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 "opt_hwpmc_hooks.h"
38 #include "opt_hwt_hooks.h"
39 #include "opt_sched.h"
40
41 #include <sys/systm.h>
42 #include <sys/cpuset.h>
43 #include <sys/kernel.h>
44 #include <sys/ktr.h>
45 #include <sys/lock.h>
46 #include <sys/kthread.h>
47 #include <sys/mutex.h>
48 #include <sys/proc.h>
49 #include <sys/resourcevar.h>
50 #include <sys/runq.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 HWT_HOOKS
66 #include <dev/hwt/hwt_hook.h>
67 #endif
68
69 /*
70 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
71 * the range 100-256 Hz (approximately).
72 */
73 #ifdef SMP
74 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
75 #else
76 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
77 #endif
78 #define NICE_WEIGHT 1 /* Priorities per nice level. */
79 #define ESTCPULIM(e) \
80 min((e), INVERSE_ESTCPU_WEIGHT * \
81 (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) + \
82 PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) \
83 + INVERSE_ESTCPU_WEIGHT - 1)
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 #define TDP_RESCHED TDP_SCHED1 /* Reschedule due to maybe_resched(). */
112
113 /* flags kept in ts_flags */
114 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
115
116 #define SKE_RUNQ_PCPU(ts) \
117 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
118
119 #define THREAD_CAN_SCHED(td, cpu) \
120 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
121
122 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
123 sizeof(struct thread0_storage),
124 "increase struct thread0_storage.t0st_sched size");
125
126 static struct mtx sched_lock;
127
128 static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */
129 static int sched_tdcnt; /* Total runnable threads in the system. */
130 static int sched_slice = 12; /* Thread run time before rescheduling. */
131
132 static void setup_runqs(void);
133 static void schedcpu(void);
134 static void schedcpu_thread(void);
135 static void sched_priority(struct thread *td, u_char prio);
136 static void maybe_resched(struct thread *td);
137 static void updatepri(struct thread *td);
138 static void resetpriority(struct thread *td);
139 static void resetpriority_thread(struct thread *td);
140 #ifdef SMP
141 static int sched_pickcpu(struct thread *td);
142 static int forward_wakeup(int cpunum);
143 static void kick_other_cpu(int pri, int cpuid);
144 #endif
145
146 static struct kproc_desc sched_kp = {
147 "schedcpu",
148 schedcpu_thread,
149 NULL
150 };
151
152 static void
sched_4bsd_schedcpu(void)153 sched_4bsd_schedcpu(void)
154 {
155 kproc_start(&sched_kp);
156 }
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_4bsd_quantum(SYSCTL_HANDLER_ARGS)193 sysctl_kern_4bsd_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_sched, OID_AUTO, 4bsd, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
211 "4BSD Scheduler");
212
213 SYSCTL_PROC(_kern_sched_4bsd, OID_AUTO, quantum,
214 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
215 sysctl_kern_4bsd_quantum, "I",
216 "Quantum for timeshare threads in microseconds");
217 SYSCTL_INT(_kern_sched_4bsd, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
218 "Quantum for timeshare threads in stathz ticks");
219 #ifdef SMP
220 /* Enable forwarding of wakeups to all other cpus */
221 static SYSCTL_NODE(_kern_sched_4bsd, OID_AUTO, ipiwakeup,
222 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
223 "Kernel SMP");
224
225 static int runq_fuzz = 1;
226 SYSCTL_INT(_kern_sched_4bsd, OID_AUTO, runq_fuzz, CTLFLAG_RW,
227 &runq_fuzz, 0, "");
228
229 static int forward_wakeup_enabled = 1;
230 SYSCTL_INT(_kern_sched_4bsd_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
231 &forward_wakeup_enabled, 0,
232 "Forwarding of wakeup to idle CPUs");
233
234 static int forward_wakeups_requested = 0;
235 SYSCTL_INT(_kern_sched_4bsd_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
236 &forward_wakeups_requested, 0,
237 "Requests for Forwarding of wakeup to idle CPUs");
238
239 static int forward_wakeups_delivered = 0;
240 SYSCTL_INT(_kern_sched_4bsd_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
241 &forward_wakeups_delivered, 0,
242 "Completed Forwarding of wakeup to idle CPUs");
243
244 static int forward_wakeup_use_mask = 1;
245 SYSCTL_INT(_kern_sched_4bsd_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
246 &forward_wakeup_use_mask, 0,
247 "Use the mask of idle cpus");
248
249 static int forward_wakeup_use_loop = 0;
250 SYSCTL_INT(_kern_sched_4bsd_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
251 &forward_wakeup_use_loop, 0,
252 "Use a loop to find idle cpus");
253
254 #endif
255 #if 0
256 static int sched_followon = 0;
257 SYSCTL_INT(_kern_sched_4bsd, OID_AUTO, followon, CTLFLAG_RW,
258 &sched_followon, 0,
259 "allow threads to share a quantum");
260 #endif
261
262 static __inline void
sched_load_add(void)263 sched_load_add(void)
264 {
265
266 sched_tdcnt++;
267 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
268 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
269 }
270
271 static __inline void
sched_load_rem(void)272 sched_load_rem(void)
273 {
274
275 sched_tdcnt--;
276 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
277 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
278 }
279
280 static void
maybe_resched_ast(struct thread * td,int tda)281 maybe_resched_ast(struct thread *td, int tda)
282 {
283 MPASS(td == curthread); /* We are AST */
284 if ((td->td_pflags & TDP_RESCHED) != 0) {
285 td->td_pflags &= ~TDP_RESCHED;
286 ast_scheduler(td, tda);
287 }
288 }
289
290 /*
291 * Arrange to reschedule if necessary, taking the priorities and
292 * schedulers into account.
293 */
294 static void
maybe_resched(struct thread * td)295 maybe_resched(struct thread *td)
296 {
297 struct thread *ctd;
298
299 ctd = curthread;
300 THREAD_LOCK_ASSERT(td, MA_OWNED);
301 if (td->td_priority < ctd->td_priority)
302 ctd->td_pflags |= TDP_RESCHED;
303 }
304
305 /*
306 * This function is called when a thread is about to be put on run queue
307 * because it has been made runnable or its priority has been adjusted. It
308 * determines if the new thread should preempt the current thread. If so,
309 * it sets td_owepreempt to request a preemption.
310 */
311 static int
maybe_preempt(struct thread * td)312 maybe_preempt(struct thread *td)
313 {
314 #ifdef PREEMPTION
315 struct thread *ctd;
316 int cpri, pri;
317
318 /*
319 * The new thread should not preempt the current thread if any of the
320 * following conditions are true:
321 *
322 * - The kernel is in the throes of crashing (panicstr).
323 * - The current thread has a higher (numerically lower) or
324 * equivalent priority. Note that this prevents curthread from
325 * trying to preempt to itself.
326 * - The current thread has an inhibitor set or is in the process of
327 * exiting. In this case, the current thread is about to switch
328 * out anyways, so there's no point in preempting. If we did,
329 * the current thread would not be properly resumed as well, so
330 * just avoid that whole landmine.
331 * - If the new thread's priority is not a realtime priority and
332 * the current thread's priority is not an idle priority and
333 * FULL_PREEMPTION is disabled.
334 *
335 * If all of these conditions are false, but the current thread is in
336 * a nested critical section, then we have to defer the preemption
337 * until we exit the critical section. Otherwise, switch immediately
338 * to the new thread.
339 */
340 ctd = curthread;
341 THREAD_LOCK_ASSERT(td, MA_OWNED);
342 KASSERT((td->td_inhibitors == 0),
343 ("maybe_preempt: trying to run inhibited thread"));
344 pri = td->td_priority;
345 cpri = ctd->td_priority;
346 if (KERNEL_PANICKED() || pri >= cpri /* || dumping */ ||
347 TD_IS_INHIBITED(ctd))
348 return (0);
349 #ifndef FULL_PREEMPTION
350 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
351 return (0);
352 #endif
353
354 CTR0(KTR_PROC, "maybe_preempt: scheduling preemption");
355 ctd->td_owepreempt = 1;
356 return (1);
357 #else
358 return (0);
359 #endif
360 }
361
362 /*
363 * Constants for digital decay and forget:
364 * 90% of (ts_estcpu) usage in 5 * loadav time
365 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
366 * Note that, as ps(1) mentions, this can let percentages
367 * total over 100% (I've seen 137.9% for 3 processes).
368 *
369 * Note that schedclock() updates ts_estcpu and p_cpticks asynchronously.
370 *
371 * We wish to decay away 90% of ts_estcpu in (5 * loadavg) seconds.
372 * That is, the system wants to compute a value of decay such
373 * that the following for loop:
374 * for (i = 0; i < (5 * loadavg); i++)
375 * ts_estcpu *= decay;
376 * will compute
377 * ts_estcpu *= 0.1;
378 * for all values of loadavg:
379 *
380 * Mathematically this loop can be expressed by saying:
381 * decay ** (5 * loadavg) ~= .1
382 *
383 * The system computes decay as:
384 * decay = (2 * loadavg) / (2 * loadavg + 1)
385 *
386 * We wish to prove that the system's computation of decay
387 * will always fulfill the equation:
388 * decay ** (5 * loadavg) ~= .1
389 *
390 * If we compute b as:
391 * b = 2 * loadavg
392 * then
393 * decay = b / (b + 1)
394 *
395 * We now need to prove two things:
396 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
397 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
398 *
399 * Facts:
400 * For x close to zero, exp(x) =~ 1 + x, since
401 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
402 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
403 * For x close to zero, ln(1+x) =~ x, since
404 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
405 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
406 * ln(.1) =~ -2.30
407 *
408 * Proof of (1):
409 * Solve (factor)**(power) =~ .1 given power (5*loadav):
410 * solving for factor,
411 * ln(factor) =~ (-2.30/5*loadav), or
412 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
413 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
414 *
415 * Proof of (2):
416 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
417 * solving for power,
418 * power*ln(b/(b+1)) =~ -2.30, or
419 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
420 *
421 * Actual power values for the implemented algorithm are as follows:
422 * loadav: 1 2 3 4
423 * power: 5.68 10.32 14.94 19.55
424 */
425
426 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
427 #define loadfactor(loadav) (2 * (loadav))
428 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
429
430 extern fixpt_t ccpu;
431
432 /*
433 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
434 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
435 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
436 *
437 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
438 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
439 *
440 * If you don't want to bother with the faster/more-accurate formula, you
441 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
442 * (more general) method of calculating the %age of CPU used by a process.
443 */
444 #define CCPU_SHIFT 11
445
446 /*
447 * Recompute process priorities, every hz ticks.
448 * MP-safe, called without the Giant mutex.
449 */
450 /* ARGSUSED */
451 static void
schedcpu(void)452 schedcpu(void)
453 {
454 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
455 struct thread *td;
456 struct proc *p;
457 struct td_sched *ts;
458 int awake;
459
460 sx_slock(&allproc_lock);
461 FOREACH_PROC_IN_SYSTEM(p) {
462 PROC_LOCK(p);
463 if (p->p_state == PRS_NEW) {
464 PROC_UNLOCK(p);
465 continue;
466 }
467 FOREACH_THREAD_IN_PROC(p, td) {
468 awake = 0;
469 ts = td_get_sched(td);
470 thread_lock(td);
471 /*
472 * Increment sleep time (if sleeping). We
473 * ignore overflow, as above.
474 */
475 /*
476 * The td_sched slptimes are not touched in wakeup
477 * because the thread may not HAVE everything in
478 * memory? XXX I think this is out of date.
479 */
480 if (TD_ON_RUNQ(td)) {
481 awake = 1;
482 td->td_flags &= ~TDF_DIDRUN;
483 } else if (TD_IS_RUNNING(td)) {
484 awake = 1;
485 /* Do not clear TDF_DIDRUN */
486 } else if (td->td_flags & TDF_DIDRUN) {
487 awake = 1;
488 td->td_flags &= ~TDF_DIDRUN;
489 }
490
491 /*
492 * ts_pctcpu is only for ps and ttyinfo().
493 */
494 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
495 /*
496 * If the td_sched has been idle the entire second,
497 * stop recalculating its priority until
498 * it wakes up.
499 */
500 if (ts->ts_cpticks != 0) {
501 #if (FSHIFT >= CCPU_SHIFT)
502 ts->ts_pctcpu += (realstathz == 100)
503 ? ((fixpt_t) ts->ts_cpticks) <<
504 (FSHIFT - CCPU_SHIFT) :
505 100 * (((fixpt_t) ts->ts_cpticks)
506 << (FSHIFT - CCPU_SHIFT)) / realstathz;
507 #else
508 ts->ts_pctcpu += ((FSCALE - ccpu) *
509 (ts->ts_cpticks *
510 FSCALE / realstathz)) >> FSHIFT;
511 #endif
512 ts->ts_cpticks = 0;
513 }
514 /*
515 * If there are ANY running threads in this process,
516 * then don't count it as sleeping.
517 * XXX: this is broken.
518 */
519 if (awake) {
520 if (ts->ts_slptime > 1) {
521 /*
522 * In an ideal world, this should not
523 * happen, because whoever woke us
524 * up from the long sleep should have
525 * unwound the slptime and reset our
526 * priority before we run at the stale
527 * priority. Should KASSERT at some
528 * point when all the cases are fixed.
529 */
530 updatepri(td);
531 }
532 ts->ts_slptime = 0;
533 } else
534 ts->ts_slptime++;
535 if (ts->ts_slptime > 1) {
536 thread_unlock(td);
537 continue;
538 }
539 ts->ts_estcpu = decay_cpu(loadfac, ts->ts_estcpu);
540 resetpriority(td);
541 resetpriority_thread(td);
542 thread_unlock(td);
543 }
544 PROC_UNLOCK(p);
545 }
546 sx_sunlock(&allproc_lock);
547 }
548
549 /*
550 * Main loop for a kthread that executes schedcpu once a second.
551 */
552 static void
schedcpu_thread(void)553 schedcpu_thread(void)
554 {
555
556 for (;;) {
557 schedcpu();
558 pause("-", hz);
559 }
560 }
561
562 /*
563 * Recalculate the priority of a process after it has slept for a while.
564 * For all load averages >= 1 and max ts_estcpu of 255, sleeping for at
565 * least six times the loadfactor will decay ts_estcpu to zero.
566 */
567 static void
updatepri(struct thread * td)568 updatepri(struct thread *td)
569 {
570 struct td_sched *ts;
571 fixpt_t loadfac;
572 unsigned int newcpu;
573
574 ts = td_get_sched(td);
575 loadfac = loadfactor(averunnable.ldavg[0]);
576 if (ts->ts_slptime > 5 * loadfac)
577 ts->ts_estcpu = 0;
578 else {
579 newcpu = ts->ts_estcpu;
580 ts->ts_slptime--; /* was incremented in schedcpu() */
581 while (newcpu && --ts->ts_slptime)
582 newcpu = decay_cpu(loadfac, newcpu);
583 ts->ts_estcpu = newcpu;
584 }
585 }
586
587 /*
588 * Compute the priority of a process when running in user mode.
589 * Arrange to reschedule if the resulting priority is better
590 * than that of the current process.
591 */
592 static void
resetpriority(struct thread * td)593 resetpriority(struct thread *td)
594 {
595 u_int newpriority;
596
597 if (td->td_pri_class != PRI_TIMESHARE)
598 return;
599 newpriority = PUSER +
600 td_get_sched(td)->ts_estcpu / INVERSE_ESTCPU_WEIGHT +
601 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
602 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
603 PRI_MAX_TIMESHARE);
604 sched_user_prio(td, newpriority);
605 }
606
607 /*
608 * Update the thread's priority when the associated process's user
609 * priority changes.
610 */
611 static void
resetpriority_thread(struct thread * td)612 resetpriority_thread(struct thread *td)
613 {
614
615 /* Only change threads with a time sharing user priority. */
616 if (td->td_priority < PRI_MIN_TIMESHARE ||
617 td->td_priority > PRI_MAX_TIMESHARE)
618 return;
619
620 /* XXX the whole needresched thing is broken, but not silly. */
621 maybe_resched(td);
622
623 sched_prio(td, td->td_user_pri);
624 }
625
626 static void
sched_4bsd_setup(void)627 sched_4bsd_setup(void)
628 {
629 /*
630 * Decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT
631 * before changing.
632 */
633 ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
634
635 setup_runqs();
636
637 /* Account for thread0. */
638 sched_load_add();
639
640 ast_register(TDA_SCHED_PRIV, ASTR_UNCOND, 0, maybe_resched_ast);
641 }
642
643 /*
644 * This routine determines time constants after stathz and hz are setup.
645 */
646 static void
sched_4bsd_initticks(void)647 sched_4bsd_initticks(void)
648 {
649
650 realstathz = stathz ? stathz : hz;
651 sched_slice = realstathz / 10; /* ~100ms */
652 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
653 realstathz);
654 }
655
656 /* External interfaces start here */
657
658 /*
659 * Very early in the boot some setup of scheduler-specific
660 * parts of proc0 and of some scheduler resources needs to be done.
661 * Called from:
662 * proc0_init()
663 */
664 static void
sched_4bsd_init(void)665 sched_4bsd_init(void)
666 {
667
668 /*
669 * Set up the scheduler specific parts of thread0.
670 */
671 thread0.td_lock = &sched_lock;
672 td_get_sched(&thread0)->ts_slice = sched_slice;
673 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN);
674 }
675
676 static void
sched_4bsd_init_ap(void)677 sched_4bsd_init_ap(void)
678 {
679
680 /* Nothing needed. */
681 }
682
683 static bool
sched_4bsd_runnable(void)684 sched_4bsd_runnable(void)
685 {
686 #ifdef SMP
687 return (runq_not_empty(&runq) ||
688 runq_not_empty(&runq_pcpu[PCPU_GET(cpuid)]));
689 #else
690 return (runq_not_empty(&runq));
691 #endif
692 }
693
694 static int
sched_4bsd_rr_interval(void)695 sched_4bsd_rr_interval(void)
696 {
697
698 /* Convert sched_slice from stathz to hz. */
699 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
700 }
701
702 /*
703 * We adjust the priority of the current process. The priority of a
704 * process gets worse as it accumulates CPU time. The cpu usage
705 * estimator (ts_estcpu) is increased here. resetpriority() will
706 * compute a different priority each time ts_estcpu increases by
707 * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached). The
708 * cpu usage estimator ramps up quite quickly when the process is
709 * running (linearly), and decays away exponentially, at a rate which
710 * is proportionally slower when the system is busy. The basic
711 * principle is that the system will 90% forget that the process used
712 * a lot of CPU time in 5 * loadav seconds. This causes the system to
713 * favor processes which haven't run much recently, and to round-robin
714 * among other processes.
715 */
716 static void
sched_clock_tick(struct thread * td)717 sched_clock_tick(struct thread *td)
718 {
719 struct pcpuidlestat *stat;
720 struct td_sched *ts;
721
722 THREAD_LOCK_ASSERT(td, MA_OWNED);
723 ts = td_get_sched(td);
724
725 ts->ts_cpticks++;
726 ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1);
727 if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
728 resetpriority(td);
729 resetpriority_thread(td);
730 }
731
732 /*
733 * Force a context switch if the current thread has used up a full
734 * time slice (default is 100ms).
735 */
736 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
737 ts->ts_slice = sched_slice;
738
739 /*
740 * If an ithread uses a full quantum, demote its
741 * priority and preempt it.
742 */
743 if (PRI_BASE(td->td_pri_class) == PRI_ITHD) {
744 SCHED_STAT_INC(ithread_preemptions);
745 td->td_owepreempt = 1;
746 if (td->td_base_pri + RQ_PPQ < PRI_MAX_ITHD) {
747 SCHED_STAT_INC(ithread_demotions);
748 sched_prio(td, td->td_base_pri + RQ_PPQ);
749 }
750 } else {
751 td->td_flags |= TDF_SLICEEND;
752 ast_sched_locked(td, TDA_SCHED);
753 }
754 }
755
756 stat = DPCPU_PTR(idlestat);
757 stat->oldidlecalls = stat->idlecalls;
758 stat->idlecalls = 0;
759 }
760
761 static void
sched_4bsd_clock(struct thread * td,int cnt)762 sched_4bsd_clock(struct thread *td, int cnt)
763 {
764
765 for ( ; cnt > 0; cnt--)
766 sched_clock_tick(td);
767 }
768
769 /*
770 * Charge child's scheduling CPU usage to parent.
771 */
772 static void
sched_4bsd_exit(struct proc * p,struct thread * td)773 sched_4bsd_exit(struct proc *p, struct thread *td)
774 {
775
776 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
777 "prio:%d", td->td_priority);
778
779 PROC_LOCK_ASSERT(p, MA_OWNED);
780 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
781 }
782
783 static void
sched_4bsd_exit_thread(struct thread * td,struct thread * child)784 sched_4bsd_exit_thread(struct thread *td, struct thread *child)
785 {
786
787 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
788 "prio:%d", child->td_priority);
789 thread_lock(td);
790 td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu +
791 td_get_sched(child)->ts_estcpu);
792 thread_unlock(td);
793 thread_lock(child);
794 if ((child->td_flags & TDF_NOLOAD) == 0)
795 sched_load_rem();
796 thread_unlock(child);
797 }
798
799 static void
sched_4bsd_fork(struct thread * td,struct thread * childtd)800 sched_4bsd_fork(struct thread *td, struct thread *childtd)
801 {
802 sched_fork_thread(td, childtd);
803 }
804
805 static void
sched_4bsd_fork_thread(struct thread * td,struct thread * childtd)806 sched_4bsd_fork_thread(struct thread *td, struct thread *childtd)
807 {
808 struct td_sched *ts, *tsc;
809
810 childtd->td_oncpu = NOCPU;
811 childtd->td_lastcpu = NOCPU;
812 childtd->td_lock = &sched_lock;
813 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
814 childtd->td_domain.dr_policy = td->td_cpuset->cs_domain;
815 childtd->td_priority = childtd->td_base_pri;
816 ts = td_get_sched(childtd);
817 bzero(ts, sizeof(*ts));
818 tsc = td_get_sched(td);
819 ts->ts_estcpu = tsc->ts_estcpu;
820 ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY);
821 ts->ts_slice = 1;
822 }
823
824 static void
sched_4bsd_nice(struct proc * p,int nice)825 sched_4bsd_nice(struct proc *p, int nice)
826 {
827 struct thread *td;
828
829 PROC_LOCK_ASSERT(p, MA_OWNED);
830 p->p_nice = nice;
831 FOREACH_THREAD_IN_PROC(p, td) {
832 thread_lock(td);
833 resetpriority(td);
834 resetpriority_thread(td);
835 thread_unlock(td);
836 }
837 }
838
839 static void
sched_4bsd_class(struct thread * td,int class)840 sched_4bsd_class(struct thread *td, int class)
841 {
842 THREAD_LOCK_ASSERT(td, MA_OWNED);
843 td->td_pri_class = class;
844 }
845
846 /*
847 * Adjust the priority of a thread.
848 */
849 static void
sched_priority(struct thread * td,u_char prio)850 sched_priority(struct thread *td, u_char prio)
851 {
852
853 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
854 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
855 sched_tdname(curthread));
856 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
857 if (td != curthread && prio > td->td_priority) {
858 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
859 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
860 prio, KTR_ATTR_LINKED, sched_tdname(td));
861 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
862 curthread);
863 }
864 THREAD_LOCK_ASSERT(td, MA_OWNED);
865 if (td->td_priority == prio)
866 return;
867 td->td_priority = prio;
868 if (TD_ON_RUNQ(td) && td->td_rqindex != RQ_PRI_TO_QUEUE_IDX(prio)) {
869 sched_rem(td);
870 sched_add(td, SRQ_BORING | SRQ_HOLDTD);
871 }
872 }
873
874 /*
875 * Update a thread's priority when it is lent another thread's
876 * priority.
877 */
878 static void
sched_4bsd_lend_prio(struct thread * td,u_char prio)879 sched_4bsd_lend_prio(struct thread *td, u_char prio)
880 {
881
882 td->td_flags |= TDF_BORROWING;
883 sched_priority(td, prio);
884 }
885
886 /*
887 * Restore a thread's priority when priority propagation is
888 * over. The prio argument is the minimum priority the thread
889 * needs to have to satisfy other possible priority lending
890 * requests. If the thread's regulary priority is less
891 * important than prio the thread will keep a priority boost
892 * of prio.
893 */
894 static void
sched_4bsd_unlend_prio(struct thread * td,u_char prio)895 sched_4bsd_unlend_prio(struct thread *td, u_char prio)
896 {
897 u_char base_pri;
898
899 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
900 td->td_base_pri <= PRI_MAX_TIMESHARE)
901 base_pri = td->td_user_pri;
902 else
903 base_pri = td->td_base_pri;
904 if (prio >= base_pri) {
905 td->td_flags &= ~TDF_BORROWING;
906 sched_prio(td, base_pri);
907 } else
908 sched_lend_prio(td, prio);
909 }
910
911 static void
sched_4bsd_prio(struct thread * td,u_char prio)912 sched_4bsd_prio(struct thread *td, u_char prio)
913 {
914 u_char oldprio;
915
916 /* First, update the base priority. */
917 td->td_base_pri = prio;
918
919 /*
920 * If the thread is borrowing another thread's priority, don't ever
921 * lower the priority.
922 */
923 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
924 return;
925
926 /* Change the real priority. */
927 oldprio = td->td_priority;
928 sched_priority(td, prio);
929
930 /*
931 * If the thread is on a turnstile, then let the turnstile update
932 * its state.
933 */
934 if (TD_ON_LOCK(td) && oldprio != prio)
935 turnstile_adjust(td, oldprio);
936 }
937
938 static void
sched_4bsd_ithread_prio(struct thread * td,u_char prio)939 sched_4bsd_ithread_prio(struct thread *td, u_char prio)
940 {
941 THREAD_LOCK_ASSERT(td, MA_OWNED);
942 MPASS(td->td_pri_class == PRI_ITHD);
943 td->td_base_ithread_pri = prio;
944 sched_prio(td, prio);
945 }
946
947 static void
sched_4bsd_user_prio(struct thread * td,u_char prio)948 sched_4bsd_user_prio(struct thread *td, u_char prio)
949 {
950
951 THREAD_LOCK_ASSERT(td, MA_OWNED);
952 td->td_base_user_pri = prio;
953 if (td->td_lend_user_pri <= prio)
954 return;
955 td->td_user_pri = prio;
956 }
957
958 static void
sched_4bsd_lend_user_prio(struct thread * td,u_char prio)959 sched_4bsd_lend_user_prio(struct thread *td, u_char prio)
960 {
961
962 THREAD_LOCK_ASSERT(td, MA_OWNED);
963 td->td_lend_user_pri = prio;
964 td->td_user_pri = min(prio, td->td_base_user_pri);
965 if (td->td_priority > td->td_user_pri)
966 sched_prio(td, td->td_user_pri);
967 else if (td->td_priority != td->td_user_pri)
968 ast_sched_locked(td, TDA_SCHED);
969 }
970
971 /*
972 * Like the above but first check if there is anything to do.
973 */
974 static void
sched_4bsd_lend_user_prio_cond(struct thread * td,u_char prio)975 sched_4bsd_lend_user_prio_cond(struct thread *td, u_char prio)
976 {
977
978 if (td->td_lend_user_pri == prio)
979 return;
980
981 thread_lock(td);
982 sched_lend_user_prio(td, prio);
983 thread_unlock(td);
984 }
985
986 static void
sched_4bsd_sleep(struct thread * td,int pri)987 sched_4bsd_sleep(struct thread *td, int pri)
988 {
989
990 THREAD_LOCK_ASSERT(td, MA_OWNED);
991 td->td_slptick = ticks;
992 td_get_sched(td)->ts_slptime = 0;
993 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
994 sched_prio(td, pri);
995 }
996
997 static void
sched_4bsd_sswitch(struct thread * td,int flags)998 sched_4bsd_sswitch(struct thread *td, int flags)
999 {
1000 struct thread *newtd;
1001 struct mtx *tmtx;
1002 int preempted;
1003
1004 tmtx = &sched_lock;
1005
1006 THREAD_LOCK_ASSERT(td, MA_OWNED);
1007
1008 td->td_lastcpu = td->td_oncpu;
1009 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
1010 (flags & SW_PREEMPT) != 0;
1011 td->td_flags &= ~TDF_SLICEEND;
1012 ast_unsched_locked(td, TDA_SCHED);
1013 td->td_owepreempt = 0;
1014 td->td_oncpu = NOCPU;
1015
1016 /*
1017 * At the last moment, if this thread is still marked RUNNING,
1018 * then put it back on the run queue as it has not been suspended
1019 * or stopped or any thing else similar. We never put the idle
1020 * threads on the run queue, however.
1021 */
1022 if (td->td_flags & TDF_IDLETD) {
1023 TD_SET_CAN_RUN(td);
1024 #ifdef SMP
1025 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
1026 #endif
1027 } else {
1028 if (TD_IS_RUNNING(td)) {
1029 /* Put us back on the run queue. */
1030 sched_add(td, SRQ_HOLDTD | SRQ_OURSELF | SRQ_YIELDING |
1031 (preempted ? SRQ_PREEMPTED : 0));
1032 }
1033 }
1034
1035 /*
1036 * Switch to the sched lock to fix things up and pick
1037 * a new thread. Block the td_lock in order to avoid
1038 * breaking the critical path.
1039 */
1040 if (td->td_lock != &sched_lock) {
1041 mtx_lock_spin(&sched_lock);
1042 tmtx = thread_lock_block(td);
1043 mtx_unlock_spin(tmtx);
1044 }
1045
1046 if ((td->td_flags & TDF_NOLOAD) == 0)
1047 sched_load_rem();
1048
1049 newtd = choosethread();
1050 MPASS(newtd->td_lock == &sched_lock);
1051
1052 #if (KTR_COMPILE & KTR_SCHED) != 0
1053 if (TD_IS_IDLETHREAD(td))
1054 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1055 "prio:%d", td->td_priority);
1056 else
1057 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1058 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1059 "lockname:\"%s\"", td->td_lockname);
1060 #endif
1061
1062 if (td != newtd) {
1063 #ifdef HWPMC_HOOKS
1064 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1065 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1066 #endif
1067
1068 #ifdef HWT_HOOKS
1069 HWT_CALL_HOOK(td, HWT_SWITCH_OUT, NULL);
1070 HWT_CALL_HOOK(newtd, HWT_SWITCH_IN, NULL);
1071 #endif
1072
1073 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1074
1075 /* I feel sleepy */
1076 lock_profile_release_lock(&sched_lock.lock_object, true);
1077 #ifdef KDTRACE_HOOKS
1078 /*
1079 * If DTrace has set the active vtime enum to anything
1080 * other than INACTIVE (0), then it should have set the
1081 * function to call.
1082 */
1083 if (dtrace_vtime_active)
1084 (*dtrace_vtime_switch_func)(newtd);
1085 #endif
1086
1087 cpu_switch(td, newtd, tmtx);
1088 lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1089 0, 0, __FILE__, __LINE__);
1090 /*
1091 * Where am I? What year is it?
1092 * We are in the same thread that went to sleep above,
1093 * but any amount of time may have passed. All our context
1094 * will still be available as will local variables.
1095 * PCPU values however may have changed as we may have
1096 * changed CPU so don't trust cached values of them.
1097 * New threads will go to fork_exit() instead of here
1098 * so if you change things here you may need to change
1099 * things there too.
1100 *
1101 * If the thread above was exiting it will never wake
1102 * up again here, so either it has saved everything it
1103 * needed to, or the thread_wait() or wait() will
1104 * need to reap it.
1105 */
1106
1107 SDT_PROBE0(sched, , , on__cpu);
1108 #ifdef HWPMC_HOOKS
1109 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1110 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1111 #endif
1112 } else {
1113 td->td_lock = &sched_lock;
1114 SDT_PROBE0(sched, , , remain__cpu);
1115 }
1116
1117 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1118 "prio:%d", td->td_priority);
1119
1120 #ifdef SMP
1121 if (td->td_flags & TDF_IDLETD)
1122 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1123 #endif
1124 sched_lock.mtx_lock = (uintptr_t)td;
1125 td->td_oncpu = PCPU_GET(cpuid);
1126 spinlock_enter();
1127 mtx_unlock_spin(&sched_lock);
1128 }
1129
1130 static void
sched_4bsd_wakeup(struct thread * td,int srqflags)1131 sched_4bsd_wakeup(struct thread *td, int srqflags)
1132 {
1133 struct td_sched *ts;
1134
1135 THREAD_LOCK_ASSERT(td, MA_OWNED);
1136 ts = td_get_sched(td);
1137 if (ts->ts_slptime > 1) {
1138 updatepri(td);
1139 resetpriority(td);
1140 }
1141 td->td_slptick = 0;
1142 ts->ts_slptime = 0;
1143 ts->ts_slice = sched_slice;
1144
1145 /*
1146 * When resuming an idle ithread, restore its base ithread
1147 * priority.
1148 */
1149 if (PRI_BASE(td->td_pri_class) == PRI_ITHD &&
1150 td->td_base_pri != td->td_base_ithread_pri)
1151 sched_prio(td, td->td_base_ithread_pri);
1152
1153 sched_add(td, srqflags);
1154 }
1155
1156 #ifdef SMP
1157 static int
forward_wakeup(int cpunum)1158 forward_wakeup(int cpunum)
1159 {
1160 struct pcpu *pc;
1161 cpuset_t dontuse, map, map2;
1162 u_int id, me;
1163 int iscpuset;
1164
1165 mtx_assert(&sched_lock, MA_OWNED);
1166
1167 CTR0(KTR_RUNQ, "forward_wakeup()");
1168
1169 if ((!forward_wakeup_enabled) ||
1170 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1171 return (0);
1172 if (!smp_started || KERNEL_PANICKED())
1173 return (0);
1174
1175 forward_wakeups_requested++;
1176
1177 /*
1178 * Check the idle mask we received against what we calculated
1179 * before in the old version.
1180 */
1181 me = PCPU_GET(cpuid);
1182
1183 /* Don't bother if we should be doing it ourself. */
1184 if (CPU_ISSET(me, &idle_cpus_mask) &&
1185 (cpunum == NOCPU || me == cpunum))
1186 return (0);
1187
1188 CPU_SETOF(me, &dontuse);
1189 CPU_OR(&dontuse, &dontuse, &stopped_cpus);
1190 CPU_OR(&dontuse, &dontuse, &hlt_cpus_mask);
1191 CPU_ZERO(&map2);
1192 if (forward_wakeup_use_loop) {
1193 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1194 id = pc->pc_cpuid;
1195 if (!CPU_ISSET(id, &dontuse) &&
1196 pc->pc_curthread == pc->pc_idlethread) {
1197 CPU_SET(id, &map2);
1198 }
1199 }
1200 }
1201
1202 if (forward_wakeup_use_mask) {
1203 map = idle_cpus_mask;
1204 CPU_ANDNOT(&map, &map, &dontuse);
1205
1206 /* If they are both on, compare and use loop if different. */
1207 if (forward_wakeup_use_loop) {
1208 if (CPU_CMP(&map, &map2)) {
1209 printf("map != map2, loop method preferred\n");
1210 map = map2;
1211 }
1212 }
1213 } else {
1214 map = map2;
1215 }
1216
1217 /* If we only allow a specific CPU, then mask off all the others. */
1218 if (cpunum != NOCPU) {
1219 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1220 iscpuset = CPU_ISSET(cpunum, &map);
1221 if (iscpuset == 0)
1222 CPU_ZERO(&map);
1223 else
1224 CPU_SETOF(cpunum, &map);
1225 }
1226 if (!CPU_EMPTY(&map)) {
1227 forward_wakeups_delivered++;
1228 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1229 id = pc->pc_cpuid;
1230 if (!CPU_ISSET(id, &map))
1231 continue;
1232 if (cpu_idle_wakeup(pc->pc_cpuid))
1233 CPU_CLR(id, &map);
1234 }
1235 if (!CPU_EMPTY(&map))
1236 ipi_selected(map, IPI_AST);
1237 return (1);
1238 }
1239 if (cpunum == NOCPU)
1240 printf("forward_wakeup: Idle processor not found\n");
1241 return (0);
1242 }
1243
1244 static void
kick_other_cpu(int pri,int cpuid)1245 kick_other_cpu(int pri, int cpuid)
1246 {
1247 struct pcpu *pcpu;
1248 int cpri;
1249
1250 pcpu = pcpu_find(cpuid);
1251 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1252 forward_wakeups_delivered++;
1253 if (!cpu_idle_wakeup(cpuid))
1254 ipi_cpu(cpuid, IPI_AST);
1255 return;
1256 }
1257
1258 cpri = pcpu->pc_curthread->td_priority;
1259 if (pri >= cpri)
1260 return;
1261
1262 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1263 #if !defined(FULL_PREEMPTION)
1264 if (pri <= PRI_MAX_ITHD)
1265 #endif /* ! FULL_PREEMPTION */
1266 {
1267 ipi_cpu(cpuid, IPI_PREEMPT);
1268 return;
1269 }
1270 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1271
1272 if (pcpu->pc_curthread->td_lock == &sched_lock) {
1273 ast_sched_locked(pcpu->pc_curthread, TDA_SCHED);
1274 ipi_cpu(cpuid, IPI_AST);
1275 }
1276 }
1277 #endif /* SMP */
1278
1279 #ifdef SMP
1280 static int
sched_pickcpu(struct thread * td)1281 sched_pickcpu(struct thread *td)
1282 {
1283 int best, cpu;
1284
1285 mtx_assert(&sched_lock, MA_OWNED);
1286
1287 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu))
1288 best = td->td_lastcpu;
1289 else
1290 best = NOCPU;
1291 CPU_FOREACH(cpu) {
1292 if (!THREAD_CAN_SCHED(td, cpu))
1293 continue;
1294
1295 if (best == NOCPU)
1296 best = cpu;
1297 else if (runq_length[cpu] < runq_length[best])
1298 best = cpu;
1299 }
1300 KASSERT(best != NOCPU, ("no valid CPUs"));
1301
1302 return (best);
1303 }
1304 #endif
1305
1306 static void
sched_4bsd_add(struct thread * td,int flags)1307 sched_4bsd_add(struct thread *td, int flags)
1308 #ifdef SMP
1309 {
1310 cpuset_t tidlemsk;
1311 struct td_sched *ts;
1312 u_int cpu, cpuid;
1313 int forwarded = 0;
1314 int single_cpu = 0;
1315
1316 ts = td_get_sched(td);
1317 THREAD_LOCK_ASSERT(td, MA_OWNED);
1318 KASSERT((td->td_inhibitors == 0),
1319 ("sched_add: trying to run inhibited thread"));
1320 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1321 ("sched_add: bad thread state"));
1322 KASSERT(td->td_flags & TDF_INMEM,
1323 ("sched_add: thread swapped out"));
1324
1325 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1326 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1327 sched_tdname(curthread));
1328 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1329 KTR_ATTR_LINKED, sched_tdname(td));
1330 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1331 flags & SRQ_PREEMPTED);
1332
1333 /*
1334 * Now that the thread is moving to the run-queue, set the lock
1335 * to the scheduler's lock.
1336 */
1337 if (td->td_lock != &sched_lock) {
1338 mtx_lock_spin(&sched_lock);
1339 if ((flags & SRQ_HOLD) != 0)
1340 td->td_lock = &sched_lock;
1341 else
1342 thread_lock_set(td, &sched_lock);
1343 }
1344 TD_SET_RUNQ(td);
1345
1346 /*
1347 * If SMP is started and the thread is pinned or otherwise limited to
1348 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1349 * Otherwise, queue the thread to the global run queue.
1350 *
1351 * If SMP has not yet been started we must use the global run queue
1352 * as per-CPU state may not be initialized yet and we may crash if we
1353 * try to access the per-CPU run queues.
1354 */
1355 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1356 ts->ts_flags & TSF_AFFINITY)) {
1357 if (td->td_pinned != 0)
1358 cpu = td->td_lastcpu;
1359 else if (td->td_flags & TDF_BOUND) {
1360 /* Find CPU from bound runq. */
1361 KASSERT(SKE_RUNQ_PCPU(ts),
1362 ("sched_add: bound td_sched not on cpu runq"));
1363 cpu = ts->ts_runq - &runq_pcpu[0];
1364 } else
1365 /* Find a valid CPU for our cpuset */
1366 cpu = sched_pickcpu(td);
1367 ts->ts_runq = &runq_pcpu[cpu];
1368 single_cpu = 1;
1369 CTR3(KTR_RUNQ,
1370 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1371 cpu);
1372 } else {
1373 CTR2(KTR_RUNQ,
1374 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1375 td);
1376 cpu = NOCPU;
1377 ts->ts_runq = &runq;
1378 }
1379
1380 if ((td->td_flags & TDF_NOLOAD) == 0)
1381 sched_load_add();
1382 runq_add(ts->ts_runq, td, flags);
1383 if (cpu != NOCPU)
1384 runq_length[cpu]++;
1385
1386 cpuid = PCPU_GET(cpuid);
1387 if (single_cpu && cpu != cpuid) {
1388 kick_other_cpu(td->td_priority, cpu);
1389 } else {
1390 if (!single_cpu) {
1391 tidlemsk = idle_cpus_mask;
1392 CPU_ANDNOT(&tidlemsk, &tidlemsk, &hlt_cpus_mask);
1393 CPU_CLR(cpuid, &tidlemsk);
1394
1395 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1396 ((flags & SRQ_INTR) == 0) &&
1397 !CPU_EMPTY(&tidlemsk))
1398 forwarded = forward_wakeup(cpu);
1399 }
1400
1401 if (!forwarded) {
1402 if (!maybe_preempt(td))
1403 maybe_resched(td);
1404 }
1405 }
1406 if ((flags & SRQ_HOLDTD) == 0)
1407 thread_unlock(td);
1408 }
1409 #else /* SMP */
1410 {
1411 struct td_sched *ts;
1412
1413 ts = td_get_sched(td);
1414 THREAD_LOCK_ASSERT(td, MA_OWNED);
1415 KASSERT((td->td_inhibitors == 0),
1416 ("sched_add: trying to run inhibited thread"));
1417 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1418 ("sched_add: bad thread state"));
1419 KASSERT(td->td_flags & TDF_INMEM,
1420 ("sched_add: thread swapped out"));
1421 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1422 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1423 sched_tdname(curthread));
1424 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1425 KTR_ATTR_LINKED, sched_tdname(td));
1426 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1427 flags & SRQ_PREEMPTED);
1428
1429 /*
1430 * Now that the thread is moving to the run-queue, set the lock
1431 * to the scheduler's lock.
1432 */
1433 if (td->td_lock != &sched_lock) {
1434 mtx_lock_spin(&sched_lock);
1435 if ((flags & SRQ_HOLD) != 0)
1436 td->td_lock = &sched_lock;
1437 else
1438 thread_lock_set(td, &sched_lock);
1439 }
1440 TD_SET_RUNQ(td);
1441 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1442 ts->ts_runq = &runq;
1443
1444 if ((td->td_flags & TDF_NOLOAD) == 0)
1445 sched_load_add();
1446 runq_add(ts->ts_runq, td, flags);
1447 if (!maybe_preempt(td))
1448 maybe_resched(td);
1449 if ((flags & SRQ_HOLDTD) == 0)
1450 thread_unlock(td);
1451 }
1452 #endif /* SMP */
1453
1454 static void
sched_4bsd_rem(struct thread * td)1455 sched_4bsd_rem(struct thread *td)
1456 {
1457 struct td_sched *ts;
1458
1459 ts = td_get_sched(td);
1460 KASSERT(td->td_flags & TDF_INMEM,
1461 ("sched_rem: thread swapped out"));
1462 KASSERT(TD_ON_RUNQ(td),
1463 ("sched_rem: thread not on run queue"));
1464 mtx_assert(&sched_lock, MA_OWNED);
1465 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1466 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1467 sched_tdname(curthread));
1468 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1469
1470 if ((td->td_flags & TDF_NOLOAD) == 0)
1471 sched_load_rem();
1472 #ifdef SMP
1473 if (ts->ts_runq != &runq)
1474 runq_length[ts->ts_runq - runq_pcpu]--;
1475 #endif
1476 runq_remove(ts->ts_runq, td);
1477 TD_SET_CAN_RUN(td);
1478 }
1479
1480 /*
1481 * Select threads to run. Note that running threads still consume a
1482 * slot.
1483 */
1484 static struct thread *
sched_4bsd_choose(void)1485 sched_4bsd_choose(void)
1486 {
1487 struct thread *td;
1488 struct runq *rq;
1489
1490 mtx_assert(&sched_lock, MA_OWNED);
1491 #ifdef SMP
1492 struct thread *tdcpu;
1493
1494 rq = &runq;
1495 td = runq_choose_fuzz(&runq, runq_fuzz);
1496 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1497
1498 if (td == NULL ||
1499 (tdcpu != NULL &&
1500 tdcpu->td_priority < td->td_priority)) {
1501 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1502 PCPU_GET(cpuid));
1503 td = tdcpu;
1504 rq = &runq_pcpu[PCPU_GET(cpuid)];
1505 } else {
1506 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1507 }
1508
1509 #else
1510 rq = &runq;
1511 td = runq_choose(&runq);
1512 #endif
1513
1514 if (td) {
1515 #ifdef SMP
1516 if (td == tdcpu)
1517 runq_length[PCPU_GET(cpuid)]--;
1518 #endif
1519 runq_remove(rq, td);
1520 td->td_flags |= TDF_DIDRUN;
1521
1522 KASSERT(td->td_flags & TDF_INMEM,
1523 ("sched_choose: thread swapped out"));
1524 return (td);
1525 }
1526 return (PCPU_GET(idlethread));
1527 }
1528
1529 static void
sched_4bsd_preempt(struct thread * td)1530 sched_4bsd_preempt(struct thread *td)
1531 {
1532 int flags;
1533
1534 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1535 if (td->td_critnest > 1) {
1536 td->td_owepreempt = 1;
1537 } else {
1538 thread_lock(td);
1539 flags = SW_INVOL | SW_PREEMPT;
1540 flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
1541 SWT_REMOTEPREEMPT;
1542 mi_switch(flags);
1543 }
1544 }
1545
1546 static void
sched_4bsd_userret_slowpath(struct thread * td)1547 sched_4bsd_userret_slowpath(struct thread *td)
1548 {
1549
1550 thread_lock(td);
1551 td->td_priority = td->td_user_pri;
1552 td->td_base_pri = td->td_user_pri;
1553 thread_unlock(td);
1554 }
1555
1556 static void
sched_4bsd_bind(struct thread * td,int cpu)1557 sched_4bsd_bind(struct thread *td, int cpu)
1558 {
1559 #ifdef SMP
1560 struct td_sched *ts = td_get_sched(td);
1561 #endif
1562
1563 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1564 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1565
1566 td->td_flags |= TDF_BOUND;
1567 #ifdef SMP
1568 ts->ts_runq = &runq_pcpu[cpu];
1569 if (PCPU_GET(cpuid) == cpu)
1570 return;
1571
1572 mi_switch(SW_VOL | SWT_BIND);
1573 thread_lock(td);
1574 #endif
1575 }
1576
1577 static void
sched_4bsd_unbind(struct thread * td)1578 sched_4bsd_unbind(struct thread* td)
1579 {
1580 THREAD_LOCK_ASSERT(td, MA_OWNED);
1581 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1582 td->td_flags &= ~TDF_BOUND;
1583 }
1584
1585 static int
sched_4bsd_is_bound(struct thread * td)1586 sched_4bsd_is_bound(struct thread *td)
1587 {
1588 THREAD_LOCK_ASSERT(td, MA_OWNED);
1589 return (td->td_flags & TDF_BOUND);
1590 }
1591
1592 static void
sched_4bsd_relinquish(struct thread * td)1593 sched_4bsd_relinquish(struct thread *td)
1594 {
1595 thread_lock(td);
1596 mi_switch(SW_VOL | SWT_RELINQUISH);
1597 }
1598
1599 static int
sched_4bsd_load(void)1600 sched_4bsd_load(void)
1601 {
1602 return (sched_tdcnt);
1603 }
1604
1605 static int
sched_4bsd_sizeof_proc(void)1606 sched_4bsd_sizeof_proc(void)
1607 {
1608 return (sizeof(struct proc));
1609 }
1610
1611 static int
sched_4bsd_sizeof_thread(void)1612 sched_4bsd_sizeof_thread(void)
1613 {
1614 return (sizeof(struct thread) + sizeof(struct td_sched));
1615 }
1616
1617 static fixpt_t
sched_4bsd_pctcpu(struct thread * td)1618 sched_4bsd_pctcpu(struct thread *td)
1619 {
1620 struct td_sched *ts;
1621
1622 THREAD_LOCK_ASSERT(td, MA_OWNED);
1623 ts = td_get_sched(td);
1624 return (ts->ts_pctcpu);
1625 }
1626
1627 static u_int
sched_4bsd_estcpu(struct thread * td)1628 sched_4bsd_estcpu(struct thread *td)
1629 {
1630
1631 return (td_get_sched(td)->ts_estcpu);
1632 }
1633
1634 /*
1635 * The actual idle process.
1636 */
1637 static void
sched_4bsd_idletd(void * dummy)1638 sched_4bsd_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()) {
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 static void
sched_throw_tail(struct thread * td)1658 sched_throw_tail(struct thread *td)
1659 {
1660 struct thread *newtd;
1661
1662 mtx_assert(&sched_lock, MA_OWNED);
1663 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1664
1665 newtd = choosethread();
1666
1667 #ifdef HWT_HOOKS
1668 if (td)
1669 HWT_CALL_HOOK(td, HWT_SWITCH_OUT, NULL);
1670 HWT_CALL_HOOK(newtd, HWT_SWITCH_IN, NULL);
1671 #endif
1672
1673 cpu_throw(td, newtd); /* doesn't return */
1674 }
1675
1676 /*
1677 * A CPU is entering for the first time.
1678 */
1679 static void
sched_4bsd_ap_entry(void)1680 sched_4bsd_ap_entry(void)
1681 {
1682
1683 /*
1684 * Correct spinlock nesting. The idle thread context that we are
1685 * borrowing was created so that it would start out with a single
1686 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1687 * explicitly acquired locks in this function, the nesting count
1688 * is now 2 rather than 1. Since we are nested, calling
1689 * spinlock_exit() will simply adjust the counts without allowing
1690 * spin lock using code to interrupt us.
1691 */
1692 mtx_lock_spin(&sched_lock);
1693 spinlock_exit();
1694 PCPU_SET(switchtime, cpu_ticks());
1695 PCPU_SET(switchticks, ticks);
1696
1697 sched_throw_tail(NULL);
1698 }
1699
1700 /*
1701 * A thread is exiting.
1702 */
1703 static void
sched_4bsd_throw(struct thread * td)1704 sched_4bsd_throw(struct thread *td)
1705 {
1706
1707 MPASS(td != NULL);
1708 MPASS(td->td_lock == &sched_lock);
1709
1710 lock_profile_release_lock(&sched_lock.lock_object, true);
1711 td->td_lastcpu = td->td_oncpu;
1712 td->td_oncpu = NOCPU;
1713
1714 sched_throw_tail(td);
1715 }
1716
1717 static void
sched_4bsd_fork_exit(struct thread * td)1718 sched_4bsd_fork_exit(struct thread *td)
1719 {
1720
1721 /*
1722 * Finish setting up thread glue so that it begins execution in a
1723 * non-nested critical section with sched_lock held but not recursed.
1724 */
1725 td->td_oncpu = PCPU_GET(cpuid);
1726 sched_lock.mtx_lock = (uintptr_t)td;
1727 lock_profile_obtain_lock_success(&sched_lock.lock_object, true,
1728 0, 0, __FILE__, __LINE__);
1729 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1730
1731 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1732 "prio:%d", td->td_priority);
1733 SDT_PROBE0(sched, , , on__cpu);
1734 }
1735
1736 static char *
sched_4bsd_tdname(struct thread * td)1737 sched_4bsd_tdname(struct thread *td)
1738 {
1739 #ifdef KTR
1740 struct td_sched *ts;
1741
1742 ts = td_get_sched(td);
1743 if (ts->ts_name[0] == '\0')
1744 snprintf(ts->ts_name, sizeof(ts->ts_name),
1745 "%s tid %d", td->td_name, td->td_tid);
1746 return (ts->ts_name);
1747 #else
1748 return (td->td_name);
1749 #endif
1750 }
1751
1752 static void
sched_4bsd_clear_tdname(struct thread * td)1753 sched_4bsd_clear_tdname(struct thread *td)
1754 {
1755 #ifdef KTR
1756 struct td_sched *ts;
1757
1758 ts = td_get_sched(td);
1759 ts->ts_name[0] = '\0';
1760 #endif
1761 }
1762
1763 static void
sched_4bsd_affinity(struct thread * td)1764 sched_4bsd_affinity(struct thread *td)
1765 {
1766 #ifdef SMP
1767 struct td_sched *ts;
1768 int cpu;
1769
1770 THREAD_LOCK_ASSERT(td, MA_OWNED);
1771
1772 /*
1773 * Set the TSF_AFFINITY flag if there is at least one CPU this
1774 * thread can't run on.
1775 */
1776 ts = td_get_sched(td);
1777 ts->ts_flags &= ~TSF_AFFINITY;
1778 CPU_FOREACH(cpu) {
1779 if (!THREAD_CAN_SCHED(td, cpu)) {
1780 ts->ts_flags |= TSF_AFFINITY;
1781 break;
1782 }
1783 }
1784
1785 /*
1786 * If this thread can run on all CPUs, nothing else to do.
1787 */
1788 if (!(ts->ts_flags & TSF_AFFINITY))
1789 return;
1790
1791 /* Pinned threads and bound threads should be left alone. */
1792 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1793 return;
1794
1795 switch (TD_GET_STATE(td)) {
1796 case TDS_RUNQ:
1797 /*
1798 * If we are on a per-CPU runqueue that is in the set,
1799 * then nothing needs to be done.
1800 */
1801 if (ts->ts_runq != &runq &&
1802 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1803 return;
1804
1805 /* Put this thread on a valid per-CPU runqueue. */
1806 sched_rem(td);
1807 sched_add(td, SRQ_HOLDTD | SRQ_BORING);
1808 break;
1809 case TDS_RUNNING:
1810 /*
1811 * See if our current CPU is in the set. If not, force a
1812 * context switch.
1813 */
1814 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1815 return;
1816
1817 ast_sched_locked(td, TDA_SCHED);
1818 if (td != curthread)
1819 ipi_cpu(cpu, IPI_AST);
1820 break;
1821 default:
1822 break;
1823 }
1824 #endif
1825 }
1826
1827 static bool
sched_4bsd_do_timer_accounting(void)1828 sched_4bsd_do_timer_accounting(void)
1829 {
1830 #ifdef SMP
1831 /*
1832 * Don't do any accounting for the disabled HTT cores, since it
1833 * will provide misleading numbers for the userland.
1834 *
1835 * No locking is necessary here, since even if we lose the race
1836 * when hlt_cpus_mask changes it is not a big deal, really.
1837 *
1838 * Don't do that for ULE, since ULE doesn't consider hlt_cpus_mask
1839 * and unlike other schedulers it actually schedules threads to
1840 * those CPUs.
1841 */
1842 return (!CPU_ISSET(PCPU_GET(cpuid), &hlt_cpus_mask));
1843 #else
1844 return (true);
1845 #endif
1846 }
1847
1848 static int
sched_4bsd_find_l2_neighbor(int cpu)1849 sched_4bsd_find_l2_neighbor(int cpu)
1850 {
1851 return (-1);
1852 }
1853
1854 struct sched_instance sched_4bsd_instance = {
1855 #define SLOT(name) .name = sched_4bsd_##name
1856 SLOT(load),
1857 SLOT(rr_interval),
1858 SLOT(runnable),
1859 SLOT(exit),
1860 SLOT(fork),
1861 SLOT(fork_exit),
1862 SLOT(class),
1863 SLOT(nice),
1864 SLOT(ap_entry),
1865 SLOT(exit_thread),
1866 SLOT(estcpu),
1867 SLOT(fork_thread),
1868 SLOT(ithread_prio),
1869 SLOT(lend_prio),
1870 SLOT(lend_user_prio),
1871 SLOT(lend_user_prio_cond),
1872 SLOT(pctcpu),
1873 SLOT(prio),
1874 SLOT(sleep),
1875 SLOT(sswitch),
1876 SLOT(throw),
1877 SLOT(unlend_prio),
1878 SLOT(user_prio),
1879 SLOT(userret_slowpath),
1880 SLOT(add),
1881 SLOT(choose),
1882 SLOT(clock),
1883 SLOT(idletd),
1884 SLOT(preempt),
1885 SLOT(relinquish),
1886 SLOT(rem),
1887 SLOT(wakeup),
1888 SLOT(bind),
1889 SLOT(unbind),
1890 SLOT(is_bound),
1891 SLOT(affinity),
1892 SLOT(sizeof_proc),
1893 SLOT(sizeof_thread),
1894 SLOT(tdname),
1895 SLOT(clear_tdname),
1896 SLOT(do_timer_accounting),
1897 SLOT(find_l2_neighbor),
1898 SLOT(init),
1899 SLOT(init_ap),
1900 SLOT(setup),
1901 SLOT(initticks),
1902 SLOT(schedcpu),
1903 #undef SLOT
1904 };
1905 DECLARE_SCHEDULER(fourbsd_sched_selector, "4BSD", &sched_4bsd_instance);
1906