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 ast_sched_locked(curthread, TDA_SCHED);
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 SCHED_STAT_DEFINE(ithread_demotions, "Interrupt thread priority demotions");
705 SCHED_STAT_DEFINE(ithread_preemptions,
706 "Interrupt thread preemptions due to time-sharing");
707
708 /*
709 * We adjust the priority of the current process. The priority of a
710 * process gets worse as it accumulates CPU time. The cpu usage
711 * estimator (ts_estcpu) is increased here. resetpriority() will
712 * compute a different priority each time ts_estcpu increases by
713 * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached). The
714 * cpu usage estimator ramps up quite quickly when the process is
715 * running (linearly), and decays away exponentially, at a rate which
716 * is proportionally slower when the system is busy. The basic
717 * principle is that the system will 90% forget that the process used
718 * a lot of CPU time in 5 * loadav seconds. This causes the system to
719 * favor processes which haven't run much recently, and to round-robin
720 * among other processes.
721 */
722 static void
sched_clock_tick(struct thread * td)723 sched_clock_tick(struct thread *td)
724 {
725 struct pcpuidlestat *stat;
726 struct td_sched *ts;
727
728 THREAD_LOCK_ASSERT(td, MA_OWNED);
729 ts = td_get_sched(td);
730
731 ts->ts_cpticks++;
732 ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1);
733 if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
734 resetpriority(td);
735 resetpriority_thread(td);
736 }
737
738 /*
739 * Force a context switch if the current thread has used up a full
740 * time slice (default is 100ms).
741 */
742 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
743 ts->ts_slice = sched_slice;
744
745 /*
746 * If an ithread uses a full quantum, demote its
747 * priority and preempt it.
748 */
749 if (PRI_BASE(td->td_pri_class) == PRI_ITHD) {
750 SCHED_STAT_INC(ithread_preemptions);
751 td->td_owepreempt = 1;
752 if (td->td_base_pri + RQ_PPQ < PRI_MAX_ITHD) {
753 SCHED_STAT_INC(ithread_demotions);
754 sched_prio(td, td->td_base_pri + RQ_PPQ);
755 }
756 } else {
757 td->td_flags |= TDF_SLICEEND;
758 ast_sched_locked(td, TDA_SCHED);
759 }
760 }
761
762 stat = DPCPU_PTR(idlestat);
763 stat->oldidlecalls = stat->idlecalls;
764 stat->idlecalls = 0;
765 }
766
767 void
sched_clock(struct thread * td,int cnt)768 sched_clock(struct thread *td, int cnt)
769 {
770
771 for ( ; cnt > 0; cnt--)
772 sched_clock_tick(td);
773 }
774
775 /*
776 * Charge child's scheduling CPU usage to parent.
777 */
778 void
sched_exit(struct proc * p,struct thread * td)779 sched_exit(struct proc *p, struct thread *td)
780 {
781
782 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
783 "prio:%d", td->td_priority);
784
785 PROC_LOCK_ASSERT(p, MA_OWNED);
786 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
787 }
788
789 void
sched_exit_thread(struct thread * td,struct thread * child)790 sched_exit_thread(struct thread *td, struct thread *child)
791 {
792
793 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
794 "prio:%d", child->td_priority);
795 thread_lock(td);
796 td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu +
797 td_get_sched(child)->ts_estcpu);
798 thread_unlock(td);
799 thread_lock(child);
800 if ((child->td_flags & TDF_NOLOAD) == 0)
801 sched_load_rem();
802 thread_unlock(child);
803 }
804
805 void
sched_fork(struct thread * td,struct thread * childtd)806 sched_fork(struct thread *td, struct thread *childtd)
807 {
808 sched_fork_thread(td, childtd);
809 }
810
811 void
sched_fork_thread(struct thread * td,struct thread * childtd)812 sched_fork_thread(struct thread *td, struct thread *childtd)
813 {
814 struct td_sched *ts, *tsc;
815
816 childtd->td_oncpu = NOCPU;
817 childtd->td_lastcpu = NOCPU;
818 childtd->td_lock = &sched_lock;
819 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
820 childtd->td_domain.dr_policy = td->td_cpuset->cs_domain;
821 childtd->td_priority = childtd->td_base_pri;
822 ts = td_get_sched(childtd);
823 bzero(ts, sizeof(*ts));
824 tsc = td_get_sched(td);
825 ts->ts_estcpu = tsc->ts_estcpu;
826 ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY);
827 ts->ts_slice = 1;
828 }
829
830 void
sched_nice(struct proc * p,int nice)831 sched_nice(struct proc *p, int nice)
832 {
833 struct thread *td;
834
835 PROC_LOCK_ASSERT(p, MA_OWNED);
836 p->p_nice = nice;
837 FOREACH_THREAD_IN_PROC(p, td) {
838 thread_lock(td);
839 resetpriority(td);
840 resetpriority_thread(td);
841 thread_unlock(td);
842 }
843 }
844
845 void
sched_class(struct thread * td,int class)846 sched_class(struct thread *td, int class)
847 {
848 THREAD_LOCK_ASSERT(td, MA_OWNED);
849 td->td_pri_class = class;
850 }
851
852 /*
853 * Adjust the priority of a thread.
854 */
855 static void
sched_priority(struct thread * td,u_char prio)856 sched_priority(struct thread *td, u_char prio)
857 {
858
859 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
860 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
861 sched_tdname(curthread));
862 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
863 if (td != curthread && prio > td->td_priority) {
864 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
865 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
866 prio, KTR_ATTR_LINKED, sched_tdname(td));
867 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
868 curthread);
869 }
870 THREAD_LOCK_ASSERT(td, MA_OWNED);
871 if (td->td_priority == prio)
872 return;
873 td->td_priority = prio;
874 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
875 sched_rem(td);
876 sched_add(td, SRQ_BORING | SRQ_HOLDTD);
877 }
878 }
879
880 /*
881 * Update a thread's priority when it is lent another thread's
882 * priority.
883 */
884 void
sched_lend_prio(struct thread * td,u_char prio)885 sched_lend_prio(struct thread *td, u_char prio)
886 {
887
888 td->td_flags |= TDF_BORROWING;
889 sched_priority(td, prio);
890 }
891
892 /*
893 * Restore a thread's priority when priority propagation is
894 * over. The prio argument is the minimum priority the thread
895 * needs to have to satisfy other possible priority lending
896 * requests. If the thread's regulary priority is less
897 * important than prio the thread will keep a priority boost
898 * of prio.
899 */
900 void
sched_unlend_prio(struct thread * td,u_char prio)901 sched_unlend_prio(struct thread *td, u_char prio)
902 {
903 u_char base_pri;
904
905 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
906 td->td_base_pri <= PRI_MAX_TIMESHARE)
907 base_pri = td->td_user_pri;
908 else
909 base_pri = td->td_base_pri;
910 if (prio >= base_pri) {
911 td->td_flags &= ~TDF_BORROWING;
912 sched_prio(td, base_pri);
913 } else
914 sched_lend_prio(td, prio);
915 }
916
917 void
sched_prio(struct thread * td,u_char prio)918 sched_prio(struct thread *td, u_char prio)
919 {
920 u_char oldprio;
921
922 /* First, update the base priority. */
923 td->td_base_pri = prio;
924
925 /*
926 * If the thread is borrowing another thread's priority, don't ever
927 * lower the priority.
928 */
929 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
930 return;
931
932 /* Change the real priority. */
933 oldprio = td->td_priority;
934 sched_priority(td, prio);
935
936 /*
937 * If the thread is on a turnstile, then let the turnstile update
938 * its state.
939 */
940 if (TD_ON_LOCK(td) && oldprio != prio)
941 turnstile_adjust(td, oldprio);
942 }
943
944 void
sched_ithread_prio(struct thread * td,u_char prio)945 sched_ithread_prio(struct thread *td, u_char prio)
946 {
947 THREAD_LOCK_ASSERT(td, MA_OWNED);
948 MPASS(td->td_pri_class == PRI_ITHD);
949 td->td_base_ithread_pri = prio;
950 sched_prio(td, prio);
951 }
952
953 void
sched_user_prio(struct thread * td,u_char prio)954 sched_user_prio(struct thread *td, u_char prio)
955 {
956
957 THREAD_LOCK_ASSERT(td, MA_OWNED);
958 td->td_base_user_pri = prio;
959 if (td->td_lend_user_pri <= prio)
960 return;
961 td->td_user_pri = prio;
962 }
963
964 void
sched_lend_user_prio(struct thread * td,u_char prio)965 sched_lend_user_prio(struct thread *td, u_char prio)
966 {
967
968 THREAD_LOCK_ASSERT(td, MA_OWNED);
969 td->td_lend_user_pri = prio;
970 td->td_user_pri = min(prio, td->td_base_user_pri);
971 if (td->td_priority > td->td_user_pri)
972 sched_prio(td, td->td_user_pri);
973 else if (td->td_priority != td->td_user_pri)
974 ast_sched_locked(td, TDA_SCHED);
975 }
976
977 /*
978 * Like the above but first check if there is anything to do.
979 */
980 void
sched_lend_user_prio_cond(struct thread * td,u_char prio)981 sched_lend_user_prio_cond(struct thread *td, u_char prio)
982 {
983
984 if (td->td_lend_user_pri == prio)
985 return;
986
987 thread_lock(td);
988 sched_lend_user_prio(td, prio);
989 thread_unlock(td);
990 }
991
992 void
sched_sleep(struct thread * td,int pri)993 sched_sleep(struct thread *td, int pri)
994 {
995
996 THREAD_LOCK_ASSERT(td, MA_OWNED);
997 td->td_slptick = ticks;
998 td_get_sched(td)->ts_slptime = 0;
999 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
1000 sched_prio(td, pri);
1001 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
1002 td->td_flags |= TDF_CANSWAP;
1003 }
1004
1005 void
sched_switch(struct thread * td,int flags)1006 sched_switch(struct thread *td, int flags)
1007 {
1008 struct thread *newtd;
1009 struct mtx *tmtx;
1010 int preempted;
1011
1012 tmtx = &sched_lock;
1013
1014 THREAD_LOCK_ASSERT(td, MA_OWNED);
1015
1016 td->td_lastcpu = td->td_oncpu;
1017 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
1018 (flags & SW_PREEMPT) != 0;
1019 td->td_flags &= ~TDF_SLICEEND;
1020 ast_unsched_locked(td, TDA_SCHED);
1021 td->td_owepreempt = 0;
1022 td->td_oncpu = NOCPU;
1023
1024 /*
1025 * At the last moment, if this thread is still marked RUNNING,
1026 * then put it back on the run queue as it has not been suspended
1027 * or stopped or any thing else similar. We never put the idle
1028 * threads on the run queue, however.
1029 */
1030 if (td->td_flags & TDF_IDLETD) {
1031 TD_SET_CAN_RUN(td);
1032 #ifdef SMP
1033 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
1034 #endif
1035 } else {
1036 if (TD_IS_RUNNING(td)) {
1037 /* Put us back on the run queue. */
1038 sched_add(td, SRQ_HOLDTD | SRQ_OURSELF | SRQ_YIELDING |
1039 (preempted ? SRQ_PREEMPTED : 0));
1040 }
1041 }
1042
1043 /*
1044 * Switch to the sched lock to fix things up and pick
1045 * a new thread. Block the td_lock in order to avoid
1046 * breaking the critical path.
1047 */
1048 if (td->td_lock != &sched_lock) {
1049 mtx_lock_spin(&sched_lock);
1050 tmtx = thread_lock_block(td);
1051 mtx_unlock_spin(tmtx);
1052 }
1053
1054 if ((td->td_flags & TDF_NOLOAD) == 0)
1055 sched_load_rem();
1056
1057 newtd = choosethread();
1058 MPASS(newtd->td_lock == &sched_lock);
1059
1060 #if (KTR_COMPILE & KTR_SCHED) != 0
1061 if (TD_IS_IDLETHREAD(td))
1062 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1063 "prio:%d", td->td_priority);
1064 else
1065 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1066 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1067 "lockname:\"%s\"", td->td_lockname);
1068 #endif
1069
1070 if (td != newtd) {
1071 #ifdef HWPMC_HOOKS
1072 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1073 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1074 #endif
1075
1076 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1077
1078 /* I feel sleepy */
1079 lock_profile_release_lock(&sched_lock.lock_object, true);
1080 #ifdef KDTRACE_HOOKS
1081 /*
1082 * If DTrace has set the active vtime enum to anything
1083 * other than INACTIVE (0), then it should have set the
1084 * function to call.
1085 */
1086 if (dtrace_vtime_active)
1087 (*dtrace_vtime_switch_func)(newtd);
1088 #endif
1089
1090 cpu_switch(td, newtd, tmtx);
1091 lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1092 0, 0, __FILE__, __LINE__);
1093 /*
1094 * Where am I? What year is it?
1095 * We are in the same thread that went to sleep above,
1096 * but any amount of time may have passed. All our context
1097 * will still be available as will local variables.
1098 * PCPU values however may have changed as we may have
1099 * changed CPU so don't trust cached values of them.
1100 * New threads will go to fork_exit() instead of here
1101 * so if you change things here you may need to change
1102 * things there too.
1103 *
1104 * If the thread above was exiting it will never wake
1105 * up again here, so either it has saved everything it
1106 * needed to, or the thread_wait() or wait() will
1107 * need to reap it.
1108 */
1109
1110 SDT_PROBE0(sched, , , on__cpu);
1111 #ifdef HWPMC_HOOKS
1112 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1113 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1114 #endif
1115 } else {
1116 td->td_lock = &sched_lock;
1117 SDT_PROBE0(sched, , , remain__cpu);
1118 }
1119
1120 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1121 "prio:%d", td->td_priority);
1122
1123 #ifdef SMP
1124 if (td->td_flags & TDF_IDLETD)
1125 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1126 #endif
1127 sched_lock.mtx_lock = (uintptr_t)td;
1128 td->td_oncpu = PCPU_GET(cpuid);
1129 spinlock_enter();
1130 mtx_unlock_spin(&sched_lock);
1131 }
1132
1133 void
sched_wakeup(struct thread * td,int srqflags)1134 sched_wakeup(struct thread *td, int srqflags)
1135 {
1136 struct td_sched *ts;
1137
1138 THREAD_LOCK_ASSERT(td, MA_OWNED);
1139 ts = td_get_sched(td);
1140 td->td_flags &= ~TDF_CANSWAP;
1141 if (ts->ts_slptime > 1) {
1142 updatepri(td);
1143 resetpriority(td);
1144 }
1145 td->td_slptick = 0;
1146 ts->ts_slptime = 0;
1147 ts->ts_slice = sched_slice;
1148
1149 /*
1150 * When resuming an idle ithread, restore its base ithread
1151 * priority.
1152 */
1153 if (PRI_BASE(td->td_pri_class) == PRI_ITHD &&
1154 td->td_base_pri != td->td_base_ithread_pri)
1155 sched_prio(td, td->td_base_ithread_pri);
1156
1157 sched_add(td, srqflags);
1158 }
1159
1160 #ifdef SMP
1161 static int
forward_wakeup(int cpunum)1162 forward_wakeup(int cpunum)
1163 {
1164 struct pcpu *pc;
1165 cpuset_t dontuse, map, map2;
1166 u_int id, me;
1167 int iscpuset;
1168
1169 mtx_assert(&sched_lock, MA_OWNED);
1170
1171 CTR0(KTR_RUNQ, "forward_wakeup()");
1172
1173 if ((!forward_wakeup_enabled) ||
1174 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1175 return (0);
1176 if (!smp_started || KERNEL_PANICKED())
1177 return (0);
1178
1179 forward_wakeups_requested++;
1180
1181 /*
1182 * Check the idle mask we received against what we calculated
1183 * before in the old version.
1184 */
1185 me = PCPU_GET(cpuid);
1186
1187 /* Don't bother if we should be doing it ourself. */
1188 if (CPU_ISSET(me, &idle_cpus_mask) &&
1189 (cpunum == NOCPU || me == cpunum))
1190 return (0);
1191
1192 CPU_SETOF(me, &dontuse);
1193 CPU_OR(&dontuse, &dontuse, &stopped_cpus);
1194 CPU_OR(&dontuse, &dontuse, &hlt_cpus_mask);
1195 CPU_ZERO(&map2);
1196 if (forward_wakeup_use_loop) {
1197 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1198 id = pc->pc_cpuid;
1199 if (!CPU_ISSET(id, &dontuse) &&
1200 pc->pc_curthread == pc->pc_idlethread) {
1201 CPU_SET(id, &map2);
1202 }
1203 }
1204 }
1205
1206 if (forward_wakeup_use_mask) {
1207 map = idle_cpus_mask;
1208 CPU_ANDNOT(&map, &map, &dontuse);
1209
1210 /* If they are both on, compare and use loop if different. */
1211 if (forward_wakeup_use_loop) {
1212 if (CPU_CMP(&map, &map2)) {
1213 printf("map != map2, loop method preferred\n");
1214 map = map2;
1215 }
1216 }
1217 } else {
1218 map = map2;
1219 }
1220
1221 /* If we only allow a specific CPU, then mask off all the others. */
1222 if (cpunum != NOCPU) {
1223 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1224 iscpuset = CPU_ISSET(cpunum, &map);
1225 if (iscpuset == 0)
1226 CPU_ZERO(&map);
1227 else
1228 CPU_SETOF(cpunum, &map);
1229 }
1230 if (!CPU_EMPTY(&map)) {
1231 forward_wakeups_delivered++;
1232 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1233 id = pc->pc_cpuid;
1234 if (!CPU_ISSET(id, &map))
1235 continue;
1236 if (cpu_idle_wakeup(pc->pc_cpuid))
1237 CPU_CLR(id, &map);
1238 }
1239 if (!CPU_EMPTY(&map))
1240 ipi_selected(map, IPI_AST);
1241 return (1);
1242 }
1243 if (cpunum == NOCPU)
1244 printf("forward_wakeup: Idle processor not found\n");
1245 return (0);
1246 }
1247
1248 static void
kick_other_cpu(int pri,int cpuid)1249 kick_other_cpu(int pri, int cpuid)
1250 {
1251 struct pcpu *pcpu;
1252 int cpri;
1253
1254 pcpu = pcpu_find(cpuid);
1255 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1256 forward_wakeups_delivered++;
1257 if (!cpu_idle_wakeup(cpuid))
1258 ipi_cpu(cpuid, IPI_AST);
1259 return;
1260 }
1261
1262 cpri = pcpu->pc_curthread->td_priority;
1263 if (pri >= cpri)
1264 return;
1265
1266 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1267 #if !defined(FULL_PREEMPTION)
1268 if (pri <= PRI_MAX_ITHD)
1269 #endif /* ! FULL_PREEMPTION */
1270 {
1271 ipi_cpu(cpuid, IPI_PREEMPT);
1272 return;
1273 }
1274 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1275
1276 if (pcpu->pc_curthread->td_lock == &sched_lock) {
1277 ast_sched_locked(pcpu->pc_curthread, TDA_SCHED);
1278 ipi_cpu(cpuid, IPI_AST);
1279 }
1280 }
1281 #endif /* SMP */
1282
1283 #ifdef SMP
1284 static int
sched_pickcpu(struct thread * td)1285 sched_pickcpu(struct thread *td)
1286 {
1287 int best, cpu;
1288
1289 mtx_assert(&sched_lock, MA_OWNED);
1290
1291 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu))
1292 best = td->td_lastcpu;
1293 else
1294 best = NOCPU;
1295 CPU_FOREACH(cpu) {
1296 if (!THREAD_CAN_SCHED(td, cpu))
1297 continue;
1298
1299 if (best == NOCPU)
1300 best = cpu;
1301 else if (runq_length[cpu] < runq_length[best])
1302 best = cpu;
1303 }
1304 KASSERT(best != NOCPU, ("no valid CPUs"));
1305
1306 return (best);
1307 }
1308 #endif
1309
1310 void
sched_add(struct thread * td,int flags)1311 sched_add(struct thread *td, int flags)
1312 #ifdef SMP
1313 {
1314 cpuset_t tidlemsk;
1315 struct td_sched *ts;
1316 u_int cpu, cpuid;
1317 int forwarded = 0;
1318 int single_cpu = 0;
1319
1320 ts = td_get_sched(td);
1321 THREAD_LOCK_ASSERT(td, MA_OWNED);
1322 KASSERT((td->td_inhibitors == 0),
1323 ("sched_add: trying to run inhibited thread"));
1324 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1325 ("sched_add: bad thread state"));
1326 KASSERT(td->td_flags & TDF_INMEM,
1327 ("sched_add: thread swapped out"));
1328
1329 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1330 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1331 sched_tdname(curthread));
1332 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1333 KTR_ATTR_LINKED, sched_tdname(td));
1334 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1335 flags & SRQ_PREEMPTED);
1336
1337 /*
1338 * Now that the thread is moving to the run-queue, set the lock
1339 * to the scheduler's lock.
1340 */
1341 if (td->td_lock != &sched_lock) {
1342 mtx_lock_spin(&sched_lock);
1343 if ((flags & SRQ_HOLD) != 0)
1344 td->td_lock = &sched_lock;
1345 else
1346 thread_lock_set(td, &sched_lock);
1347 }
1348 TD_SET_RUNQ(td);
1349
1350 /*
1351 * If SMP is started and the thread is pinned or otherwise limited to
1352 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1353 * Otherwise, queue the thread to the global run queue.
1354 *
1355 * If SMP has not yet been started we must use the global run queue
1356 * as per-CPU state may not be initialized yet and we may crash if we
1357 * try to access the per-CPU run queues.
1358 */
1359 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1360 ts->ts_flags & TSF_AFFINITY)) {
1361 if (td->td_pinned != 0)
1362 cpu = td->td_lastcpu;
1363 else if (td->td_flags & TDF_BOUND) {
1364 /* Find CPU from bound runq. */
1365 KASSERT(SKE_RUNQ_PCPU(ts),
1366 ("sched_add: bound td_sched not on cpu runq"));
1367 cpu = ts->ts_runq - &runq_pcpu[0];
1368 } else
1369 /* Find a valid CPU for our cpuset */
1370 cpu = sched_pickcpu(td);
1371 ts->ts_runq = &runq_pcpu[cpu];
1372 single_cpu = 1;
1373 CTR3(KTR_RUNQ,
1374 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1375 cpu);
1376 } else {
1377 CTR2(KTR_RUNQ,
1378 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1379 td);
1380 cpu = NOCPU;
1381 ts->ts_runq = &runq;
1382 }
1383
1384 if ((td->td_flags & TDF_NOLOAD) == 0)
1385 sched_load_add();
1386 runq_add(ts->ts_runq, td, flags);
1387 if (cpu != NOCPU)
1388 runq_length[cpu]++;
1389
1390 cpuid = PCPU_GET(cpuid);
1391 if (single_cpu && cpu != cpuid) {
1392 kick_other_cpu(td->td_priority, cpu);
1393 } else {
1394 if (!single_cpu) {
1395 tidlemsk = idle_cpus_mask;
1396 CPU_ANDNOT(&tidlemsk, &tidlemsk, &hlt_cpus_mask);
1397 CPU_CLR(cpuid, &tidlemsk);
1398
1399 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1400 ((flags & SRQ_INTR) == 0) &&
1401 !CPU_EMPTY(&tidlemsk))
1402 forwarded = forward_wakeup(cpu);
1403 }
1404
1405 if (!forwarded) {
1406 if (!maybe_preempt(td))
1407 maybe_resched(td);
1408 }
1409 }
1410 if ((flags & SRQ_HOLDTD) == 0)
1411 thread_unlock(td);
1412 }
1413 #else /* SMP */
1414 {
1415 struct td_sched *ts;
1416
1417 ts = td_get_sched(td);
1418 THREAD_LOCK_ASSERT(td, MA_OWNED);
1419 KASSERT((td->td_inhibitors == 0),
1420 ("sched_add: trying to run inhibited thread"));
1421 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1422 ("sched_add: bad thread state"));
1423 KASSERT(td->td_flags & TDF_INMEM,
1424 ("sched_add: thread swapped out"));
1425 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1426 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1427 sched_tdname(curthread));
1428 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1429 KTR_ATTR_LINKED, sched_tdname(td));
1430 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1431 flags & SRQ_PREEMPTED);
1432
1433 /*
1434 * Now that the thread is moving to the run-queue, set the lock
1435 * to the scheduler's lock.
1436 */
1437 if (td->td_lock != &sched_lock) {
1438 mtx_lock_spin(&sched_lock);
1439 if ((flags & SRQ_HOLD) != 0)
1440 td->td_lock = &sched_lock;
1441 else
1442 thread_lock_set(td, &sched_lock);
1443 }
1444 TD_SET_RUNQ(td);
1445 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1446 ts->ts_runq = &runq;
1447
1448 if ((td->td_flags & TDF_NOLOAD) == 0)
1449 sched_load_add();
1450 runq_add(ts->ts_runq, td, flags);
1451 if (!maybe_preempt(td))
1452 maybe_resched(td);
1453 if ((flags & SRQ_HOLDTD) == 0)
1454 thread_unlock(td);
1455 }
1456 #endif /* SMP */
1457
1458 void
sched_rem(struct thread * td)1459 sched_rem(struct thread *td)
1460 {
1461 struct td_sched *ts;
1462
1463 ts = td_get_sched(td);
1464 KASSERT(td->td_flags & TDF_INMEM,
1465 ("sched_rem: thread swapped out"));
1466 KASSERT(TD_ON_RUNQ(td),
1467 ("sched_rem: thread not on run queue"));
1468 mtx_assert(&sched_lock, MA_OWNED);
1469 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1470 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1471 sched_tdname(curthread));
1472 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1473
1474 if ((td->td_flags & TDF_NOLOAD) == 0)
1475 sched_load_rem();
1476 #ifdef SMP
1477 if (ts->ts_runq != &runq)
1478 runq_length[ts->ts_runq - runq_pcpu]--;
1479 #endif
1480 runq_remove(ts->ts_runq, td);
1481 TD_SET_CAN_RUN(td);
1482 }
1483
1484 /*
1485 * Select threads to run. Note that running threads still consume a
1486 * slot.
1487 */
1488 struct thread *
sched_choose(void)1489 sched_choose(void)
1490 {
1491 struct thread *td;
1492 struct runq *rq;
1493
1494 mtx_assert(&sched_lock, MA_OWNED);
1495 #ifdef SMP
1496 struct thread *tdcpu;
1497
1498 rq = &runq;
1499 td = runq_choose_fuzz(&runq, runq_fuzz);
1500 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1501
1502 if (td == NULL ||
1503 (tdcpu != NULL &&
1504 tdcpu->td_priority < td->td_priority)) {
1505 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1506 PCPU_GET(cpuid));
1507 td = tdcpu;
1508 rq = &runq_pcpu[PCPU_GET(cpuid)];
1509 } else {
1510 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1511 }
1512
1513 #else
1514 rq = &runq;
1515 td = runq_choose(&runq);
1516 #endif
1517
1518 if (td) {
1519 #ifdef SMP
1520 if (td == tdcpu)
1521 runq_length[PCPU_GET(cpuid)]--;
1522 #endif
1523 runq_remove(rq, td);
1524 td->td_flags |= TDF_DIDRUN;
1525
1526 KASSERT(td->td_flags & TDF_INMEM,
1527 ("sched_choose: thread swapped out"));
1528 return (td);
1529 }
1530 return (PCPU_GET(idlethread));
1531 }
1532
1533 void
sched_preempt(struct thread * td)1534 sched_preempt(struct thread *td)
1535 {
1536 int flags;
1537
1538 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1539 if (td->td_critnest > 1) {
1540 td->td_owepreempt = 1;
1541 } else {
1542 thread_lock(td);
1543 flags = SW_INVOL | SW_PREEMPT;
1544 flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
1545 SWT_REMOTEPREEMPT;
1546 mi_switch(flags);
1547 }
1548 }
1549
1550 void
sched_userret_slowpath(struct thread * td)1551 sched_userret_slowpath(struct thread *td)
1552 {
1553
1554 thread_lock(td);
1555 td->td_priority = td->td_user_pri;
1556 td->td_base_pri = td->td_user_pri;
1557 thread_unlock(td);
1558 }
1559
1560 void
sched_bind(struct thread * td,int cpu)1561 sched_bind(struct thread *td, int cpu)
1562 {
1563 #ifdef SMP
1564 struct td_sched *ts = td_get_sched(td);
1565 #endif
1566
1567 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1568 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1569
1570 td->td_flags |= TDF_BOUND;
1571 #ifdef SMP
1572 ts->ts_runq = &runq_pcpu[cpu];
1573 if (PCPU_GET(cpuid) == cpu)
1574 return;
1575
1576 mi_switch(SW_VOL | SWT_BIND);
1577 thread_lock(td);
1578 #endif
1579 }
1580
1581 void
sched_unbind(struct thread * td)1582 sched_unbind(struct thread* td)
1583 {
1584 THREAD_LOCK_ASSERT(td, MA_OWNED);
1585 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1586 td->td_flags &= ~TDF_BOUND;
1587 }
1588
1589 int
sched_is_bound(struct thread * td)1590 sched_is_bound(struct thread *td)
1591 {
1592 THREAD_LOCK_ASSERT(td, MA_OWNED);
1593 return (td->td_flags & TDF_BOUND);
1594 }
1595
1596 void
sched_relinquish(struct thread * td)1597 sched_relinquish(struct thread *td)
1598 {
1599 thread_lock(td);
1600 mi_switch(SW_VOL | SWT_RELINQUISH);
1601 }
1602
1603 int
sched_load(void)1604 sched_load(void)
1605 {
1606 return (sched_tdcnt);
1607 }
1608
1609 int
sched_sizeof_proc(void)1610 sched_sizeof_proc(void)
1611 {
1612 return (sizeof(struct proc));
1613 }
1614
1615 int
sched_sizeof_thread(void)1616 sched_sizeof_thread(void)
1617 {
1618 return (sizeof(struct thread) + sizeof(struct td_sched));
1619 }
1620
1621 fixpt_t
sched_pctcpu(struct thread * td)1622 sched_pctcpu(struct thread *td)
1623 {
1624 struct td_sched *ts;
1625
1626 THREAD_LOCK_ASSERT(td, MA_OWNED);
1627 ts = td_get_sched(td);
1628 return (ts->ts_pctcpu);
1629 }
1630
1631 #ifdef RACCT
1632 /*
1633 * Calculates the contribution to the thread cpu usage for the latest
1634 * (unfinished) second.
1635 */
1636 fixpt_t
sched_pctcpu_delta(struct thread * td)1637 sched_pctcpu_delta(struct thread *td)
1638 {
1639 struct td_sched *ts;
1640 fixpt_t delta;
1641 int realstathz;
1642
1643 THREAD_LOCK_ASSERT(td, MA_OWNED);
1644 ts = td_get_sched(td);
1645 delta = 0;
1646 realstathz = stathz ? stathz : hz;
1647 if (ts->ts_cpticks != 0) {
1648 #if (FSHIFT >= CCPU_SHIFT)
1649 delta = (realstathz == 100)
1650 ? ((fixpt_t) ts->ts_cpticks) <<
1651 (FSHIFT - CCPU_SHIFT) :
1652 100 * (((fixpt_t) ts->ts_cpticks)
1653 << (FSHIFT - CCPU_SHIFT)) / realstathz;
1654 #else
1655 delta = ((FSCALE - ccpu) *
1656 (ts->ts_cpticks *
1657 FSCALE / realstathz)) >> FSHIFT;
1658 #endif
1659 }
1660
1661 return (delta);
1662 }
1663 #endif
1664
1665 u_int
sched_estcpu(struct thread * td)1666 sched_estcpu(struct thread *td)
1667 {
1668
1669 return (td_get_sched(td)->ts_estcpu);
1670 }
1671
1672 /*
1673 * The actual idle process.
1674 */
1675 void
sched_idletd(void * dummy)1676 sched_idletd(void *dummy)
1677 {
1678 struct pcpuidlestat *stat;
1679
1680 THREAD_NO_SLEEPING();
1681 stat = DPCPU_PTR(idlestat);
1682 for (;;) {
1683 mtx_assert(&Giant, MA_NOTOWNED);
1684
1685 while (sched_runnable() == 0) {
1686 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1687 stat->idlecalls++;
1688 }
1689
1690 mtx_lock_spin(&sched_lock);
1691 mi_switch(SW_VOL | SWT_IDLE);
1692 }
1693 }
1694
1695 static void
sched_throw_tail(struct thread * td)1696 sched_throw_tail(struct thread *td)
1697 {
1698
1699 mtx_assert(&sched_lock, MA_OWNED);
1700 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1701 cpu_throw(td, choosethread()); /* doesn't return */
1702 }
1703
1704 /*
1705 * A CPU is entering for the first time.
1706 */
1707 void
sched_ap_entry(void)1708 sched_ap_entry(void)
1709 {
1710
1711 /*
1712 * Correct spinlock nesting. The idle thread context that we are
1713 * borrowing was created so that it would start out with a single
1714 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1715 * explicitly acquired locks in this function, the nesting count
1716 * is now 2 rather than 1. Since we are nested, calling
1717 * spinlock_exit() will simply adjust the counts without allowing
1718 * spin lock using code to interrupt us.
1719 */
1720 mtx_lock_spin(&sched_lock);
1721 spinlock_exit();
1722 PCPU_SET(switchtime, cpu_ticks());
1723 PCPU_SET(switchticks, ticks);
1724
1725 sched_throw_tail(NULL);
1726 }
1727
1728 /*
1729 * A thread is exiting.
1730 */
1731 void
sched_throw(struct thread * td)1732 sched_throw(struct thread *td)
1733 {
1734
1735 MPASS(td != NULL);
1736 MPASS(td->td_lock == &sched_lock);
1737
1738 lock_profile_release_lock(&sched_lock.lock_object, true);
1739 td->td_lastcpu = td->td_oncpu;
1740 td->td_oncpu = NOCPU;
1741
1742 sched_throw_tail(td);
1743 }
1744
1745 void
sched_fork_exit(struct thread * td)1746 sched_fork_exit(struct thread *td)
1747 {
1748
1749 /*
1750 * Finish setting up thread glue so that it begins execution in a
1751 * non-nested critical section with sched_lock held but not recursed.
1752 */
1753 td->td_oncpu = PCPU_GET(cpuid);
1754 sched_lock.mtx_lock = (uintptr_t)td;
1755 lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1756 0, 0, __FILE__, __LINE__);
1757 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1758
1759 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1760 "prio:%d", td->td_priority);
1761 SDT_PROBE0(sched, , , on__cpu);
1762 }
1763
1764 char *
sched_tdname(struct thread * td)1765 sched_tdname(struct thread *td)
1766 {
1767 #ifdef KTR
1768 struct td_sched *ts;
1769
1770 ts = td_get_sched(td);
1771 if (ts->ts_name[0] == '\0')
1772 snprintf(ts->ts_name, sizeof(ts->ts_name),
1773 "%s tid %d", td->td_name, td->td_tid);
1774 return (ts->ts_name);
1775 #else
1776 return (td->td_name);
1777 #endif
1778 }
1779
1780 #ifdef KTR
1781 void
sched_clear_tdname(struct thread * td)1782 sched_clear_tdname(struct thread *td)
1783 {
1784 struct td_sched *ts;
1785
1786 ts = td_get_sched(td);
1787 ts->ts_name[0] = '\0';
1788 }
1789 #endif
1790
1791 void
sched_affinity(struct thread * td)1792 sched_affinity(struct thread *td)
1793 {
1794 #ifdef SMP
1795 struct td_sched *ts;
1796 int cpu;
1797
1798 THREAD_LOCK_ASSERT(td, MA_OWNED);
1799
1800 /*
1801 * Set the TSF_AFFINITY flag if there is at least one CPU this
1802 * thread can't run on.
1803 */
1804 ts = td_get_sched(td);
1805 ts->ts_flags &= ~TSF_AFFINITY;
1806 CPU_FOREACH(cpu) {
1807 if (!THREAD_CAN_SCHED(td, cpu)) {
1808 ts->ts_flags |= TSF_AFFINITY;
1809 break;
1810 }
1811 }
1812
1813 /*
1814 * If this thread can run on all CPUs, nothing else to do.
1815 */
1816 if (!(ts->ts_flags & TSF_AFFINITY))
1817 return;
1818
1819 /* Pinned threads and bound threads should be left alone. */
1820 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1821 return;
1822
1823 switch (TD_GET_STATE(td)) {
1824 case TDS_RUNQ:
1825 /*
1826 * If we are on a per-CPU runqueue that is in the set,
1827 * then nothing needs to be done.
1828 */
1829 if (ts->ts_runq != &runq &&
1830 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1831 return;
1832
1833 /* Put this thread on a valid per-CPU runqueue. */
1834 sched_rem(td);
1835 sched_add(td, SRQ_HOLDTD | SRQ_BORING);
1836 break;
1837 case TDS_RUNNING:
1838 /*
1839 * See if our current CPU is in the set. If not, force a
1840 * context switch.
1841 */
1842 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1843 return;
1844
1845 ast_sched_locked(td, TDA_SCHED);
1846 if (td != curthread)
1847 ipi_cpu(cpu, IPI_AST);
1848 break;
1849 default:
1850 break;
1851 }
1852 #endif
1853 }
1854