1 /*-
2 * Copyright (c) 1982, 1986, 1990, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 */
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37
38 #include "opt_hwpmc_hooks.h"
39 #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_slice; /* Remaining part of time slice. */
98 int ts_flags;
99 struct runq *ts_runq; /* runq the thread is currently on */
100 #ifdef KTR
101 char ts_name[TS_NAME_LEN];
102 #endif
103 };
104
105 /* flags kept in td_flags */
106 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
107 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
108 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
109
110 /* flags kept in ts_flags */
111 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
112
113 #define SKE_RUNQ_PCPU(ts) \
114 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
115
116 #define THREAD_CAN_SCHED(td, cpu) \
117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
118
119 static struct td_sched td_sched0;
120 struct mtx sched_lock;
121
122 static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */
123 static int sched_tdcnt; /* Total runnable threads in the system. */
124 static int sched_slice = 12; /* Thread run time before rescheduling. */
125
126 static void setup_runqs(void);
127 static void schedcpu(void);
128 static void schedcpu_thread(void);
129 static void sched_priority(struct thread *td, u_char prio);
130 static void sched_setup(void *dummy);
131 static void maybe_resched(struct thread *td);
132 static void updatepri(struct thread *td);
133 static void resetpriority(struct thread *td);
134 static void resetpriority_thread(struct thread *td);
135 #ifdef SMP
136 static int sched_pickcpu(struct thread *td);
137 static int forward_wakeup(int cpunum);
138 static void kick_other_cpu(int pri, int cpuid);
139 #endif
140
141 static struct kproc_desc sched_kp = {
142 "schedcpu",
143 schedcpu_thread,
144 NULL
145 };
146 SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start,
147 &sched_kp);
148 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
149
150 static void sched_initticks(void *dummy);
151 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
152 NULL);
153
154 /*
155 * Global run queue.
156 */
157 static struct runq runq;
158
159 #ifdef SMP
160 /*
161 * Per-CPU run queues
162 */
163 static struct runq runq_pcpu[MAXCPU];
164 long runq_length[MAXCPU];
165
166 static cpuset_t idle_cpus_mask;
167 #endif
168
169 struct pcpuidlestat {
170 u_int idlecalls;
171 u_int oldidlecalls;
172 };
173 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
174
175 static void
setup_runqs(void)176 setup_runqs(void)
177 {
178 #ifdef SMP
179 int i;
180
181 for (i = 0; i < MAXCPU; ++i)
182 runq_init(&runq_pcpu[i]);
183 #endif
184
185 runq_init(&runq);
186 }
187
188 static int
sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)189 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
190 {
191 int error, new_val, period;
192
193 period = 1000000 / realstathz;
194 new_val = period * sched_slice;
195 error = sysctl_handle_int(oidp, &new_val, 0, req);
196 if (error != 0 || req->newptr == NULL)
197 return (error);
198 if (new_val <= 0)
199 return (EINVAL);
200 sched_slice = imax(1, (new_val + period / 2) / period);
201 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
202 realstathz);
203 return (0);
204 }
205
206 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
207
208 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
209 "Scheduler name");
210 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
211 NULL, 0, sysctl_kern_quantum, "I",
212 "Quantum for timeshare threads in microseconds");
213 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
214 "Quantum for timeshare threads in stathz ticks");
215 #ifdef SMP
216 /* Enable forwarding of wakeups to all other cpus */
217 static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL,
218 "Kernel SMP");
219
220 static int runq_fuzz = 1;
221 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
222
223 static int forward_wakeup_enabled = 1;
224 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
225 &forward_wakeup_enabled, 0,
226 "Forwarding of wakeup to idle CPUs");
227
228 static int forward_wakeups_requested = 0;
229 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
230 &forward_wakeups_requested, 0,
231 "Requests for Forwarding of wakeup to idle CPUs");
232
233 static int forward_wakeups_delivered = 0;
234 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
235 &forward_wakeups_delivered, 0,
236 "Completed Forwarding of wakeup to idle CPUs");
237
238 static int forward_wakeup_use_mask = 1;
239 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
240 &forward_wakeup_use_mask, 0,
241 "Use the mask of idle cpus");
242
243 static int forward_wakeup_use_loop = 0;
244 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
245 &forward_wakeup_use_loop, 0,
246 "Use a loop to find idle cpus");
247
248 #endif
249 #if 0
250 static int sched_followon = 0;
251 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
252 &sched_followon, 0,
253 "allow threads to share a quantum");
254 #endif
255
256 SDT_PROVIDER_DEFINE(sched);
257
258 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
259 "struct proc *", "uint8_t");
260 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
261 "struct proc *", "void *");
262 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
263 "struct proc *", "void *", "int");
264 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
265 "struct proc *", "uint8_t", "struct thread *");
266 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
267 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
268 "struct proc *");
269 SDT_PROBE_DEFINE(sched, , , on__cpu);
270 SDT_PROBE_DEFINE(sched, , , remain__cpu);
271 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
272 "struct proc *");
273
274 static __inline void
sched_load_add(void)275 sched_load_add(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 static __inline void
sched_load_rem(void)284 sched_load_rem(void)
285 {
286
287 sched_tdcnt--;
288 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
289 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
290 }
291 /*
292 * Arrange to reschedule if necessary, taking the priorities and
293 * schedulers into account.
294 */
295 static void
maybe_resched(struct thread * td)296 maybe_resched(struct thread *td)
297 {
298
299 THREAD_LOCK_ASSERT(td, MA_OWNED);
300 if (td->td_priority < curthread->td_priority)
301 curthread->td_flags |= TDF_NEEDRESCHED;
302 }
303
304 /*
305 * This function is called when a thread is about to be put on run queue
306 * because it has been made runnable or its priority has been adjusted. It
307 * determines if the new thread should be immediately preempted to. If so,
308 * it switches to it and eventually returns true. If not, it returns false
309 * so that the caller may place the thread on an appropriate run queue.
310 */
311 int
maybe_preempt(struct thread * td)312 maybe_preempt(struct thread *td)
313 {
314 #ifdef PREEMPTION
315 struct thread *ctd;
316 int cpri, pri;
317
318 /*
319 * The new thread should not preempt the current thread if any of the
320 * following conditions are true:
321 *
322 * - The kernel is in the throes of crashing (panicstr).
323 * - The current thread has a higher (numerically lower) or
324 * equivalent priority. Note that this prevents curthread from
325 * trying to preempt to itself.
326 * - It is too early in the boot for context switches (cold is set).
327 * - The current thread has an inhibitor set or is in the process of
328 * exiting. In this case, the current thread is about to switch
329 * out anyways, so there's no point in preempting. If we did,
330 * the current thread would not be properly resumed as well, so
331 * just avoid that whole landmine.
332 * - If the new thread's priority is not a realtime priority and
333 * the current thread's priority is not an idle priority and
334 * FULL_PREEMPTION is disabled.
335 *
336 * If all of these conditions are false, but the current thread is in
337 * a nested critical section, then we have to defer the preemption
338 * until we exit the critical section. Otherwise, switch immediately
339 * to the new thread.
340 */
341 ctd = curthread;
342 THREAD_LOCK_ASSERT(td, MA_OWNED);
343 KASSERT((td->td_inhibitors == 0),
344 ("maybe_preempt: trying to run inhibited thread"));
345 pri = td->td_priority;
346 cpri = ctd->td_priority;
347 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
348 TD_IS_INHIBITED(ctd))
349 return (0);
350 #ifndef FULL_PREEMPTION
351 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
352 return (0);
353 #endif
354
355 if (ctd->td_critnest > 1) {
356 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
357 ctd->td_critnest);
358 ctd->td_owepreempt = 1;
359 return (0);
360 }
361 /*
362 * Thread is runnable but not yet put on system run queue.
363 */
364 MPASS(ctd->td_lock == td->td_lock);
365 MPASS(TD_ON_RUNQ(td));
366 TD_SET_RUNNING(td);
367 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
368 td->td_proc->p_pid, td->td_name);
369 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
370 /*
371 * td's lock pointer may have changed. We have to return with it
372 * locked.
373 */
374 spinlock_enter();
375 thread_unlock(ctd);
376 thread_lock(td);
377 spinlock_exit();
378 return (1);
379 #else
380 return (0);
381 #endif
382 }
383
384 /*
385 * Constants for digital decay and forget:
386 * 90% of (td_estcpu) usage in 5 * loadav time
387 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
388 * Note that, as ps(1) mentions, this can let percentages
389 * total over 100% (I've seen 137.9% for 3 processes).
390 *
391 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
392 *
393 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
394 * That is, the system wants to compute a value of decay such
395 * that the following for loop:
396 * for (i = 0; i < (5 * loadavg); i++)
397 * td_estcpu *= decay;
398 * will compute
399 * td_estcpu *= 0.1;
400 * for all values of loadavg:
401 *
402 * Mathematically this loop can be expressed by saying:
403 * decay ** (5 * loadavg) ~= .1
404 *
405 * The system computes decay as:
406 * decay = (2 * loadavg) / (2 * loadavg + 1)
407 *
408 * We wish to prove that the system's computation of decay
409 * will always fulfill the equation:
410 * decay ** (5 * loadavg) ~= .1
411 *
412 * If we compute b as:
413 * b = 2 * loadavg
414 * then
415 * decay = b / (b + 1)
416 *
417 * We now need to prove two things:
418 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
419 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
420 *
421 * Facts:
422 * For x close to zero, exp(x) =~ 1 + x, since
423 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
424 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
425 * For x close to zero, ln(1+x) =~ x, since
426 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
427 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
428 * ln(.1) =~ -2.30
429 *
430 * Proof of (1):
431 * Solve (factor)**(power) =~ .1 given power (5*loadav):
432 * solving for factor,
433 * ln(factor) =~ (-2.30/5*loadav), or
434 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
435 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
436 *
437 * Proof of (2):
438 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
439 * solving for power,
440 * power*ln(b/(b+1)) =~ -2.30, or
441 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
442 *
443 * Actual power values for the implemented algorithm are as follows:
444 * loadav: 1 2 3 4
445 * power: 5.68 10.32 14.94 19.55
446 */
447
448 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
449 #define loadfactor(loadav) (2 * (loadav))
450 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
451
452 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
453 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
454 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
455
456 /*
457 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
458 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
459 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
460 *
461 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
462 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
463 *
464 * If you don't want to bother with the faster/more-accurate formula, you
465 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
466 * (more general) method of calculating the %age of CPU used by a process.
467 */
468 #define CCPU_SHIFT 11
469
470 /*
471 * Recompute process priorities, every hz ticks.
472 * MP-safe, called without the Giant mutex.
473 */
474 /* ARGSUSED */
475 static void
schedcpu(void)476 schedcpu(void)
477 {
478 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
479 struct thread *td;
480 struct proc *p;
481 struct td_sched *ts;
482 int awake;
483
484 sx_slock(&allproc_lock);
485 FOREACH_PROC_IN_SYSTEM(p) {
486 PROC_LOCK(p);
487 if (p->p_state == PRS_NEW) {
488 PROC_UNLOCK(p);
489 continue;
490 }
491 FOREACH_THREAD_IN_PROC(p, td) {
492 awake = 0;
493 thread_lock(td);
494 ts = td->td_sched;
495 /*
496 * Increment sleep time (if sleeping). We
497 * ignore overflow, as above.
498 */
499 /*
500 * The td_sched slptimes are not touched in wakeup
501 * because the thread may not HAVE everything in
502 * memory? XXX I think this is out of date.
503 */
504 if (TD_ON_RUNQ(td)) {
505 awake = 1;
506 td->td_flags &= ~TDF_DIDRUN;
507 } else if (TD_IS_RUNNING(td)) {
508 awake = 1;
509 /* Do not clear TDF_DIDRUN */
510 } else if (td->td_flags & TDF_DIDRUN) {
511 awake = 1;
512 td->td_flags &= ~TDF_DIDRUN;
513 }
514
515 /*
516 * ts_pctcpu is only for ps and ttyinfo().
517 */
518 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
519 /*
520 * If the td_sched has been idle the entire second,
521 * stop recalculating its priority until
522 * it wakes up.
523 */
524 if (ts->ts_cpticks != 0) {
525 #if (FSHIFT >= CCPU_SHIFT)
526 ts->ts_pctcpu += (realstathz == 100)
527 ? ((fixpt_t) ts->ts_cpticks) <<
528 (FSHIFT - CCPU_SHIFT) :
529 100 * (((fixpt_t) ts->ts_cpticks)
530 << (FSHIFT - CCPU_SHIFT)) / realstathz;
531 #else
532 ts->ts_pctcpu += ((FSCALE - ccpu) *
533 (ts->ts_cpticks *
534 FSCALE / realstathz)) >> FSHIFT;
535 #endif
536 ts->ts_cpticks = 0;
537 }
538 /*
539 * If there are ANY running threads in this process,
540 * then don't count it as sleeping.
541 * XXX: this is broken.
542 */
543 if (awake) {
544 if (ts->ts_slptime > 1) {
545 /*
546 * In an ideal world, this should not
547 * happen, because whoever woke us
548 * up from the long sleep should have
549 * unwound the slptime and reset our
550 * priority before we run at the stale
551 * priority. Should KASSERT at some
552 * point when all the cases are fixed.
553 */
554 updatepri(td);
555 }
556 ts->ts_slptime = 0;
557 } else
558 ts->ts_slptime++;
559 if (ts->ts_slptime > 1) {
560 thread_unlock(td);
561 continue;
562 }
563 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
564 resetpriority(td);
565 resetpriority_thread(td);
566 thread_unlock(td);
567 }
568 PROC_UNLOCK(p);
569 }
570 sx_sunlock(&allproc_lock);
571 }
572
573 /*
574 * Main loop for a kthread that executes schedcpu once a second.
575 */
576 static void
schedcpu_thread(void)577 schedcpu_thread(void)
578 {
579
580 for (;;) {
581 schedcpu();
582 pause("-", hz);
583 }
584 }
585
586 /*
587 * Recalculate the priority of a process after it has slept for a while.
588 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
589 * least six times the loadfactor will decay td_estcpu to zero.
590 */
591 static void
updatepri(struct thread * td)592 updatepri(struct thread *td)
593 {
594 struct td_sched *ts;
595 fixpt_t loadfac;
596 unsigned int newcpu;
597
598 ts = td->td_sched;
599 loadfac = loadfactor(averunnable.ldavg[0]);
600 if (ts->ts_slptime > 5 * loadfac)
601 td->td_estcpu = 0;
602 else {
603 newcpu = td->td_estcpu;
604 ts->ts_slptime--; /* was incremented in schedcpu() */
605 while (newcpu && --ts->ts_slptime)
606 newcpu = decay_cpu(loadfac, newcpu);
607 td->td_estcpu = newcpu;
608 }
609 }
610
611 /*
612 * Compute the priority of a process when running in user mode.
613 * Arrange to reschedule if the resulting priority is better
614 * than that of the current process.
615 */
616 static void
resetpriority(struct thread * td)617 resetpriority(struct thread *td)
618 {
619 register unsigned int newpriority;
620
621 if (td->td_pri_class == PRI_TIMESHARE) {
622 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
623 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
624 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
625 PRI_MAX_TIMESHARE);
626 sched_user_prio(td, newpriority);
627 }
628 }
629
630 /*
631 * Update the thread's priority when the associated process's user
632 * priority changes.
633 */
634 static void
resetpriority_thread(struct thread * td)635 resetpriority_thread(struct thread *td)
636 {
637
638 /* Only change threads with a time sharing user priority. */
639 if (td->td_priority < PRI_MIN_TIMESHARE ||
640 td->td_priority > PRI_MAX_TIMESHARE)
641 return;
642
643 /* XXX the whole needresched thing is broken, but not silly. */
644 maybe_resched(td);
645
646 sched_prio(td, td->td_user_pri);
647 }
648
649 /* ARGSUSED */
650 static void
sched_setup(void * dummy)651 sched_setup(void *dummy)
652 {
653
654 setup_runqs();
655
656 /* Account for thread0. */
657 sched_load_add();
658 }
659
660 /*
661 * This routine determines time constants after stathz and hz are setup.
662 */
663 static void
sched_initticks(void * dummy)664 sched_initticks(void *dummy)
665 {
666
667 realstathz = stathz ? stathz : hz;
668 sched_slice = realstathz / 10; /* ~100ms */
669 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
670 realstathz);
671 }
672
673 /* External interfaces start here */
674
675 /*
676 * Very early in the boot some setup of scheduler-specific
677 * parts of proc0 and of some scheduler resources needs to be done.
678 * Called from:
679 * proc0_init()
680 */
681 void
schedinit(void)682 schedinit(void)
683 {
684 /*
685 * Set up the scheduler specific parts of proc0.
686 */
687 proc0.p_sched = NULL; /* XXX */
688 thread0.td_sched = &td_sched0;
689 thread0.td_lock = &sched_lock;
690 td_sched0.ts_slice = sched_slice;
691 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
692 }
693
694 int
sched_runnable(void)695 sched_runnable(void)
696 {
697 #ifdef SMP
698 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
699 #else
700 return runq_check(&runq);
701 #endif
702 }
703
704 int
sched_rr_interval(void)705 sched_rr_interval(void)
706 {
707
708 /* Convert sched_slice from stathz to hz. */
709 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
710 }
711
712 /*
713 * We adjust the priority of the current process. The priority of
714 * a process gets worse as it accumulates CPU time. The cpu usage
715 * estimator (td_estcpu) is increased here. resetpriority() will
716 * compute a different priority each time td_estcpu increases by
717 * INVERSE_ESTCPU_WEIGHT
718 * (until MAXPRI is reached). The cpu usage estimator ramps up
719 * quite quickly when the process is running (linearly), and decays
720 * away exponentially, at a rate which is proportionally slower when
721 * the system is busy. The basic principle is that the system will
722 * 90% forget that the process used a lot of CPU time in 5 * loadav
723 * seconds. This causes the system to favor processes which haven't
724 * run much recently, and to round-robin among other processes.
725 */
726 void
sched_clock(struct thread * td)727 sched_clock(struct thread *td)
728 {
729 struct pcpuidlestat *stat;
730 struct td_sched *ts;
731
732 THREAD_LOCK_ASSERT(td, MA_OWNED);
733 ts = td->td_sched;
734
735 ts->ts_cpticks++;
736 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
737 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
738 resetpriority(td);
739 resetpriority_thread(td);
740 }
741
742 /*
743 * Force a context switch if the current thread has used up a full
744 * time slice (default is 100ms).
745 */
746 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
747 ts->ts_slice = sched_slice;
748 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
749 }
750
751 stat = DPCPU_PTR(idlestat);
752 stat->oldidlecalls = stat->idlecalls;
753 stat->idlecalls = 0;
754 }
755
756 /*
757 * Charge child's scheduling CPU usage to parent.
758 */
759 void
sched_exit(struct proc * p,struct thread * td)760 sched_exit(struct proc *p, struct thread *td)
761 {
762
763 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
764 "prio:%d", td->td_priority);
765
766 PROC_LOCK_ASSERT(p, MA_OWNED);
767 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
768 }
769
770 void
sched_exit_thread(struct thread * td,struct thread * child)771 sched_exit_thread(struct thread *td, struct thread *child)
772 {
773
774 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
775 "prio:%d", child->td_priority);
776 thread_lock(td);
777 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
778 thread_unlock(td);
779 thread_lock(child);
780 if ((child->td_flags & TDF_NOLOAD) == 0)
781 sched_load_rem();
782 thread_unlock(child);
783 }
784
785 void
sched_fork(struct thread * td,struct thread * childtd)786 sched_fork(struct thread *td, struct thread *childtd)
787 {
788 sched_fork_thread(td, childtd);
789 }
790
791 void
sched_fork_thread(struct thread * td,struct thread * childtd)792 sched_fork_thread(struct thread *td, struct thread *childtd)
793 {
794 struct td_sched *ts;
795
796 childtd->td_estcpu = td->td_estcpu;
797 childtd->td_lock = &sched_lock;
798 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
799 childtd->td_priority = childtd->td_base_pri;
800 ts = childtd->td_sched;
801 bzero(ts, sizeof(*ts));
802 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
803 ts->ts_slice = 1;
804 }
805
806 void
sched_nice(struct proc * p,int nice)807 sched_nice(struct proc *p, int nice)
808 {
809 struct thread *td;
810
811 PROC_LOCK_ASSERT(p, MA_OWNED);
812 p->p_nice = nice;
813 FOREACH_THREAD_IN_PROC(p, td) {
814 thread_lock(td);
815 resetpriority(td);
816 resetpriority_thread(td);
817 thread_unlock(td);
818 }
819 }
820
821 void
sched_class(struct thread * td,int class)822 sched_class(struct thread *td, int class)
823 {
824 THREAD_LOCK_ASSERT(td, MA_OWNED);
825 td->td_pri_class = class;
826 }
827
828 /*
829 * Adjust the priority of a thread.
830 */
831 static void
sched_priority(struct thread * td,u_char prio)832 sched_priority(struct thread *td, u_char prio)
833 {
834
835
836 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
837 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
838 sched_tdname(curthread));
839 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
840 if (td != curthread && prio > td->td_priority) {
841 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
842 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
843 prio, KTR_ATTR_LINKED, sched_tdname(td));
844 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
845 curthread);
846 }
847 THREAD_LOCK_ASSERT(td, MA_OWNED);
848 if (td->td_priority == prio)
849 return;
850 td->td_priority = prio;
851 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
852 sched_rem(td);
853 sched_add(td, SRQ_BORING);
854 }
855 }
856
857 /*
858 * Update a thread's priority when it is lent another thread's
859 * priority.
860 */
861 void
sched_lend_prio(struct thread * td,u_char prio)862 sched_lend_prio(struct thread *td, u_char prio)
863 {
864
865 td->td_flags |= TDF_BORROWING;
866 sched_priority(td, prio);
867 }
868
869 /*
870 * Restore a thread's priority when priority propagation is
871 * over. The prio argument is the minimum priority the thread
872 * needs to have to satisfy other possible priority lending
873 * requests. If the thread's regulary priority is less
874 * important than prio the thread will keep a priority boost
875 * of prio.
876 */
877 void
sched_unlend_prio(struct thread * td,u_char prio)878 sched_unlend_prio(struct thread *td, u_char prio)
879 {
880 u_char base_pri;
881
882 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
883 td->td_base_pri <= PRI_MAX_TIMESHARE)
884 base_pri = td->td_user_pri;
885 else
886 base_pri = td->td_base_pri;
887 if (prio >= base_pri) {
888 td->td_flags &= ~TDF_BORROWING;
889 sched_prio(td, base_pri);
890 } else
891 sched_lend_prio(td, prio);
892 }
893
894 void
sched_prio(struct thread * td,u_char prio)895 sched_prio(struct thread *td, u_char prio)
896 {
897 u_char oldprio;
898
899 /* First, update the base priority. */
900 td->td_base_pri = prio;
901
902 /*
903 * If the thread is borrowing another thread's priority, don't ever
904 * lower the priority.
905 */
906 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
907 return;
908
909 /* Change the real priority. */
910 oldprio = td->td_priority;
911 sched_priority(td, prio);
912
913 /*
914 * If the thread is on a turnstile, then let the turnstile update
915 * its state.
916 */
917 if (TD_ON_LOCK(td) && oldprio != prio)
918 turnstile_adjust(td, oldprio);
919 }
920
921 void
sched_user_prio(struct thread * td,u_char prio)922 sched_user_prio(struct thread *td, u_char prio)
923 {
924
925 THREAD_LOCK_ASSERT(td, MA_OWNED);
926 td->td_base_user_pri = prio;
927 if (td->td_lend_user_pri <= prio)
928 return;
929 td->td_user_pri = prio;
930 }
931
932 void
sched_lend_user_prio(struct thread * td,u_char prio)933 sched_lend_user_prio(struct thread *td, u_char prio)
934 {
935
936 THREAD_LOCK_ASSERT(td, MA_OWNED);
937 td->td_lend_user_pri = prio;
938 td->td_user_pri = min(prio, td->td_base_user_pri);
939 if (td->td_priority > td->td_user_pri)
940 sched_prio(td, td->td_user_pri);
941 else if (td->td_priority != td->td_user_pri)
942 td->td_flags |= TDF_NEEDRESCHED;
943 }
944
945 void
sched_sleep(struct thread * td,int pri)946 sched_sleep(struct thread *td, int pri)
947 {
948
949 THREAD_LOCK_ASSERT(td, MA_OWNED);
950 td->td_slptick = ticks;
951 td->td_sched->ts_slptime = 0;
952 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
953 sched_prio(td, pri);
954 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
955 td->td_flags |= TDF_CANSWAP;
956 }
957
958 void
sched_switch(struct thread * td,struct thread * newtd,int flags)959 sched_switch(struct thread *td, struct thread *newtd, int flags)
960 {
961 struct mtx *tmtx;
962 struct td_sched *ts;
963 struct proc *p;
964 int preempted;
965
966 tmtx = NULL;
967 ts = td->td_sched;
968 p = td->td_proc;
969
970 THREAD_LOCK_ASSERT(td, MA_OWNED);
971
972 /*
973 * Switch to the sched lock to fix things up and pick
974 * a new thread.
975 * Block the td_lock in order to avoid breaking the critical path.
976 */
977 if (td->td_lock != &sched_lock) {
978 mtx_lock_spin(&sched_lock);
979 tmtx = thread_lock_block(td);
980 }
981
982 if ((td->td_flags & TDF_NOLOAD) == 0)
983 sched_load_rem();
984
985 td->td_lastcpu = td->td_oncpu;
986 preempted = !((td->td_flags & TDF_SLICEEND) ||
987 (flags & SWT_RELINQUISH));
988 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
989 td->td_owepreempt = 0;
990 td->td_oncpu = NOCPU;
991
992 /*
993 * At the last moment, if this thread is still marked RUNNING,
994 * then put it back on the run queue as it has not been suspended
995 * or stopped or any thing else similar. We never put the idle
996 * threads on the run queue, however.
997 */
998 if (td->td_flags & TDF_IDLETD) {
999 TD_SET_CAN_RUN(td);
1000 #ifdef SMP
1001 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
1002 #endif
1003 } else {
1004 if (TD_IS_RUNNING(td)) {
1005 /* Put us back on the run queue. */
1006 sched_add(td, preempted ?
1007 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1008 SRQ_OURSELF|SRQ_YIELDING);
1009 }
1010 }
1011 if (newtd) {
1012 /*
1013 * The thread we are about to run needs to be counted
1014 * as if it had been added to the run queue and selected.
1015 * It came from:
1016 * * A preemption
1017 * * An upcall
1018 * * A followon
1019 */
1020 KASSERT((newtd->td_inhibitors == 0),
1021 ("trying to run inhibited thread"));
1022 newtd->td_flags |= TDF_DIDRUN;
1023 TD_SET_RUNNING(newtd);
1024 if ((newtd->td_flags & TDF_NOLOAD) == 0)
1025 sched_load_add();
1026 } else {
1027 newtd = choosethread();
1028 MPASS(newtd->td_lock == &sched_lock);
1029 }
1030
1031 if (td != newtd) {
1032 #ifdef HWPMC_HOOKS
1033 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1034 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1035 #endif
1036
1037 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1038
1039 /* I feel sleepy */
1040 lock_profile_release_lock(&sched_lock.lock_object);
1041 #ifdef KDTRACE_HOOKS
1042 /*
1043 * If DTrace has set the active vtime enum to anything
1044 * other than INACTIVE (0), then it should have set the
1045 * function to call.
1046 */
1047 if (dtrace_vtime_active)
1048 (*dtrace_vtime_switch_func)(newtd);
1049 #endif
1050
1051 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1052 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1053 0, 0, __FILE__, __LINE__);
1054 /*
1055 * Where am I? What year is it?
1056 * We are in the same thread that went to sleep above,
1057 * but any amount of time may have passed. All our context
1058 * will still be available as will local variables.
1059 * PCPU values however may have changed as we may have
1060 * changed CPU so don't trust cached values of them.
1061 * New threads will go to fork_exit() instead of here
1062 * so if you change things here you may need to change
1063 * things there too.
1064 *
1065 * If the thread above was exiting it will never wake
1066 * up again here, so either it has saved everything it
1067 * needed to, or the thread_wait() or wait() will
1068 * need to reap it.
1069 */
1070
1071 SDT_PROBE0(sched, , , on__cpu);
1072 #ifdef HWPMC_HOOKS
1073 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1074 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1075 #endif
1076 } else
1077 SDT_PROBE0(sched, , , remain__cpu);
1078
1079 #ifdef SMP
1080 if (td->td_flags & TDF_IDLETD)
1081 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1082 #endif
1083 sched_lock.mtx_lock = (uintptr_t)td;
1084 td->td_oncpu = PCPU_GET(cpuid);
1085 MPASS(td->td_lock == &sched_lock);
1086 }
1087
1088 void
sched_wakeup(struct thread * td)1089 sched_wakeup(struct thread *td)
1090 {
1091 struct td_sched *ts;
1092
1093 THREAD_LOCK_ASSERT(td, MA_OWNED);
1094 ts = td->td_sched;
1095 td->td_flags &= ~TDF_CANSWAP;
1096 if (ts->ts_slptime > 1) {
1097 updatepri(td);
1098 resetpriority(td);
1099 }
1100 td->td_slptick = 0;
1101 ts->ts_slptime = 0;
1102 ts->ts_slice = sched_slice;
1103 sched_add(td, SRQ_BORING);
1104 }
1105
1106 #ifdef SMP
1107 static int
forward_wakeup(int cpunum)1108 forward_wakeup(int cpunum)
1109 {
1110 struct pcpu *pc;
1111 cpuset_t dontuse, map, map2;
1112 u_int id, me;
1113 int iscpuset;
1114
1115 mtx_assert(&sched_lock, MA_OWNED);
1116
1117 CTR0(KTR_RUNQ, "forward_wakeup()");
1118
1119 if ((!forward_wakeup_enabled) ||
1120 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1121 return (0);
1122 if (!smp_started || cold || panicstr)
1123 return (0);
1124
1125 forward_wakeups_requested++;
1126
1127 /*
1128 * Check the idle mask we received against what we calculated
1129 * before in the old version.
1130 */
1131 me = PCPU_GET(cpuid);
1132
1133 /* Don't bother if we should be doing it ourself. */
1134 if (CPU_ISSET(me, &idle_cpus_mask) &&
1135 (cpunum == NOCPU || me == cpunum))
1136 return (0);
1137
1138 CPU_SETOF(me, &dontuse);
1139 CPU_OR(&dontuse, &stopped_cpus);
1140 CPU_OR(&dontuse, &hlt_cpus_mask);
1141 CPU_ZERO(&map2);
1142 if (forward_wakeup_use_loop) {
1143 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1144 id = pc->pc_cpuid;
1145 if (!CPU_ISSET(id, &dontuse) &&
1146 pc->pc_curthread == pc->pc_idlethread) {
1147 CPU_SET(id, &map2);
1148 }
1149 }
1150 }
1151
1152 if (forward_wakeup_use_mask) {
1153 map = idle_cpus_mask;
1154 CPU_NAND(&map, &dontuse);
1155
1156 /* If they are both on, compare and use loop if different. */
1157 if (forward_wakeup_use_loop) {
1158 if (CPU_CMP(&map, &map2)) {
1159 printf("map != map2, loop method preferred\n");
1160 map = map2;
1161 }
1162 }
1163 } else {
1164 map = map2;
1165 }
1166
1167 /* If we only allow a specific CPU, then mask off all the others. */
1168 if (cpunum != NOCPU) {
1169 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1170 iscpuset = CPU_ISSET(cpunum, &map);
1171 if (iscpuset == 0)
1172 CPU_ZERO(&map);
1173 else
1174 CPU_SETOF(cpunum, &map);
1175 }
1176 if (!CPU_EMPTY(&map)) {
1177 forward_wakeups_delivered++;
1178 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1179 id = pc->pc_cpuid;
1180 if (!CPU_ISSET(id, &map))
1181 continue;
1182 if (cpu_idle_wakeup(pc->pc_cpuid))
1183 CPU_CLR(id, &map);
1184 }
1185 if (!CPU_EMPTY(&map))
1186 ipi_selected(map, IPI_AST);
1187 return (1);
1188 }
1189 if (cpunum == NOCPU)
1190 printf("forward_wakeup: Idle processor not found\n");
1191 return (0);
1192 }
1193
1194 static void
kick_other_cpu(int pri,int cpuid)1195 kick_other_cpu(int pri, int cpuid)
1196 {
1197 struct pcpu *pcpu;
1198 int cpri;
1199
1200 pcpu = pcpu_find(cpuid);
1201 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1202 forward_wakeups_delivered++;
1203 if (!cpu_idle_wakeup(cpuid))
1204 ipi_cpu(cpuid, IPI_AST);
1205 return;
1206 }
1207
1208 cpri = pcpu->pc_curthread->td_priority;
1209 if (pri >= cpri)
1210 return;
1211
1212 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1213 #if !defined(FULL_PREEMPTION)
1214 if (pri <= PRI_MAX_ITHD)
1215 #endif /* ! FULL_PREEMPTION */
1216 {
1217 ipi_cpu(cpuid, IPI_PREEMPT);
1218 return;
1219 }
1220 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1221
1222 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1223 ipi_cpu(cpuid, IPI_AST);
1224 return;
1225 }
1226 #endif /* SMP */
1227
1228 #ifdef SMP
1229 static int
sched_pickcpu(struct thread * td)1230 sched_pickcpu(struct thread *td)
1231 {
1232 int best, cpu;
1233
1234 mtx_assert(&sched_lock, MA_OWNED);
1235
1236 if (THREAD_CAN_SCHED(td, td->td_lastcpu))
1237 best = td->td_lastcpu;
1238 else
1239 best = NOCPU;
1240 CPU_FOREACH(cpu) {
1241 if (!THREAD_CAN_SCHED(td, cpu))
1242 continue;
1243
1244 if (best == NOCPU)
1245 best = cpu;
1246 else if (runq_length[cpu] < runq_length[best])
1247 best = cpu;
1248 }
1249 KASSERT(best != NOCPU, ("no valid CPUs"));
1250
1251 return (best);
1252 }
1253 #endif
1254
1255 void
sched_add(struct thread * td,int flags)1256 sched_add(struct thread *td, int flags)
1257 #ifdef SMP
1258 {
1259 cpuset_t tidlemsk;
1260 struct td_sched *ts;
1261 u_int cpu, cpuid;
1262 int forwarded = 0;
1263 int single_cpu = 0;
1264
1265 ts = td->td_sched;
1266 THREAD_LOCK_ASSERT(td, MA_OWNED);
1267 KASSERT((td->td_inhibitors == 0),
1268 ("sched_add: trying to run inhibited thread"));
1269 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1270 ("sched_add: bad thread state"));
1271 KASSERT(td->td_flags & TDF_INMEM,
1272 ("sched_add: thread swapped out"));
1273
1274 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1275 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1276 sched_tdname(curthread));
1277 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1278 KTR_ATTR_LINKED, sched_tdname(td));
1279 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1280 flags & SRQ_PREEMPTED);
1281
1282
1283 /*
1284 * Now that the thread is moving to the run-queue, set the lock
1285 * to the scheduler's lock.
1286 */
1287 if (td->td_lock != &sched_lock) {
1288 mtx_lock_spin(&sched_lock);
1289 thread_lock_set(td, &sched_lock);
1290 }
1291 TD_SET_RUNQ(td);
1292
1293 /*
1294 * If SMP is started and the thread is pinned or otherwise limited to
1295 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1296 * Otherwise, queue the thread to the global run queue.
1297 *
1298 * If SMP has not yet been started we must use the global run queue
1299 * as per-CPU state may not be initialized yet and we may crash if we
1300 * try to access the per-CPU run queues.
1301 */
1302 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1303 ts->ts_flags & TSF_AFFINITY)) {
1304 if (td->td_pinned != 0)
1305 cpu = td->td_lastcpu;
1306 else if (td->td_flags & TDF_BOUND) {
1307 /* Find CPU from bound runq. */
1308 KASSERT(SKE_RUNQ_PCPU(ts),
1309 ("sched_add: bound td_sched not on cpu runq"));
1310 cpu = ts->ts_runq - &runq_pcpu[0];
1311 } else
1312 /* Find a valid CPU for our cpuset */
1313 cpu = sched_pickcpu(td);
1314 ts->ts_runq = &runq_pcpu[cpu];
1315 single_cpu = 1;
1316 CTR3(KTR_RUNQ,
1317 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1318 cpu);
1319 } else {
1320 CTR2(KTR_RUNQ,
1321 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1322 td);
1323 cpu = NOCPU;
1324 ts->ts_runq = &runq;
1325 }
1326
1327 cpuid = PCPU_GET(cpuid);
1328 if (single_cpu && cpu != cpuid) {
1329 kick_other_cpu(td->td_priority, cpu);
1330 } else {
1331 if (!single_cpu) {
1332 tidlemsk = idle_cpus_mask;
1333 CPU_NAND(&tidlemsk, &hlt_cpus_mask);
1334 CPU_CLR(cpuid, &tidlemsk);
1335
1336 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1337 ((flags & SRQ_INTR) == 0) &&
1338 !CPU_EMPTY(&tidlemsk))
1339 forwarded = forward_wakeup(cpu);
1340 }
1341
1342 if (!forwarded) {
1343 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
1344 return;
1345 else
1346 maybe_resched(td);
1347 }
1348 }
1349
1350 if ((td->td_flags & TDF_NOLOAD) == 0)
1351 sched_load_add();
1352 runq_add(ts->ts_runq, td, flags);
1353 if (cpu != NOCPU)
1354 runq_length[cpu]++;
1355 }
1356 #else /* SMP */
1357 {
1358 struct td_sched *ts;
1359
1360 ts = td->td_sched;
1361 THREAD_LOCK_ASSERT(td, MA_OWNED);
1362 KASSERT((td->td_inhibitors == 0),
1363 ("sched_add: trying to run inhibited thread"));
1364 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1365 ("sched_add: bad thread state"));
1366 KASSERT(td->td_flags & TDF_INMEM,
1367 ("sched_add: thread swapped out"));
1368 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1369 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1370 sched_tdname(curthread));
1371 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1372 KTR_ATTR_LINKED, sched_tdname(td));
1373 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1374 flags & SRQ_PREEMPTED);
1375
1376 /*
1377 * Now that the thread is moving to the run-queue, set the lock
1378 * to the scheduler's lock.
1379 */
1380 if (td->td_lock != &sched_lock) {
1381 mtx_lock_spin(&sched_lock);
1382 thread_lock_set(td, &sched_lock);
1383 }
1384 TD_SET_RUNQ(td);
1385 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1386 ts->ts_runq = &runq;
1387
1388 /*
1389 * If we are yielding (on the way out anyhow) or the thread
1390 * being saved is US, then don't try be smart about preemption
1391 * or kicking off another CPU as it won't help and may hinder.
1392 * In the YIEDLING case, we are about to run whoever is being
1393 * put in the queue anyhow, and in the OURSELF case, we are
1394 * puting ourself on the run queue which also only happens
1395 * when we are about to yield.
1396 */
1397 if ((flags & SRQ_YIELDING) == 0) {
1398 if (maybe_preempt(td))
1399 return;
1400 }
1401 if ((td->td_flags & TDF_NOLOAD) == 0)
1402 sched_load_add();
1403 runq_add(ts->ts_runq, td, flags);
1404 maybe_resched(td);
1405 }
1406 #endif /* SMP */
1407
1408 void
sched_rem(struct thread * td)1409 sched_rem(struct thread *td)
1410 {
1411 struct td_sched *ts;
1412
1413 ts = td->td_sched;
1414 KASSERT(td->td_flags & TDF_INMEM,
1415 ("sched_rem: thread swapped out"));
1416 KASSERT(TD_ON_RUNQ(td),
1417 ("sched_rem: thread not on run queue"));
1418 mtx_assert(&sched_lock, MA_OWNED);
1419 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1420 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1421 sched_tdname(curthread));
1422 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1423
1424 if ((td->td_flags & TDF_NOLOAD) == 0)
1425 sched_load_rem();
1426 #ifdef SMP
1427 if (ts->ts_runq != &runq)
1428 runq_length[ts->ts_runq - runq_pcpu]--;
1429 #endif
1430 runq_remove(ts->ts_runq, td);
1431 TD_SET_CAN_RUN(td);
1432 }
1433
1434 /*
1435 * Select threads to run. Note that running threads still consume a
1436 * slot.
1437 */
1438 struct thread *
sched_choose(void)1439 sched_choose(void)
1440 {
1441 struct thread *td;
1442 struct runq *rq;
1443
1444 mtx_assert(&sched_lock, MA_OWNED);
1445 #ifdef SMP
1446 struct thread *tdcpu;
1447
1448 rq = &runq;
1449 td = runq_choose_fuzz(&runq, runq_fuzz);
1450 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1451
1452 if (td == NULL ||
1453 (tdcpu != NULL &&
1454 tdcpu->td_priority < td->td_priority)) {
1455 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1456 PCPU_GET(cpuid));
1457 td = tdcpu;
1458 rq = &runq_pcpu[PCPU_GET(cpuid)];
1459 } else {
1460 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1461 }
1462
1463 #else
1464 rq = &runq;
1465 td = runq_choose(&runq);
1466 #endif
1467
1468 if (td) {
1469 #ifdef SMP
1470 if (td == tdcpu)
1471 runq_length[PCPU_GET(cpuid)]--;
1472 #endif
1473 runq_remove(rq, td);
1474 td->td_flags |= TDF_DIDRUN;
1475
1476 KASSERT(td->td_flags & TDF_INMEM,
1477 ("sched_choose: thread swapped out"));
1478 return (td);
1479 }
1480 return (PCPU_GET(idlethread));
1481 }
1482
1483 void
sched_preempt(struct thread * td)1484 sched_preempt(struct thread *td)
1485 {
1486
1487 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1488 thread_lock(td);
1489 if (td->td_critnest > 1)
1490 td->td_owepreempt = 1;
1491 else
1492 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1493 thread_unlock(td);
1494 }
1495
1496 void
sched_userret(struct thread * td)1497 sched_userret(struct thread *td)
1498 {
1499 /*
1500 * XXX we cheat slightly on the locking here to avoid locking in
1501 * the usual case. Setting td_priority here is essentially an
1502 * incomplete workaround for not setting it properly elsewhere.
1503 * Now that some interrupt handlers are threads, not setting it
1504 * properly elsewhere can clobber it in the window between setting
1505 * it here and returning to user mode, so don't waste time setting
1506 * it perfectly here.
1507 */
1508 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1509 ("thread with borrowed priority returning to userland"));
1510 if (td->td_priority != td->td_user_pri) {
1511 thread_lock(td);
1512 td->td_priority = td->td_user_pri;
1513 td->td_base_pri = td->td_user_pri;
1514 thread_unlock(td);
1515 }
1516 }
1517
1518 void
sched_bind(struct thread * td,int cpu)1519 sched_bind(struct thread *td, int cpu)
1520 {
1521 struct td_sched *ts;
1522
1523 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1524 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1525
1526 ts = td->td_sched;
1527
1528 td->td_flags |= TDF_BOUND;
1529 #ifdef SMP
1530 ts->ts_runq = &runq_pcpu[cpu];
1531 if (PCPU_GET(cpuid) == cpu)
1532 return;
1533
1534 mi_switch(SW_VOL, NULL);
1535 #endif
1536 }
1537
1538 void
sched_unbind(struct thread * td)1539 sched_unbind(struct thread* td)
1540 {
1541 THREAD_LOCK_ASSERT(td, MA_OWNED);
1542 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1543 td->td_flags &= ~TDF_BOUND;
1544 }
1545
1546 int
sched_is_bound(struct thread * td)1547 sched_is_bound(struct thread *td)
1548 {
1549 THREAD_LOCK_ASSERT(td, MA_OWNED);
1550 return (td->td_flags & TDF_BOUND);
1551 }
1552
1553 void
sched_relinquish(struct thread * td)1554 sched_relinquish(struct thread *td)
1555 {
1556 thread_lock(td);
1557 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1558 thread_unlock(td);
1559 }
1560
1561 int
sched_load(void)1562 sched_load(void)
1563 {
1564 return (sched_tdcnt);
1565 }
1566
1567 int
sched_sizeof_proc(void)1568 sched_sizeof_proc(void)
1569 {
1570 return (sizeof(struct proc));
1571 }
1572
1573 int
sched_sizeof_thread(void)1574 sched_sizeof_thread(void)
1575 {
1576 return (sizeof(struct thread) + sizeof(struct td_sched));
1577 }
1578
1579 fixpt_t
sched_pctcpu(struct thread * td)1580 sched_pctcpu(struct thread *td)
1581 {
1582 struct td_sched *ts;
1583
1584 THREAD_LOCK_ASSERT(td, MA_OWNED);
1585 ts = td->td_sched;
1586 return (ts->ts_pctcpu);
1587 }
1588
1589 #ifdef RACCT
1590 /*
1591 * Calculates the contribution to the thread cpu usage for the latest
1592 * (unfinished) second.
1593 */
1594 fixpt_t
sched_pctcpu_delta(struct thread * td)1595 sched_pctcpu_delta(struct thread *td)
1596 {
1597 struct td_sched *ts;
1598 fixpt_t delta;
1599 int realstathz;
1600
1601 THREAD_LOCK_ASSERT(td, MA_OWNED);
1602 ts = td->td_sched;
1603 delta = 0;
1604 realstathz = stathz ? stathz : hz;
1605 if (ts->ts_cpticks != 0) {
1606 #if (FSHIFT >= CCPU_SHIFT)
1607 delta = (realstathz == 100)
1608 ? ((fixpt_t) ts->ts_cpticks) <<
1609 (FSHIFT - CCPU_SHIFT) :
1610 100 * (((fixpt_t) ts->ts_cpticks)
1611 << (FSHIFT - CCPU_SHIFT)) / realstathz;
1612 #else
1613 delta = ((FSCALE - ccpu) *
1614 (ts->ts_cpticks *
1615 FSCALE / realstathz)) >> FSHIFT;
1616 #endif
1617 }
1618
1619 return (delta);
1620 }
1621 #endif
1622
1623 void
sched_tick(int cnt)1624 sched_tick(int cnt)
1625 {
1626 }
1627
1628 /*
1629 * The actual idle process.
1630 */
1631 void
sched_idletd(void * dummy)1632 sched_idletd(void *dummy)
1633 {
1634 struct pcpuidlestat *stat;
1635
1636 THREAD_NO_SLEEPING();
1637 stat = DPCPU_PTR(idlestat);
1638 for (;;) {
1639 mtx_assert(&Giant, MA_NOTOWNED);
1640
1641 while (sched_runnable() == 0) {
1642 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1643 stat->idlecalls++;
1644 }
1645
1646 mtx_lock_spin(&sched_lock);
1647 mi_switch(SW_VOL | SWT_IDLE, NULL);
1648 mtx_unlock_spin(&sched_lock);
1649 }
1650 }
1651
1652 /*
1653 * A CPU is entering for the first time or a thread is exiting.
1654 */
1655 void
sched_throw(struct thread * td)1656 sched_throw(struct thread *td)
1657 {
1658 /*
1659 * Correct spinlock nesting. The idle thread context that we are
1660 * borrowing was created so that it would start out with a single
1661 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1662 * explicitly acquired locks in this function, the nesting count
1663 * is now 2 rather than 1. Since we are nested, calling
1664 * spinlock_exit() will simply adjust the counts without allowing
1665 * spin lock using code to interrupt us.
1666 */
1667 if (td == NULL) {
1668 mtx_lock_spin(&sched_lock);
1669 spinlock_exit();
1670 PCPU_SET(switchtime, cpu_ticks());
1671 PCPU_SET(switchticks, ticks);
1672 } else {
1673 lock_profile_release_lock(&sched_lock.lock_object);
1674 MPASS(td->td_lock == &sched_lock);
1675 }
1676 mtx_assert(&sched_lock, MA_OWNED);
1677 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1678 cpu_throw(td, choosethread()); /* doesn't return */
1679 }
1680
1681 void
sched_fork_exit(struct thread * td)1682 sched_fork_exit(struct thread *td)
1683 {
1684
1685 /*
1686 * Finish setting up thread glue so that it begins execution in a
1687 * non-nested critical section with sched_lock held but not recursed.
1688 */
1689 td->td_oncpu = PCPU_GET(cpuid);
1690 sched_lock.mtx_lock = (uintptr_t)td;
1691 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1692 0, 0, __FILE__, __LINE__);
1693 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1694 }
1695
1696 char *
sched_tdname(struct thread * td)1697 sched_tdname(struct thread *td)
1698 {
1699 #ifdef KTR
1700 struct td_sched *ts;
1701
1702 ts = td->td_sched;
1703 if (ts->ts_name[0] == '\0')
1704 snprintf(ts->ts_name, sizeof(ts->ts_name),
1705 "%s tid %d", td->td_name, td->td_tid);
1706 return (ts->ts_name);
1707 #else
1708 return (td->td_name);
1709 #endif
1710 }
1711
1712 #ifdef KTR
1713 void
sched_clear_tdname(struct thread * td)1714 sched_clear_tdname(struct thread *td)
1715 {
1716 struct td_sched *ts;
1717
1718 ts = td->td_sched;
1719 ts->ts_name[0] = '\0';
1720 }
1721 #endif
1722
1723 void
sched_affinity(struct thread * td)1724 sched_affinity(struct thread *td)
1725 {
1726 #ifdef SMP
1727 struct td_sched *ts;
1728 int cpu;
1729
1730 THREAD_LOCK_ASSERT(td, MA_OWNED);
1731
1732 /*
1733 * Set the TSF_AFFINITY flag if there is at least one CPU this
1734 * thread can't run on.
1735 */
1736 ts = td->td_sched;
1737 ts->ts_flags &= ~TSF_AFFINITY;
1738 CPU_FOREACH(cpu) {
1739 if (!THREAD_CAN_SCHED(td, cpu)) {
1740 ts->ts_flags |= TSF_AFFINITY;
1741 break;
1742 }
1743 }
1744
1745 /*
1746 * If this thread can run on all CPUs, nothing else to do.
1747 */
1748 if (!(ts->ts_flags & TSF_AFFINITY))
1749 return;
1750
1751 /* Pinned threads and bound threads should be left alone. */
1752 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1753 return;
1754
1755 switch (td->td_state) {
1756 case TDS_RUNQ:
1757 /*
1758 * If we are on a per-CPU runqueue that is in the set,
1759 * then nothing needs to be done.
1760 */
1761 if (ts->ts_runq != &runq &&
1762 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1763 return;
1764
1765 /* Put this thread on a valid per-CPU runqueue. */
1766 sched_rem(td);
1767 sched_add(td, SRQ_BORING);
1768 break;
1769 case TDS_RUNNING:
1770 /*
1771 * See if our current CPU is in the set. If not, force a
1772 * context switch.
1773 */
1774 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1775 return;
1776
1777 td->td_flags |= TDF_NEEDRESCHED;
1778 if (td != curthread)
1779 ipi_cpu(cpu, IPI_AST);
1780 break;
1781 default:
1782 break;
1783 }
1784 #endif
1785 }
1786