1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1982, 1986, 1990, 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 */
36
37 #include <sys/cdefs.h>
38 #include "opt_hwpmc_hooks.h"
39 #include "opt_sched.h"
40
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/cpuset.h>
44 #include <sys/kernel.h>
45 #include <sys/ktr.h>
46 #include <sys/lock.h>
47 #include <sys/kthread.h>
48 #include <sys/mutex.h>
49 #include <sys/proc.h>
50 #include <sys/resourcevar.h>
51 #include <sys/sched.h>
52 #include <sys/sdt.h>
53 #include <sys/smp.h>
54 #include <sys/sysctl.h>
55 #include <sys/sx.h>
56 #include <sys/turnstile.h>
57 #include <sys/umtxvar.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>
67 int __read_mostly dtrace_vtime_active;
68 dtrace_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 * All fields are protected by the scheduler lock.
92 */
93 struct td_sched {
94 fixpt_t ts_pctcpu; /* %cpu during p_swtime. */
95 u_int ts_estcpu; /* Estimated cpu utilization. */
96 int ts_cpticks; /* Ticks of cpu time. */
97 int ts_slptime; /* Seconds !RUNNING. */
98 int ts_slice; /* Remaining part of time slice. */
99 int ts_flags;
100 struct runq *ts_runq; /* runq the thread is currently on */
101 #ifdef KTR
102 char ts_name[TS_NAME_LEN];
103 #endif
104 };
105
106 /* flags kept in td_flags */
107 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
108 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
109 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
110
111 /* flags kept in ts_flags */
112 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
113
114 #define SKE_RUNQ_PCPU(ts) \
115 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
116
117 #define THREAD_CAN_SCHED(td, cpu) \
118 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
119
120 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
121 sizeof(struct thread0_storage),
122 "increase struct thread0_storage.t0st_sched size");
123
124 static struct mtx sched_lock;
125
126 static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */
127 static int sched_tdcnt; /* Total runnable threads in the system. */
128 static int sched_slice = 12; /* Thread run time before rescheduling. */
129
130 static void setup_runqs(void);
131 static void schedcpu(void);
132 static void schedcpu_thread(void);
133 static void sched_priority(struct thread *td, u_char prio);
134 static void sched_setup(void *dummy);
135 static void maybe_resched(struct thread *td);
136 static void updatepri(struct thread *td);
137 static void resetpriority(struct thread *td);
138 static void resetpriority_thread(struct thread *td);
139 #ifdef SMP
140 static int sched_pickcpu(struct thread *td);
141 static int forward_wakeup(int cpunum);
142 static void kick_other_cpu(int pri, int cpuid);
143 #endif
144
145 static struct kproc_desc sched_kp = {
146 "schedcpu",
147 schedcpu_thread,
148 NULL
149 };
150 SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start,
151 &sched_kp);
152 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
153
154 static void sched_initticks(void *dummy);
155 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
156 NULL);
157
158 /*
159 * Global run queue.
160 */
161 static struct runq runq;
162
163 #ifdef SMP
164 /*
165 * Per-CPU run queues
166 */
167 static struct runq runq_pcpu[MAXCPU];
168 long runq_length[MAXCPU];
169
170 static cpuset_t idle_cpus_mask;
171 #endif
172
173 struct pcpuidlestat {
174 u_int idlecalls;
175 u_int oldidlecalls;
176 };
177 DPCPU_DEFINE_STATIC(struct pcpuidlestat, idlestat);
178
179 static void
setup_runqs(void)180 setup_runqs(void)
181 {
182 #ifdef SMP
183 int i;
184
185 for (i = 0; i < MAXCPU; ++i)
186 runq_init(&runq_pcpu[i]);
187 #endif
188
189 runq_init(&runq);
190 }
191
192 static int
sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)193 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
194 {
195 int error, new_val, period;
196
197 period = 1000000 / realstathz;
198 new_val = period * sched_slice;
199 error = sysctl_handle_int(oidp, &new_val, 0, req);
200 if (error != 0 || req->newptr == NULL)
201 return (error);
202 if (new_val <= 0)
203 return (EINVAL);
204 sched_slice = imax(1, (new_val + period / 2) / period);
205 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
206 realstathz);
207 return (0);
208 }
209
210 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
211 "Scheduler");
212
213 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
214 "Scheduler name");
215 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum,
216 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
217 sysctl_kern_quantum, "I",
218 "Quantum for timeshare threads in microseconds");
219 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
220 "Quantum for timeshare threads in stathz ticks");
221 #ifdef SMP
222 /* Enable forwarding of wakeups to all other cpus */
223 static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup,
224 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
225 "Kernel SMP");
226
227 static int runq_fuzz = 1;
228 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
229
230 static int forward_wakeup_enabled = 1;
231 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
232 &forward_wakeup_enabled, 0,
233 "Forwarding of wakeup to idle CPUs");
234
235 static int forward_wakeups_requested = 0;
236 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
237 &forward_wakeups_requested, 0,
238 "Requests for Forwarding of wakeup to idle CPUs");
239
240 static int forward_wakeups_delivered = 0;
241 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
242 &forward_wakeups_delivered, 0,
243 "Completed Forwarding of wakeup to idle CPUs");
244
245 static int forward_wakeup_use_mask = 1;
246 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
247 &forward_wakeup_use_mask, 0,
248 "Use the mask of idle cpus");
249
250 static int forward_wakeup_use_loop = 0;
251 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
252 &forward_wakeup_use_loop, 0,
253 "Use a loop to find idle cpus");
254
255 #endif
256 #if 0
257 static int sched_followon = 0;
258 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
259 &sched_followon, 0,
260 "allow threads to share a quantum");
261 #endif
262
263 SDT_PROVIDER_DEFINE(sched);
264
265 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
266 "struct proc *", "uint8_t");
267 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
268 "struct proc *", "void *");
269 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
270 "struct proc *", "void *", "int");
271 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
272 "struct proc *", "uint8_t", "struct thread *");
273 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
274 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
275 "struct proc *");
276 SDT_PROBE_DEFINE(sched, , , on__cpu);
277 SDT_PROBE_DEFINE(sched, , , remain__cpu);
278 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
279 "struct proc *");
280
281 static __inline void
sched_load_add(void)282 sched_load_add(void)
283 {
284
285 sched_tdcnt++;
286 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
287 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
288 }
289
290 static __inline void
sched_load_rem(void)291 sched_load_rem(void)
292 {
293
294 sched_tdcnt--;
295 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
296 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
297 }
298 /*
299 * Arrange to reschedule if necessary, taking the priorities and
300 * schedulers into account.
301 */
302 static void
maybe_resched(struct thread * td)303 maybe_resched(struct thread *td)
304 {
305
306 THREAD_LOCK_ASSERT(td, MA_OWNED);
307 if (td->td_priority < curthread->td_priority)
308 curthread->td_flags |= TDF_NEEDRESCHED;
309 }
310
311 /*
312 * This function is called when a thread is about to be put on run queue
313 * because it has been made runnable or its priority has been adjusted. It
314 * determines if the new thread should preempt the current thread. If so,
315 * it sets td_owepreempt to request a preemption.
316 */
317 int
maybe_preempt(struct thread * td)318 maybe_preempt(struct thread *td)
319 {
320 #ifdef PREEMPTION
321 struct thread *ctd;
322 int cpri, pri;
323
324 /*
325 * The new thread should not preempt the current thread if any of the
326 * following conditions are true:
327 *
328 * - The kernel is in the throes of crashing (panicstr).
329 * - The current thread has a higher (numerically lower) or
330 * equivalent priority. Note that this prevents curthread from
331 * trying to preempt to itself.
332 * - The current thread has an inhibitor set or is in the process of
333 * exiting. In this case, the current thread is about to switch
334 * out anyways, so there's no point in preempting. If we did,
335 * the current thread would not be properly resumed as well, so
336 * just avoid that whole landmine.
337 * - If the new thread's priority is not a realtime priority and
338 * the current thread's priority is not an idle priority and
339 * FULL_PREEMPTION is disabled.
340 *
341 * If all of these conditions are false, but the current thread is in
342 * a nested critical section, then we have to defer the preemption
343 * until we exit the critical section. Otherwise, switch immediately
344 * to the new thread.
345 */
346 ctd = curthread;
347 THREAD_LOCK_ASSERT(td, MA_OWNED);
348 KASSERT((td->td_inhibitors == 0),
349 ("maybe_preempt: trying to run inhibited thread"));
350 pri = td->td_priority;
351 cpri = ctd->td_priority;
352 if (KERNEL_PANICKED() || pri >= cpri /* || dumping */ ||
353 TD_IS_INHIBITED(ctd))
354 return (0);
355 #ifndef FULL_PREEMPTION
356 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
357 return (0);
358 #endif
359
360 CTR0(KTR_PROC, "maybe_preempt: scheduling preemption");
361 ctd->td_owepreempt = 1;
362 return (1);
363 #else
364 return (0);
365 #endif
366 }
367
368 /*
369 * Constants for digital decay and forget:
370 * 90% of (ts_estcpu) usage in 5 * loadav time
371 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
372 * Note that, as ps(1) mentions, this can let percentages
373 * total over 100% (I've seen 137.9% for 3 processes).
374 *
375 * Note that schedclock() updates ts_estcpu and p_cpticks asynchronously.
376 *
377 * We wish to decay away 90% of ts_estcpu in (5 * loadavg) seconds.
378 * That is, the system wants to compute a value of decay such
379 * that the following for loop:
380 * for (i = 0; i < (5 * loadavg); i++)
381 * ts_estcpu *= decay;
382 * will compute
383 * ts_estcpu *= 0.1;
384 * for all values of loadavg:
385 *
386 * Mathematically this loop can be expressed by saying:
387 * decay ** (5 * loadavg) ~= .1
388 *
389 * The system computes decay as:
390 * decay = (2 * loadavg) / (2 * loadavg + 1)
391 *
392 * We wish to prove that the system's computation of decay
393 * will always fulfill the equation:
394 * decay ** (5 * loadavg) ~= .1
395 *
396 * If we compute b as:
397 * b = 2 * loadavg
398 * then
399 * decay = b / (b + 1)
400 *
401 * We now need to prove two things:
402 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
403 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
404 *
405 * Facts:
406 * For x close to zero, exp(x) =~ 1 + x, since
407 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
408 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
409 * For x close to zero, ln(1+x) =~ x, since
410 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
411 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
412 * ln(.1) =~ -2.30
413 *
414 * Proof of (1):
415 * Solve (factor)**(power) =~ .1 given power (5*loadav):
416 * solving for factor,
417 * ln(factor) =~ (-2.30/5*loadav), or
418 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
419 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
420 *
421 * Proof of (2):
422 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
423 * solving for power,
424 * power*ln(b/(b+1)) =~ -2.30, or
425 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
426 *
427 * Actual power values for the implemented algorithm are as follows:
428 * loadav: 1 2 3 4
429 * power: 5.68 10.32 14.94 19.55
430 */
431
432 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
433 #define loadfactor(loadav) (2 * (loadav))
434 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
435
436 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
437 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
438 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0,
439 "Decay factor used for updating %CPU");
440
441 /*
442 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
443 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
444 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
445 *
446 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
447 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
448 *
449 * If you don't want to bother with the faster/more-accurate formula, you
450 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
451 * (more general) method of calculating the %age of CPU used by a process.
452 */
453 #define CCPU_SHIFT 11
454
455 /*
456 * Recompute process priorities, every hz ticks.
457 * MP-safe, called without the Giant mutex.
458 */
459 /* ARGSUSED */
460 static void
schedcpu(void)461 schedcpu(void)
462 {
463 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
464 struct thread *td;
465 struct proc *p;
466 struct td_sched *ts;
467 int awake;
468
469 sx_slock(&allproc_lock);
470 FOREACH_PROC_IN_SYSTEM(p) {
471 PROC_LOCK(p);
472 if (p->p_state == PRS_NEW) {
473 PROC_UNLOCK(p);
474 continue;
475 }
476 FOREACH_THREAD_IN_PROC(p, td) {
477 awake = 0;
478 ts = td_get_sched(td);
479 thread_lock(td);
480 /*
481 * Increment sleep time (if sleeping). We
482 * ignore overflow, as above.
483 */
484 /*
485 * The td_sched slptimes are not touched in wakeup
486 * because the thread may not HAVE everything in
487 * memory? XXX I think this is out of date.
488 */
489 if (TD_ON_RUNQ(td)) {
490 awake = 1;
491 td->td_flags &= ~TDF_DIDRUN;
492 } else if (TD_IS_RUNNING(td)) {
493 awake = 1;
494 /* Do not clear TDF_DIDRUN */
495 } else if (td->td_flags & TDF_DIDRUN) {
496 awake = 1;
497 td->td_flags &= ~TDF_DIDRUN;
498 }
499
500 /*
501 * ts_pctcpu is only for ps and ttyinfo().
502 */
503 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
504 /*
505 * If the td_sched has been idle the entire second,
506 * stop recalculating its priority until
507 * it wakes up.
508 */
509 if (ts->ts_cpticks != 0) {
510 #if (FSHIFT >= CCPU_SHIFT)
511 ts->ts_pctcpu += (realstathz == 100)
512 ? ((fixpt_t) ts->ts_cpticks) <<
513 (FSHIFT - CCPU_SHIFT) :
514 100 * (((fixpt_t) ts->ts_cpticks)
515 << (FSHIFT - CCPU_SHIFT)) / realstathz;
516 #else
517 ts->ts_pctcpu += ((FSCALE - ccpu) *
518 (ts->ts_cpticks *
519 FSCALE / realstathz)) >> FSHIFT;
520 #endif
521 ts->ts_cpticks = 0;
522 }
523 /*
524 * If there are ANY running threads in this process,
525 * then don't count it as sleeping.
526 * XXX: this is broken.
527 */
528 if (awake) {
529 if (ts->ts_slptime > 1) {
530 /*
531 * In an ideal world, this should not
532 * happen, because whoever woke us
533 * up from the long sleep should have
534 * unwound the slptime and reset our
535 * priority before we run at the stale
536 * priority. Should KASSERT at some
537 * point when all the cases are fixed.
538 */
539 updatepri(td);
540 }
541 ts->ts_slptime = 0;
542 } else
543 ts->ts_slptime++;
544 if (ts->ts_slptime > 1) {
545 thread_unlock(td);
546 continue;
547 }
548 ts->ts_estcpu = decay_cpu(loadfac, ts->ts_estcpu);
549 resetpriority(td);
550 resetpriority_thread(td);
551 thread_unlock(td);
552 }
553 PROC_UNLOCK(p);
554 }
555 sx_sunlock(&allproc_lock);
556 }
557
558 /*
559 * Main loop for a kthread that executes schedcpu once a second.
560 */
561 static void
schedcpu_thread(void)562 schedcpu_thread(void)
563 {
564
565 for (;;) {
566 schedcpu();
567 pause("-", hz);
568 }
569 }
570
571 /*
572 * Recalculate the priority of a process after it has slept for a while.
573 * For all load averages >= 1 and max ts_estcpu of 255, sleeping for at
574 * least six times the loadfactor will decay ts_estcpu to zero.
575 */
576 static void
updatepri(struct thread * td)577 updatepri(struct thread *td)
578 {
579 struct td_sched *ts;
580 fixpt_t loadfac;
581 unsigned int newcpu;
582
583 ts = td_get_sched(td);
584 loadfac = loadfactor(averunnable.ldavg[0]);
585 if (ts->ts_slptime > 5 * loadfac)
586 ts->ts_estcpu = 0;
587 else {
588 newcpu = ts->ts_estcpu;
589 ts->ts_slptime--; /* was incremented in schedcpu() */
590 while (newcpu && --ts->ts_slptime)
591 newcpu = decay_cpu(loadfac, newcpu);
592 ts->ts_estcpu = newcpu;
593 }
594 }
595
596 /*
597 * Compute the priority of a process when running in user mode.
598 * Arrange to reschedule if the resulting priority is better
599 * than that of the current process.
600 */
601 static void
resetpriority(struct thread * td)602 resetpriority(struct thread *td)
603 {
604 u_int newpriority;
605
606 if (td->td_pri_class != PRI_TIMESHARE)
607 return;
608 newpriority = PUSER +
609 td_get_sched(td)->ts_estcpu / INVERSE_ESTCPU_WEIGHT +
610 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
611 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
612 PRI_MAX_TIMESHARE);
613 sched_user_prio(td, newpriority);
614 }
615
616 /*
617 * Update the thread's priority when the associated process's user
618 * priority changes.
619 */
620 static void
resetpriority_thread(struct thread * td)621 resetpriority_thread(struct thread *td)
622 {
623
624 /* Only change threads with a time sharing user priority. */
625 if (td->td_priority < PRI_MIN_TIMESHARE ||
626 td->td_priority > PRI_MAX_TIMESHARE)
627 return;
628
629 /* XXX the whole needresched thing is broken, but not silly. */
630 maybe_resched(td);
631
632 sched_prio(td, td->td_user_pri);
633 }
634
635 /* ARGSUSED */
636 static void
sched_setup(void * dummy)637 sched_setup(void *dummy)
638 {
639
640 setup_runqs();
641
642 /* Account for thread0. */
643 sched_load_add();
644 }
645
646 /*
647 * This routine determines time constants after stathz and hz are setup.
648 */
649 static void
sched_initticks(void * dummy)650 sched_initticks(void *dummy)
651 {
652
653 realstathz = stathz ? stathz : hz;
654 sched_slice = realstathz / 10; /* ~100ms */
655 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
656 realstathz);
657 }
658
659 /* External interfaces start here */
660
661 /*
662 * Very early in the boot some setup of scheduler-specific
663 * parts of proc0 and of some scheduler resources needs to be done.
664 * Called from:
665 * proc0_init()
666 */
667 void
schedinit(void)668 schedinit(void)
669 {
670
671 /*
672 * Set up the scheduler specific parts of thread0.
673 */
674 thread0.td_lock = &sched_lock;
675 td_get_sched(&thread0)->ts_slice = sched_slice;
676 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN);
677 }
678
679 void
schedinit_ap(void)680 schedinit_ap(void)
681 {
682
683 /* Nothing needed. */
684 }
685
686 int
sched_runnable(void)687 sched_runnable(void)
688 {
689 #ifdef SMP
690 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
691 #else
692 return runq_check(&runq);
693 #endif
694 }
695
696 int
sched_rr_interval(void)697 sched_rr_interval(void)
698 {
699
700 /* Convert sched_slice from stathz to hz. */
701 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
702 }
703
704 /*
705 * We adjust the priority of the current process. The priority of a
706 * process gets worse as it accumulates CPU time. The cpu usage
707 * estimator (ts_estcpu) is increased here. resetpriority() will
708 * compute a different priority each time ts_estcpu increases by
709 * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached). The
710 * cpu usage estimator ramps up quite quickly when the process is
711 * running (linearly), and decays away exponentially, at a rate which
712 * is proportionally slower when the system is busy. The basic
713 * principle is that the system will 90% forget that the process used
714 * a lot of CPU time in 5 * loadav seconds. This causes the system to
715 * favor processes which haven't run much recently, and to round-robin
716 * among other processes.
717 */
718 static void
sched_clock_tick(struct thread * td)719 sched_clock_tick(struct thread *td)
720 {
721 struct pcpuidlestat *stat;
722 struct td_sched *ts;
723
724 THREAD_LOCK_ASSERT(td, MA_OWNED);
725 ts = td_get_sched(td);
726
727 ts->ts_cpticks++;
728 ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1);
729 if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
730 resetpriority(td);
731 resetpriority_thread(td);
732 }
733
734 /*
735 * Force a context switch if the current thread has used up a full
736 * time slice (default is 100ms).
737 */
738 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
739 ts->ts_slice = sched_slice;
740 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
741 }
742
743 stat = DPCPU_PTR(idlestat);
744 stat->oldidlecalls = stat->idlecalls;
745 stat->idlecalls = 0;
746 }
747
748 void
sched_clock(struct thread * td,int cnt)749 sched_clock(struct thread *td, int cnt)
750 {
751
752 for ( ; cnt > 0; cnt--)
753 sched_clock_tick(td);
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_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu +
778 td_get_sched(child)->ts_estcpu);
779 thread_unlock(td);
780 thread_lock(child);
781 if ((child->td_flags & TDF_NOLOAD) == 0)
782 sched_load_rem();
783 thread_unlock(child);
784 }
785
786 void
sched_fork(struct thread * td,struct thread * childtd)787 sched_fork(struct thread *td, struct thread *childtd)
788 {
789 sched_fork_thread(td, childtd);
790 }
791
792 void
sched_fork_thread(struct thread * td,struct thread * childtd)793 sched_fork_thread(struct thread *td, struct thread *childtd)
794 {
795 struct td_sched *ts, *tsc;
796
797 childtd->td_oncpu = NOCPU;
798 childtd->td_lastcpu = NOCPU;
799 childtd->td_lock = &sched_lock;
800 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
801 childtd->td_domain.dr_policy = td->td_cpuset->cs_domain;
802 childtd->td_priority = childtd->td_base_pri;
803 ts = td_get_sched(childtd);
804 bzero(ts, sizeof(*ts));
805 tsc = td_get_sched(td);
806 ts->ts_estcpu = tsc->ts_estcpu;
807 ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY);
808 ts->ts_slice = 1;
809 }
810
811 void
sched_nice(struct proc * p,int nice)812 sched_nice(struct proc *p, int nice)
813 {
814 struct thread *td;
815
816 PROC_LOCK_ASSERT(p, MA_OWNED);
817 p->p_nice = nice;
818 FOREACH_THREAD_IN_PROC(p, td) {
819 thread_lock(td);
820 resetpriority(td);
821 resetpriority_thread(td);
822 thread_unlock(td);
823 }
824 }
825
826 void
sched_class(struct thread * td,int class)827 sched_class(struct thread *td, int class)
828 {
829 THREAD_LOCK_ASSERT(td, MA_OWNED);
830 td->td_pri_class = class;
831 }
832
833 /*
834 * Adjust the priority of a thread.
835 */
836 static void
sched_priority(struct thread * td,u_char prio)837 sched_priority(struct thread *td, u_char prio)
838 {
839
840 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
841 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
842 sched_tdname(curthread));
843 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
844 if (td != curthread && prio > td->td_priority) {
845 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
846 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
847 prio, KTR_ATTR_LINKED, sched_tdname(td));
848 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
849 curthread);
850 }
851 THREAD_LOCK_ASSERT(td, MA_OWNED);
852 if (td->td_priority == prio)
853 return;
854 td->td_priority = prio;
855 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
856 sched_rem(td);
857 sched_add(td, SRQ_BORING | SRQ_HOLDTD);
858 }
859 }
860
861 /*
862 * Update a thread's priority when it is lent another thread's
863 * priority.
864 */
865 void
sched_lend_prio(struct thread * td,u_char prio)866 sched_lend_prio(struct thread *td, u_char prio)
867 {
868
869 td->td_flags |= TDF_BORROWING;
870 sched_priority(td, prio);
871 }
872
873 /*
874 * Restore a thread's priority when priority propagation is
875 * over. The prio argument is the minimum priority the thread
876 * needs to have to satisfy other possible priority lending
877 * requests. If the thread's regulary priority is less
878 * important than prio the thread will keep a priority boost
879 * of prio.
880 */
881 void
sched_unlend_prio(struct thread * td,u_char prio)882 sched_unlend_prio(struct thread *td, u_char prio)
883 {
884 u_char base_pri;
885
886 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
887 td->td_base_pri <= PRI_MAX_TIMESHARE)
888 base_pri = td->td_user_pri;
889 else
890 base_pri = td->td_base_pri;
891 if (prio >= base_pri) {
892 td->td_flags &= ~TDF_BORROWING;
893 sched_prio(td, base_pri);
894 } else
895 sched_lend_prio(td, prio);
896 }
897
898 void
sched_prio(struct thread * td,u_char prio)899 sched_prio(struct thread *td, u_char prio)
900 {
901 u_char oldprio;
902
903 /* First, update the base priority. */
904 td->td_base_pri = prio;
905
906 /*
907 * If the thread is borrowing another thread's priority, don't ever
908 * lower the priority.
909 */
910 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
911 return;
912
913 /* Change the real priority. */
914 oldprio = td->td_priority;
915 sched_priority(td, prio);
916
917 /*
918 * If the thread is on a turnstile, then let the turnstile update
919 * its state.
920 */
921 if (TD_ON_LOCK(td) && oldprio != prio)
922 turnstile_adjust(td, oldprio);
923 }
924
925 void
sched_user_prio(struct thread * td,u_char prio)926 sched_user_prio(struct thread *td, u_char prio)
927 {
928
929 THREAD_LOCK_ASSERT(td, MA_OWNED);
930 td->td_base_user_pri = prio;
931 if (td->td_lend_user_pri <= prio)
932 return;
933 td->td_user_pri = prio;
934 }
935
936 void
sched_lend_user_prio(struct thread * td,u_char prio)937 sched_lend_user_prio(struct thread *td, u_char prio)
938 {
939
940 THREAD_LOCK_ASSERT(td, MA_OWNED);
941 td->td_lend_user_pri = prio;
942 td->td_user_pri = min(prio, td->td_base_user_pri);
943 if (td->td_priority > td->td_user_pri)
944 sched_prio(td, td->td_user_pri);
945 else if (td->td_priority != td->td_user_pri)
946 td->td_flags |= TDF_NEEDRESCHED;
947 }
948
949 /*
950 * Like the above but first check if there is anything to do.
951 */
952 void
sched_lend_user_prio_cond(struct thread * td,u_char prio)953 sched_lend_user_prio_cond(struct thread *td, u_char prio)
954 {
955
956 if (td->td_lend_user_pri == prio)
957 return;
958
959 thread_lock(td);
960 sched_lend_user_prio(td, prio);
961 thread_unlock(td);
962 }
963
964 void
sched_sleep(struct thread * td,int pri)965 sched_sleep(struct thread *td, int pri)
966 {
967
968 THREAD_LOCK_ASSERT(td, MA_OWNED);
969 td->td_slptick = ticks;
970 td_get_sched(td)->ts_slptime = 0;
971 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
972 sched_prio(td, pri);
973 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
974 td->td_flags |= TDF_CANSWAP;
975 }
976
977 void
sched_switch(struct thread * td,int flags)978 sched_switch(struct thread *td, int flags)
979 {
980 struct thread *newtd;
981 struct mtx *tmtx;
982 struct td_sched *ts;
983 struct proc *p;
984 int preempted;
985
986 tmtx = &sched_lock;
987 ts = td_get_sched(td);
988 p = td->td_proc;
989
990 THREAD_LOCK_ASSERT(td, MA_OWNED);
991
992 td->td_lastcpu = td->td_oncpu;
993 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
994 (flags & SW_PREEMPT) != 0;
995 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
996 td->td_owepreempt = 0;
997 td->td_oncpu = NOCPU;
998
999 /*
1000 * At the last moment, if this thread is still marked RUNNING,
1001 * then put it back on the run queue as it has not been suspended
1002 * or stopped or any thing else similar. We never put the idle
1003 * threads on the run queue, however.
1004 */
1005 if (td->td_flags & TDF_IDLETD) {
1006 TD_SET_CAN_RUN(td);
1007 #ifdef SMP
1008 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
1009 #endif
1010 } else {
1011 if (TD_IS_RUNNING(td)) {
1012 /* Put us back on the run queue. */
1013 sched_add(td, SRQ_HOLDTD | SRQ_OURSELF | SRQ_YIELDING |
1014 (preempted ? SRQ_PREEMPTED : 0));
1015 }
1016 }
1017
1018 /*
1019 * Switch to the sched lock to fix things up and pick
1020 * a new thread. Block the td_lock in order to avoid
1021 * breaking the critical path.
1022 */
1023 if (td->td_lock != &sched_lock) {
1024 mtx_lock_spin(&sched_lock);
1025 tmtx = thread_lock_block(td);
1026 mtx_unlock_spin(tmtx);
1027 }
1028
1029 if ((td->td_flags & TDF_NOLOAD) == 0)
1030 sched_load_rem();
1031
1032 newtd = choosethread();
1033 MPASS(newtd->td_lock == &sched_lock);
1034
1035 #if (KTR_COMPILE & KTR_SCHED) != 0
1036 if (TD_IS_IDLETHREAD(td))
1037 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1038 "prio:%d", td->td_priority);
1039 else
1040 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1041 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1042 "lockname:\"%s\"", td->td_lockname);
1043 #endif
1044
1045 if (td != newtd) {
1046 #ifdef HWPMC_HOOKS
1047 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1048 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1049 #endif
1050
1051 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1052
1053 /* I feel sleepy */
1054 lock_profile_release_lock(&sched_lock.lock_object, true);
1055 #ifdef KDTRACE_HOOKS
1056 /*
1057 * If DTrace has set the active vtime enum to anything
1058 * other than INACTIVE (0), then it should have set the
1059 * function to call.
1060 */
1061 if (dtrace_vtime_active)
1062 (*dtrace_vtime_switch_func)(newtd);
1063 #endif
1064
1065 cpu_switch(td, newtd, tmtx);
1066 lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1067 0, 0, __FILE__, __LINE__);
1068 /*
1069 * Where am I? What year is it?
1070 * We are in the same thread that went to sleep above,
1071 * but any amount of time may have passed. All our context
1072 * will still be available as will local variables.
1073 * PCPU values however may have changed as we may have
1074 * changed CPU so don't trust cached values of them.
1075 * New threads will go to fork_exit() instead of here
1076 * so if you change things here you may need to change
1077 * things there too.
1078 *
1079 * If the thread above was exiting it will never wake
1080 * up again here, so either it has saved everything it
1081 * needed to, or the thread_wait() or wait() will
1082 * need to reap it.
1083 */
1084
1085 SDT_PROBE0(sched, , , on__cpu);
1086 #ifdef HWPMC_HOOKS
1087 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1088 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1089 #endif
1090 } else {
1091 td->td_lock = &sched_lock;
1092 SDT_PROBE0(sched, , , remain__cpu);
1093 }
1094
1095 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1096 "prio:%d", td->td_priority);
1097
1098 #ifdef SMP
1099 if (td->td_flags & TDF_IDLETD)
1100 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1101 #endif
1102 sched_lock.mtx_lock = (uintptr_t)td;
1103 td->td_oncpu = PCPU_GET(cpuid);
1104 spinlock_enter();
1105 mtx_unlock_spin(&sched_lock);
1106 }
1107
1108 void
sched_wakeup(struct thread * td,int srqflags)1109 sched_wakeup(struct thread *td, int srqflags)
1110 {
1111 struct td_sched *ts;
1112
1113 THREAD_LOCK_ASSERT(td, MA_OWNED);
1114 ts = td_get_sched(td);
1115 td->td_flags &= ~TDF_CANSWAP;
1116 if (ts->ts_slptime > 1) {
1117 updatepri(td);
1118 resetpriority(td);
1119 }
1120 td->td_slptick = 0;
1121 ts->ts_slptime = 0;
1122 ts->ts_slice = sched_slice;
1123 sched_add(td, srqflags);
1124 }
1125
1126 #ifdef SMP
1127 static int
forward_wakeup(int cpunum)1128 forward_wakeup(int cpunum)
1129 {
1130 struct pcpu *pc;
1131 cpuset_t dontuse, map, map2;
1132 u_int id, me;
1133 int iscpuset;
1134
1135 mtx_assert(&sched_lock, MA_OWNED);
1136
1137 CTR0(KTR_RUNQ, "forward_wakeup()");
1138
1139 if ((!forward_wakeup_enabled) ||
1140 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1141 return (0);
1142 if (!smp_started || KERNEL_PANICKED())
1143 return (0);
1144
1145 forward_wakeups_requested++;
1146
1147 /*
1148 * Check the idle mask we received against what we calculated
1149 * before in the old version.
1150 */
1151 me = PCPU_GET(cpuid);
1152
1153 /* Don't bother if we should be doing it ourself. */
1154 if (CPU_ISSET(me, &idle_cpus_mask) &&
1155 (cpunum == NOCPU || me == cpunum))
1156 return (0);
1157
1158 CPU_SETOF(me, &dontuse);
1159 CPU_OR(&dontuse, &dontuse, &stopped_cpus);
1160 CPU_OR(&dontuse, &dontuse, &hlt_cpus_mask);
1161 CPU_ZERO(&map2);
1162 if (forward_wakeup_use_loop) {
1163 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1164 id = pc->pc_cpuid;
1165 if (!CPU_ISSET(id, &dontuse) &&
1166 pc->pc_curthread == pc->pc_idlethread) {
1167 CPU_SET(id, &map2);
1168 }
1169 }
1170 }
1171
1172 if (forward_wakeup_use_mask) {
1173 map = idle_cpus_mask;
1174 CPU_ANDNOT(&map, &map, &dontuse);
1175
1176 /* If they are both on, compare and use loop if different. */
1177 if (forward_wakeup_use_loop) {
1178 if (CPU_CMP(&map, &map2)) {
1179 printf("map != map2, loop method preferred\n");
1180 map = map2;
1181 }
1182 }
1183 } else {
1184 map = map2;
1185 }
1186
1187 /* If we only allow a specific CPU, then mask off all the others. */
1188 if (cpunum != NOCPU) {
1189 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1190 iscpuset = CPU_ISSET(cpunum, &map);
1191 if (iscpuset == 0)
1192 CPU_ZERO(&map);
1193 else
1194 CPU_SETOF(cpunum, &map);
1195 }
1196 if (!CPU_EMPTY(&map)) {
1197 forward_wakeups_delivered++;
1198 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1199 id = pc->pc_cpuid;
1200 if (!CPU_ISSET(id, &map))
1201 continue;
1202 if (cpu_idle_wakeup(pc->pc_cpuid))
1203 CPU_CLR(id, &map);
1204 }
1205 if (!CPU_EMPTY(&map))
1206 ipi_selected(map, IPI_AST);
1207 return (1);
1208 }
1209 if (cpunum == NOCPU)
1210 printf("forward_wakeup: Idle processor not found\n");
1211 return (0);
1212 }
1213
1214 static void
kick_other_cpu(int pri,int cpuid)1215 kick_other_cpu(int pri, int cpuid)
1216 {
1217 struct pcpu *pcpu;
1218 int cpri;
1219
1220 pcpu = pcpu_find(cpuid);
1221 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1222 forward_wakeups_delivered++;
1223 if (!cpu_idle_wakeup(cpuid))
1224 ipi_cpu(cpuid, IPI_AST);
1225 return;
1226 }
1227
1228 cpri = pcpu->pc_curthread->td_priority;
1229 if (pri >= cpri)
1230 return;
1231
1232 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1233 #if !defined(FULL_PREEMPTION)
1234 if (pri <= PRI_MAX_ITHD)
1235 #endif /* ! FULL_PREEMPTION */
1236 {
1237 ipi_cpu(cpuid, IPI_PREEMPT);
1238 return;
1239 }
1240 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1241
1242 if (pcpu->pc_curthread->td_lock == &sched_lock) {
1243 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1244 ipi_cpu(cpuid, IPI_AST);
1245 }
1246 }
1247 #endif /* SMP */
1248
1249 #ifdef SMP
1250 static int
sched_pickcpu(struct thread * td)1251 sched_pickcpu(struct thread *td)
1252 {
1253 int best, cpu;
1254
1255 mtx_assert(&sched_lock, MA_OWNED);
1256
1257 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu))
1258 best = td->td_lastcpu;
1259 else
1260 best = NOCPU;
1261 CPU_FOREACH(cpu) {
1262 if (!THREAD_CAN_SCHED(td, cpu))
1263 continue;
1264
1265 if (best == NOCPU)
1266 best = cpu;
1267 else if (runq_length[cpu] < runq_length[best])
1268 best = cpu;
1269 }
1270 KASSERT(best != NOCPU, ("no valid CPUs"));
1271
1272 return (best);
1273 }
1274 #endif
1275
1276 void
sched_add(struct thread * td,int flags)1277 sched_add(struct thread *td, int flags)
1278 #ifdef SMP
1279 {
1280 cpuset_t tidlemsk;
1281 struct td_sched *ts;
1282 u_int cpu, cpuid;
1283 int forwarded = 0;
1284 int single_cpu = 0;
1285
1286 ts = td_get_sched(td);
1287 THREAD_LOCK_ASSERT(td, MA_OWNED);
1288 KASSERT((td->td_inhibitors == 0),
1289 ("sched_add: trying to run inhibited thread"));
1290 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1291 ("sched_add: bad thread state"));
1292 KASSERT(td->td_flags & TDF_INMEM,
1293 ("sched_add: thread swapped out"));
1294
1295 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1296 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1297 sched_tdname(curthread));
1298 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1299 KTR_ATTR_LINKED, sched_tdname(td));
1300 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1301 flags & SRQ_PREEMPTED);
1302
1303 /*
1304 * Now that the thread is moving to the run-queue, set the lock
1305 * to the scheduler's lock.
1306 */
1307 if (td->td_lock != &sched_lock) {
1308 mtx_lock_spin(&sched_lock);
1309 if ((flags & SRQ_HOLD) != 0)
1310 td->td_lock = &sched_lock;
1311 else
1312 thread_lock_set(td, &sched_lock);
1313 }
1314 TD_SET_RUNQ(td);
1315
1316 /*
1317 * If SMP is started and the thread is pinned or otherwise limited to
1318 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1319 * Otherwise, queue the thread to the global run queue.
1320 *
1321 * If SMP has not yet been started we must use the global run queue
1322 * as per-CPU state may not be initialized yet and we may crash if we
1323 * try to access the per-CPU run queues.
1324 */
1325 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1326 ts->ts_flags & TSF_AFFINITY)) {
1327 if (td->td_pinned != 0)
1328 cpu = td->td_lastcpu;
1329 else if (td->td_flags & TDF_BOUND) {
1330 /* Find CPU from bound runq. */
1331 KASSERT(SKE_RUNQ_PCPU(ts),
1332 ("sched_add: bound td_sched not on cpu runq"));
1333 cpu = ts->ts_runq - &runq_pcpu[0];
1334 } else
1335 /* Find a valid CPU for our cpuset */
1336 cpu = sched_pickcpu(td);
1337 ts->ts_runq = &runq_pcpu[cpu];
1338 single_cpu = 1;
1339 CTR3(KTR_RUNQ,
1340 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1341 cpu);
1342 } else {
1343 CTR2(KTR_RUNQ,
1344 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1345 td);
1346 cpu = NOCPU;
1347 ts->ts_runq = &runq;
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 cpuid = PCPU_GET(cpuid);
1357 if (single_cpu && cpu != cpuid) {
1358 kick_other_cpu(td->td_priority, cpu);
1359 } else {
1360 if (!single_cpu) {
1361 tidlemsk = idle_cpus_mask;
1362 CPU_ANDNOT(&tidlemsk, &tidlemsk, &hlt_cpus_mask);
1363 CPU_CLR(cpuid, &tidlemsk);
1364
1365 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1366 ((flags & SRQ_INTR) == 0) &&
1367 !CPU_EMPTY(&tidlemsk))
1368 forwarded = forward_wakeup(cpu);
1369 }
1370
1371 if (!forwarded) {
1372 if (!maybe_preempt(td))
1373 maybe_resched(td);
1374 }
1375 }
1376 if ((flags & SRQ_HOLDTD) == 0)
1377 thread_unlock(td);
1378 }
1379 #else /* SMP */
1380 {
1381 struct td_sched *ts;
1382
1383 ts = td_get_sched(td);
1384 THREAD_LOCK_ASSERT(td, MA_OWNED);
1385 KASSERT((td->td_inhibitors == 0),
1386 ("sched_add: trying to run inhibited thread"));
1387 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1388 ("sched_add: bad thread state"));
1389 KASSERT(td->td_flags & TDF_INMEM,
1390 ("sched_add: thread swapped out"));
1391 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1392 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1393 sched_tdname(curthread));
1394 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1395 KTR_ATTR_LINKED, sched_tdname(td));
1396 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1397 flags & SRQ_PREEMPTED);
1398
1399 /*
1400 * Now that the thread is moving to the run-queue, set the lock
1401 * to the scheduler's lock.
1402 */
1403 if (td->td_lock != &sched_lock) {
1404 mtx_lock_spin(&sched_lock);
1405 if ((flags & SRQ_HOLD) != 0)
1406 td->td_lock = &sched_lock;
1407 else
1408 thread_lock_set(td, &sched_lock);
1409 }
1410 TD_SET_RUNQ(td);
1411 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1412 ts->ts_runq = &runq;
1413
1414 if ((td->td_flags & TDF_NOLOAD) == 0)
1415 sched_load_add();
1416 runq_add(ts->ts_runq, td, flags);
1417 if (!maybe_preempt(td))
1418 maybe_resched(td);
1419 if ((flags & SRQ_HOLDTD) == 0)
1420 thread_unlock(td);
1421 }
1422 #endif /* SMP */
1423
1424 void
sched_rem(struct thread * td)1425 sched_rem(struct thread *td)
1426 {
1427 struct td_sched *ts;
1428
1429 ts = td_get_sched(td);
1430 KASSERT(td->td_flags & TDF_INMEM,
1431 ("sched_rem: thread swapped out"));
1432 KASSERT(TD_ON_RUNQ(td),
1433 ("sched_rem: thread not on run queue"));
1434 mtx_assert(&sched_lock, MA_OWNED);
1435 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1436 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1437 sched_tdname(curthread));
1438 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1439
1440 if ((td->td_flags & TDF_NOLOAD) == 0)
1441 sched_load_rem();
1442 #ifdef SMP
1443 if (ts->ts_runq != &runq)
1444 runq_length[ts->ts_runq - runq_pcpu]--;
1445 #endif
1446 runq_remove(ts->ts_runq, td);
1447 TD_SET_CAN_RUN(td);
1448 }
1449
1450 /*
1451 * Select threads to run. Note that running threads still consume a
1452 * slot.
1453 */
1454 struct thread *
sched_choose(void)1455 sched_choose(void)
1456 {
1457 struct thread *td;
1458 struct runq *rq;
1459
1460 mtx_assert(&sched_lock, MA_OWNED);
1461 #ifdef SMP
1462 struct thread *tdcpu;
1463
1464 rq = &runq;
1465 td = runq_choose_fuzz(&runq, runq_fuzz);
1466 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1467
1468 if (td == NULL ||
1469 (tdcpu != NULL &&
1470 tdcpu->td_priority < td->td_priority)) {
1471 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1472 PCPU_GET(cpuid));
1473 td = tdcpu;
1474 rq = &runq_pcpu[PCPU_GET(cpuid)];
1475 } else {
1476 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1477 }
1478
1479 #else
1480 rq = &runq;
1481 td = runq_choose(&runq);
1482 #endif
1483
1484 if (td) {
1485 #ifdef SMP
1486 if (td == tdcpu)
1487 runq_length[PCPU_GET(cpuid)]--;
1488 #endif
1489 runq_remove(rq, td);
1490 td->td_flags |= TDF_DIDRUN;
1491
1492 KASSERT(td->td_flags & TDF_INMEM,
1493 ("sched_choose: thread swapped out"));
1494 return (td);
1495 }
1496 return (PCPU_GET(idlethread));
1497 }
1498
1499 void
sched_preempt(struct thread * td)1500 sched_preempt(struct thread *td)
1501 {
1502
1503 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1504 if (td->td_critnest > 1) {
1505 td->td_owepreempt = 1;
1506 } else {
1507 thread_lock(td);
1508 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT);
1509 }
1510 }
1511
1512 void
sched_userret_slowpath(struct thread * td)1513 sched_userret_slowpath(struct thread *td)
1514 {
1515
1516 thread_lock(td);
1517 td->td_priority = td->td_user_pri;
1518 td->td_base_pri = td->td_user_pri;
1519 thread_unlock(td);
1520 }
1521
1522 void
sched_bind(struct thread * td,int cpu)1523 sched_bind(struct thread *td, int cpu)
1524 {
1525 struct td_sched *ts;
1526
1527 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1528 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1529
1530 ts = td_get_sched(td);
1531
1532 td->td_flags |= TDF_BOUND;
1533 #ifdef SMP
1534 ts->ts_runq = &runq_pcpu[cpu];
1535 if (PCPU_GET(cpuid) == cpu)
1536 return;
1537
1538 mi_switch(SW_VOL);
1539 thread_lock(td);
1540 #endif
1541 }
1542
1543 void
sched_unbind(struct thread * td)1544 sched_unbind(struct thread* td)
1545 {
1546 THREAD_LOCK_ASSERT(td, MA_OWNED);
1547 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1548 td->td_flags &= ~TDF_BOUND;
1549 }
1550
1551 int
sched_is_bound(struct thread * td)1552 sched_is_bound(struct thread *td)
1553 {
1554 THREAD_LOCK_ASSERT(td, MA_OWNED);
1555 return (td->td_flags & TDF_BOUND);
1556 }
1557
1558 void
sched_relinquish(struct thread * td)1559 sched_relinquish(struct thread *td)
1560 {
1561 thread_lock(td);
1562 mi_switch(SW_VOL | SWT_RELINQUISH);
1563 }
1564
1565 int
sched_load(void)1566 sched_load(void)
1567 {
1568 return (sched_tdcnt);
1569 }
1570
1571 int
sched_sizeof_proc(void)1572 sched_sizeof_proc(void)
1573 {
1574 return (sizeof(struct proc));
1575 }
1576
1577 int
sched_sizeof_thread(void)1578 sched_sizeof_thread(void)
1579 {
1580 return (sizeof(struct thread) + sizeof(struct td_sched));
1581 }
1582
1583 fixpt_t
sched_pctcpu(struct thread * td)1584 sched_pctcpu(struct thread *td)
1585 {
1586 struct td_sched *ts;
1587
1588 THREAD_LOCK_ASSERT(td, MA_OWNED);
1589 ts = td_get_sched(td);
1590 return (ts->ts_pctcpu);
1591 }
1592
1593 #ifdef RACCT
1594 /*
1595 * Calculates the contribution to the thread cpu usage for the latest
1596 * (unfinished) second.
1597 */
1598 fixpt_t
sched_pctcpu_delta(struct thread * td)1599 sched_pctcpu_delta(struct thread *td)
1600 {
1601 struct td_sched *ts;
1602 fixpt_t delta;
1603 int realstathz;
1604
1605 THREAD_LOCK_ASSERT(td, MA_OWNED);
1606 ts = td_get_sched(td);
1607 delta = 0;
1608 realstathz = stathz ? stathz : hz;
1609 if (ts->ts_cpticks != 0) {
1610 #if (FSHIFT >= CCPU_SHIFT)
1611 delta = (realstathz == 100)
1612 ? ((fixpt_t) ts->ts_cpticks) <<
1613 (FSHIFT - CCPU_SHIFT) :
1614 100 * (((fixpt_t) ts->ts_cpticks)
1615 << (FSHIFT - CCPU_SHIFT)) / realstathz;
1616 #else
1617 delta = ((FSCALE - ccpu) *
1618 (ts->ts_cpticks *
1619 FSCALE / realstathz)) >> FSHIFT;
1620 #endif
1621 }
1622
1623 return (delta);
1624 }
1625 #endif
1626
1627 u_int
sched_estcpu(struct thread * td)1628 sched_estcpu(struct thread *td)
1629 {
1630
1631 return (td_get_sched(td)->ts_estcpu);
1632 }
1633
1634 /*
1635 * The actual idle process.
1636 */
1637 void
sched_idletd(void * dummy)1638 sched_idletd(void *dummy)
1639 {
1640 struct pcpuidlestat *stat;
1641
1642 THREAD_NO_SLEEPING();
1643 stat = DPCPU_PTR(idlestat);
1644 for (;;) {
1645 mtx_assert(&Giant, MA_NOTOWNED);
1646
1647 while (sched_runnable() == 0) {
1648 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1649 stat->idlecalls++;
1650 }
1651
1652 mtx_lock_spin(&sched_lock);
1653 mi_switch(SW_VOL | SWT_IDLE);
1654 }
1655 }
1656
1657 /*
1658 * A CPU is entering for the first time or a thread is exiting.
1659 */
1660 void
sched_throw(struct thread * td)1661 sched_throw(struct thread *td)
1662 {
1663 /*
1664 * Correct spinlock nesting. The idle thread context that we are
1665 * borrowing was created so that it would start out with a single
1666 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1667 * explicitly acquired locks in this function, the nesting count
1668 * is now 2 rather than 1. Since we are nested, calling
1669 * spinlock_exit() will simply adjust the counts without allowing
1670 * spin lock using code to interrupt us.
1671 */
1672 if (td == NULL) {
1673 mtx_lock_spin(&sched_lock);
1674 spinlock_exit();
1675 PCPU_SET(switchtime, cpu_ticks());
1676 PCPU_SET(switchticks, ticks);
1677 } else {
1678 lock_profile_release_lock(&sched_lock.lock_object, true);
1679 MPASS(td->td_lock == &sched_lock);
1680 td->td_lastcpu = td->td_oncpu;
1681 td->td_oncpu = NOCPU;
1682 }
1683 mtx_assert(&sched_lock, MA_OWNED);
1684 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1685 cpu_throw(td, choosethread()); /* doesn't return */
1686 }
1687
1688 void
sched_fork_exit(struct thread * td)1689 sched_fork_exit(struct thread *td)
1690 {
1691
1692 /*
1693 * Finish setting up thread glue so that it begins execution in a
1694 * non-nested critical section with sched_lock held but not recursed.
1695 */
1696 td->td_oncpu = PCPU_GET(cpuid);
1697 sched_lock.mtx_lock = (uintptr_t)td;
1698 lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1699 0, 0, __FILE__, __LINE__);
1700 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1701
1702 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1703 "prio:%d", td->td_priority);
1704 SDT_PROBE0(sched, , , on__cpu);
1705 }
1706
1707 char *
sched_tdname(struct thread * td)1708 sched_tdname(struct thread *td)
1709 {
1710 #ifdef KTR
1711 struct td_sched *ts;
1712
1713 ts = td_get_sched(td);
1714 if (ts->ts_name[0] == '\0')
1715 snprintf(ts->ts_name, sizeof(ts->ts_name),
1716 "%s tid %d", td->td_name, td->td_tid);
1717 return (ts->ts_name);
1718 #else
1719 return (td->td_name);
1720 #endif
1721 }
1722
1723 #ifdef KTR
1724 void
sched_clear_tdname(struct thread * td)1725 sched_clear_tdname(struct thread *td)
1726 {
1727 struct td_sched *ts;
1728
1729 ts = td_get_sched(td);
1730 ts->ts_name[0] = '\0';
1731 }
1732 #endif
1733
1734 void
sched_affinity(struct thread * td)1735 sched_affinity(struct thread *td)
1736 {
1737 #ifdef SMP
1738 struct td_sched *ts;
1739 int cpu;
1740
1741 THREAD_LOCK_ASSERT(td, MA_OWNED);
1742
1743 /*
1744 * Set the TSF_AFFINITY flag if there is at least one CPU this
1745 * thread can't run on.
1746 */
1747 ts = td_get_sched(td);
1748 ts->ts_flags &= ~TSF_AFFINITY;
1749 CPU_FOREACH(cpu) {
1750 if (!THREAD_CAN_SCHED(td, cpu)) {
1751 ts->ts_flags |= TSF_AFFINITY;
1752 break;
1753 }
1754 }
1755
1756 /*
1757 * If this thread can run on all CPUs, nothing else to do.
1758 */
1759 if (!(ts->ts_flags & TSF_AFFINITY))
1760 return;
1761
1762 /* Pinned threads and bound threads should be left alone. */
1763 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1764 return;
1765
1766 switch (td->td_state) {
1767 case TDS_RUNQ:
1768 /*
1769 * If we are on a per-CPU runqueue that is in the set,
1770 * then nothing needs to be done.
1771 */
1772 if (ts->ts_runq != &runq &&
1773 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1774 return;
1775
1776 /* Put this thread on a valid per-CPU runqueue. */
1777 sched_rem(td);
1778 sched_add(td, SRQ_HOLDTD | SRQ_BORING);
1779 break;
1780 case TDS_RUNNING:
1781 /*
1782 * See if our current CPU is in the set. If not, force a
1783 * context switch.
1784 */
1785 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1786 return;
1787
1788 td->td_flags |= TDF_NEEDRESCHED;
1789 if (td != curthread)
1790 ipi_cpu(cpu, IPI_AST);
1791 break;
1792 default:
1793 break;
1794 }
1795 #endif
1796 }
1797