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: releng/9.1/sys/kern/sched_4bsd.c 236344 2012-05-30 23:22:52Z rstone $");
37
38 #include "opt_hwpmc_hooks.h"
39 #include "opt_sched.h"
40 #include "opt_kdtrace.h"
41
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/cpuset.h>
45 #include <sys/kernel.h>
46 #include <sys/ktr.h>
47 #include <sys/lock.h>
48 #include <sys/kthread.h>
49 #include <sys/mutex.h>
50 #include <sys/proc.h>
51 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
53 #include <sys/sdt.h>
54 #include <sys/smp.h>
55 #include <sys/sysctl.h>
56 #include <sys/sx.h>
57 #include <sys/turnstile.h>
58 #include <sys/umtx.h>
59 #include <machine/pcb.h>
60 #include <machine/smp.h>
61
62 #ifdef HWPMC_HOOKS
63 #include <sys/pmckern.h>
64 #endif
65
66 #ifdef KDTRACE_HOOKS
67 #include <sys/dtrace_bsd.h>
68 int dtrace_vtime_active;
69 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
70 #endif
71
72 /*
73 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
74 * the range 100-256 Hz (approximately).
75 */
76 #define ESTCPULIM(e) \
77 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
78 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
79 #ifdef SMP
80 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
81 #else
82 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
83 #endif
84 #define NICE_WEIGHT 1 /* Priorities per nice level. */
85
86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
87
88 /*
89 * The schedulable entity that runs a context.
90 * This is an extension to the thread structure and is tailored to
91 * the requirements of this scheduler
92 */
93 struct td_sched {
94 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
95 int ts_cpticks; /* (j) Ticks of cpu time. */
96 int ts_slptime; /* (j) Seconds !RUNNING. */
97 int ts_flags;
98 struct runq *ts_runq; /* runq the thread is currently on */
99 #ifdef KTR
100 char ts_name[TS_NAME_LEN];
101 #endif
102 };
103
104 /* flags kept in td_flags */
105 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
106 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
107
108 /* flags kept in ts_flags */
109 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
110
111 #define SKE_RUNQ_PCPU(ts) \
112 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
113
114 #define THREAD_CAN_SCHED(td, cpu) \
115 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
116
117 static struct td_sched td_sched0;
118 struct mtx sched_lock;
119
120 static int sched_tdcnt; /* Total runnable threads in the system. */
121 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
122 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
123
124 static void setup_runqs(void);
125 static void schedcpu(void);
126 static void schedcpu_thread(void);
127 static void sched_priority(struct thread *td, u_char prio);
128 static void sched_setup(void *dummy);
129 static void maybe_resched(struct thread *td);
130 static void updatepri(struct thread *td);
131 static void resetpriority(struct thread *td);
132 static void resetpriority_thread(struct thread *td);
133 #ifdef SMP
134 static int sched_pickcpu(struct thread *td);
135 static int forward_wakeup(int cpunum);
136 static void kick_other_cpu(int pri, int cpuid);
137 #endif
138
139 static struct kproc_desc sched_kp = {
140 "schedcpu",
141 schedcpu_thread,
142 NULL
143 };
144 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
145 &sched_kp);
146 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
147
148 /*
149 * Global run queue.
150 */
151 static struct runq runq;
152
153 #ifdef SMP
154 /*
155 * Per-CPU run queues
156 */
157 static struct runq runq_pcpu[MAXCPU];
158 long runq_length[MAXCPU];
159
160 static cpuset_t idle_cpus_mask;
161 #endif
162
163 struct pcpuidlestat {
164 u_int idlecalls;
165 u_int oldidlecalls;
166 };
167 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
168
169 static void
170 setup_runqs(void)
171 {
172 #ifdef SMP
173 int i;
174
175 for (i = 0; i < MAXCPU; ++i)
176 runq_init(&runq_pcpu[i]);
177 #endif
178
179 runq_init(&runq);
180 }
181
182 static int
183 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
184 {
185 int error, new_val;
186
187 new_val = sched_quantum * tick;
188 error = sysctl_handle_int(oidp, &new_val, 0, req);
189 if (error != 0 || req->newptr == NULL)
190 return (error);
191 if (new_val < tick)
192 return (EINVAL);
193 sched_quantum = new_val / tick;
194 hogticks = 2 * sched_quantum;
195 return (0);
196 }
197
198 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
199
200 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
201 "Scheduler name");
202
203 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
204 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
205 "Roundrobin scheduling quantum in microseconds");
206
207 #ifdef SMP
208 /* Enable forwarding of wakeups to all other cpus */
209 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
210
211 static int runq_fuzz = 1;
212 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
213
214 static int forward_wakeup_enabled = 1;
215 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
216 &forward_wakeup_enabled, 0,
217 "Forwarding of wakeup to idle CPUs");
218
219 static int forward_wakeups_requested = 0;
220 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
221 &forward_wakeups_requested, 0,
222 "Requests for Forwarding of wakeup to idle CPUs");
223
224 static int forward_wakeups_delivered = 0;
225 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
226 &forward_wakeups_delivered, 0,
227 "Completed Forwarding of wakeup to idle CPUs");
228
229 static int forward_wakeup_use_mask = 1;
230 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
231 &forward_wakeup_use_mask, 0,
232 "Use the mask of idle cpus");
233
234 static int forward_wakeup_use_loop = 0;
235 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
236 &forward_wakeup_use_loop, 0,
237 "Use a loop to find idle cpus");
238
239 #endif
240 #if 0
241 static int sched_followon = 0;
242 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
243 &sched_followon, 0,
244 "allow threads to share a quantum");
245 #endif
246
247 SDT_PROVIDER_DEFINE(sched);
248
249 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
250 "struct proc *", "uint8_t");
251 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
252 "struct proc *", "void *");
253 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
254 "struct proc *", "void *", "int");
255 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
256 "struct proc *", "uint8_t", "struct thread *");
257 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
258 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
259 "struct proc *");
260 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
261 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
262 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
263 "struct proc *");
264
265 static __inline void
266 sched_load_add(void)
267 {
268
269 sched_tdcnt++;
270 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
271 SDT_PROBE2(sched, , , load_change, NOCPU, sched_tdcnt);
272 }
273
274 static __inline void
275 sched_load_rem(void)
276 {
277
278 sched_tdcnt--;
279 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
280 SDT_PROBE2(sched, , , load_change, NOCPU, sched_tdcnt);
281 }
282 /*
283 * Arrange to reschedule if necessary, taking the priorities and
284 * schedulers into account.
285 */
286 static void
287 maybe_resched(struct thread *td)
288 {
289
290 THREAD_LOCK_ASSERT(td, MA_OWNED);
291 if (td->td_priority < curthread->td_priority)
292 curthread->td_flags |= TDF_NEEDRESCHED;
293 }
294
295 /*
296 * This function is called when a thread is about to be put on run queue
297 * because it has been made runnable or its priority has been adjusted. It
298 * determines if the new thread should be immediately preempted to. If so,
299 * it switches to it and eventually returns true. If not, it returns false
300 * so that the caller may place the thread on an appropriate run queue.
301 */
302 int
303 maybe_preempt(struct thread *td)
304 {
305 #ifdef PREEMPTION
306 struct thread *ctd;
307 int cpri, pri;
308
309 /*
310 * The new thread should not preempt the current thread if any of the
311 * following conditions are true:
312 *
313 * - The kernel is in the throes of crashing (panicstr).
314 * - The current thread has a higher (numerically lower) or
315 * equivalent priority. Note that this prevents curthread from
316 * trying to preempt to itself.
317 * - It is too early in the boot for context switches (cold is set).
318 * - The current thread has an inhibitor set or is in the process of
319 * exiting. In this case, the current thread is about to switch
320 * out anyways, so there's no point in preempting. If we did,
321 * the current thread would not be properly resumed as well, so
322 * just avoid that whole landmine.
323 * - If the new thread's priority is not a realtime priority and
324 * the current thread's priority is not an idle priority and
325 * FULL_PREEMPTION is disabled.
326 *
327 * If all of these conditions are false, but the current thread is in
328 * a nested critical section, then we have to defer the preemption
329 * until we exit the critical section. Otherwise, switch immediately
330 * to the new thread.
331 */
332 ctd = curthread;
333 THREAD_LOCK_ASSERT(td, MA_OWNED);
334 KASSERT((td->td_inhibitors == 0),
335 ("maybe_preempt: trying to run inhibited thread"));
336 pri = td->td_priority;
337 cpri = ctd->td_priority;
338 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
339 TD_IS_INHIBITED(ctd))
340 return (0);
341 #ifndef FULL_PREEMPTION
342 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
343 return (0);
344 #endif
345
346 if (ctd->td_critnest > 1) {
347 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
348 ctd->td_critnest);
349 ctd->td_owepreempt = 1;
350 return (0);
351 }
352 /*
353 * Thread is runnable but not yet put on system run queue.
354 */
355 MPASS(ctd->td_lock == td->td_lock);
356 MPASS(TD_ON_RUNQ(td));
357 TD_SET_RUNNING(td);
358 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
359 td->td_proc->p_pid, td->td_name);
360 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
361 /*
362 * td's lock pointer may have changed. We have to return with it
363 * locked.
364 */
365 spinlock_enter();
366 thread_unlock(ctd);
367 thread_lock(td);
368 spinlock_exit();
369 return (1);
370 #else
371 return (0);
372 #endif
373 }
374
375 /*
376 * Constants for digital decay and forget:
377 * 90% of (td_estcpu) usage in 5 * loadav time
378 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
379 * Note that, as ps(1) mentions, this can let percentages
380 * total over 100% (I've seen 137.9% for 3 processes).
381 *
382 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
383 *
384 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
385 * That is, the system wants to compute a value of decay such
386 * that the following for loop:
387 * for (i = 0; i < (5 * loadavg); i++)
388 * td_estcpu *= decay;
389 * will compute
390 * td_estcpu *= 0.1;
391 * for all values of loadavg:
392 *
393 * Mathematically this loop can be expressed by saying:
394 * decay ** (5 * loadavg) ~= .1
395 *
396 * The system computes decay as:
397 * decay = (2 * loadavg) / (2 * loadavg + 1)
398 *
399 * We wish to prove that the system's computation of decay
400 * will always fulfill the equation:
401 * decay ** (5 * loadavg) ~= .1
402 *
403 * If we compute b as:
404 * b = 2 * loadavg
405 * then
406 * decay = b / (b + 1)
407 *
408 * We now need to prove two things:
409 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
410 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
411 *
412 * Facts:
413 * For x close to zero, exp(x) =~ 1 + x, since
414 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
415 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
416 * For x close to zero, ln(1+x) =~ x, since
417 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
418 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
419 * ln(.1) =~ -2.30
420 *
421 * Proof of (1):
422 * Solve (factor)**(power) =~ .1 given power (5*loadav):
423 * solving for factor,
424 * ln(factor) =~ (-2.30/5*loadav), or
425 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
426 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
427 *
428 * Proof of (2):
429 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
430 * solving for power,
431 * power*ln(b/(b+1)) =~ -2.30, or
432 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
433 *
434 * Actual power values for the implemented algorithm are as follows:
435 * loadav: 1 2 3 4
436 * power: 5.68 10.32 14.94 19.55
437 */
438
439 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
440 #define loadfactor(loadav) (2 * (loadav))
441 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
442
443 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
444 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
445 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
446
447 /*
448 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
449 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
450 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
451 *
452 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
453 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
454 *
455 * If you don't want to bother with the faster/more-accurate formula, you
456 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
457 * (more general) method of calculating the %age of CPU used by a process.
458 */
459 #define CCPU_SHIFT 11
460
461 /*
462 * Recompute process priorities, every hz ticks.
463 * MP-safe, called without the Giant mutex.
464 */
465 /* ARGSUSED */
466 static void
467 schedcpu(void)
468 {
469 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
470 struct thread *td;
471 struct proc *p;
472 struct td_sched *ts;
473 int awake, realstathz;
474
475 realstathz = stathz ? stathz : hz;
476 sx_slock(&allproc_lock);
477 FOREACH_PROC_IN_SYSTEM(p) {
478 PROC_LOCK(p);
479 if (p->p_state == PRS_NEW) {
480 PROC_UNLOCK(p);
481 continue;
482 }
483 FOREACH_THREAD_IN_PROC(p, td) {
484 awake = 0;
485 thread_lock(td);
486 ts = td->td_sched;
487 /*
488 * Increment sleep time (if sleeping). We
489 * ignore overflow, as above.
490 */
491 /*
492 * The td_sched slptimes are not touched in wakeup
493 * because the thread may not HAVE everything in
494 * memory? XXX I think this is out of date.
495 */
496 if (TD_ON_RUNQ(td)) {
497 awake = 1;
498 td->td_flags &= ~TDF_DIDRUN;
499 } else if (TD_IS_RUNNING(td)) {
500 awake = 1;
501 /* Do not clear TDF_DIDRUN */
502 } else if (td->td_flags & TDF_DIDRUN) {
503 awake = 1;
504 td->td_flags &= ~TDF_DIDRUN;
505 }
506
507 /*
508 * ts_pctcpu is only for ps and ttyinfo().
509 */
510 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
511 /*
512 * If the td_sched has been idle the entire second,
513 * stop recalculating its priority until
514 * it wakes up.
515 */
516 if (ts->ts_cpticks != 0) {
517 #if (FSHIFT >= CCPU_SHIFT)
518 ts->ts_pctcpu += (realstathz == 100)
519 ? ((fixpt_t) ts->ts_cpticks) <<
520 (FSHIFT - CCPU_SHIFT) :
521 100 * (((fixpt_t) ts->ts_cpticks)
522 << (FSHIFT - CCPU_SHIFT)) / realstathz;
523 #else
524 ts->ts_pctcpu += ((FSCALE - ccpu) *
525 (ts->ts_cpticks *
526 FSCALE / realstathz)) >> FSHIFT;
527 #endif
528 ts->ts_cpticks = 0;
529 }
530 /*
531 * If there are ANY running threads in this process,
532 * then don't count it as sleeping.
533 * XXX: this is broken.
534 */
535 if (awake) {
536 if (ts->ts_slptime > 1) {
537 /*
538 * In an ideal world, this should not
539 * happen, because whoever woke us
540 * up from the long sleep should have
541 * unwound the slptime and reset our
542 * priority before we run at the stale
543 * priority. Should KASSERT at some
544 * point when all the cases are fixed.
545 */
546 updatepri(td);
547 }
548 ts->ts_slptime = 0;
549 } else
550 ts->ts_slptime++;
551 if (ts->ts_slptime > 1) {
552 thread_unlock(td);
553 continue;
554 }
555 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
556 resetpriority(td);
557 resetpriority_thread(td);
558 thread_unlock(td);
559 }
560 PROC_UNLOCK(p);
561 }
562 sx_sunlock(&allproc_lock);
563 }
564
565 /*
566 * Main loop for a kthread that executes schedcpu once a second.
567 */
568 static void
569 schedcpu_thread(void)
570 {
571
572 for (;;) {
573 schedcpu();
574 pause("-", hz);
575 }
576 }
577
578 /*
579 * Recalculate the priority of a process after it has slept for a while.
580 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
581 * least six times the loadfactor will decay td_estcpu to zero.
582 */
583 static void
584 updatepri(struct thread *td)
585 {
586 struct td_sched *ts;
587 fixpt_t loadfac;
588 unsigned int newcpu;
589
590 ts = td->td_sched;
591 loadfac = loadfactor(averunnable.ldavg[0]);
592 if (ts->ts_slptime > 5 * loadfac)
593 td->td_estcpu = 0;
594 else {
595 newcpu = td->td_estcpu;
596 ts->ts_slptime--; /* was incremented in schedcpu() */
597 while (newcpu && --ts->ts_slptime)
598 newcpu = decay_cpu(loadfac, newcpu);
599 td->td_estcpu = newcpu;
600 }
601 }
602
603 /*
604 * Compute the priority of a process when running in user mode.
605 * Arrange to reschedule if the resulting priority is better
606 * than that of the current process.
607 */
608 static void
609 resetpriority(struct thread *td)
610 {
611 register unsigned int newpriority;
612
613 if (td->td_pri_class == PRI_TIMESHARE) {
614 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
615 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
616 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
617 PRI_MAX_TIMESHARE);
618 sched_user_prio(td, newpriority);
619 }
620 }
621
622 /*
623 * Update the thread's priority when the associated process's user
624 * priority changes.
625 */
626 static void
627 resetpriority_thread(struct thread *td)
628 {
629
630 /* Only change threads with a time sharing user priority. */
631 if (td->td_priority < PRI_MIN_TIMESHARE ||
632 td->td_priority > PRI_MAX_TIMESHARE)
633 return;
634
635 /* XXX the whole needresched thing is broken, but not silly. */
636 maybe_resched(td);
637
638 sched_prio(td, td->td_user_pri);
639 }
640
641 /* ARGSUSED */
642 static void
643 sched_setup(void *dummy)
644 {
645 setup_runqs();
646
647 if (sched_quantum == 0)
648 sched_quantum = SCHED_QUANTUM;
649 hogticks = 2 * sched_quantum;
650
651 /* Account for thread0. */
652 sched_load_add();
653 }
654
655 /* External interfaces start here */
656
657 /*
658 * Very early in the boot some setup of scheduler-specific
659 * parts of proc0 and of some scheduler resources needs to be done.
660 * Called from:
661 * proc0_init()
662 */
663 void
664 schedinit(void)
665 {
666 /*
667 * Set up the scheduler specific parts of proc0.
668 */
669 proc0.p_sched = NULL; /* XXX */
670 thread0.td_sched = &td_sched0;
671 thread0.td_lock = &sched_lock;
672 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
673 }
674
675 int
676 sched_runnable(void)
677 {
678 #ifdef SMP
679 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
680 #else
681 return runq_check(&runq);
682 #endif
683 }
684
685 int
686 sched_rr_interval(void)
687 {
688 if (sched_quantum == 0)
689 sched_quantum = SCHED_QUANTUM;
690 return (sched_quantum);
691 }
692
693 /*
694 * We adjust the priority of the current process. The priority of
695 * a process gets worse as it accumulates CPU time. The cpu usage
696 * estimator (td_estcpu) is increased here. resetpriority() will
697 * compute a different priority each time td_estcpu increases by
698 * INVERSE_ESTCPU_WEIGHT
699 * (until MAXPRI is reached). The cpu usage estimator ramps up
700 * quite quickly when the process is running (linearly), and decays
701 * away exponentially, at a rate which is proportionally slower when
702 * the system is busy. The basic principle is that the system will
703 * 90% forget that the process used a lot of CPU time in 5 * loadav
704 * seconds. This causes the system to favor processes which haven't
705 * run much recently, and to round-robin among other processes.
706 */
707 void
708 sched_clock(struct thread *td)
709 {
710 struct pcpuidlestat *stat;
711 struct td_sched *ts;
712
713 THREAD_LOCK_ASSERT(td, MA_OWNED);
714 ts = td->td_sched;
715
716 ts->ts_cpticks++;
717 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
718 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
719 resetpriority(td);
720 resetpriority_thread(td);
721 }
722
723 /*
724 * Force a context switch if the current thread has used up a full
725 * quantum (default quantum is 100ms).
726 */
727 if (!TD_IS_IDLETHREAD(td) &&
728 ticks - PCPU_GET(switchticks) >= sched_quantum)
729 td->td_flags |= TDF_NEEDRESCHED;
730
731 stat = DPCPU_PTR(idlestat);
732 stat->oldidlecalls = stat->idlecalls;
733 stat->idlecalls = 0;
734 }
735
736 /*
737 * Charge child's scheduling CPU usage to parent.
738 */
739 void
740 sched_exit(struct proc *p, struct thread *td)
741 {
742
743 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
744 "prio:%d", td->td_priority);
745
746 PROC_LOCK_ASSERT(p, MA_OWNED);
747 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
748 }
749
750 void
751 sched_exit_thread(struct thread *td, struct thread *child)
752 {
753
754 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
755 "prio:%d", child->td_priority);
756 thread_lock(td);
757 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
758 thread_unlock(td);
759 thread_lock(child);
760 if ((child->td_flags & TDF_NOLOAD) == 0)
761 sched_load_rem();
762 thread_unlock(child);
763 }
764
765 void
766 sched_fork(struct thread *td, struct thread *childtd)
767 {
768 sched_fork_thread(td, childtd);
769 }
770
771 void
772 sched_fork_thread(struct thread *td, struct thread *childtd)
773 {
774 struct td_sched *ts;
775
776 childtd->td_estcpu = td->td_estcpu;
777 childtd->td_lock = &sched_lock;
778 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
779 childtd->td_priority = childtd->td_base_pri;
780 ts = childtd->td_sched;
781 bzero(ts, sizeof(*ts));
782 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
783 }
784
785 void
786 sched_nice(struct proc *p, int nice)
787 {
788 struct thread *td;
789
790 PROC_LOCK_ASSERT(p, MA_OWNED);
791 p->p_nice = nice;
792 FOREACH_THREAD_IN_PROC(p, td) {
793 thread_lock(td);
794 resetpriority(td);
795 resetpriority_thread(td);
796 thread_unlock(td);
797 }
798 }
799
800 void
801 sched_class(struct thread *td, int class)
802 {
803 THREAD_LOCK_ASSERT(td, MA_OWNED);
804 td->td_pri_class = class;
805 }
806
807 /*
808 * Adjust the priority of a thread.
809 */
810 static void
811 sched_priority(struct thread *td, u_char prio)
812 {
813
814
815 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
816 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
817 sched_tdname(curthread));
818 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
819 if (td != curthread && prio > td->td_priority) {
820 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
821 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
822 prio, KTR_ATTR_LINKED, sched_tdname(td));
823 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
824 curthread);
825 }
826 THREAD_LOCK_ASSERT(td, MA_OWNED);
827 if (td->td_priority == prio)
828 return;
829 td->td_priority = prio;
830 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
831 sched_rem(td);
832 sched_add(td, SRQ_BORING);
833 }
834 }
835
836 /*
837 * Update a thread's priority when it is lent another thread's
838 * priority.
839 */
840 void
841 sched_lend_prio(struct thread *td, u_char prio)
842 {
843
844 td->td_flags |= TDF_BORROWING;
845 sched_priority(td, prio);
846 }
847
848 /*
849 * Restore a thread's priority when priority propagation is
850 * over. The prio argument is the minimum priority the thread
851 * needs to have to satisfy other possible priority lending
852 * requests. If the thread's regulary priority is less
853 * important than prio the thread will keep a priority boost
854 * of prio.
855 */
856 void
857 sched_unlend_prio(struct thread *td, u_char prio)
858 {
859 u_char base_pri;
860
861 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
862 td->td_base_pri <= PRI_MAX_TIMESHARE)
863 base_pri = td->td_user_pri;
864 else
865 base_pri = td->td_base_pri;
866 if (prio >= base_pri) {
867 td->td_flags &= ~TDF_BORROWING;
868 sched_prio(td, base_pri);
869 } else
870 sched_lend_prio(td, prio);
871 }
872
873 void
874 sched_prio(struct thread *td, u_char prio)
875 {
876 u_char oldprio;
877
878 /* First, update the base priority. */
879 td->td_base_pri = prio;
880
881 /*
882 * If the thread is borrowing another thread's priority, don't ever
883 * lower the priority.
884 */
885 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
886 return;
887
888 /* Change the real priority. */
889 oldprio = td->td_priority;
890 sched_priority(td, prio);
891
892 /*
893 * If the thread is on a turnstile, then let the turnstile update
894 * its state.
895 */
896 if (TD_ON_LOCK(td) && oldprio != prio)
897 turnstile_adjust(td, oldprio);
898 }
899
900 void
901 sched_user_prio(struct thread *td, u_char prio)
902 {
903
904 THREAD_LOCK_ASSERT(td, MA_OWNED);
905 td->td_base_user_pri = prio;
906 if (td->td_lend_user_pri <= prio)
907 return;
908 td->td_user_pri = prio;
909 }
910
911 void
912 sched_lend_user_prio(struct thread *td, u_char prio)
913 {
914
915 THREAD_LOCK_ASSERT(td, MA_OWNED);
916 td->td_lend_user_pri = prio;
917 td->td_user_pri = min(prio, td->td_base_user_pri);
918 if (td->td_priority > td->td_user_pri)
919 sched_prio(td, td->td_user_pri);
920 else if (td->td_priority != td->td_user_pri)
921 td->td_flags |= TDF_NEEDRESCHED;
922 }
923
924 void
925 sched_sleep(struct thread *td, int pri)
926 {
927
928 THREAD_LOCK_ASSERT(td, MA_OWNED);
929 td->td_slptick = ticks;
930 td->td_sched->ts_slptime = 0;
931 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
932 sched_prio(td, pri);
933 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
934 td->td_flags |= TDF_CANSWAP;
935 }
936
937 void
938 sched_switch(struct thread *td, struct thread *newtd, int flags)
939 {
940 struct mtx *tmtx;
941 struct td_sched *ts;
942 struct proc *p;
943
944 tmtx = NULL;
945 ts = td->td_sched;
946 p = td->td_proc;
947
948 THREAD_LOCK_ASSERT(td, MA_OWNED);
949
950 /*
951 * Switch to the sched lock to fix things up and pick
952 * a new thread.
953 * Block the td_lock in order to avoid breaking the critical path.
954 */
955 if (td->td_lock != &sched_lock) {
956 mtx_lock_spin(&sched_lock);
957 tmtx = thread_lock_block(td);
958 }
959
960 if ((td->td_flags & TDF_NOLOAD) == 0)
961 sched_load_rem();
962
963 td->td_lastcpu = td->td_oncpu;
964 if (!(flags & SW_PREEMPT))
965 td->td_flags &= ~TDF_NEEDRESCHED;
966 td->td_owepreempt = 0;
967 td->td_oncpu = NOCPU;
968
969 /*
970 * At the last moment, if this thread is still marked RUNNING,
971 * then put it back on the run queue as it has not been suspended
972 * or stopped or any thing else similar. We never put the idle
973 * threads on the run queue, however.
974 */
975 if (td->td_flags & TDF_IDLETD) {
976 TD_SET_CAN_RUN(td);
977 #ifdef SMP
978 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
979 #endif
980 } else {
981 if (TD_IS_RUNNING(td)) {
982 /* Put us back on the run queue. */
983 sched_add(td, (flags & SW_PREEMPT) ?
984 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
985 SRQ_OURSELF|SRQ_YIELDING);
986 }
987 }
988 if (newtd) {
989 /*
990 * The thread we are about to run needs to be counted
991 * as if it had been added to the run queue and selected.
992 * It came from:
993 * * A preemption
994 * * An upcall
995 * * A followon
996 */
997 KASSERT((newtd->td_inhibitors == 0),
998 ("trying to run inhibited thread"));
999 newtd->td_flags |= TDF_DIDRUN;
1000 TD_SET_RUNNING(newtd);
1001 if ((newtd->td_flags & TDF_NOLOAD) == 0)
1002 sched_load_add();
1003 } else {
1004 newtd = choosethread();
1005 MPASS(newtd->td_lock == &sched_lock);
1006 }
1007
1008 if (td != newtd) {
1009 #ifdef HWPMC_HOOKS
1010 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1011 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1012 #endif
1013
1014 SDT_PROBE2(sched, , , off_cpu, td, td->td_proc);
1015
1016 /* I feel sleepy */
1017 lock_profile_release_lock(&sched_lock.lock_object);
1018 #ifdef KDTRACE_HOOKS
1019 /*
1020 * If DTrace has set the active vtime enum to anything
1021 * other than INACTIVE (0), then it should have set the
1022 * function to call.
1023 */
1024 if (dtrace_vtime_active)
1025 (*dtrace_vtime_switch_func)(newtd);
1026 #endif
1027
1028 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1029 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1030 0, 0, __FILE__, __LINE__);
1031 /*
1032 * Where am I? What year is it?
1033 * We are in the same thread that went to sleep above,
1034 * but any amount of time may have passed. All our context
1035 * will still be available as will local variables.
1036 * PCPU values however may have changed as we may have
1037 * changed CPU so don't trust cached values of them.
1038 * New threads will go to fork_exit() instead of here
1039 * so if you change things here you may need to change
1040 * things there too.
1041 *
1042 * If the thread above was exiting it will never wake
1043 * up again here, so either it has saved everything it
1044 * needed to, or the thread_wait() or wait() will
1045 * need to reap it.
1046 */
1047
1048 SDT_PROBE0(sched, , , on_cpu);
1049 #ifdef HWPMC_HOOKS
1050 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1051 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1052 #endif
1053 } else
1054 SDT_PROBE0(sched, , , remain_cpu);
1055
1056 #ifdef SMP
1057 if (td->td_flags & TDF_IDLETD)
1058 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1059 #endif
1060 sched_lock.mtx_lock = (uintptr_t)td;
1061 td->td_oncpu = PCPU_GET(cpuid);
1062 MPASS(td->td_lock == &sched_lock);
1063 }
1064
1065 void
1066 sched_wakeup(struct thread *td)
1067 {
1068 struct td_sched *ts;
1069
1070 THREAD_LOCK_ASSERT(td, MA_OWNED);
1071 ts = td->td_sched;
1072 td->td_flags &= ~TDF_CANSWAP;
1073 if (ts->ts_slptime > 1) {
1074 updatepri(td);
1075 resetpriority(td);
1076 }
1077 td->td_slptick = 0;
1078 ts->ts_slptime = 0;
1079 sched_add(td, SRQ_BORING);
1080 }
1081
1082 #ifdef SMP
1083 static int
1084 forward_wakeup(int cpunum)
1085 {
1086 struct pcpu *pc;
1087 cpuset_t dontuse, map, map2;
1088 u_int id, me;
1089 int iscpuset;
1090
1091 mtx_assert(&sched_lock, MA_OWNED);
1092
1093 CTR0(KTR_RUNQ, "forward_wakeup()");
1094
1095 if ((!forward_wakeup_enabled) ||
1096 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1097 return (0);
1098 if (!smp_started || cold || panicstr)
1099 return (0);
1100
1101 forward_wakeups_requested++;
1102
1103 /*
1104 * Check the idle mask we received against what we calculated
1105 * before in the old version.
1106 */
1107 me = PCPU_GET(cpuid);
1108
1109 /* Don't bother if we should be doing it ourself. */
1110 if (CPU_ISSET(me, &idle_cpus_mask) &&
1111 (cpunum == NOCPU || me == cpunum))
1112 return (0);
1113
1114 CPU_SETOF(me, &dontuse);
1115 CPU_OR(&dontuse, &stopped_cpus);
1116 CPU_OR(&dontuse, &hlt_cpus_mask);
1117 CPU_ZERO(&map2);
1118 if (forward_wakeup_use_loop) {
1119 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1120 id = pc->pc_cpuid;
1121 if (!CPU_ISSET(id, &dontuse) &&
1122 pc->pc_curthread == pc->pc_idlethread) {
1123 CPU_SET(id, &map2);
1124 }
1125 }
1126 }
1127
1128 if (forward_wakeup_use_mask) {
1129 map = idle_cpus_mask;
1130 CPU_NAND(&map, &dontuse);
1131
1132 /* If they are both on, compare and use loop if different. */
1133 if (forward_wakeup_use_loop) {
1134 if (CPU_CMP(&map, &map2)) {
1135 printf("map != map2, loop method preferred\n");
1136 map = map2;
1137 }
1138 }
1139 } else {
1140 map = map2;
1141 }
1142
1143 /* If we only allow a specific CPU, then mask off all the others. */
1144 if (cpunum != NOCPU) {
1145 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1146 iscpuset = CPU_ISSET(cpunum, &map);
1147 if (iscpuset == 0)
1148 CPU_ZERO(&map);
1149 else
1150 CPU_SETOF(cpunum, &map);
1151 }
1152 if (!CPU_EMPTY(&map)) {
1153 forward_wakeups_delivered++;
1154 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1155 id = pc->pc_cpuid;
1156 if (!CPU_ISSET(id, &map))
1157 continue;
1158 if (cpu_idle_wakeup(pc->pc_cpuid))
1159 CPU_CLR(id, &map);
1160 }
1161 if (!CPU_EMPTY(&map))
1162 ipi_selected(map, IPI_AST);
1163 return (1);
1164 }
1165 if (cpunum == NOCPU)
1166 printf("forward_wakeup: Idle processor not found\n");
1167 return (0);
1168 }
1169
1170 static void
1171 kick_other_cpu(int pri, int cpuid)
1172 {
1173 struct pcpu *pcpu;
1174 int cpri;
1175
1176 pcpu = pcpu_find(cpuid);
1177 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1178 forward_wakeups_delivered++;
1179 if (!cpu_idle_wakeup(cpuid))
1180 ipi_cpu(cpuid, IPI_AST);
1181 return;
1182 }
1183
1184 cpri = pcpu->pc_curthread->td_priority;
1185 if (pri >= cpri)
1186 return;
1187
1188 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1189 #if !defined(FULL_PREEMPTION)
1190 if (pri <= PRI_MAX_ITHD)
1191 #endif /* ! FULL_PREEMPTION */
1192 {
1193 ipi_cpu(cpuid, IPI_PREEMPT);
1194 return;
1195 }
1196 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1197
1198 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1199 ipi_cpu(cpuid, IPI_AST);
1200 return;
1201 }
1202 #endif /* SMP */
1203
1204 #ifdef SMP
1205 static int
1206 sched_pickcpu(struct thread *td)
1207 {
1208 int best, cpu;
1209
1210 mtx_assert(&sched_lock, MA_OWNED);
1211
1212 if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1213 best = td->td_lastcpu;
1214 else
1215 best = NOCPU;
1216 CPU_FOREACH(cpu) {
1217 if (!THREAD_CAN_SCHED(td, cpu))
1218 continue;
1219
1220 if (best == NOCPU)
1221 best = cpu;
1222 else if (runq_length[cpu] < runq_length[best])
1223 best = cpu;
1224 }
1225 KASSERT(best != NOCPU, ("no valid CPUs"));
1226
1227 return (best);
1228 }
1229 #endif
1230
1231 void
1232 sched_add(struct thread *td, int flags)
1233 #ifdef SMP
1234 {
1235 cpuset_t tidlemsk;
1236 struct td_sched *ts;
1237 u_int cpu, cpuid;
1238 int forwarded = 0;
1239 int single_cpu = 0;
1240
1241 ts = td->td_sched;
1242 THREAD_LOCK_ASSERT(td, MA_OWNED);
1243 KASSERT((td->td_inhibitors == 0),
1244 ("sched_add: trying to run inhibited thread"));
1245 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1246 ("sched_add: bad thread state"));
1247 KASSERT(td->td_flags & TDF_INMEM,
1248 ("sched_add: thread swapped out"));
1249
1250 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1251 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1252 sched_tdname(curthread));
1253 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1254 KTR_ATTR_LINKED, sched_tdname(td));
1255 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1256 flags & SRQ_PREEMPTED);
1257
1258
1259 /*
1260 * Now that the thread is moving to the run-queue, set the lock
1261 * to the scheduler's lock.
1262 */
1263 if (td->td_lock != &sched_lock) {
1264 mtx_lock_spin(&sched_lock);
1265 thread_lock_set(td, &sched_lock);
1266 }
1267 TD_SET_RUNQ(td);
1268
1269 /*
1270 * If SMP is started and the thread is pinned or otherwise limited to
1271 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1272 * Otherwise, queue the thread to the global run queue.
1273 *
1274 * If SMP has not yet been started we must use the global run queue
1275 * as per-CPU state may not be initialized yet and we may crash if we
1276 * try to access the per-CPU run queues.
1277 */
1278 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1279 ts->ts_flags & TSF_AFFINITY)) {
1280 if (td->td_pinned != 0)
1281 cpu = td->td_lastcpu;
1282 else if (td->td_flags & TDF_BOUND) {
1283 /* Find CPU from bound runq. */
1284 KASSERT(SKE_RUNQ_PCPU(ts),
1285 ("sched_add: bound td_sched not on cpu runq"));
1286 cpu = ts->ts_runq - &runq_pcpu[0];
1287 } else
1288 /* Find a valid CPU for our cpuset */
1289 cpu = sched_pickcpu(td);
1290 ts->ts_runq = &runq_pcpu[cpu];
1291 single_cpu = 1;
1292 CTR3(KTR_RUNQ,
1293 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1294 cpu);
1295 } else {
1296 CTR2(KTR_RUNQ,
1297 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1298 td);
1299 cpu = NOCPU;
1300 ts->ts_runq = &runq;
1301 }
1302
1303 cpuid = PCPU_GET(cpuid);
1304 if (single_cpu && cpu != cpuid) {
1305 kick_other_cpu(td->td_priority, cpu);
1306 } else {
1307 if (!single_cpu) {
1308 tidlemsk = idle_cpus_mask;
1309 CPU_NAND(&tidlemsk, &hlt_cpus_mask);
1310 CPU_CLR(cpuid, &tidlemsk);
1311
1312 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1313 ((flags & SRQ_INTR) == 0) &&
1314 !CPU_EMPTY(&tidlemsk))
1315 forwarded = forward_wakeup(cpu);
1316 }
1317
1318 if (!forwarded) {
1319 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1320 return;
1321 else
1322 maybe_resched(td);
1323 }
1324 }
1325
1326 if ((td->td_flags & TDF_NOLOAD) == 0)
1327 sched_load_add();
1328 runq_add(ts->ts_runq, td, flags);
1329 if (cpu != NOCPU)
1330 runq_length[cpu]++;
1331 }
1332 #else /* SMP */
1333 {
1334 struct td_sched *ts;
1335
1336 ts = td->td_sched;
1337 THREAD_LOCK_ASSERT(td, MA_OWNED);
1338 KASSERT((td->td_inhibitors == 0),
1339 ("sched_add: trying to run inhibited thread"));
1340 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1341 ("sched_add: bad thread state"));
1342 KASSERT(td->td_flags & TDF_INMEM,
1343 ("sched_add: thread swapped out"));
1344 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1345 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1346 sched_tdname(curthread));
1347 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1348 KTR_ATTR_LINKED, sched_tdname(td));
1349 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1350 flags & SRQ_PREEMPTED);
1351
1352 /*
1353 * Now that the thread is moving to the run-queue, set the lock
1354 * to the scheduler's lock.
1355 */
1356 if (td->td_lock != &sched_lock) {
1357 mtx_lock_spin(&sched_lock);
1358 thread_lock_set(td, &sched_lock);
1359 }
1360 TD_SET_RUNQ(td);
1361 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1362 ts->ts_runq = &runq;
1363
1364 /*
1365 * If we are yielding (on the way out anyhow) or the thread
1366 * being saved is US, then don't try be smart about preemption
1367 * or kicking off another CPU as it won't help and may hinder.
1368 * In the YIEDLING case, we are about to run whoever is being
1369 * put in the queue anyhow, and in the OURSELF case, we are
1370 * puting ourself on the run queue which also only happens
1371 * when we are about to yield.
1372 */
1373 if ((flags & SRQ_YIELDING) == 0) {
1374 if (maybe_preempt(td))
1375 return;
1376 }
1377 if ((td->td_flags & TDF_NOLOAD) == 0)
1378 sched_load_add();
1379 runq_add(ts->ts_runq, td, flags);
1380 maybe_resched(td);
1381 }
1382 #endif /* SMP */
1383
1384 void
1385 sched_rem(struct thread *td)
1386 {
1387 struct td_sched *ts;
1388
1389 ts = td->td_sched;
1390 KASSERT(td->td_flags & TDF_INMEM,
1391 ("sched_rem: thread swapped out"));
1392 KASSERT(TD_ON_RUNQ(td),
1393 ("sched_rem: thread not on run queue"));
1394 mtx_assert(&sched_lock, MA_OWNED);
1395 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1396 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1397 sched_tdname(curthread));
1398 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1399
1400 if ((td->td_flags & TDF_NOLOAD) == 0)
1401 sched_load_rem();
1402 #ifdef SMP
1403 if (ts->ts_runq != &runq)
1404 runq_length[ts->ts_runq - runq_pcpu]--;
1405 #endif
1406 runq_remove(ts->ts_runq, td);
1407 TD_SET_CAN_RUN(td);
1408 }
1409
1410 /*
1411 * Select threads to run. Note that running threads still consume a
1412 * slot.
1413 */
1414 struct thread *
1415 sched_choose(void)
1416 {
1417 struct thread *td;
1418 struct runq *rq;
1419
1420 mtx_assert(&sched_lock, MA_OWNED);
1421 #ifdef SMP
1422 struct thread *tdcpu;
1423
1424 rq = &runq;
1425 td = runq_choose_fuzz(&runq, runq_fuzz);
1426 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1427
1428 if (td == NULL ||
1429 (tdcpu != NULL &&
1430 tdcpu->td_priority < td->td_priority)) {
1431 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1432 PCPU_GET(cpuid));
1433 td = tdcpu;
1434 rq = &runq_pcpu[PCPU_GET(cpuid)];
1435 } else {
1436 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1437 }
1438
1439 #else
1440 rq = &runq;
1441 td = runq_choose(&runq);
1442 #endif
1443
1444 if (td) {
1445 #ifdef SMP
1446 if (td == tdcpu)
1447 runq_length[PCPU_GET(cpuid)]--;
1448 #endif
1449 runq_remove(rq, td);
1450 td->td_flags |= TDF_DIDRUN;
1451
1452 KASSERT(td->td_flags & TDF_INMEM,
1453 ("sched_choose: thread swapped out"));
1454 return (td);
1455 }
1456 return (PCPU_GET(idlethread));
1457 }
1458
1459 void
1460 sched_preempt(struct thread *td)
1461 {
1462
1463 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1464 thread_lock(td);
1465 if (td->td_critnest > 1)
1466 td->td_owepreempt = 1;
1467 else
1468 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1469 thread_unlock(td);
1470 }
1471
1472 void
1473 sched_userret(struct thread *td)
1474 {
1475 /*
1476 * XXX we cheat slightly on the locking here to avoid locking in
1477 * the usual case. Setting td_priority here is essentially an
1478 * incomplete workaround for not setting it properly elsewhere.
1479 * Now that some interrupt handlers are threads, not setting it
1480 * properly elsewhere can clobber it in the window between setting
1481 * it here and returning to user mode, so don't waste time setting
1482 * it perfectly here.
1483 */
1484 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1485 ("thread with borrowed priority returning to userland"));
1486 if (td->td_priority != td->td_user_pri) {
1487 thread_lock(td);
1488 td->td_priority = td->td_user_pri;
1489 td->td_base_pri = td->td_user_pri;
1490 thread_unlock(td);
1491 }
1492 }
1493
1494 void
1495 sched_bind(struct thread *td, int cpu)
1496 {
1497 struct td_sched *ts;
1498
1499 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1500 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1501
1502 ts = td->td_sched;
1503
1504 td->td_flags |= TDF_BOUND;
1505 #ifdef SMP
1506 ts->ts_runq = &runq_pcpu[cpu];
1507 if (PCPU_GET(cpuid) == cpu)
1508 return;
1509
1510 mi_switch(SW_VOL, NULL);
1511 #endif
1512 }
1513
1514 void
1515 sched_unbind(struct thread* td)
1516 {
1517 THREAD_LOCK_ASSERT(td, MA_OWNED);
1518 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1519 td->td_flags &= ~TDF_BOUND;
1520 }
1521
1522 int
1523 sched_is_bound(struct thread *td)
1524 {
1525 THREAD_LOCK_ASSERT(td, MA_OWNED);
1526 return (td->td_flags & TDF_BOUND);
1527 }
1528
1529 void
1530 sched_relinquish(struct thread *td)
1531 {
1532 thread_lock(td);
1533 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1534 thread_unlock(td);
1535 }
1536
1537 int
1538 sched_load(void)
1539 {
1540 return (sched_tdcnt);
1541 }
1542
1543 int
1544 sched_sizeof_proc(void)
1545 {
1546 return (sizeof(struct proc));
1547 }
1548
1549 int
1550 sched_sizeof_thread(void)
1551 {
1552 return (sizeof(struct thread) + sizeof(struct td_sched));
1553 }
1554
1555 fixpt_t
1556 sched_pctcpu(struct thread *td)
1557 {
1558 struct td_sched *ts;
1559
1560 THREAD_LOCK_ASSERT(td, MA_OWNED);
1561 ts = td->td_sched;
1562 return (ts->ts_pctcpu);
1563 }
1564
1565 void
1566 sched_tick(int cnt)
1567 {
1568 }
1569
1570 /*
1571 * The actual idle process.
1572 */
1573 void
1574 sched_idletd(void *dummy)
1575 {
1576 struct pcpuidlestat *stat;
1577
1578 stat = DPCPU_PTR(idlestat);
1579 for (;;) {
1580 mtx_assert(&Giant, MA_NOTOWNED);
1581
1582 while (sched_runnable() == 0) {
1583 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1584 stat->idlecalls++;
1585 }
1586
1587 mtx_lock_spin(&sched_lock);
1588 mi_switch(SW_VOL | SWT_IDLE, NULL);
1589 mtx_unlock_spin(&sched_lock);
1590 }
1591 }
1592
1593 /*
1594 * A CPU is entering for the first time or a thread is exiting.
1595 */
1596 void
1597 sched_throw(struct thread *td)
1598 {
1599 /*
1600 * Correct spinlock nesting. The idle thread context that we are
1601 * borrowing was created so that it would start out with a single
1602 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1603 * explicitly acquired locks in this function, the nesting count
1604 * is now 2 rather than 1. Since we are nested, calling
1605 * spinlock_exit() will simply adjust the counts without allowing
1606 * spin lock using code to interrupt us.
1607 */
1608 if (td == NULL) {
1609 mtx_lock_spin(&sched_lock);
1610 spinlock_exit();
1611 PCPU_SET(switchtime, cpu_ticks());
1612 PCPU_SET(switchticks, ticks);
1613 } else {
1614 lock_profile_release_lock(&sched_lock.lock_object);
1615 MPASS(td->td_lock == &sched_lock);
1616 }
1617 mtx_assert(&sched_lock, MA_OWNED);
1618 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1619 cpu_throw(td, choosethread()); /* doesn't return */
1620 }
1621
1622 void
1623 sched_fork_exit(struct thread *td)
1624 {
1625
1626 /*
1627 * Finish setting up thread glue so that it begins execution in a
1628 * non-nested critical section with sched_lock held but not recursed.
1629 */
1630 td->td_oncpu = PCPU_GET(cpuid);
1631 sched_lock.mtx_lock = (uintptr_t)td;
1632 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1633 0, 0, __FILE__, __LINE__);
1634 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1635 }
1636
1637 char *
1638 sched_tdname(struct thread *td)
1639 {
1640 #ifdef KTR
1641 struct td_sched *ts;
1642
1643 ts = td->td_sched;
1644 if (ts->ts_name[0] == '\0')
1645 snprintf(ts->ts_name, sizeof(ts->ts_name),
1646 "%s tid %d", td->td_name, td->td_tid);
1647 return (ts->ts_name);
1648 #else
1649 return (td->td_name);
1650 #endif
1651 }
1652
1653 #ifdef KTR
1654 void
1655 sched_clear_tdname(struct thread *td)
1656 {
1657 struct td_sched *ts;
1658
1659 ts = td->td_sched;
1660 ts->ts_name[0] = '\0';
1661 }
1662 #endif
1663
1664 void
1665 sched_affinity(struct thread *td)
1666 {
1667 #ifdef SMP
1668 struct td_sched *ts;
1669 int cpu;
1670
1671 THREAD_LOCK_ASSERT(td, MA_OWNED);
1672
1673 /*
1674 * Set the TSF_AFFINITY flag if there is at least one CPU this
1675 * thread can't run on.
1676 */
1677 ts = td->td_sched;
1678 ts->ts_flags &= ~TSF_AFFINITY;
1679 CPU_FOREACH(cpu) {
1680 if (!THREAD_CAN_SCHED(td, cpu)) {
1681 ts->ts_flags |= TSF_AFFINITY;
1682 break;
1683 }
1684 }
1685
1686 /*
1687 * If this thread can run on all CPUs, nothing else to do.
1688 */
1689 if (!(ts->ts_flags & TSF_AFFINITY))
1690 return;
1691
1692 /* Pinned threads and bound threads should be left alone. */
1693 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1694 return;
1695
1696 switch (td->td_state) {
1697 case TDS_RUNQ:
1698 /*
1699 * If we are on a per-CPU runqueue that is in the set,
1700 * then nothing needs to be done.
1701 */
1702 if (ts->ts_runq != &runq &&
1703 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1704 return;
1705
1706 /* Put this thread on a valid per-CPU runqueue. */
1707 sched_rem(td);
1708 sched_add(td, SRQ_BORING);
1709 break;
1710 case TDS_RUNNING:
1711 /*
1712 * See if our current CPU is in the set. If not, force a
1713 * context switch.
1714 */
1715 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1716 return;
1717
1718 td->td_flags |= TDF_NEEDRESCHED;
1719 if (td != curthread)
1720 ipi_cpu(cpu, IPI_AST);
1721 break;
1722 default:
1723 break;
1724 }
1725 #endif
1726 }
Cache object: dbc03ad52ea1eb6e6ed690f4043547a8
|