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/smp.h>
54#include <sys/sysctl.h>
55#include <sys/sx.h>
56#include <sys/turnstile.h>
57#include <sys/umtx.h>
58#include <machine/pcb.h>
59#include <machine/smp.h>
60
61#ifdef HWPMC_HOOKS
62#include <sys/pmckern.h>
63#endif
64
65#ifdef KDTRACE_HOOKS
66#include <sys/dtrace_bsd.h>
67int dtrace_vtime_active;
68dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
69#endif
70
71/*
72 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
73 * the range 100-256 Hz (approximately).
74 */
75#define ESTCPULIM(e) \
76 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
77 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
78#ifdef SMP
79#define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
80#else
81#define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
82#endif
83#define NICE_WEIGHT 1 /* Priorities per nice level. */
84
85#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
86
87/*
88 * The schedulable entity that runs a context.
89 * This is an extension to the thread structure and is tailored to
90 * the requirements of this scheduler
91 */
92struct td_sched {
93 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
94 int ts_cpticks; /* (j) Ticks of cpu time. */
95 int ts_slptime; /* (j) Seconds !RUNNING. */
96 int ts_flags;
97 struct runq *ts_runq; /* runq the thread is currently on */
98#ifdef KTR
99 char ts_name[TS_NAME_LEN];
100#endif
101};
102
103/* flags kept in td_flags */
104#define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
105#define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
106
107/* flags kept in ts_flags */
108#define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
109
110#define SKE_RUNQ_PCPU(ts) \
111 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
112
113#define THREAD_CAN_SCHED(td, cpu) \
114 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
115
116static struct td_sched td_sched0;
117struct mtx sched_lock;
118
119static int sched_tdcnt; /* Total runnable threads in the system. */
120static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
121#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
122
123static void setup_runqs(void);
124static void schedcpu(void);
125static void schedcpu_thread(void);
126static void sched_priority(struct thread *td, u_char prio);
127static void sched_setup(void *dummy);
128static void maybe_resched(struct thread *td);
129static void updatepri(struct thread *td);
130static void resetpriority(struct thread *td);
131static void resetpriority_thread(struct thread *td);
132#ifdef SMP
133static int sched_pickcpu(struct thread *td);
134static int forward_wakeup(int cpunum);
135static void kick_other_cpu(int pri, int cpuid);
136#endif
137
138static struct kproc_desc sched_kp = {
139 "schedcpu",
140 schedcpu_thread,
141 NULL
142};
143SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
144 &sched_kp);
145SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
146
147/*
148 * Global run queue.
149 */
150static struct runq runq;
151
152#ifdef SMP
153/*
154 * Per-CPU run queues
155 */
156static struct runq runq_pcpu[MAXCPU];
157long runq_length[MAXCPU];
158#endif
159
160static void
161setup_runqs(void)
162{
163#ifdef SMP
164 int i;
165
166 for (i = 0; i < MAXCPU; ++i)
167 runq_init(&runq_pcpu[i]);
168#endif
169
170 runq_init(&runq);
171}
172
173static int
174sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
175{
176 int error, new_val;
177
178 new_val = sched_quantum * tick;
179 error = sysctl_handle_int(oidp, &new_val, 0, req);
180 if (error != 0 || req->newptr == NULL)
181 return (error);
182 if (new_val < tick)
183 return (EINVAL);
184 sched_quantum = new_val / tick;
185 hogticks = 2 * sched_quantum;
186 return (0);
187}
188
189SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
190
191SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
192 "Scheduler name");
193
194SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
195 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
196 "Roundrobin scheduling quantum in microseconds");
197
198#ifdef SMP
199/* Enable forwarding of wakeups to all other cpus */
200SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
201
202static int runq_fuzz = 1;
203SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
204
205static int forward_wakeup_enabled = 1;
206SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
207 &forward_wakeup_enabled, 0,
208 "Forwarding of wakeup to idle CPUs");
209
210static int forward_wakeups_requested = 0;
211SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
212 &forward_wakeups_requested, 0,
213 "Requests for Forwarding of wakeup to idle CPUs");
214
215static int forward_wakeups_delivered = 0;
216SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
217 &forward_wakeups_delivered, 0,
218 "Completed Forwarding of wakeup to idle CPUs");
219
220static int forward_wakeup_use_mask = 1;
221SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
222 &forward_wakeup_use_mask, 0,
223 "Use the mask of idle cpus");
224
225static int forward_wakeup_use_loop = 0;
226SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
227 &forward_wakeup_use_loop, 0,
228 "Use a loop to find idle cpus");
229
230static int forward_wakeup_use_single = 0;
231SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
232 &forward_wakeup_use_single, 0,
233 "Only signal one idle cpu");
234
235static int forward_wakeup_use_htt = 0;
236SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
237 &forward_wakeup_use_htt, 0,
238 "account for htt");
239
240#endif
241#if 0
242static int sched_followon = 0;
243SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
244 &sched_followon, 0,
245 "allow threads to share a quantum");
246#endif
247
248static __inline void
249sched_load_add(void)
250{
251
252 sched_tdcnt++;
253 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
254}
255
256static __inline void
257sched_load_rem(void)
258{
259
260 sched_tdcnt--;
261 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
262}
263/*
264 * Arrange to reschedule if necessary, taking the priorities and
265 * schedulers into account.
266 */
267static void
268maybe_resched(struct thread *td)
269{
270
271 THREAD_LOCK_ASSERT(td, MA_OWNED);
272 if (td->td_priority < curthread->td_priority)
273 curthread->td_flags |= TDF_NEEDRESCHED;
274}
275
276/*
277 * This function is called when a thread is about to be put on run queue
278 * because it has been made runnable or its priority has been adjusted. It
279 * determines if the new thread should be immediately preempted to. If so,
280 * it switches to it and eventually returns true. If not, it returns false
281 * so that the caller may place the thread on an appropriate run queue.
282 */
283int
284maybe_preempt(struct thread *td)
285{
286#ifdef PREEMPTION
287 struct thread *ctd;
288 int cpri, pri;
289
290 /*
291 * The new thread should not preempt the current thread if any of the
292 * following conditions are true:
293 *
294 * - The kernel is in the throes of crashing (panicstr).
295 * - The current thread has a higher (numerically lower) or
296 * equivalent priority. Note that this prevents curthread from
297 * trying to preempt to itself.
298 * - It is too early in the boot for context switches (cold is set).
299 * - The current thread has an inhibitor set or is in the process of
300 * exiting. In this case, the current thread is about to switch
301 * out anyways, so there's no point in preempting. If we did,
302 * the current thread would not be properly resumed as well, so
303 * just avoid that whole landmine.
304 * - If the new thread's priority is not a realtime priority and
305 * the current thread's priority is not an idle priority and
306 * FULL_PREEMPTION is disabled.
307 *
308 * If all of these conditions are false, but the current thread is in
309 * a nested critical section, then we have to defer the preemption
310 * until we exit the critical section. Otherwise, switch immediately
311 * to the new thread.
312 */
313 ctd = curthread;
314 THREAD_LOCK_ASSERT(td, MA_OWNED);
315 KASSERT((td->td_inhibitors == 0),
316 ("maybe_preempt: trying to run inhibited thread"));
317 pri = td->td_priority;
318 cpri = ctd->td_priority;
319 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
320 TD_IS_INHIBITED(ctd))
321 return (0);
322#ifndef FULL_PREEMPTION
323 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
324 return (0);
325#endif
326
327 if (ctd->td_critnest > 1) {
328 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
329 ctd->td_critnest);
330 ctd->td_owepreempt = 1;
331 return (0);
332 }
333 /*
334 * Thread is runnable but not yet put on system run queue.
335 */
336 MPASS(ctd->td_lock == td->td_lock);
337 MPASS(TD_ON_RUNQ(td));
338 TD_SET_RUNNING(td);
339 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
340 td->td_proc->p_pid, td->td_name);
341 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
342 /*
343 * td's lock pointer may have changed. We have to return with it
344 * locked.
345 */
346 spinlock_enter();
347 thread_unlock(ctd);
348 thread_lock(td);
349 spinlock_exit();
350 return (1);
351#else
352 return (0);
353#endif
354}
355
356/*
357 * Constants for digital decay and forget:
358 * 90% of (td_estcpu) usage in 5 * loadav time
359 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
360 * Note that, as ps(1) mentions, this can let percentages
361 * total over 100% (I've seen 137.9% for 3 processes).
362 *
363 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
364 *
365 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
366 * That is, the system wants to compute a value of decay such
367 * that the following for loop:
368 * for (i = 0; i < (5 * loadavg); i++)
369 * td_estcpu *= decay;
370 * will compute
371 * td_estcpu *= 0.1;
372 * for all values of loadavg:
373 *
374 * Mathematically this loop can be expressed by saying:
375 * decay ** (5 * loadavg) ~= .1
376 *
377 * The system computes decay as:
378 * decay = (2 * loadavg) / (2 * loadavg + 1)
379 *
380 * We wish to prove that the system's computation of decay
381 * will always fulfill the equation:
382 * decay ** (5 * loadavg) ~= .1
383 *
384 * If we compute b as:
385 * b = 2 * loadavg
386 * then
387 * decay = b / (b + 1)
388 *
389 * We now need to prove two things:
390 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
391 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
392 *
393 * Facts:
394 * For x close to zero, exp(x) =~ 1 + x, since
395 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
396 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
397 * For x close to zero, ln(1+x) =~ x, since
398 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
399 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
400 * ln(.1) =~ -2.30
401 *
402 * Proof of (1):
403 * Solve (factor)**(power) =~ .1 given power (5*loadav):
404 * solving for factor,
405 * ln(factor) =~ (-2.30/5*loadav), or
406 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
407 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
408 *
409 * Proof of (2):
410 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
411 * solving for power,
412 * power*ln(b/(b+1)) =~ -2.30, or
413 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
414 *
415 * Actual power values for the implemented algorithm are as follows:
416 * loadav: 1 2 3 4
417 * power: 5.68 10.32 14.94 19.55
418 */
419
420/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
421#define loadfactor(loadav) (2 * (loadav))
422#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
423
424/* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
425static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
426SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
427
428/*
429 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
430 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
431 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
432 *
433 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
434 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
435 *
436 * If you don't want to bother with the faster/more-accurate formula, you
437 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
438 * (more general) method of calculating the %age of CPU used by a process.
439 */
440#define CCPU_SHIFT 11
441
442/*
443 * Recompute process priorities, every hz ticks.
444 * MP-safe, called without the Giant mutex.
445 */
446/* ARGSUSED */
447static void
448schedcpu(void)
449{
450 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
451 struct thread *td;
452 struct proc *p;
453 struct td_sched *ts;
454 int awake, realstathz;
455
456 realstathz = stathz ? stathz : hz;
457 sx_slock(&allproc_lock);
458 FOREACH_PROC_IN_SYSTEM(p) {
459 PROC_LOCK(p);
460 FOREACH_THREAD_IN_PROC(p, td) {
461 awake = 0;
462 thread_lock(td);
463 ts = td->td_sched;
464 /*
465 * Increment sleep time (if sleeping). We
466 * ignore overflow, as above.
467 */
468 /*
469 * The td_sched slptimes are not touched in wakeup
470 * because the thread may not HAVE everything in
471 * memory? XXX I think this is out of date.
472 */
473 if (TD_ON_RUNQ(td)) {
474 awake = 1;
475 td->td_flags &= ~TDF_DIDRUN;
476 } else if (TD_IS_RUNNING(td)) {
477 awake = 1;
478 /* Do not clear TDF_DIDRUN */
479 } else if (td->td_flags & TDF_DIDRUN) {
480 awake = 1;
481 td->td_flags &= ~TDF_DIDRUN;
482 }
483
484 /*
485 * ts_pctcpu is only for ps and ttyinfo().
486 */
487 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
488 /*
489 * If the td_sched has been idle the entire second,
490 * stop recalculating its priority until
491 * it wakes up.
492 */
493 if (ts->ts_cpticks != 0) {
494#if (FSHIFT >= CCPU_SHIFT)
495 ts->ts_pctcpu += (realstathz == 100)
496 ? ((fixpt_t) ts->ts_cpticks) <<
497 (FSHIFT - CCPU_SHIFT) :
498 100 * (((fixpt_t) ts->ts_cpticks)
499 << (FSHIFT - CCPU_SHIFT)) / realstathz;
500#else
501 ts->ts_pctcpu += ((FSCALE - ccpu) *
502 (ts->ts_cpticks *
503 FSCALE / realstathz)) >> FSHIFT;
504#endif
505 ts->ts_cpticks = 0;
506 }
507 /*
508 * If there are ANY running threads in this process,
509 * then don't count it as sleeping.
510 * XXX: this is broken.
511 */
512 if (awake) {
513 if (ts->ts_slptime > 1) {
514 /*
515 * In an ideal world, this should not
516 * happen, because whoever woke us
517 * up from the long sleep should have
518 * unwound the slptime and reset our
519 * priority before we run at the stale
520 * priority. Should KASSERT at some
521 * point when all the cases are fixed.
522 */
523 updatepri(td);
524 }
525 ts->ts_slptime = 0;
526 } else
527 ts->ts_slptime++;
528 if (ts->ts_slptime > 1) {
529 thread_unlock(td);
530 continue;
531 }
532 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
533 resetpriority(td);
534 resetpriority_thread(td);
535 thread_unlock(td);
536 }
537 PROC_UNLOCK(p);
538 }
539 sx_sunlock(&allproc_lock);
540}
541
542/*
543 * Main loop for a kthread that executes schedcpu once a second.
544 */
545static void
546schedcpu_thread(void)
547{
548
549 for (;;) {
550 schedcpu();
551 pause("-", hz);
552 }
553}
554
555/*
556 * Recalculate the priority of a process after it has slept for a while.
557 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
558 * least six times the loadfactor will decay td_estcpu to zero.
559 */
560static void
561updatepri(struct thread *td)
562{
563 struct td_sched *ts;
564 fixpt_t loadfac;
565 unsigned int newcpu;
566
567 ts = td->td_sched;
568 loadfac = loadfactor(averunnable.ldavg[0]);
569 if (ts->ts_slptime > 5 * loadfac)
570 td->td_estcpu = 0;
571 else {
572 newcpu = td->td_estcpu;
573 ts->ts_slptime--; /* was incremented in schedcpu() */
574 while (newcpu && --ts->ts_slptime)
575 newcpu = decay_cpu(loadfac, newcpu);
576 td->td_estcpu = newcpu;
577 }
578}
579
580/*
581 * Compute the priority of a process when running in user mode.
582 * Arrange to reschedule if the resulting priority is better
583 * than that of the current process.
584 */
585static void
586resetpriority(struct thread *td)
587{
588 register unsigned int newpriority;
589
590 if (td->td_pri_class == PRI_TIMESHARE) {
591 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
592 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
593 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
594 PRI_MAX_TIMESHARE);
595 sched_user_prio(td, newpriority);
596 }
597}
598
599/*
600 * Update the thread's priority when the associated process's user
601 * priority changes.
602 */
603static void
604resetpriority_thread(struct thread *td)
605{
606
607 /* Only change threads with a time sharing user priority. */
608 if (td->td_priority < PRI_MIN_TIMESHARE ||
609 td->td_priority > PRI_MAX_TIMESHARE)
610 return;
611
612 /* XXX the whole needresched thing is broken, but not silly. */
613 maybe_resched(td);
614
615 sched_prio(td, td->td_user_pri);
616}
617
618/* ARGSUSED */
619static void
620sched_setup(void *dummy)
621{
622 setup_runqs();
623
624 if (sched_quantum == 0)
625 sched_quantum = SCHED_QUANTUM;
626 hogticks = 2 * sched_quantum;
627
628 /* Account for thread0. */
629 sched_load_add();
630}
631
632/* External interfaces start here */
633
634/*
635 * Very early in the boot some setup of scheduler-specific
636 * parts of proc0 and of some scheduler resources needs to be done.
637 * Called from:
638 * proc0_init()
639 */
640void
641schedinit(void)
642{
643 /*
644 * Set up the scheduler specific parts of proc0.
645 */
646 proc0.p_sched = NULL; /* XXX */
647 thread0.td_sched = &td_sched0;
648 thread0.td_lock = &sched_lock;
649 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
650}
651
652int
653sched_runnable(void)
654{
655#ifdef SMP
656 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
657#else
658 return runq_check(&runq);
659#endif
660}
661
662int
663sched_rr_interval(void)
664{
665 if (sched_quantum == 0)
666 sched_quantum = SCHED_QUANTUM;
667 return (sched_quantum);
668}
669
670/*
671 * We adjust the priority of the current process. The priority of
672 * a process gets worse as it accumulates CPU time. The cpu usage
673 * estimator (td_estcpu) is increased here. resetpriority() will
674 * compute a different priority each time td_estcpu increases by
675 * INVERSE_ESTCPU_WEIGHT
676 * (until MAXPRI is reached). The cpu usage estimator ramps up
677 * quite quickly when the process is running (linearly), and decays
678 * away exponentially, at a rate which is proportionally slower when
679 * the system is busy. The basic principle is that the system will
680 * 90% forget that the process used a lot of CPU time in 5 * loadav
681 * seconds. This causes the system to favor processes which haven't
682 * run much recently, and to round-robin among other processes.
683 */
684void
685sched_clock(struct thread *td)
686{
687 struct td_sched *ts;
688
689 THREAD_LOCK_ASSERT(td, MA_OWNED);
690 ts = td->td_sched;
691
692 ts->ts_cpticks++;
693 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
694 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
695 resetpriority(td);
696 resetpriority_thread(td);
697 }
698
699 /*
700 * Force a context switch if the current thread has used up a full
701 * quantum (default quantum is 100ms).
702 */
703 if (!TD_IS_IDLETHREAD(td) &&
704 ticks - PCPU_GET(switchticks) >= sched_quantum)
705 td->td_flags |= TDF_NEEDRESCHED;
706}
707
708/*
709 * Charge child's scheduling CPU usage to parent.
710 */
711void
712sched_exit(struct proc *p, struct thread *td)
713{
714
715 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
716 "prio:td", td->td_priority);
717
718 PROC_LOCK_ASSERT(p, MA_OWNED);
719 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
720}
721
722void
723sched_exit_thread(struct thread *td, struct thread *child)
724{
725
726 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
727 "prio:td", child->td_priority);
728 thread_lock(td);
729 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
730 thread_unlock(td);
731 thread_lock(child);
732 if ((child->td_flags & TDF_NOLOAD) == 0)
733 sched_load_rem();
734 thread_unlock(child);
735}
736
737void
738sched_fork(struct thread *td, struct thread *childtd)
739{
740 sched_fork_thread(td, childtd);
741}
742
743void
744sched_fork_thread(struct thread *td, struct thread *childtd)
745{
746 struct td_sched *ts;
747
748 childtd->td_estcpu = td->td_estcpu;
749 childtd->td_lock = &sched_lock;
750 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
751 ts = childtd->td_sched;
752 bzero(ts, sizeof(*ts));
753 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
754}
755
756void
757sched_nice(struct proc *p, int nice)
758{
759 struct thread *td;
760
761 PROC_LOCK_ASSERT(p, MA_OWNED);
762 p->p_nice = nice;
763 FOREACH_THREAD_IN_PROC(p, td) {
764 thread_lock(td);
765 resetpriority(td);
766 resetpriority_thread(td);
767 thread_unlock(td);
768 }
769}
770
771void
772sched_class(struct thread *td, int class)
773{
774 THREAD_LOCK_ASSERT(td, MA_OWNED);
775 td->td_pri_class = class;
776}
777
778/*
779 * Adjust the priority of a thread.
780 */
781static void
782sched_priority(struct thread *td, u_char prio)
783{
784
785
786 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
787 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
788 sched_tdname(curthread));
789 if (td != curthread && prio > td->td_priority) {
790 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
791 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
792 prio, KTR_ATTR_LINKED, sched_tdname(td));
793 }
794 THREAD_LOCK_ASSERT(td, MA_OWNED);
795 if (td->td_priority == prio)
796 return;
797 td->td_priority = prio;
798 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
799 sched_rem(td);
800 sched_add(td, SRQ_BORING);
801 }
802}
803
804/*
805 * Update a thread's priority when it is lent another thread's
806 * priority.
807 */
808void
809sched_lend_prio(struct thread *td, u_char prio)
810{
811
812 td->td_flags |= TDF_BORROWING;
813 sched_priority(td, prio);
814}
815
816/*
817 * Restore a thread's priority when priority propagation is
818 * over. The prio argument is the minimum priority the thread
819 * needs to have to satisfy other possible priority lending
820 * requests. If the thread's regulary priority is less
821 * important than prio the thread will keep a priority boost
822 * of prio.
823 */
824void
825sched_unlend_prio(struct thread *td, u_char prio)
826{
827 u_char base_pri;
828
829 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
830 td->td_base_pri <= PRI_MAX_TIMESHARE)
831 base_pri = td->td_user_pri;
832 else
833 base_pri = td->td_base_pri;
834 if (prio >= base_pri) {
835 td->td_flags &= ~TDF_BORROWING;
836 sched_prio(td, base_pri);
837 } else
838 sched_lend_prio(td, prio);
839}
840
841void
842sched_prio(struct thread *td, u_char prio)
843{
844 u_char oldprio;
845
846 /* First, update the base priority. */
847 td->td_base_pri = prio;
848
849 /*
850 * If the thread is borrowing another thread's priority, don't ever
851 * lower the priority.
852 */
853 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
854 return;
855
856 /* Change the real priority. */
857 oldprio = td->td_priority;
858 sched_priority(td, prio);
859
860 /*
861 * If the thread is on a turnstile, then let the turnstile update
862 * its state.
863 */
864 if (TD_ON_LOCK(td) && oldprio != prio)
865 turnstile_adjust(td, oldprio);
866}
867
868void
869sched_user_prio(struct thread *td, u_char prio)
870{
871 u_char oldprio;
872
873 THREAD_LOCK_ASSERT(td, MA_OWNED);
874 td->td_base_user_pri = prio;
875 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
876 return;
877 oldprio = td->td_user_pri;
878 td->td_user_pri = prio;
879}
880
881void
882sched_lend_user_prio(struct thread *td, u_char prio)
883{
884 u_char oldprio;
885
886 THREAD_LOCK_ASSERT(td, MA_OWNED);
887 td->td_flags |= TDF_UBORROWING;
888 oldprio = td->td_user_pri;
889 td->td_user_pri = prio;
890}
891
892void
893sched_unlend_user_prio(struct thread *td, u_char prio)
894{
895 u_char base_pri;
896
897 THREAD_LOCK_ASSERT(td, MA_OWNED);
898 base_pri = td->td_base_user_pri;
899 if (prio >= base_pri) {
900 td->td_flags &= ~TDF_UBORROWING;
901 sched_user_prio(td, base_pri);
902 } else {
903 sched_lend_user_prio(td, prio);
904 }
905}
906
907void
908sched_sleep(struct thread *td, int pri)
909{
910
911 THREAD_LOCK_ASSERT(td, MA_OWNED);
912 td->td_slptick = ticks;
913 td->td_sched->ts_slptime = 0;
914 if (pri)
915 sched_prio(td, pri);
916 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
917 td->td_flags |= TDF_CANSWAP;
918}
919
920void
921sched_switch(struct thread *td, struct thread *newtd, int flags)
922{
923 struct mtx *tmtx;
924 struct td_sched *ts;
925 struct proc *p;
926
927 tmtx = NULL;
928 ts = td->td_sched;
929 p = td->td_proc;
930
931 THREAD_LOCK_ASSERT(td, MA_OWNED);
932
933 /*
934 * Switch to the sched lock to fix things up and pick
935 * a new thread.
936 * Block the td_lock in order to avoid breaking the critical path.
937 */
938 if (td->td_lock != &sched_lock) {
939 mtx_lock_spin(&sched_lock);
940 tmtx = thread_lock_block(td);
941 }
942
943 if ((td->td_flags & TDF_NOLOAD) == 0)
944 sched_load_rem();
945
946 if (newtd) {
947 MPASS(newtd->td_lock == &sched_lock);
948 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
949 }
950
951 td->td_lastcpu = td->td_oncpu;
952 td->td_flags &= ~TDF_NEEDRESCHED;
953 td->td_owepreempt = 0;
954 td->td_oncpu = NOCPU;
955
956 /*
957 * At the last moment, if this thread is still marked RUNNING,
958 * then put it back on the run queue as it has not been suspended
959 * or stopped or any thing else similar. We never put the idle
960 * threads on the run queue, however.
961 */
962 if (td->td_flags & TDF_IDLETD) {
963 TD_SET_CAN_RUN(td);
964#ifdef SMP
965 idle_cpus_mask &= ~PCPU_GET(cpumask);
966#endif
967 } else {
968 if (TD_IS_RUNNING(td)) {
969 /* Put us back on the run queue. */
970 sched_add(td, (flags & SW_PREEMPT) ?
971 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
972 SRQ_OURSELF|SRQ_YIELDING);
973 }
974 }
975 if (newtd) {
976 /*
977 * The thread we are about to run needs to be counted
978 * as if it had been added to the run queue and selected.
979 * It came from:
980 * * A preemption
981 * * An upcall
982 * * A followon
983 */
984 KASSERT((newtd->td_inhibitors == 0),
985 ("trying to run inhibited thread"));
986 newtd->td_flags |= TDF_DIDRUN;
987 TD_SET_RUNNING(newtd);
988 if ((newtd->td_flags & TDF_NOLOAD) == 0)
989 sched_load_add();
990 } else {
991 newtd = choosethread();
992 MPASS(newtd->td_lock == &sched_lock);
993 }
994
995 if (td != newtd) {
996#ifdef HWPMC_HOOKS
997 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
998 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
999#endif
1000 /* I feel sleepy */
1001 lock_profile_release_lock(&sched_lock.lock_object);
1002#ifdef KDTRACE_HOOKS
1003 /*
1004 * If DTrace has set the active vtime enum to anything
1005 * other than INACTIVE (0), then it should have set the
1006 * function to call.
1007 */
1008 if (dtrace_vtime_active)
1009 (*dtrace_vtime_switch_func)(newtd);
1010#endif
1011
1012 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1013 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1014 0, 0, __FILE__, __LINE__);
1015 /*
1016 * Where am I? What year is it?
1017 * We are in the same thread that went to sleep above,
1018 * but any amount of time may have passed. All our context
1019 * will still be available as will local variables.
1020 * PCPU values however may have changed as we may have
1021 * changed CPU so don't trust cached values of them.
1022 * New threads will go to fork_exit() instead of here
1023 * so if you change things here you may need to change
1024 * things there too.
1025 *
1026 * If the thread above was exiting it will never wake
1027 * up again here, so either it has saved everything it
1028 * needed to, or the thread_wait() or wait() will
1029 * need to reap it.
1030 */
1031#ifdef HWPMC_HOOKS
1032 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1033 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1034#endif
1035 }
1036
1037#ifdef SMP
1038 if (td->td_flags & TDF_IDLETD)
1039 idle_cpus_mask |= PCPU_GET(cpumask);
1040#endif
1041 sched_lock.mtx_lock = (uintptr_t)td;
1042 td->td_oncpu = PCPU_GET(cpuid);
1043 MPASS(td->td_lock == &sched_lock);
1044}
1045
1046void
1047sched_wakeup(struct thread *td)
1048{
1049 struct td_sched *ts;
1050
1051 THREAD_LOCK_ASSERT(td, MA_OWNED);
1052 ts = td->td_sched;
1053 td->td_flags &= ~TDF_CANSWAP;
1054 if (ts->ts_slptime > 1) {
1055 updatepri(td);
1056 resetpriority(td);
1057 }
1058 td->td_slptick = 0;
1059 ts->ts_slptime = 0;
1060 sched_add(td, SRQ_BORING);
1061}
1062
1063#ifdef SMP
1064static int
1065forward_wakeup(int cpunum)
1066{
1067 struct pcpu *pc;
1068 cpumask_t dontuse, id, map, map2, map3, me;
1069
1070 mtx_assert(&sched_lock, MA_OWNED);
1071
1072 CTR0(KTR_RUNQ, "forward_wakeup()");
1073
1074 if ((!forward_wakeup_enabled) ||
1075 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1076 return (0);
1077 if (!smp_started || cold || panicstr)
1078 return (0);
1079
1080 forward_wakeups_requested++;
1081
1082 /*
1083 * Check the idle mask we received against what we calculated
1084 * before in the old version.
1085 */
1086 me = PCPU_GET(cpumask);
1087
1088 /* Don't bother if we should be doing it ourself. */
1089 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
1090 return (0);
1091
1092 dontuse = me | stopped_cpus | hlt_cpus_mask;
1093 map3 = 0;
1094 if (forward_wakeup_use_loop) {
1095 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1096 id = pc->pc_cpumask;
1097 if ((id & dontuse) == 0 &&
1098 pc->pc_curthread == pc->pc_idlethread) {
1099 map3 |= id;
1100 }
1101 }
1102 }
1103
1104 if (forward_wakeup_use_mask) {
1105 map = 0;
1106 map = idle_cpus_mask & ~dontuse;
1107
1108 /* If they are both on, compare and use loop if different. */
1109 if (forward_wakeup_use_loop) {
1110 if (map != map3) {
1111 printf("map (%02X) != map3 (%02X)\n", map,
1112 map3);
1113 map = map3;
1114 }
1115 }
1116 } else {
1117 map = map3;
1118 }
1119
1120 /* If we only allow a specific CPU, then mask off all the others. */
1121 if (cpunum != NOCPU) {
1122 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1123 map &= (1 << cpunum);
1124 } else {
1125 /* Try choose an idle die. */
1126 if (forward_wakeup_use_htt) {
1127 map2 = (map & (map >> 1)) & 0x5555;
1128 if (map2) {
1129 map = map2;
1130 }
1131 }
1132
1133 /* Set only one bit. */
1134 if (forward_wakeup_use_single) {
1135 map = map & ((~map) + 1);
1136 }
1137 }
1138 if (map) {
1139 forward_wakeups_delivered++;
1140 ipi_selected(map, IPI_AST);
1141 return (1);
1142 }
1143 if (cpunum == NOCPU)
1144 printf("forward_wakeup: Idle processor not found\n");
1145 return (0);
1146}
1147
1148static void
1149kick_other_cpu(int pri, int cpuid)
1150{
1151 struct pcpu *pcpu;
1152 int cpri;
1153
1154 pcpu = pcpu_find(cpuid);
1155 if (idle_cpus_mask & pcpu->pc_cpumask) {
1156 forward_wakeups_delivered++;
1157 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1158 return;
1159 }
1160
1161 cpri = pcpu->pc_curthread->td_priority;
1162 if (pri >= cpri)
1163 return;
1164
1165#if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1166#if !defined(FULL_PREEMPTION)
1167 if (pri <= PRI_MAX_ITHD)
1168#endif /* ! FULL_PREEMPTION */
1169 {
1170 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1171 return;
1172 }
1173#endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1174
1175 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1176 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1177 return;
1178}
1179#endif /* SMP */
1180
1181#ifdef SMP
1182static int
1183sched_pickcpu(struct thread *td)
1184{
1185 int best, cpu;
1186
1187 mtx_assert(&sched_lock, MA_OWNED);
1188
1189 if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1190 best = td->td_lastcpu;
1191 else
1192 best = NOCPU;
| 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/smp.h>
54#include <sys/sysctl.h>
55#include <sys/sx.h>
56#include <sys/turnstile.h>
57#include <sys/umtx.h>
58#include <machine/pcb.h>
59#include <machine/smp.h>
60
61#ifdef HWPMC_HOOKS
62#include <sys/pmckern.h>
63#endif
64
65#ifdef KDTRACE_HOOKS
66#include <sys/dtrace_bsd.h>
67int dtrace_vtime_active;
68dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
69#endif
70
71/*
72 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
73 * the range 100-256 Hz (approximately).
74 */
75#define ESTCPULIM(e) \
76 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
77 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
78#ifdef SMP
79#define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
80#else
81#define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
82#endif
83#define NICE_WEIGHT 1 /* Priorities per nice level. */
84
85#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
86
87/*
88 * The schedulable entity that runs a context.
89 * This is an extension to the thread structure and is tailored to
90 * the requirements of this scheduler
91 */
92struct td_sched {
93 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
94 int ts_cpticks; /* (j) Ticks of cpu time. */
95 int ts_slptime; /* (j) Seconds !RUNNING. */
96 int ts_flags;
97 struct runq *ts_runq; /* runq the thread is currently on */
98#ifdef KTR
99 char ts_name[TS_NAME_LEN];
100#endif
101};
102
103/* flags kept in td_flags */
104#define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
105#define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
106
107/* flags kept in ts_flags */
108#define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
109
110#define SKE_RUNQ_PCPU(ts) \
111 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
112
113#define THREAD_CAN_SCHED(td, cpu) \
114 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
115
116static struct td_sched td_sched0;
117struct mtx sched_lock;
118
119static int sched_tdcnt; /* Total runnable threads in the system. */
120static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
121#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
122
123static void setup_runqs(void);
124static void schedcpu(void);
125static void schedcpu_thread(void);
126static void sched_priority(struct thread *td, u_char prio);
127static void sched_setup(void *dummy);
128static void maybe_resched(struct thread *td);
129static void updatepri(struct thread *td);
130static void resetpriority(struct thread *td);
131static void resetpriority_thread(struct thread *td);
132#ifdef SMP
133static int sched_pickcpu(struct thread *td);
134static int forward_wakeup(int cpunum);
135static void kick_other_cpu(int pri, int cpuid);
136#endif
137
138static struct kproc_desc sched_kp = {
139 "schedcpu",
140 schedcpu_thread,
141 NULL
142};
143SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
144 &sched_kp);
145SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
146
147/*
148 * Global run queue.
149 */
150static struct runq runq;
151
152#ifdef SMP
153/*
154 * Per-CPU run queues
155 */
156static struct runq runq_pcpu[MAXCPU];
157long runq_length[MAXCPU];
158#endif
159
160static void
161setup_runqs(void)
162{
163#ifdef SMP
164 int i;
165
166 for (i = 0; i < MAXCPU; ++i)
167 runq_init(&runq_pcpu[i]);
168#endif
169
170 runq_init(&runq);
171}
172
173static int
174sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
175{
176 int error, new_val;
177
178 new_val = sched_quantum * tick;
179 error = sysctl_handle_int(oidp, &new_val, 0, req);
180 if (error != 0 || req->newptr == NULL)
181 return (error);
182 if (new_val < tick)
183 return (EINVAL);
184 sched_quantum = new_val / tick;
185 hogticks = 2 * sched_quantum;
186 return (0);
187}
188
189SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
190
191SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
192 "Scheduler name");
193
194SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
195 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
196 "Roundrobin scheduling quantum in microseconds");
197
198#ifdef SMP
199/* Enable forwarding of wakeups to all other cpus */
200SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
201
202static int runq_fuzz = 1;
203SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
204
205static int forward_wakeup_enabled = 1;
206SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
207 &forward_wakeup_enabled, 0,
208 "Forwarding of wakeup to idle CPUs");
209
210static int forward_wakeups_requested = 0;
211SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
212 &forward_wakeups_requested, 0,
213 "Requests for Forwarding of wakeup to idle CPUs");
214
215static int forward_wakeups_delivered = 0;
216SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
217 &forward_wakeups_delivered, 0,
218 "Completed Forwarding of wakeup to idle CPUs");
219
220static int forward_wakeup_use_mask = 1;
221SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
222 &forward_wakeup_use_mask, 0,
223 "Use the mask of idle cpus");
224
225static int forward_wakeup_use_loop = 0;
226SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
227 &forward_wakeup_use_loop, 0,
228 "Use a loop to find idle cpus");
229
230static int forward_wakeup_use_single = 0;
231SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
232 &forward_wakeup_use_single, 0,
233 "Only signal one idle cpu");
234
235static int forward_wakeup_use_htt = 0;
236SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
237 &forward_wakeup_use_htt, 0,
238 "account for htt");
239
240#endif
241#if 0
242static int sched_followon = 0;
243SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
244 &sched_followon, 0,
245 "allow threads to share a quantum");
246#endif
247
248static __inline void
249sched_load_add(void)
250{
251
252 sched_tdcnt++;
253 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
254}
255
256static __inline void
257sched_load_rem(void)
258{
259
260 sched_tdcnt--;
261 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
262}
263/*
264 * Arrange to reschedule if necessary, taking the priorities and
265 * schedulers into account.
266 */
267static void
268maybe_resched(struct thread *td)
269{
270
271 THREAD_LOCK_ASSERT(td, MA_OWNED);
272 if (td->td_priority < curthread->td_priority)
273 curthread->td_flags |= TDF_NEEDRESCHED;
274}
275
276/*
277 * This function is called when a thread is about to be put on run queue
278 * because it has been made runnable or its priority has been adjusted. It
279 * determines if the new thread should be immediately preempted to. If so,
280 * it switches to it and eventually returns true. If not, it returns false
281 * so that the caller may place the thread on an appropriate run queue.
282 */
283int
284maybe_preempt(struct thread *td)
285{
286#ifdef PREEMPTION
287 struct thread *ctd;
288 int cpri, pri;
289
290 /*
291 * The new thread should not preempt the current thread if any of the
292 * following conditions are true:
293 *
294 * - The kernel is in the throes of crashing (panicstr).
295 * - The current thread has a higher (numerically lower) or
296 * equivalent priority. Note that this prevents curthread from
297 * trying to preempt to itself.
298 * - It is too early in the boot for context switches (cold is set).
299 * - The current thread has an inhibitor set or is in the process of
300 * exiting. In this case, the current thread is about to switch
301 * out anyways, so there's no point in preempting. If we did,
302 * the current thread would not be properly resumed as well, so
303 * just avoid that whole landmine.
304 * - If the new thread's priority is not a realtime priority and
305 * the current thread's priority is not an idle priority and
306 * FULL_PREEMPTION is disabled.
307 *
308 * If all of these conditions are false, but the current thread is in
309 * a nested critical section, then we have to defer the preemption
310 * until we exit the critical section. Otherwise, switch immediately
311 * to the new thread.
312 */
313 ctd = curthread;
314 THREAD_LOCK_ASSERT(td, MA_OWNED);
315 KASSERT((td->td_inhibitors == 0),
316 ("maybe_preempt: trying to run inhibited thread"));
317 pri = td->td_priority;
318 cpri = ctd->td_priority;
319 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
320 TD_IS_INHIBITED(ctd))
321 return (0);
322#ifndef FULL_PREEMPTION
323 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
324 return (0);
325#endif
326
327 if (ctd->td_critnest > 1) {
328 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
329 ctd->td_critnest);
330 ctd->td_owepreempt = 1;
331 return (0);
332 }
333 /*
334 * Thread is runnable but not yet put on system run queue.
335 */
336 MPASS(ctd->td_lock == td->td_lock);
337 MPASS(TD_ON_RUNQ(td));
338 TD_SET_RUNNING(td);
339 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
340 td->td_proc->p_pid, td->td_name);
341 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
342 /*
343 * td's lock pointer may have changed. We have to return with it
344 * locked.
345 */
346 spinlock_enter();
347 thread_unlock(ctd);
348 thread_lock(td);
349 spinlock_exit();
350 return (1);
351#else
352 return (0);
353#endif
354}
355
356/*
357 * Constants for digital decay and forget:
358 * 90% of (td_estcpu) usage in 5 * loadav time
359 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
360 * Note that, as ps(1) mentions, this can let percentages
361 * total over 100% (I've seen 137.9% for 3 processes).
362 *
363 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
364 *
365 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
366 * That is, the system wants to compute a value of decay such
367 * that the following for loop:
368 * for (i = 0; i < (5 * loadavg); i++)
369 * td_estcpu *= decay;
370 * will compute
371 * td_estcpu *= 0.1;
372 * for all values of loadavg:
373 *
374 * Mathematically this loop can be expressed by saying:
375 * decay ** (5 * loadavg) ~= .1
376 *
377 * The system computes decay as:
378 * decay = (2 * loadavg) / (2 * loadavg + 1)
379 *
380 * We wish to prove that the system's computation of decay
381 * will always fulfill the equation:
382 * decay ** (5 * loadavg) ~= .1
383 *
384 * If we compute b as:
385 * b = 2 * loadavg
386 * then
387 * decay = b / (b + 1)
388 *
389 * We now need to prove two things:
390 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
391 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
392 *
393 * Facts:
394 * For x close to zero, exp(x) =~ 1 + x, since
395 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
396 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
397 * For x close to zero, ln(1+x) =~ x, since
398 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
399 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
400 * ln(.1) =~ -2.30
401 *
402 * Proof of (1):
403 * Solve (factor)**(power) =~ .1 given power (5*loadav):
404 * solving for factor,
405 * ln(factor) =~ (-2.30/5*loadav), or
406 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
407 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
408 *
409 * Proof of (2):
410 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
411 * solving for power,
412 * power*ln(b/(b+1)) =~ -2.30, or
413 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
414 *
415 * Actual power values for the implemented algorithm are as follows:
416 * loadav: 1 2 3 4
417 * power: 5.68 10.32 14.94 19.55
418 */
419
420/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
421#define loadfactor(loadav) (2 * (loadav))
422#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
423
424/* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
425static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
426SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
427
428/*
429 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
430 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
431 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
432 *
433 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
434 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
435 *
436 * If you don't want to bother with the faster/more-accurate formula, you
437 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
438 * (more general) method of calculating the %age of CPU used by a process.
439 */
440#define CCPU_SHIFT 11
441
442/*
443 * Recompute process priorities, every hz ticks.
444 * MP-safe, called without the Giant mutex.
445 */
446/* ARGSUSED */
447static void
448schedcpu(void)
449{
450 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
451 struct thread *td;
452 struct proc *p;
453 struct td_sched *ts;
454 int awake, realstathz;
455
456 realstathz = stathz ? stathz : hz;
457 sx_slock(&allproc_lock);
458 FOREACH_PROC_IN_SYSTEM(p) {
459 PROC_LOCK(p);
460 FOREACH_THREAD_IN_PROC(p, td) {
461 awake = 0;
462 thread_lock(td);
463 ts = td->td_sched;
464 /*
465 * Increment sleep time (if sleeping). We
466 * ignore overflow, as above.
467 */
468 /*
469 * The td_sched slptimes are not touched in wakeup
470 * because the thread may not HAVE everything in
471 * memory? XXX I think this is out of date.
472 */
473 if (TD_ON_RUNQ(td)) {
474 awake = 1;
475 td->td_flags &= ~TDF_DIDRUN;
476 } else if (TD_IS_RUNNING(td)) {
477 awake = 1;
478 /* Do not clear TDF_DIDRUN */
479 } else if (td->td_flags & TDF_DIDRUN) {
480 awake = 1;
481 td->td_flags &= ~TDF_DIDRUN;
482 }
483
484 /*
485 * ts_pctcpu is only for ps and ttyinfo().
486 */
487 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
488 /*
489 * If the td_sched has been idle the entire second,
490 * stop recalculating its priority until
491 * it wakes up.
492 */
493 if (ts->ts_cpticks != 0) {
494#if (FSHIFT >= CCPU_SHIFT)
495 ts->ts_pctcpu += (realstathz == 100)
496 ? ((fixpt_t) ts->ts_cpticks) <<
497 (FSHIFT - CCPU_SHIFT) :
498 100 * (((fixpt_t) ts->ts_cpticks)
499 << (FSHIFT - CCPU_SHIFT)) / realstathz;
500#else
501 ts->ts_pctcpu += ((FSCALE - ccpu) *
502 (ts->ts_cpticks *
503 FSCALE / realstathz)) >> FSHIFT;
504#endif
505 ts->ts_cpticks = 0;
506 }
507 /*
508 * If there are ANY running threads in this process,
509 * then don't count it as sleeping.
510 * XXX: this is broken.
511 */
512 if (awake) {
513 if (ts->ts_slptime > 1) {
514 /*
515 * In an ideal world, this should not
516 * happen, because whoever woke us
517 * up from the long sleep should have
518 * unwound the slptime and reset our
519 * priority before we run at the stale
520 * priority. Should KASSERT at some
521 * point when all the cases are fixed.
522 */
523 updatepri(td);
524 }
525 ts->ts_slptime = 0;
526 } else
527 ts->ts_slptime++;
528 if (ts->ts_slptime > 1) {
529 thread_unlock(td);
530 continue;
531 }
532 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
533 resetpriority(td);
534 resetpriority_thread(td);
535 thread_unlock(td);
536 }
537 PROC_UNLOCK(p);
538 }
539 sx_sunlock(&allproc_lock);
540}
541
542/*
543 * Main loop for a kthread that executes schedcpu once a second.
544 */
545static void
546schedcpu_thread(void)
547{
548
549 for (;;) {
550 schedcpu();
551 pause("-", hz);
552 }
553}
554
555/*
556 * Recalculate the priority of a process after it has slept for a while.
557 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
558 * least six times the loadfactor will decay td_estcpu to zero.
559 */
560static void
561updatepri(struct thread *td)
562{
563 struct td_sched *ts;
564 fixpt_t loadfac;
565 unsigned int newcpu;
566
567 ts = td->td_sched;
568 loadfac = loadfactor(averunnable.ldavg[0]);
569 if (ts->ts_slptime > 5 * loadfac)
570 td->td_estcpu = 0;
571 else {
572 newcpu = td->td_estcpu;
573 ts->ts_slptime--; /* was incremented in schedcpu() */
574 while (newcpu && --ts->ts_slptime)
575 newcpu = decay_cpu(loadfac, newcpu);
576 td->td_estcpu = newcpu;
577 }
578}
579
580/*
581 * Compute the priority of a process when running in user mode.
582 * Arrange to reschedule if the resulting priority is better
583 * than that of the current process.
584 */
585static void
586resetpriority(struct thread *td)
587{
588 register unsigned int newpriority;
589
590 if (td->td_pri_class == PRI_TIMESHARE) {
591 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
592 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
593 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
594 PRI_MAX_TIMESHARE);
595 sched_user_prio(td, newpriority);
596 }
597}
598
599/*
600 * Update the thread's priority when the associated process's user
601 * priority changes.
602 */
603static void
604resetpriority_thread(struct thread *td)
605{
606
607 /* Only change threads with a time sharing user priority. */
608 if (td->td_priority < PRI_MIN_TIMESHARE ||
609 td->td_priority > PRI_MAX_TIMESHARE)
610 return;
611
612 /* XXX the whole needresched thing is broken, but not silly. */
613 maybe_resched(td);
614
615 sched_prio(td, td->td_user_pri);
616}
617
618/* ARGSUSED */
619static void
620sched_setup(void *dummy)
621{
622 setup_runqs();
623
624 if (sched_quantum == 0)
625 sched_quantum = SCHED_QUANTUM;
626 hogticks = 2 * sched_quantum;
627
628 /* Account for thread0. */
629 sched_load_add();
630}
631
632/* External interfaces start here */
633
634/*
635 * Very early in the boot some setup of scheduler-specific
636 * parts of proc0 and of some scheduler resources needs to be done.
637 * Called from:
638 * proc0_init()
639 */
640void
641schedinit(void)
642{
643 /*
644 * Set up the scheduler specific parts of proc0.
645 */
646 proc0.p_sched = NULL; /* XXX */
647 thread0.td_sched = &td_sched0;
648 thread0.td_lock = &sched_lock;
649 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
650}
651
652int
653sched_runnable(void)
654{
655#ifdef SMP
656 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
657#else
658 return runq_check(&runq);
659#endif
660}
661
662int
663sched_rr_interval(void)
664{
665 if (sched_quantum == 0)
666 sched_quantum = SCHED_QUANTUM;
667 return (sched_quantum);
668}
669
670/*
671 * We adjust the priority of the current process. The priority of
672 * a process gets worse as it accumulates CPU time. The cpu usage
673 * estimator (td_estcpu) is increased here. resetpriority() will
674 * compute a different priority each time td_estcpu increases by
675 * INVERSE_ESTCPU_WEIGHT
676 * (until MAXPRI is reached). The cpu usage estimator ramps up
677 * quite quickly when the process is running (linearly), and decays
678 * away exponentially, at a rate which is proportionally slower when
679 * the system is busy. The basic principle is that the system will
680 * 90% forget that the process used a lot of CPU time in 5 * loadav
681 * seconds. This causes the system to favor processes which haven't
682 * run much recently, and to round-robin among other processes.
683 */
684void
685sched_clock(struct thread *td)
686{
687 struct td_sched *ts;
688
689 THREAD_LOCK_ASSERT(td, MA_OWNED);
690 ts = td->td_sched;
691
692 ts->ts_cpticks++;
693 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
694 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
695 resetpriority(td);
696 resetpriority_thread(td);
697 }
698
699 /*
700 * Force a context switch if the current thread has used up a full
701 * quantum (default quantum is 100ms).
702 */
703 if (!TD_IS_IDLETHREAD(td) &&
704 ticks - PCPU_GET(switchticks) >= sched_quantum)
705 td->td_flags |= TDF_NEEDRESCHED;
706}
707
708/*
709 * Charge child's scheduling CPU usage to parent.
710 */
711void
712sched_exit(struct proc *p, struct thread *td)
713{
714
715 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
716 "prio:td", td->td_priority);
717
718 PROC_LOCK_ASSERT(p, MA_OWNED);
719 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
720}
721
722void
723sched_exit_thread(struct thread *td, struct thread *child)
724{
725
726 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
727 "prio:td", child->td_priority);
728 thread_lock(td);
729 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
730 thread_unlock(td);
731 thread_lock(child);
732 if ((child->td_flags & TDF_NOLOAD) == 0)
733 sched_load_rem();
734 thread_unlock(child);
735}
736
737void
738sched_fork(struct thread *td, struct thread *childtd)
739{
740 sched_fork_thread(td, childtd);
741}
742
743void
744sched_fork_thread(struct thread *td, struct thread *childtd)
745{
746 struct td_sched *ts;
747
748 childtd->td_estcpu = td->td_estcpu;
749 childtd->td_lock = &sched_lock;
750 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
751 ts = childtd->td_sched;
752 bzero(ts, sizeof(*ts));
753 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
754}
755
756void
757sched_nice(struct proc *p, int nice)
758{
759 struct thread *td;
760
761 PROC_LOCK_ASSERT(p, MA_OWNED);
762 p->p_nice = nice;
763 FOREACH_THREAD_IN_PROC(p, td) {
764 thread_lock(td);
765 resetpriority(td);
766 resetpriority_thread(td);
767 thread_unlock(td);
768 }
769}
770
771void
772sched_class(struct thread *td, int class)
773{
774 THREAD_LOCK_ASSERT(td, MA_OWNED);
775 td->td_pri_class = class;
776}
777
778/*
779 * Adjust the priority of a thread.
780 */
781static void
782sched_priority(struct thread *td, u_char prio)
783{
784
785
786 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
787 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
788 sched_tdname(curthread));
789 if (td != curthread && prio > td->td_priority) {
790 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
791 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
792 prio, KTR_ATTR_LINKED, sched_tdname(td));
793 }
794 THREAD_LOCK_ASSERT(td, MA_OWNED);
795 if (td->td_priority == prio)
796 return;
797 td->td_priority = prio;
798 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
799 sched_rem(td);
800 sched_add(td, SRQ_BORING);
801 }
802}
803
804/*
805 * Update a thread's priority when it is lent another thread's
806 * priority.
807 */
808void
809sched_lend_prio(struct thread *td, u_char prio)
810{
811
812 td->td_flags |= TDF_BORROWING;
813 sched_priority(td, prio);
814}
815
816/*
817 * Restore a thread's priority when priority propagation is
818 * over. The prio argument is the minimum priority the thread
819 * needs to have to satisfy other possible priority lending
820 * requests. If the thread's regulary priority is less
821 * important than prio the thread will keep a priority boost
822 * of prio.
823 */
824void
825sched_unlend_prio(struct thread *td, u_char prio)
826{
827 u_char base_pri;
828
829 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
830 td->td_base_pri <= PRI_MAX_TIMESHARE)
831 base_pri = td->td_user_pri;
832 else
833 base_pri = td->td_base_pri;
834 if (prio >= base_pri) {
835 td->td_flags &= ~TDF_BORROWING;
836 sched_prio(td, base_pri);
837 } else
838 sched_lend_prio(td, prio);
839}
840
841void
842sched_prio(struct thread *td, u_char prio)
843{
844 u_char oldprio;
845
846 /* First, update the base priority. */
847 td->td_base_pri = prio;
848
849 /*
850 * If the thread is borrowing another thread's priority, don't ever
851 * lower the priority.
852 */
853 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
854 return;
855
856 /* Change the real priority. */
857 oldprio = td->td_priority;
858 sched_priority(td, prio);
859
860 /*
861 * If the thread is on a turnstile, then let the turnstile update
862 * its state.
863 */
864 if (TD_ON_LOCK(td) && oldprio != prio)
865 turnstile_adjust(td, oldprio);
866}
867
868void
869sched_user_prio(struct thread *td, u_char prio)
870{
871 u_char oldprio;
872
873 THREAD_LOCK_ASSERT(td, MA_OWNED);
874 td->td_base_user_pri = prio;
875 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
876 return;
877 oldprio = td->td_user_pri;
878 td->td_user_pri = prio;
879}
880
881void
882sched_lend_user_prio(struct thread *td, u_char prio)
883{
884 u_char oldprio;
885
886 THREAD_LOCK_ASSERT(td, MA_OWNED);
887 td->td_flags |= TDF_UBORROWING;
888 oldprio = td->td_user_pri;
889 td->td_user_pri = prio;
890}
891
892void
893sched_unlend_user_prio(struct thread *td, u_char prio)
894{
895 u_char base_pri;
896
897 THREAD_LOCK_ASSERT(td, MA_OWNED);
898 base_pri = td->td_base_user_pri;
899 if (prio >= base_pri) {
900 td->td_flags &= ~TDF_UBORROWING;
901 sched_user_prio(td, base_pri);
902 } else {
903 sched_lend_user_prio(td, prio);
904 }
905}
906
907void
908sched_sleep(struct thread *td, int pri)
909{
910
911 THREAD_LOCK_ASSERT(td, MA_OWNED);
912 td->td_slptick = ticks;
913 td->td_sched->ts_slptime = 0;
914 if (pri)
915 sched_prio(td, pri);
916 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
917 td->td_flags |= TDF_CANSWAP;
918}
919
920void
921sched_switch(struct thread *td, struct thread *newtd, int flags)
922{
923 struct mtx *tmtx;
924 struct td_sched *ts;
925 struct proc *p;
926
927 tmtx = NULL;
928 ts = td->td_sched;
929 p = td->td_proc;
930
931 THREAD_LOCK_ASSERT(td, MA_OWNED);
932
933 /*
934 * Switch to the sched lock to fix things up and pick
935 * a new thread.
936 * Block the td_lock in order to avoid breaking the critical path.
937 */
938 if (td->td_lock != &sched_lock) {
939 mtx_lock_spin(&sched_lock);
940 tmtx = thread_lock_block(td);
941 }
942
943 if ((td->td_flags & TDF_NOLOAD) == 0)
944 sched_load_rem();
945
946 if (newtd) {
947 MPASS(newtd->td_lock == &sched_lock);
948 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
949 }
950
951 td->td_lastcpu = td->td_oncpu;
952 td->td_flags &= ~TDF_NEEDRESCHED;
953 td->td_owepreempt = 0;
954 td->td_oncpu = NOCPU;
955
956 /*
957 * At the last moment, if this thread is still marked RUNNING,
958 * then put it back on the run queue as it has not been suspended
959 * or stopped or any thing else similar. We never put the idle
960 * threads on the run queue, however.
961 */
962 if (td->td_flags & TDF_IDLETD) {
963 TD_SET_CAN_RUN(td);
964#ifdef SMP
965 idle_cpus_mask &= ~PCPU_GET(cpumask);
966#endif
967 } else {
968 if (TD_IS_RUNNING(td)) {
969 /* Put us back on the run queue. */
970 sched_add(td, (flags & SW_PREEMPT) ?
971 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
972 SRQ_OURSELF|SRQ_YIELDING);
973 }
974 }
975 if (newtd) {
976 /*
977 * The thread we are about to run needs to be counted
978 * as if it had been added to the run queue and selected.
979 * It came from:
980 * * A preemption
981 * * An upcall
982 * * A followon
983 */
984 KASSERT((newtd->td_inhibitors == 0),
985 ("trying to run inhibited thread"));
986 newtd->td_flags |= TDF_DIDRUN;
987 TD_SET_RUNNING(newtd);
988 if ((newtd->td_flags & TDF_NOLOAD) == 0)
989 sched_load_add();
990 } else {
991 newtd = choosethread();
992 MPASS(newtd->td_lock == &sched_lock);
993 }
994
995 if (td != newtd) {
996#ifdef HWPMC_HOOKS
997 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
998 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
999#endif
1000 /* I feel sleepy */
1001 lock_profile_release_lock(&sched_lock.lock_object);
1002#ifdef KDTRACE_HOOKS
1003 /*
1004 * If DTrace has set the active vtime enum to anything
1005 * other than INACTIVE (0), then it should have set the
1006 * function to call.
1007 */
1008 if (dtrace_vtime_active)
1009 (*dtrace_vtime_switch_func)(newtd);
1010#endif
1011
1012 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1013 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1014 0, 0, __FILE__, __LINE__);
1015 /*
1016 * Where am I? What year is it?
1017 * We are in the same thread that went to sleep above,
1018 * but any amount of time may have passed. All our context
1019 * will still be available as will local variables.
1020 * PCPU values however may have changed as we may have
1021 * changed CPU so don't trust cached values of them.
1022 * New threads will go to fork_exit() instead of here
1023 * so if you change things here you may need to change
1024 * things there too.
1025 *
1026 * If the thread above was exiting it will never wake
1027 * up again here, so either it has saved everything it
1028 * needed to, or the thread_wait() or wait() will
1029 * need to reap it.
1030 */
1031#ifdef HWPMC_HOOKS
1032 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1033 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1034#endif
1035 }
1036
1037#ifdef SMP
1038 if (td->td_flags & TDF_IDLETD)
1039 idle_cpus_mask |= PCPU_GET(cpumask);
1040#endif
1041 sched_lock.mtx_lock = (uintptr_t)td;
1042 td->td_oncpu = PCPU_GET(cpuid);
1043 MPASS(td->td_lock == &sched_lock);
1044}
1045
1046void
1047sched_wakeup(struct thread *td)
1048{
1049 struct td_sched *ts;
1050
1051 THREAD_LOCK_ASSERT(td, MA_OWNED);
1052 ts = td->td_sched;
1053 td->td_flags &= ~TDF_CANSWAP;
1054 if (ts->ts_slptime > 1) {
1055 updatepri(td);
1056 resetpriority(td);
1057 }
1058 td->td_slptick = 0;
1059 ts->ts_slptime = 0;
1060 sched_add(td, SRQ_BORING);
1061}
1062
1063#ifdef SMP
1064static int
1065forward_wakeup(int cpunum)
1066{
1067 struct pcpu *pc;
1068 cpumask_t dontuse, id, map, map2, map3, me;
1069
1070 mtx_assert(&sched_lock, MA_OWNED);
1071
1072 CTR0(KTR_RUNQ, "forward_wakeup()");
1073
1074 if ((!forward_wakeup_enabled) ||
1075 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1076 return (0);
1077 if (!smp_started || cold || panicstr)
1078 return (0);
1079
1080 forward_wakeups_requested++;
1081
1082 /*
1083 * Check the idle mask we received against what we calculated
1084 * before in the old version.
1085 */
1086 me = PCPU_GET(cpumask);
1087
1088 /* Don't bother if we should be doing it ourself. */
1089 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
1090 return (0);
1091
1092 dontuse = me | stopped_cpus | hlt_cpus_mask;
1093 map3 = 0;
1094 if (forward_wakeup_use_loop) {
1095 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1096 id = pc->pc_cpumask;
1097 if ((id & dontuse) == 0 &&
1098 pc->pc_curthread == pc->pc_idlethread) {
1099 map3 |= id;
1100 }
1101 }
1102 }
1103
1104 if (forward_wakeup_use_mask) {
1105 map = 0;
1106 map = idle_cpus_mask & ~dontuse;
1107
1108 /* If they are both on, compare and use loop if different. */
1109 if (forward_wakeup_use_loop) {
1110 if (map != map3) {
1111 printf("map (%02X) != map3 (%02X)\n", map,
1112 map3);
1113 map = map3;
1114 }
1115 }
1116 } else {
1117 map = map3;
1118 }
1119
1120 /* If we only allow a specific CPU, then mask off all the others. */
1121 if (cpunum != NOCPU) {
1122 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1123 map &= (1 << cpunum);
1124 } else {
1125 /* Try choose an idle die. */
1126 if (forward_wakeup_use_htt) {
1127 map2 = (map & (map >> 1)) & 0x5555;
1128 if (map2) {
1129 map = map2;
1130 }
1131 }
1132
1133 /* Set only one bit. */
1134 if (forward_wakeup_use_single) {
1135 map = map & ((~map) + 1);
1136 }
1137 }
1138 if (map) {
1139 forward_wakeups_delivered++;
1140 ipi_selected(map, IPI_AST);
1141 return (1);
1142 }
1143 if (cpunum == NOCPU)
1144 printf("forward_wakeup: Idle processor not found\n");
1145 return (0);
1146}
1147
1148static void
1149kick_other_cpu(int pri, int cpuid)
1150{
1151 struct pcpu *pcpu;
1152 int cpri;
1153
1154 pcpu = pcpu_find(cpuid);
1155 if (idle_cpus_mask & pcpu->pc_cpumask) {
1156 forward_wakeups_delivered++;
1157 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1158 return;
1159 }
1160
1161 cpri = pcpu->pc_curthread->td_priority;
1162 if (pri >= cpri)
1163 return;
1164
1165#if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1166#if !defined(FULL_PREEMPTION)
1167 if (pri <= PRI_MAX_ITHD)
1168#endif /* ! FULL_PREEMPTION */
1169 {
1170 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT);
1171 return;
1172 }
1173#endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1174
1175 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1176 ipi_selected(pcpu->pc_cpumask, IPI_AST);
1177 return;
1178}
1179#endif /* SMP */
1180
1181#ifdef SMP
1182static int
1183sched_pickcpu(struct thread *td)
1184{
1185 int best, cpu;
1186
1187 mtx_assert(&sched_lock, MA_OWNED);
1188
1189 if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1190 best = td->td_lastcpu;
1191 else
1192 best = NOCPU;
|
1196 if (!THREAD_CAN_SCHED(td, cpu))
1197 continue;
1198
1199 if (best == NOCPU)
1200 best = cpu;
1201 else if (runq_length[cpu] < runq_length[best])
1202 best = cpu;
1203 }
1204 KASSERT(best != NOCPU, ("no valid CPUs"));
1205
1206 return (best);
1207}
1208#endif
1209
1210void
1211sched_add(struct thread *td, int flags)
1212#ifdef SMP
1213{
1214 struct td_sched *ts;
1215 int forwarded = 0;
1216 int cpu;
1217 int single_cpu = 0;
1218
1219 ts = td->td_sched;
1220 THREAD_LOCK_ASSERT(td, MA_OWNED);
1221 KASSERT((td->td_inhibitors == 0),
1222 ("sched_add: trying to run inhibited thread"));
1223 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1224 ("sched_add: bad thread state"));
1225 KASSERT(td->td_flags & TDF_INMEM,
1226 ("sched_add: thread swapped out"));
1227
1228 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1229 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1230 sched_tdname(curthread));
1231 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1232 KTR_ATTR_LINKED, sched_tdname(td));
1233
1234
1235 /*
1236 * Now that the thread is moving to the run-queue, set the lock
1237 * to the scheduler's lock.
1238 */
1239 if (td->td_lock != &sched_lock) {
1240 mtx_lock_spin(&sched_lock);
1241 thread_lock_set(td, &sched_lock);
1242 }
1243 TD_SET_RUNQ(td);
1244
1245 if (td->td_pinned != 0) {
1246 cpu = td->td_lastcpu;
1247 ts->ts_runq = &runq_pcpu[cpu];
1248 single_cpu = 1;
1249 CTR3(KTR_RUNQ,
1250 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1251 cpu);
1252 } else if (td->td_flags & TDF_BOUND) {
1253 /* Find CPU from bound runq. */
1254 KASSERT(SKE_RUNQ_PCPU(ts),
1255 ("sched_add: bound td_sched not on cpu runq"));
1256 cpu = ts->ts_runq - &runq_pcpu[0];
1257 single_cpu = 1;
1258 CTR3(KTR_RUNQ,
1259 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1260 cpu);
1261 } else if (ts->ts_flags & TSF_AFFINITY) {
1262 /* Find a valid CPU for our cpuset */
1263 cpu = sched_pickcpu(td);
1264 ts->ts_runq = &runq_pcpu[cpu];
1265 single_cpu = 1;
1266 CTR3(KTR_RUNQ,
1267 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1268 cpu);
1269 } else {
1270 CTR2(KTR_RUNQ,
1271 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1272 td);
1273 cpu = NOCPU;
1274 ts->ts_runq = &runq;
1275 }
1276
1277 if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1278 kick_other_cpu(td->td_priority, cpu);
1279 } else {
1280 if (!single_cpu) {
1281 cpumask_t me = PCPU_GET(cpumask);
1282 cpumask_t idle = idle_cpus_mask & me;
1283
1284 if (!idle && ((flags & SRQ_INTR) == 0) &&
1285 (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1286 forwarded = forward_wakeup(cpu);
1287 }
1288
1289 if (!forwarded) {
1290 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1291 return;
1292 else
1293 maybe_resched(td);
1294 }
1295 }
1296
1297 if ((td->td_flags & TDF_NOLOAD) == 0)
1298 sched_load_add();
1299 runq_add(ts->ts_runq, td, flags);
1300 if (cpu != NOCPU)
1301 runq_length[cpu]++;
1302}
1303#else /* SMP */
1304{
1305 struct td_sched *ts;
1306
1307 ts = td->td_sched;
1308 THREAD_LOCK_ASSERT(td, MA_OWNED);
1309 KASSERT((td->td_inhibitors == 0),
1310 ("sched_add: trying to run inhibited thread"));
1311 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1312 ("sched_add: bad thread state"));
1313 KASSERT(td->td_flags & TDF_INMEM,
1314 ("sched_add: thread swapped out"));
1315 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1316 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1317 sched_tdname(curthread));
1318 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1319 KTR_ATTR_LINKED, sched_tdname(td));
1320
1321 /*
1322 * Now that the thread is moving to the run-queue, set the lock
1323 * to the scheduler's lock.
1324 */
1325 if (td->td_lock != &sched_lock) {
1326 mtx_lock_spin(&sched_lock);
1327 thread_lock_set(td, &sched_lock);
1328 }
1329 TD_SET_RUNQ(td);
1330 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1331 ts->ts_runq = &runq;
1332
1333 /*
1334 * If we are yielding (on the way out anyhow) or the thread
1335 * being saved is US, then don't try be smart about preemption
1336 * or kicking off another CPU as it won't help and may hinder.
1337 * In the YIEDLING case, we are about to run whoever is being
1338 * put in the queue anyhow, and in the OURSELF case, we are
1339 * puting ourself on the run queue which also only happens
1340 * when we are about to yield.
1341 */
1342 if ((flags & SRQ_YIELDING) == 0) {
1343 if (maybe_preempt(td))
1344 return;
1345 }
1346 if ((td->td_flags & TDF_NOLOAD) == 0)
1347 sched_load_add();
1348 runq_add(ts->ts_runq, td, flags);
1349 maybe_resched(td);
1350}
1351#endif /* SMP */
1352
1353void
1354sched_rem(struct thread *td)
1355{
1356 struct td_sched *ts;
1357
1358 ts = td->td_sched;
1359 KASSERT(td->td_flags & TDF_INMEM,
1360 ("sched_rem: thread swapped out"));
1361 KASSERT(TD_ON_RUNQ(td),
1362 ("sched_rem: thread not on run queue"));
1363 mtx_assert(&sched_lock, MA_OWNED);
1364 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1365 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1366 sched_tdname(curthread));
1367
1368 if ((td->td_flags & TDF_NOLOAD) == 0)
1369 sched_load_rem();
1370#ifdef SMP
1371 if (ts->ts_runq != &runq)
1372 runq_length[ts->ts_runq - runq_pcpu]--;
1373#endif
1374 runq_remove(ts->ts_runq, td);
1375 TD_SET_CAN_RUN(td);
1376}
1377
1378/*
1379 * Select threads to run. Note that running threads still consume a
1380 * slot.
1381 */
1382struct thread *
1383sched_choose(void)
1384{
1385 struct thread *td;
1386 struct runq *rq;
1387
1388 mtx_assert(&sched_lock, MA_OWNED);
1389#ifdef SMP
1390 struct thread *tdcpu;
1391
1392 rq = &runq;
1393 td = runq_choose_fuzz(&runq, runq_fuzz);
1394 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1395
1396 if (td == NULL ||
1397 (tdcpu != NULL &&
1398 tdcpu->td_priority < td->td_priority)) {
1399 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1400 PCPU_GET(cpuid));
1401 td = tdcpu;
1402 rq = &runq_pcpu[PCPU_GET(cpuid)];
1403 } else {
1404 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1405 }
1406
1407#else
1408 rq = &runq;
1409 td = runq_choose(&runq);
1410#endif
1411
1412 if (td) {
1413#ifdef SMP
1414 if (td == tdcpu)
1415 runq_length[PCPU_GET(cpuid)]--;
1416#endif
1417 runq_remove(rq, td);
1418 td->td_flags |= TDF_DIDRUN;
1419
1420 KASSERT(td->td_flags & TDF_INMEM,
1421 ("sched_choose: thread swapped out"));
1422 return (td);
1423 }
1424 return (PCPU_GET(idlethread));
1425}
1426
1427void
1428sched_preempt(struct thread *td)
1429{
1430 thread_lock(td);
1431 if (td->td_critnest > 1)
1432 td->td_owepreempt = 1;
1433 else
1434 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1435 thread_unlock(td);
1436}
1437
1438void
1439sched_userret(struct thread *td)
1440{
1441 /*
1442 * XXX we cheat slightly on the locking here to avoid locking in
1443 * the usual case. Setting td_priority here is essentially an
1444 * incomplete workaround for not setting it properly elsewhere.
1445 * Now that some interrupt handlers are threads, not setting it
1446 * properly elsewhere can clobber it in the window between setting
1447 * it here and returning to user mode, so don't waste time setting
1448 * it perfectly here.
1449 */
1450 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1451 ("thread with borrowed priority returning to userland"));
1452 if (td->td_priority != td->td_user_pri) {
1453 thread_lock(td);
1454 td->td_priority = td->td_user_pri;
1455 td->td_base_pri = td->td_user_pri;
1456 thread_unlock(td);
1457 }
1458}
1459
1460void
1461sched_bind(struct thread *td, int cpu)
1462{
1463 struct td_sched *ts;
1464
1465 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1466 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1467
1468 ts = td->td_sched;
1469
1470 td->td_flags |= TDF_BOUND;
1471#ifdef SMP
1472 ts->ts_runq = &runq_pcpu[cpu];
1473 if (PCPU_GET(cpuid) == cpu)
1474 return;
1475
1476 mi_switch(SW_VOL, NULL);
1477#endif
1478}
1479
1480void
1481sched_unbind(struct thread* td)
1482{
1483 THREAD_LOCK_ASSERT(td, MA_OWNED);
1484 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1485 td->td_flags &= ~TDF_BOUND;
1486}
1487
1488int
1489sched_is_bound(struct thread *td)
1490{
1491 THREAD_LOCK_ASSERT(td, MA_OWNED);
1492 return (td->td_flags & TDF_BOUND);
1493}
1494
1495void
1496sched_relinquish(struct thread *td)
1497{
1498 thread_lock(td);
1499 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1500 thread_unlock(td);
1501}
1502
1503int
1504sched_load(void)
1505{
1506 return (sched_tdcnt);
1507}
1508
1509int
1510sched_sizeof_proc(void)
1511{
1512 return (sizeof(struct proc));
1513}
1514
1515int
1516sched_sizeof_thread(void)
1517{
1518 return (sizeof(struct thread) + sizeof(struct td_sched));
1519}
1520
1521fixpt_t
1522sched_pctcpu(struct thread *td)
1523{
1524 struct td_sched *ts;
1525
1526 THREAD_LOCK_ASSERT(td, MA_OWNED);
1527 ts = td->td_sched;
1528 return (ts->ts_pctcpu);
1529}
1530
1531void
1532sched_tick(void)
1533{
1534}
1535
1536/*
1537 * The actual idle process.
1538 */
1539void
1540sched_idletd(void *dummy)
1541{
1542
1543 for (;;) {
1544 mtx_assert(&Giant, MA_NOTOWNED);
1545
1546 while (sched_runnable() == 0)
1547 cpu_idle(0);
1548
1549 mtx_lock_spin(&sched_lock);
1550 mi_switch(SW_VOL | SWT_IDLE, NULL);
1551 mtx_unlock_spin(&sched_lock);
1552 }
1553}
1554
1555/*
1556 * A CPU is entering for the first time or a thread is exiting.
1557 */
1558void
1559sched_throw(struct thread *td)
1560{
1561 /*
1562 * Correct spinlock nesting. The idle thread context that we are
1563 * borrowing was created so that it would start out with a single
1564 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1565 * explicitly acquired locks in this function, the nesting count
1566 * is now 2 rather than 1. Since we are nested, calling
1567 * spinlock_exit() will simply adjust the counts without allowing
1568 * spin lock using code to interrupt us.
1569 */
1570 if (td == NULL) {
1571 mtx_lock_spin(&sched_lock);
1572 spinlock_exit();
1573 } else {
1574 lock_profile_release_lock(&sched_lock.lock_object);
1575 MPASS(td->td_lock == &sched_lock);
1576 }
1577 mtx_assert(&sched_lock, MA_OWNED);
1578 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1579 PCPU_SET(switchtime, cpu_ticks());
1580 PCPU_SET(switchticks, ticks);
1581 cpu_throw(td, choosethread()); /* doesn't return */
1582}
1583
1584void
1585sched_fork_exit(struct thread *td)
1586{
1587
1588 /*
1589 * Finish setting up thread glue so that it begins execution in a
1590 * non-nested critical section with sched_lock held but not recursed.
1591 */
1592 td->td_oncpu = PCPU_GET(cpuid);
1593 sched_lock.mtx_lock = (uintptr_t)td;
1594 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1595 0, 0, __FILE__, __LINE__);
1596 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1597}
1598
1599char *
1600sched_tdname(struct thread *td)
1601{
1602#ifdef KTR
1603 struct td_sched *ts;
1604
1605 ts = td->td_sched;
1606 if (ts->ts_name[0] == '\0')
1607 snprintf(ts->ts_name, sizeof(ts->ts_name),
1608 "%s tid %d", td->td_name, td->td_tid);
1609 return (ts->ts_name);
1610#else
1611 return (td->td_name);
1612#endif
1613}
1614
1615void
1616sched_affinity(struct thread *td)
1617{
1618#ifdef SMP
1619 struct td_sched *ts;
1620 int cpu;
1621
1622 THREAD_LOCK_ASSERT(td, MA_OWNED);
1623
1624 /*
1625 * Set the TSF_AFFINITY flag if there is at least one CPU this
1626 * thread can't run on.
1627 */
1628 ts = td->td_sched;
1629 ts->ts_flags &= ~TSF_AFFINITY;
| 1194 if (!THREAD_CAN_SCHED(td, cpu))
1195 continue;
1196
1197 if (best == NOCPU)
1198 best = cpu;
1199 else if (runq_length[cpu] < runq_length[best])
1200 best = cpu;
1201 }
1202 KASSERT(best != NOCPU, ("no valid CPUs"));
1203
1204 return (best);
1205}
1206#endif
1207
1208void
1209sched_add(struct thread *td, int flags)
1210#ifdef SMP
1211{
1212 struct td_sched *ts;
1213 int forwarded = 0;
1214 int cpu;
1215 int single_cpu = 0;
1216
1217 ts = td->td_sched;
1218 THREAD_LOCK_ASSERT(td, MA_OWNED);
1219 KASSERT((td->td_inhibitors == 0),
1220 ("sched_add: trying to run inhibited thread"));
1221 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1222 ("sched_add: bad thread state"));
1223 KASSERT(td->td_flags & TDF_INMEM,
1224 ("sched_add: thread swapped out"));
1225
1226 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1227 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1228 sched_tdname(curthread));
1229 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1230 KTR_ATTR_LINKED, sched_tdname(td));
1231
1232
1233 /*
1234 * Now that the thread is moving to the run-queue, set the lock
1235 * to the scheduler's lock.
1236 */
1237 if (td->td_lock != &sched_lock) {
1238 mtx_lock_spin(&sched_lock);
1239 thread_lock_set(td, &sched_lock);
1240 }
1241 TD_SET_RUNQ(td);
1242
1243 if (td->td_pinned != 0) {
1244 cpu = td->td_lastcpu;
1245 ts->ts_runq = &runq_pcpu[cpu];
1246 single_cpu = 1;
1247 CTR3(KTR_RUNQ,
1248 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1249 cpu);
1250 } else if (td->td_flags & TDF_BOUND) {
1251 /* Find CPU from bound runq. */
1252 KASSERT(SKE_RUNQ_PCPU(ts),
1253 ("sched_add: bound td_sched not on cpu runq"));
1254 cpu = ts->ts_runq - &runq_pcpu[0];
1255 single_cpu = 1;
1256 CTR3(KTR_RUNQ,
1257 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1258 cpu);
1259 } else if (ts->ts_flags & TSF_AFFINITY) {
1260 /* Find a valid CPU for our cpuset */
1261 cpu = sched_pickcpu(td);
1262 ts->ts_runq = &runq_pcpu[cpu];
1263 single_cpu = 1;
1264 CTR3(KTR_RUNQ,
1265 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1266 cpu);
1267 } else {
1268 CTR2(KTR_RUNQ,
1269 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1270 td);
1271 cpu = NOCPU;
1272 ts->ts_runq = &runq;
1273 }
1274
1275 if (single_cpu && (cpu != PCPU_GET(cpuid))) {
1276 kick_other_cpu(td->td_priority, cpu);
1277 } else {
1278 if (!single_cpu) {
1279 cpumask_t me = PCPU_GET(cpumask);
1280 cpumask_t idle = idle_cpus_mask & me;
1281
1282 if (!idle && ((flags & SRQ_INTR) == 0) &&
1283 (idle_cpus_mask & ~(hlt_cpus_mask | me)))
1284 forwarded = forward_wakeup(cpu);
1285 }
1286
1287 if (!forwarded) {
1288 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1289 return;
1290 else
1291 maybe_resched(td);
1292 }
1293 }
1294
1295 if ((td->td_flags & TDF_NOLOAD) == 0)
1296 sched_load_add();
1297 runq_add(ts->ts_runq, td, flags);
1298 if (cpu != NOCPU)
1299 runq_length[cpu]++;
1300}
1301#else /* SMP */
1302{
1303 struct td_sched *ts;
1304
1305 ts = td->td_sched;
1306 THREAD_LOCK_ASSERT(td, MA_OWNED);
1307 KASSERT((td->td_inhibitors == 0),
1308 ("sched_add: trying to run inhibited thread"));
1309 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1310 ("sched_add: bad thread state"));
1311 KASSERT(td->td_flags & TDF_INMEM,
1312 ("sched_add: thread swapped out"));
1313 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1314 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1315 sched_tdname(curthread));
1316 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1317 KTR_ATTR_LINKED, sched_tdname(td));
1318
1319 /*
1320 * Now that the thread is moving to the run-queue, set the lock
1321 * to the scheduler's lock.
1322 */
1323 if (td->td_lock != &sched_lock) {
1324 mtx_lock_spin(&sched_lock);
1325 thread_lock_set(td, &sched_lock);
1326 }
1327 TD_SET_RUNQ(td);
1328 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1329 ts->ts_runq = &runq;
1330
1331 /*
1332 * If we are yielding (on the way out anyhow) or the thread
1333 * being saved is US, then don't try be smart about preemption
1334 * or kicking off another CPU as it won't help and may hinder.
1335 * In the YIEDLING case, we are about to run whoever is being
1336 * put in the queue anyhow, and in the OURSELF case, we are
1337 * puting ourself on the run queue which also only happens
1338 * when we are about to yield.
1339 */
1340 if ((flags & SRQ_YIELDING) == 0) {
1341 if (maybe_preempt(td))
1342 return;
1343 }
1344 if ((td->td_flags & TDF_NOLOAD) == 0)
1345 sched_load_add();
1346 runq_add(ts->ts_runq, td, flags);
1347 maybe_resched(td);
1348}
1349#endif /* SMP */
1350
1351void
1352sched_rem(struct thread *td)
1353{
1354 struct td_sched *ts;
1355
1356 ts = td->td_sched;
1357 KASSERT(td->td_flags & TDF_INMEM,
1358 ("sched_rem: thread swapped out"));
1359 KASSERT(TD_ON_RUNQ(td),
1360 ("sched_rem: thread not on run queue"));
1361 mtx_assert(&sched_lock, MA_OWNED);
1362 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1363 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1364 sched_tdname(curthread));
1365
1366 if ((td->td_flags & TDF_NOLOAD) == 0)
1367 sched_load_rem();
1368#ifdef SMP
1369 if (ts->ts_runq != &runq)
1370 runq_length[ts->ts_runq - runq_pcpu]--;
1371#endif
1372 runq_remove(ts->ts_runq, td);
1373 TD_SET_CAN_RUN(td);
1374}
1375
1376/*
1377 * Select threads to run. Note that running threads still consume a
1378 * slot.
1379 */
1380struct thread *
1381sched_choose(void)
1382{
1383 struct thread *td;
1384 struct runq *rq;
1385
1386 mtx_assert(&sched_lock, MA_OWNED);
1387#ifdef SMP
1388 struct thread *tdcpu;
1389
1390 rq = &runq;
1391 td = runq_choose_fuzz(&runq, runq_fuzz);
1392 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1393
1394 if (td == NULL ||
1395 (tdcpu != NULL &&
1396 tdcpu->td_priority < td->td_priority)) {
1397 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1398 PCPU_GET(cpuid));
1399 td = tdcpu;
1400 rq = &runq_pcpu[PCPU_GET(cpuid)];
1401 } else {
1402 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1403 }
1404
1405#else
1406 rq = &runq;
1407 td = runq_choose(&runq);
1408#endif
1409
1410 if (td) {
1411#ifdef SMP
1412 if (td == tdcpu)
1413 runq_length[PCPU_GET(cpuid)]--;
1414#endif
1415 runq_remove(rq, td);
1416 td->td_flags |= TDF_DIDRUN;
1417
1418 KASSERT(td->td_flags & TDF_INMEM,
1419 ("sched_choose: thread swapped out"));
1420 return (td);
1421 }
1422 return (PCPU_GET(idlethread));
1423}
1424
1425void
1426sched_preempt(struct thread *td)
1427{
1428 thread_lock(td);
1429 if (td->td_critnest > 1)
1430 td->td_owepreempt = 1;
1431 else
1432 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1433 thread_unlock(td);
1434}
1435
1436void
1437sched_userret(struct thread *td)
1438{
1439 /*
1440 * XXX we cheat slightly on the locking here to avoid locking in
1441 * the usual case. Setting td_priority here is essentially an
1442 * incomplete workaround for not setting it properly elsewhere.
1443 * Now that some interrupt handlers are threads, not setting it
1444 * properly elsewhere can clobber it in the window between setting
1445 * it here and returning to user mode, so don't waste time setting
1446 * it perfectly here.
1447 */
1448 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1449 ("thread with borrowed priority returning to userland"));
1450 if (td->td_priority != td->td_user_pri) {
1451 thread_lock(td);
1452 td->td_priority = td->td_user_pri;
1453 td->td_base_pri = td->td_user_pri;
1454 thread_unlock(td);
1455 }
1456}
1457
1458void
1459sched_bind(struct thread *td, int cpu)
1460{
1461 struct td_sched *ts;
1462
1463 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1464 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1465
1466 ts = td->td_sched;
1467
1468 td->td_flags |= TDF_BOUND;
1469#ifdef SMP
1470 ts->ts_runq = &runq_pcpu[cpu];
1471 if (PCPU_GET(cpuid) == cpu)
1472 return;
1473
1474 mi_switch(SW_VOL, NULL);
1475#endif
1476}
1477
1478void
1479sched_unbind(struct thread* td)
1480{
1481 THREAD_LOCK_ASSERT(td, MA_OWNED);
1482 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1483 td->td_flags &= ~TDF_BOUND;
1484}
1485
1486int
1487sched_is_bound(struct thread *td)
1488{
1489 THREAD_LOCK_ASSERT(td, MA_OWNED);
1490 return (td->td_flags & TDF_BOUND);
1491}
1492
1493void
1494sched_relinquish(struct thread *td)
1495{
1496 thread_lock(td);
1497 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1498 thread_unlock(td);
1499}
1500
1501int
1502sched_load(void)
1503{
1504 return (sched_tdcnt);
1505}
1506
1507int
1508sched_sizeof_proc(void)
1509{
1510 return (sizeof(struct proc));
1511}
1512
1513int
1514sched_sizeof_thread(void)
1515{
1516 return (sizeof(struct thread) + sizeof(struct td_sched));
1517}
1518
1519fixpt_t
1520sched_pctcpu(struct thread *td)
1521{
1522 struct td_sched *ts;
1523
1524 THREAD_LOCK_ASSERT(td, MA_OWNED);
1525 ts = td->td_sched;
1526 return (ts->ts_pctcpu);
1527}
1528
1529void
1530sched_tick(void)
1531{
1532}
1533
1534/*
1535 * The actual idle process.
1536 */
1537void
1538sched_idletd(void *dummy)
1539{
1540
1541 for (;;) {
1542 mtx_assert(&Giant, MA_NOTOWNED);
1543
1544 while (sched_runnable() == 0)
1545 cpu_idle(0);
1546
1547 mtx_lock_spin(&sched_lock);
1548 mi_switch(SW_VOL | SWT_IDLE, NULL);
1549 mtx_unlock_spin(&sched_lock);
1550 }
1551}
1552
1553/*
1554 * A CPU is entering for the first time or a thread is exiting.
1555 */
1556void
1557sched_throw(struct thread *td)
1558{
1559 /*
1560 * Correct spinlock nesting. The idle thread context that we are
1561 * borrowing was created so that it would start out with a single
1562 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1563 * explicitly acquired locks in this function, the nesting count
1564 * is now 2 rather than 1. Since we are nested, calling
1565 * spinlock_exit() will simply adjust the counts without allowing
1566 * spin lock using code to interrupt us.
1567 */
1568 if (td == NULL) {
1569 mtx_lock_spin(&sched_lock);
1570 spinlock_exit();
1571 } else {
1572 lock_profile_release_lock(&sched_lock.lock_object);
1573 MPASS(td->td_lock == &sched_lock);
1574 }
1575 mtx_assert(&sched_lock, MA_OWNED);
1576 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1577 PCPU_SET(switchtime, cpu_ticks());
1578 PCPU_SET(switchticks, ticks);
1579 cpu_throw(td, choosethread()); /* doesn't return */
1580}
1581
1582void
1583sched_fork_exit(struct thread *td)
1584{
1585
1586 /*
1587 * Finish setting up thread glue so that it begins execution in a
1588 * non-nested critical section with sched_lock held but not recursed.
1589 */
1590 td->td_oncpu = PCPU_GET(cpuid);
1591 sched_lock.mtx_lock = (uintptr_t)td;
1592 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1593 0, 0, __FILE__, __LINE__);
1594 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1595}
1596
1597char *
1598sched_tdname(struct thread *td)
1599{
1600#ifdef KTR
1601 struct td_sched *ts;
1602
1603 ts = td->td_sched;
1604 if (ts->ts_name[0] == '\0')
1605 snprintf(ts->ts_name, sizeof(ts->ts_name),
1606 "%s tid %d", td->td_name, td->td_tid);
1607 return (ts->ts_name);
1608#else
1609 return (td->td_name);
1610#endif
1611}
1612
1613void
1614sched_affinity(struct thread *td)
1615{
1616#ifdef SMP
1617 struct td_sched *ts;
1618 int cpu;
1619
1620 THREAD_LOCK_ASSERT(td, MA_OWNED);
1621
1622 /*
1623 * Set the TSF_AFFINITY flag if there is at least one CPU this
1624 * thread can't run on.
1625 */
1626 ts = td->td_sched;
1627 ts->ts_flags &= ~TSF_AFFINITY;
|