu(td->td_priority,cpu);
1127 } else {
1128
1129 if (!single_cpu) {
1130 cpumask_t me = PCPU_GET(cpumask);
1131 int idle = idle_cpus_mask & me;
1132
1133 if (!idle && ((flags & SRQ_INTR) == 0) &&
1134 (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1135 forwarded = forward_wakeup(cpu);
1136 }
1137
1138 if (!forwarded) {
1139 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1140 return;
1141 else
1142 maybe_resched(td);
1143 }
1144 }
1145
1146 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1147 sched_load_add();
1148 runq_add(ts->ts_runq, ts, flags);
1149 }
1150 #else /* SMP */
1151 {
1152 struct td_sched *ts;
1153 ts = td->td_sched;
1154 THREAD_LOCK_ASSERT(td, MA_OWNED);
1155 KASSERT((td->td_inhibitors == 0),
1156 ("sched_add: trying to run inhibited thread"));
1157 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1158 ("sched_add: bad thread state"));
1159 KASSERT(td->td_flags & TDF_INMEM,
1160 ("sched_add: thread swapped out"));
1161 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1162 td, td->td_proc->p_comm, td->td_priority, curthread,
1163 curthread->td_proc->p_comm);
1164 /*
1165 * Now that the thread is moving to the run-queue, set the lock
1166 * to the scheduler's lock.
1167 */
1168 if (td->td_lock != &sched_lock) {
1169 mtx_lock_spin(&sched_lock);
1170 thread_lock_set(td, &sched_lock);
1171 }
1172 TD_SET_RUNQ(td);
1173 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1174 ts->ts_runq = &runq;
1175
1176 /*
1177 * If we are yielding (on the way out anyhow)
1178 * or the thread being saved is US,
1179 * then don't try be smart about preemption
1180 * or kicking off another CPU
1181 * as it won't help and may hinder.
1182 * In the YIEDLING case, we are about to run whoever is
1183 * being put in the queue anyhow, and in the
1184 * OURSELF case, we are puting ourself on the run queue
1185 * which also only happens when we are about to yield.
1186 */
1187 if((flags & SRQ_YIELDING) == 0) {
1188 if (maybe_preempt(td))
1189 return;
1190 }
1191 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1192 sched_load_add();
1193 runq_add(ts->ts_runq, ts, flags);
1194 maybe_resched(td);
1195 }
1196 #endif /* SMP */
1197
1198 void
1199 sched_rem(struct thread *td)
1200 {
1201 struct td_sched *ts;
1202
1203 ts = td->td_sched;
1204 KASSERT(td->td_flags & TDF_INMEM,
1205 ("sched_rem: thread swapped out"));
1206 KASSERT(TD_ON_RUNQ(td),
1207 ("sched_rem: thread not on run queue"));
1208 mtx_assert(&sched_lock, MA_OWNED);
1209 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1210 td, td->td_proc->p_comm, td->td_priority, curthread,
1211 curthread->td_proc->p_comm);
1212
1213 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
1214 sched_load_rem();
1215 runq_remove(ts->ts_runq, ts);
1216 TD_SET_CAN_RUN(td);
1217 }
1218
1219 /*
1220 * Select threads to run.
1221 * Notice that the running threads still consume a slot.
1222 */
1223 struct thread *
1224 sched_choose(void)
1225 {
1226 struct td_sched *ts;
1227 struct runq *rq;
1228
1229 mtx_assert(&sched_lock, MA_OWNED);
1230 #ifdef SMP
1231 struct td_sched *kecpu;
1232
1233 rq = &runq;
1234 ts = runq_choose(&runq);
1235 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1236
1237 if (ts == NULL ||
1238 (kecpu != NULL &&
1239 kecpu->ts_thread->td_priority < ts->ts_thread->td_priority)) {
1240 CTR2(KTR_RUNQ, "choosing td_sched %p from pcpu runq %d", kecpu,
1241 PCPU_GET(cpuid));
1242 ts = kecpu;
1243 rq = &runq_pcpu[PCPU_GET(cpuid)];
1244 } else {
1245 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", ts);
1246 }
1247
1248 #else
1249 rq = &runq;
1250 ts = runq_choose(&runq);
1251 #endif
1252
1253 if (ts) {
1254 runq_remove(rq, ts);
1255 ts->ts_flags |= TSF_DIDRUN;
1256
1257 KASSERT(ts->ts_thread->td_flags & TDF_INMEM,
1258 ("sched_choose: thread swapped out"));
1259 return (ts->ts_thread);
1260 }
1261 return (PCPU_GET(idlethread));
1262 }
1263
1264 void
1265 sched_userret(struct thread *td)
1266 {
1267 /*
1268 * XXX we cheat slightly on the locking here to avoid locking in
1269 * the usual case. Setting td_priority here is essentially an
1270 * incomplete workaround for not setting it properly elsewhere.
1271 * Now that some interrupt handlers are threads, not setting it
1272 * properly elsewhere can clobber it in the window between setting
1273 * it here and returning to user mode, so don't waste time setting
1274 * it perfectly here.
1275 */
1276 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1277 ("thread with borrowed priority returning to userland"));
1278 if (td->td_priority != td->td_user_pri) {
1279 thread_lock(td);
1280 td->td_priority = td->td_user_pri;
1281 td->td_base_pri = td->td_user_pri;
1282 thread_unlock(td);
1283 }
1284 }
1285
1286 void
1287 sched_bind(struct thread *td, int cpu)
1288 {
1289 struct td_sched *ts;
1290
1291 THREAD_LOCK_ASSERT(td, MA_OWNED);
1292 KASSERT(TD_IS_RUNNING(td),
1293 ("sched_bind: cannot bind non-running thread"));
1294
1295 ts = td->td_sched;
1296
1297 ts->ts_flags |= TSF_BOUND;
1298 #ifdef SMP
1299 ts->ts_runq = &runq_pcpu[cpu];
1300 if (PCPU_GET(cpuid) == cpu)
1301 return;
1302
1303 mi_switch(SW_VOL, NULL);
1304 #endif
1305 }
1306
1307 void
1308 sched_unbind(struct thread* td)
1309 {
1310 THREAD_LOCK_ASSERT(td, MA_OWNED);
1311 td->td_sched->ts_flags &= ~TSF_BOUND;
1312 }
1313
1314 int
1315 sched_is_bound(struct thread *td)
1316 {
1317 THREAD_LOCK_ASSERT(td, MA_OWNED);
1318 return (td->td_sched->ts_flags & TSF_BOUND);
1319 }
1320
1321 void
1322 sched_relinquish(struct thread *td)
1323 {
1324 thread_lock(td);
1325 SCHED_STAT_INC(switch_relinquish);
1326 mi_switch(SW_VOL, NULL);
1327 thread_unlock(td);
1328 }
1329
1330 int
1331 sched_load(void)
1332 {
1333 return (sched_tdcnt);
1334 }
1335
1336 int
1337 sched_sizeof_proc(void)
1338 {
1339 return (sizeof(struct proc));
1340 }
1341
1342 int
1343 sched_sizeof_thread(void)
1344 {
1345 return (sizeof(struct thread) + sizeof(struct td_sched));
1346 }
1347
1348 fixpt_t
1349 sched_pctcpu(struct thread *td)
1350 {
1351 struct td_sched *ts;
1352
1353 ts = td->td_sched;
1354 return (ts->ts_pctcpu);
1355 }
1356
1357 void
1358 sched_tick(void)
1359 {
1360 }
1361
1362 /*
1363 * The actual idle process.
1364 */
1365 void
1366 sched_idletd(void *dummy)
1367 {
1368 struct proc *p;
1369 struct thread *td;
1370
1371 td = curthread;
1372 p = td->td_proc;
1373 for (;;) {
1374 mtx_assert(&Giant, MA_NOTOWNED);
1375
1376 while (sched_runnable() == 0)
1377 cpu_idle();
1378
1379 mtx_lock_spin(&sched_lock);
1380 mi_switch(SW_VOL, NULL);
1381 mtx_unlock_spin(&sched_lock);
1382 }
1383 }
1384
1385 /*
1386 * A CPU is entering for the first time or a thread is exiting.
1387 */
1388 void
1389 sched_throw(struct thread *td)
1390 {
1391 /*
1392 * Correct spinlock nesting. The idle thread context that we are
1393 * borrowing was created so that it would start out with a single
1394 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1395 * explicitly acquired locks in this function, the nesting count
1396 * is now 2 rather than 1. Since we are nested, calling
1397 * spinlock_exit() will simply adjust the counts without allowing
1398 * spin lock using code to interrupt us.
1399 */
1400 if (td == NULL) {
1401 mtx_lock_spin(&sched_lock);
1402 spinlock_exit();
1403 } else {
1404 MPASS(td->td_lock == &sched_lock);
1405 }
1406 mtx_assert(&sched_lock, MA_OWNED);
1407 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1408 PCPU_SET(switchtime, cpu_ticks());
1409 PCPU_SET(switchticks, ticks);
1410 cpu_throw(td, choosethread()); /* doesn't return */
1411 }
1412
1413 void
1414 sched_fork_exit(struct thread *td)
1415 {
1416
1417 /*
1418 * Finish setting up thread glue so that it begins execution in a
1419 * non-nested critical section with sched_lock held but not recursed.
1420 */
1421 td->td_oncpu = PCPU_GET(cpuid);
1422 sched_lock.mtx_lock = (uintptr_t)td;
1423 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1424 }
1425
1426 #define KERN_SWITCH_INCLUDE 1
1427 #include "kern/kern_switch.c"
Cache object: 5d39bda7cc213c592ed542ad1f329d4f
|