1 /*-
2 * Copyright (c) 1982, 1986, 1990, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 */
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD: stable/10/sys/kern/sched_4bsd.c 316841 2017-04-14 14:44:06Z avg $");
37
38 #include "opt_hwpmc_hooks.h"
39 #include "opt_sched.h"
40 #include "opt_kdtrace.h"
41
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/cpuset.h>
45 #include <sys/kernel.h>
46 #include <sys/ktr.h>
47 #include <sys/lock.h>
48 #include <sys/kthread.h>
49 #include <sys/mutex.h>
50 #include <sys/proc.h>
51 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
53 #include <sys/sdt.h>
54 #include <sys/smp.h>
55 #include <sys/sysctl.h>
56 #include <sys/sx.h>
57 #include <sys/turnstile.h>
58 #include <sys/umtx.h>
59 #include <machine/pcb.h>
60 #include <machine/smp.h>
61
62 #ifdef HWPMC_HOOKS
63 #include <sys/pmckern.h>
64 #endif
65
66 #ifdef KDTRACE_HOOKS
67 #include <sys/dtrace_bsd.h>
68 int dtrace_vtime_active;
69 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
70 #endif
71
72 /*
73 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
74 * the range 100-256 Hz (approximately).
75 */
76 #define ESTCPULIM(e) \
77 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
78 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
79 #ifdef SMP
80 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
81 #else
82 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
83 #endif
84 #define NICE_WEIGHT 1 /* Priorities per nice level. */
85
86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
87
88 /*
89 * The schedulable entity that runs a context.
90 * This is an extension to the thread structure and is tailored to
91 * the requirements of this scheduler
92 */
93 struct td_sched {
94 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */
95 int ts_cpticks; /* (j) Ticks of cpu time. */
96 int ts_slptime; /* (j) Seconds !RUNNING. */
97 int ts_slice; /* Remaining part of time slice. */
98 int ts_flags;
99 struct runq *ts_runq; /* runq the thread is currently on */
100 #ifdef KTR
101 char ts_name[TS_NAME_LEN];
102 #endif
103 };
104
105 /* flags kept in td_flags */
106 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */
107 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */
108 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
109
110 /* flags kept in ts_flags */
111 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */
112
113 #define SKE_RUNQ_PCPU(ts) \
114 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
115
116 #define THREAD_CAN_SCHED(td, cpu) \
117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
118
119 static struct td_sched td_sched0;
120 struct mtx sched_lock;
121
122 static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */
123 static int sched_tdcnt; /* Total runnable threads in the system. */
124 static int sched_slice = 12; /* Thread run time before rescheduling. */
125
126 static void setup_runqs(void);
127 static void schedcpu(void);
128 static void schedcpu_thread(void);
129 static void sched_priority(struct thread *td, u_char prio);
130 static void sched_setup(void *dummy);
131 static void maybe_resched(struct thread *td);
132 static void updatepri(struct thread *td);
133 static void resetpriority(struct thread *td);
134 static void resetpriority_thread(struct thread *td);
135 #ifdef SMP
136 static int sched_pickcpu(struct thread *td);
137 static int forward_wakeup(int cpunum);
138 static void kick_other_cpu(int pri, int cpuid);
139 #endif
140
141 static struct kproc_desc sched_kp = {
142 "schedcpu",
143 schedcpu_thread,
144 NULL
145 };
146 SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start,
147 &sched_kp);
148 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
149
150 static void sched_initticks(void *dummy);
151 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
152 NULL);
153
154 /*
155 * Global run queue.
156 */
157 static struct runq runq;
158
159 #ifdef SMP
160 /*
161 * Per-CPU run queues
162 */
163 static struct runq runq_pcpu[MAXCPU];
164 long runq_length[MAXCPU];
165
166 static cpuset_t idle_cpus_mask;
167 #endif
168
169 struct pcpuidlestat {
170 u_int idlecalls;
171 u_int oldidlecalls;
172 };
173 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
174
175 static void
setup_runqs(void)176 setup_runqs(void)
177 {
178 #ifdef SMP
179 int i;
180
181 for (i = 0; i < MAXCPU; ++i)
182 runq_init(&runq_pcpu[i]);
183 #endif
184
185 runq_init(&runq);
186 }
187
188 static int
sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)189 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
190 {
191 int error, new_val, period;
192
193 period = 1000000 / realstathz;
194 new_val = period * sched_slice;
195 error = sysctl_handle_int(oidp, &new_val, 0, req);
196 if (error != 0 || req->newptr == NULL)
197 return (error);
198 if (new_val <= 0)
199 return (EINVAL);
200 sched_slice = imax(1, (new_val + period / 2) / period);
201 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
202 realstathz);
203 return (0);
204 }
205
206 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
207
208 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
209 "Scheduler name");
210 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
211 NULL, 0, sysctl_kern_quantum, "I",
212 "Quantum for timeshare threads in microseconds");
213 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
214 "Quantum for timeshare threads in stathz ticks");
215 #ifdef SMP
216 /* Enable forwarding of wakeups to all other cpus */
217 static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL,
218 "Kernel SMP");
219
220 static int runq_fuzz = 1;
221 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
222
223 static int forward_wakeup_enabled = 1;
224 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
225 &forward_wakeup_enabled, 0,
226 "Forwarding of wakeup to idle CPUs");
227
228 static int forward_wakeups_requested = 0;
229 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
230 &forward_wakeups_requested, 0,
231 "Requests for Forwarding of wakeup to idle CPUs");
232
233 static int forward_wakeups_delivered = 0;
234 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
235 &forward_wakeups_delivered, 0,
236 "Completed Forwarding of wakeup to idle CPUs");
237
238 static int forward_wakeup_use_mask = 1;
239 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
240 &forward_wakeup_use_mask, 0,
241 "Use the mask of idle cpus");
242
243 static int forward_wakeup_use_loop = 0;
244 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
245 &forward_wakeup_use_loop, 0,
246 "Use a loop to find idle cpus");
247
248 #endif
249 #if 0
250 static int sched_followon = 0;
251 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
252 &sched_followon, 0,
253 "allow threads to share a quantum");
254 #endif
255
256 SDT_PROVIDER_DEFINE(sched);
257
258 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
259 "struct proc *", "uint8_t");
260 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
261 "struct proc *", "void *");
262 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
263 "struct proc *", "void *", "int");
264 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
265 "struct proc *", "uint8_t", "struct thread *");
266 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
267 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
268 "struct proc *");
269 SDT_PROBE_DEFINE(sched, , , on__cpu);
270 SDT_PROBE_DEFINE(sched, , , remain__cpu);
271 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
272 "struct proc *");
273
274 static __inline void
sched_load_add(void)275 sched_load_add(void)
276 {
277
278 sched_tdcnt++;
279 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
280 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
281 }
282
283 static __inline void
sched_load_rem(void)284 sched_load_rem(void)
285 {
286
287 sched_tdcnt--;
288 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
289 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
290 }
291 /*
292 * Arrange to reschedule if necessary, taking the priorities and
293 * schedulers into account.
294 */
295 static void
maybe_resched(struct thread * td)296 maybe_resched(struct thread *td)
297 {
298
299 THREAD_LOCK_ASSERT(td, MA_OWNED);
300 if (td->td_priority < curthread->td_priority)
301 curthread->td_flags |= TDF_NEEDRESCHED;
302 }
303
304 /*
305 * This function is called when a thread is about to be put on run queue
306 * because it has been made runnable or its priority has been adjusted. It
307 * determines if the new thread should preempt the current thread. If so,
308 * it sets td_owepreempt to request a preemption.
309 */
310 int
maybe_preempt(struct thread * td)311 maybe_preempt(struct thread *td)
312 {
313 #ifdef PREEMPTION
314 struct thread *ctd;
315 int cpri, pri;
316
317 /*
318 * The new thread should not preempt the current thread if any of the
319 * following conditions are true:
320 *
321 * - The kernel is in the throes of crashing (panicstr).
322 * - The current thread has a higher (numerically lower) or
323 * equivalent priority. Note that this prevents curthread from
324 * trying to preempt to itself.
325 * - It is too early in the boot for context switches (cold is set).
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 (panicstr != NULL || pri >= cpri || cold /* || 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 (td_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 td_estcpu and p_cpticks asynchronously.
370 *
371 * We wish to decay away 90% of td_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 * td_estcpu *= decay;
376 * will compute
377 * td_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 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
431 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
432 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
433
434 /*
435 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
436 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
437 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
438 *
439 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
440 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
441 *
442 * If you don't want to bother with the faster/more-accurate formula, you
443 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
444 * (more general) method of calculating the %age of CPU used by a process.
445 */
446 #define CCPU_SHIFT 11
447
448 /*
449 * Recompute process priorities, every hz ticks.
450 * MP-safe, called without the Giant mutex.
451 */
452 /* ARGSUSED */
453 static void
schedcpu(void)454 schedcpu(void)
455 {
456 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
457 struct thread *td;
458 struct proc *p;
459 struct td_sched *ts;
460 int awake;
461
462 sx_slock(&allproc_lock);
463 FOREACH_PROC_IN_SYSTEM(p) {
464 PROC_LOCK(p);
465 if (p->p_state == PRS_NEW) {
466 PROC_UNLOCK(p);
467 continue;
468 }
469 FOREACH_THREAD_IN_PROC(p, td) {
470 awake = 0;
471 thread_lock(td);
472 ts = td->td_sched;
473 /*
474 * Increment sleep time (if sleeping). We
475 * ignore overflow, as above.
476 */
477 /*
478 * The td_sched slptimes are not touched in wakeup
479 * because the thread may not HAVE everything in
480 * memory? XXX I think this is out of date.
481 */
482 if (TD_ON_RUNQ(td)) {
483 awake = 1;
484 td->td_flags &= ~TDF_DIDRUN;
485 } else if (TD_IS_RUNNING(td)) {
486 awake = 1;
487 /* Do not clear TDF_DIDRUN */
488 } else if (td->td_flags & TDF_DIDRUN) {
489 awake = 1;
490 td->td_flags &= ~TDF_DIDRUN;
491 }
492
493 /*
494 * ts_pctcpu is only for ps and ttyinfo().
495 */
496 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
497 /*
498 * If the td_sched has been idle the entire second,
499 * stop recalculating its priority until
500 * it wakes up.
501 */
502 if (ts->ts_cpticks != 0) {
503 #if (FSHIFT >= CCPU_SHIFT)
504 ts->ts_pctcpu += (realstathz == 100)
505 ? ((fixpt_t) ts->ts_cpticks) <<
506 (FSHIFT - CCPU_SHIFT) :
507 100 * (((fixpt_t) ts->ts_cpticks)
508 << (FSHIFT - CCPU_SHIFT)) / realstathz;
509 #else
510 ts->ts_pctcpu += ((FSCALE - ccpu) *
511 (ts->ts_cpticks *
512 FSCALE / realstathz)) >> FSHIFT;
513 #endif
514 ts->ts_cpticks = 0;
515 }
516 /*
517 * If there are ANY running threads in this process,
518 * then don't count it as sleeping.
519 * XXX: this is broken.
520 */
521 if (awake) {
522 if (ts->ts_slptime > 1) {
523 /*
524 * In an ideal world, this should not
525 * happen, because whoever woke us
526 * up from the long sleep should have
527 * unwound the slptime and reset our
528 * priority before we run at the stale
529 * priority. Should KASSERT at some
530 * point when all the cases are fixed.
531 */
532 updatepri(td);
533 }
534 ts->ts_slptime = 0;
535 } else
536 ts->ts_slptime++;
537 if (ts->ts_slptime > 1) {
538 thread_unlock(td);
539 continue;
540 }
541 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
542 resetpriority(td);
543 resetpriority_thread(td);
544 thread_unlock(td);
545 }
546 PROC_UNLOCK(p);
547 }
548 sx_sunlock(&allproc_lock);
549 }
550
551 /*
552 * Main loop for a kthread that executes schedcpu once a second.
553 */
554 static void
schedcpu_thread(void)555 schedcpu_thread(void)
556 {
557
558 for (;;) {
559 schedcpu();
560 pause("-", hz);
561 }
562 }
563
564 /*
565 * Recalculate the priority of a process after it has slept for a while.
566 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
567 * least six times the loadfactor will decay td_estcpu to zero.
568 */
569 static void
updatepri(struct thread * td)570 updatepri(struct thread *td)
571 {
572 struct td_sched *ts;
573 fixpt_t loadfac;
574 unsigned int newcpu;
575
576 ts = td->td_sched;
577 loadfac = loadfactor(averunnable.ldavg[0]);
578 if (ts->ts_slptime > 5 * loadfac)
579 td->td_estcpu = 0;
580 else {
581 newcpu = td->td_estcpu;
582 ts->ts_slptime--; /* was incremented in schedcpu() */
583 while (newcpu && --ts->ts_slptime)
584 newcpu = decay_cpu(loadfac, newcpu);
585 td->td_estcpu = newcpu;
586 }
587 }
588
589 /*
590 * Compute the priority of a process when running in user mode.
591 * Arrange to reschedule if the resulting priority is better
592 * than that of the current process.
593 */
594 static void
resetpriority(struct thread * td)595 resetpriority(struct thread *td)
596 {
597 register unsigned int newpriority;
598
599 if (td->td_pri_class == PRI_TIMESHARE) {
600 newpriority = PUSER + td->td_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 /*
609 * Update the thread's priority when the associated process's user
610 * priority changes.
611 */
612 static void
resetpriority_thread(struct thread * td)613 resetpriority_thread(struct thread *td)
614 {
615
616 /* Only change threads with a time sharing user priority. */
617 if (td->td_priority < PRI_MIN_TIMESHARE ||
618 td->td_priority > PRI_MAX_TIMESHARE)
619 return;
620
621 /* XXX the whole needresched thing is broken, but not silly. */
622 maybe_resched(td);
623
624 sched_prio(td, td->td_user_pri);
625 }
626
627 /* ARGSUSED */
628 static void
sched_setup(void * dummy)629 sched_setup(void *dummy)
630 {
631
632 setup_runqs();
633
634 /* Account for thread0. */
635 sched_load_add();
636 }
637
638 /*
639 * This routine determines time constants after stathz and hz are setup.
640 */
641 static void
sched_initticks(void * dummy)642 sched_initticks(void *dummy)
643 {
644
645 realstathz = stathz ? stathz : hz;
646 sched_slice = realstathz / 10; /* ~100ms */
647 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
648 realstathz);
649 }
650
651 /* External interfaces start here */
652
653 /*
654 * Very early in the boot some setup of scheduler-specific
655 * parts of proc0 and of some scheduler resources needs to be done.
656 * Called from:
657 * proc0_init()
658 */
659 void
schedinit(void)660 schedinit(void)
661 {
662 /*
663 * Set up the scheduler specific parts of proc0.
664 */
665 proc0.p_sched = NULL; /* XXX */
666 thread0.td_sched = &td_sched0;
667 thread0.td_lock = &sched_lock;
668 td_sched0.ts_slice = sched_slice;
669 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
670 }
671
672 int
sched_runnable(void)673 sched_runnable(void)
674 {
675 #ifdef SMP
676 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
677 #else
678 return runq_check(&runq);
679 #endif
680 }
681
682 int
sched_rr_interval(void)683 sched_rr_interval(void)
684 {
685
686 /* Convert sched_slice from stathz to hz. */
687 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
688 }
689
690 /*
691 * We adjust the priority of the current process. The priority of
692 * a process gets worse as it accumulates CPU time. The cpu usage
693 * estimator (td_estcpu) is increased here. resetpriority() will
694 * compute a different priority each time td_estcpu increases by
695 * INVERSE_ESTCPU_WEIGHT
696 * (until MAXPRI is reached). The cpu usage estimator ramps up
697 * quite quickly when the process is running (linearly), and decays
698 * away exponentially, at a rate which is proportionally slower when
699 * the system is busy. The basic principle is that the system will
700 * 90% forget that the process used a lot of CPU time in 5 * loadav
701 * seconds. This causes the system to favor processes which haven't
702 * run much recently, and to round-robin among other processes.
703 */
704 void
sched_clock(struct thread * td)705 sched_clock(struct thread *td)
706 {
707 struct pcpuidlestat *stat;
708 struct td_sched *ts;
709
710 THREAD_LOCK_ASSERT(td, MA_OWNED);
711 ts = td->td_sched;
712
713 ts->ts_cpticks++;
714 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
715 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
716 resetpriority(td);
717 resetpriority_thread(td);
718 }
719
720 /*
721 * Force a context switch if the current thread has used up a full
722 * time slice (default is 100ms).
723 */
724 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
725 ts->ts_slice = sched_slice;
726 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
727 }
728
729 stat = DPCPU_PTR(idlestat);
730 stat->oldidlecalls = stat->idlecalls;
731 stat->idlecalls = 0;
732 }
733
734 /*
735 * Charge child's scheduling CPU usage to parent.
736 */
737 void
sched_exit(struct proc * p,struct thread * td)738 sched_exit(struct proc *p, struct thread *td)
739 {
740
741 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
742 "prio:%d", td->td_priority);
743
744 PROC_LOCK_ASSERT(p, MA_OWNED);
745 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
746 }
747
748 void
sched_exit_thread(struct thread * td,struct thread * child)749 sched_exit_thread(struct thread *td, struct thread *child)
750 {
751
752 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
753 "prio:%d", child->td_priority);
754 thread_lock(td);
755 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
756 thread_unlock(td);
757 thread_lock(child);
758 if ((child->td_flags & TDF_NOLOAD) == 0)
759 sched_load_rem();
760 thread_unlock(child);
761 }
762
763 void
sched_fork(struct thread * td,struct thread * childtd)764 sched_fork(struct thread *td, struct thread *childtd)
765 {
766 sched_fork_thread(td, childtd);
767 }
768
769 void
sched_fork_thread(struct thread * td,struct thread * childtd)770 sched_fork_thread(struct thread *td, struct thread *childtd)
771 {
772 struct td_sched *ts;
773
774 childtd->td_oncpu = NOCPU;
775 childtd->td_lastcpu = NOCPU;
776 childtd->td_estcpu = td->td_estcpu;
777 childtd->td_lock = &sched_lock;
778 childtd->td_cpuset = cpuset_ref(td->td_cpuset);
779 childtd->td_priority = childtd->td_base_pri;
780 ts = childtd->td_sched;
781 bzero(ts, sizeof(*ts));
782 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
783 ts->ts_slice = 1;
784 }
785
786 void
sched_nice(struct proc * p,int nice)787 sched_nice(struct proc *p, int nice)
788 {
789 struct thread *td;
790
791 PROC_LOCK_ASSERT(p, MA_OWNED);
792 p->p_nice = nice;
793 FOREACH_THREAD_IN_PROC(p, td) {
794 thread_lock(td);
795 resetpriority(td);
796 resetpriority_thread(td);
797 thread_unlock(td);
798 }
799 }
800
801 void
sched_class(struct thread * td,int class)802 sched_class(struct thread *td, int class)
803 {
804 THREAD_LOCK_ASSERT(td, MA_OWNED);
805 td->td_pri_class = class;
806 }
807
808 /*
809 * Adjust the priority of a thread.
810 */
811 static void
sched_priority(struct thread * td,u_char prio)812 sched_priority(struct thread *td, u_char prio)
813 {
814
815
816 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
817 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
818 sched_tdname(curthread));
819 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
820 if (td != curthread && prio > td->td_priority) {
821 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
822 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
823 prio, KTR_ATTR_LINKED, sched_tdname(td));
824 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
825 curthread);
826 }
827 THREAD_LOCK_ASSERT(td, MA_OWNED);
828 if (td->td_priority == prio)
829 return;
830 td->td_priority = prio;
831 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
832 sched_rem(td);
833 sched_add(td, SRQ_BORING);
834 }
835 }
836
837 /*
838 * Update a thread's priority when it is lent another thread's
839 * priority.
840 */
841 void
sched_lend_prio(struct thread * td,u_char prio)842 sched_lend_prio(struct thread *td, u_char prio)
843 {
844
845 td->td_flags |= TDF_BORROWING;
846 sched_priority(td, prio);
847 }
848
849 /*
850 * Restore a thread's priority when priority propagation is
851 * over. The prio argument is the minimum priority the thread
852 * needs to have to satisfy other possible priority lending
853 * requests. If the thread's regulary priority is less
854 * important than prio the thread will keep a priority boost
855 * of prio.
856 */
857 void
sched_unlend_prio(struct thread * td,u_char prio)858 sched_unlend_prio(struct thread *td, u_char prio)
859 {
860 u_char base_pri;
861
862 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
863 td->td_base_pri <= PRI_MAX_TIMESHARE)
864 base_pri = td->td_user_pri;
865 else
866 base_pri = td->td_base_pri;
867 if (prio >= base_pri) {
868 td->td_flags &= ~TDF_BORROWING;
869 sched_prio(td, base_pri);
870 } else
871 sched_lend_prio(td, prio);
872 }
873
874 void
sched_prio(struct thread * td,u_char prio)875 sched_prio(struct thread *td, u_char prio)
876 {
877 u_char oldprio;
878
879 /* First, update the base priority. */
880 td->td_base_pri = prio;
881
882 /*
883 * If the thread is borrowing another thread's priority, don't ever
884 * lower the priority.
885 */
886 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
887 return;
888
889 /* Change the real priority. */
890 oldprio = td->td_priority;
891 sched_priority(td, prio);
892
893 /*
894 * If the thread is on a turnstile, then let the turnstile update
895 * its state.
896 */
897 if (TD_ON_LOCK(td) && oldprio != prio)
898 turnstile_adjust(td, oldprio);
899 }
900
901 void
sched_user_prio(struct thread * td,u_char prio)902 sched_user_prio(struct thread *td, u_char prio)
903 {
904
905 THREAD_LOCK_ASSERT(td, MA_OWNED);
906 td->td_base_user_pri = prio;
907 if (td->td_lend_user_pri <= prio)
908 return;
909 td->td_user_pri = prio;
910 }
911
912 void
sched_lend_user_prio(struct thread * td,u_char prio)913 sched_lend_user_prio(struct thread *td, u_char prio)
914 {
915
916 THREAD_LOCK_ASSERT(td, MA_OWNED);
917 td->td_lend_user_pri = prio;
918 td->td_user_pri = min(prio, td->td_base_user_pri);
919 if (td->td_priority > td->td_user_pri)
920 sched_prio(td, td->td_user_pri);
921 else if (td->td_priority != td->td_user_pri)
922 td->td_flags |= TDF_NEEDRESCHED;
923 }
924
925 void
sched_sleep(struct thread * td,int pri)926 sched_sleep(struct thread *td, int pri)
927 {
928
929 THREAD_LOCK_ASSERT(td, MA_OWNED);
930 td->td_slptick = ticks;
931 td->td_sched->ts_slptime = 0;
932 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
933 sched_prio(td, pri);
934 if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
935 td->td_flags |= TDF_CANSWAP;
936 }
937
938 void
sched_switch(struct thread * td,struct thread * newtd,int flags)939 sched_switch(struct thread *td, struct thread *newtd, int flags)
940 {
941 struct mtx *tmtx;
942 struct td_sched *ts;
943 struct proc *p;
944 int preempted;
945
946 tmtx = NULL;
947 ts = td->td_sched;
948 p = td->td_proc;
949
950 THREAD_LOCK_ASSERT(td, MA_OWNED);
951
952 /*
953 * Switch to the sched lock to fix things up and pick
954 * a new thread.
955 * Block the td_lock in order to avoid breaking the critical path.
956 */
957 if (td->td_lock != &sched_lock) {
958 mtx_lock_spin(&sched_lock);
959 tmtx = thread_lock_block(td);
960 }
961
962 if ((td->td_flags & TDF_NOLOAD) == 0)
963 sched_load_rem();
964
965 td->td_lastcpu = td->td_oncpu;
966 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
967 (flags & SW_PREEMPT) != 0;
968 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
969 td->td_owepreempt = 0;
970 td->td_oncpu = NOCPU;
971
972 /*
973 * At the last moment, if this thread is still marked RUNNING,
974 * then put it back on the run queue as it has not been suspended
975 * or stopped or any thing else similar. We never put the idle
976 * threads on the run queue, however.
977 */
978 if (td->td_flags & TDF_IDLETD) {
979 TD_SET_CAN_RUN(td);
980 #ifdef SMP
981 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask);
982 #endif
983 } else {
984 if (TD_IS_RUNNING(td)) {
985 /* Put us back on the run queue. */
986 sched_add(td, preempted ?
987 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
988 SRQ_OURSELF|SRQ_YIELDING);
989 }
990 }
991 if (newtd) {
992 /*
993 * The thread we are about to run needs to be counted
994 * as if it had been added to the run queue and selected.
995 * It came from:
996 * * A preemption
997 * * An upcall
998 * * A followon
999 */
1000 KASSERT((newtd->td_inhibitors == 0),
1001 ("trying to run inhibited thread"));
1002 newtd->td_flags |= TDF_DIDRUN;
1003 TD_SET_RUNNING(newtd);
1004 if ((newtd->td_flags & TDF_NOLOAD) == 0)
1005 sched_load_add();
1006 } else {
1007 newtd = choosethread();
1008 MPASS(newtd->td_lock == &sched_lock);
1009 }
1010
1011 #if (KTR_COMPILE & KTR_SCHED) != 0
1012 if (TD_IS_IDLETHREAD(td))
1013 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1014 "prio:%d", td->td_priority);
1015 else
1016 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1017 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1018 "lockname:\"%s\"", td->td_lockname);
1019 #endif
1020
1021 if (td != newtd) {
1022 #ifdef HWPMC_HOOKS
1023 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1024 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1025 #endif
1026
1027 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1028
1029 /* I feel sleepy */
1030 lock_profile_release_lock(&sched_lock.lock_object);
1031 #ifdef KDTRACE_HOOKS
1032 /*
1033 * If DTrace has set the active vtime enum to anything
1034 * other than INACTIVE (0), then it should have set the
1035 * function to call.
1036 */
1037 if (dtrace_vtime_active)
1038 (*dtrace_vtime_switch_func)(newtd);
1039 #endif
1040
1041 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
1042 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1043 0, 0, __FILE__, __LINE__);
1044 /*
1045 * Where am I? What year is it?
1046 * We are in the same thread that went to sleep above,
1047 * but any amount of time may have passed. All our context
1048 * will still be available as will local variables.
1049 * PCPU values however may have changed as we may have
1050 * changed CPU so don't trust cached values of them.
1051 * New threads will go to fork_exit() instead of here
1052 * so if you change things here you may need to change
1053 * things there too.
1054 *
1055 * If the thread above was exiting it will never wake
1056 * up again here, so either it has saved everything it
1057 * needed to, or the thread_wait() or wait() will
1058 * need to reap it.
1059 */
1060
1061 SDT_PROBE0(sched, , , on__cpu);
1062 #ifdef HWPMC_HOOKS
1063 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1064 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1065 #endif
1066 } else
1067 SDT_PROBE0(sched, , , remain__cpu);
1068
1069 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1070 "prio:%d", td->td_priority);
1071
1072 #ifdef SMP
1073 if (td->td_flags & TDF_IDLETD)
1074 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask);
1075 #endif
1076 sched_lock.mtx_lock = (uintptr_t)td;
1077 td->td_oncpu = PCPU_GET(cpuid);
1078 MPASS(td->td_lock == &sched_lock);
1079 }
1080
1081 void
sched_wakeup(struct thread * td)1082 sched_wakeup(struct thread *td)
1083 {
1084 struct td_sched *ts;
1085
1086 THREAD_LOCK_ASSERT(td, MA_OWNED);
1087 ts = td->td_sched;
1088 td->td_flags &= ~TDF_CANSWAP;
1089 if (ts->ts_slptime > 1) {
1090 updatepri(td);
1091 resetpriority(td);
1092 }
1093 td->td_slptick = 0;
1094 ts->ts_slptime = 0;
1095 ts->ts_slice = sched_slice;
1096 sched_add(td, SRQ_BORING);
1097 }
1098
1099 #ifdef SMP
1100 static int
forward_wakeup(int cpunum)1101 forward_wakeup(int cpunum)
1102 {
1103 struct pcpu *pc;
1104 cpuset_t dontuse, map, map2;
1105 u_int id, me;
1106 int iscpuset;
1107
1108 mtx_assert(&sched_lock, MA_OWNED);
1109
1110 CTR0(KTR_RUNQ, "forward_wakeup()");
1111
1112 if ((!forward_wakeup_enabled) ||
1113 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
1114 return (0);
1115 if (!smp_started || cold || panicstr)
1116 return (0);
1117
1118 forward_wakeups_requested++;
1119
1120 /*
1121 * Check the idle mask we received against what we calculated
1122 * before in the old version.
1123 */
1124 me = PCPU_GET(cpuid);
1125
1126 /* Don't bother if we should be doing it ourself. */
1127 if (CPU_ISSET(me, &idle_cpus_mask) &&
1128 (cpunum == NOCPU || me == cpunum))
1129 return (0);
1130
1131 CPU_SETOF(me, &dontuse);
1132 CPU_OR(&dontuse, &stopped_cpus);
1133 CPU_OR(&dontuse, &hlt_cpus_mask);
1134 CPU_ZERO(&map2);
1135 if (forward_wakeup_use_loop) {
1136 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1137 id = pc->pc_cpuid;
1138 if (!CPU_ISSET(id, &dontuse) &&
1139 pc->pc_curthread == pc->pc_idlethread) {
1140 CPU_SET(id, &map2);
1141 }
1142 }
1143 }
1144
1145 if (forward_wakeup_use_mask) {
1146 map = idle_cpus_mask;
1147 CPU_NAND(&map, &dontuse);
1148
1149 /* If they are both on, compare and use loop if different. */
1150 if (forward_wakeup_use_loop) {
1151 if (CPU_CMP(&map, &map2)) {
1152 printf("map != map2, loop method preferred\n");
1153 map = map2;
1154 }
1155 }
1156 } else {
1157 map = map2;
1158 }
1159
1160 /* If we only allow a specific CPU, then mask off all the others. */
1161 if (cpunum != NOCPU) {
1162 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
1163 iscpuset = CPU_ISSET(cpunum, &map);
1164 if (iscpuset == 0)
1165 CPU_ZERO(&map);
1166 else
1167 CPU_SETOF(cpunum, &map);
1168 }
1169 if (!CPU_EMPTY(&map)) {
1170 forward_wakeups_delivered++;
1171 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1172 id = pc->pc_cpuid;
1173 if (!CPU_ISSET(id, &map))
1174 continue;
1175 if (cpu_idle_wakeup(pc->pc_cpuid))
1176 CPU_CLR(id, &map);
1177 }
1178 if (!CPU_EMPTY(&map))
1179 ipi_selected(map, IPI_AST);
1180 return (1);
1181 }
1182 if (cpunum == NOCPU)
1183 printf("forward_wakeup: Idle processor not found\n");
1184 return (0);
1185 }
1186
1187 static void
kick_other_cpu(int pri,int cpuid)1188 kick_other_cpu(int pri, int cpuid)
1189 {
1190 struct pcpu *pcpu;
1191 int cpri;
1192
1193 pcpu = pcpu_find(cpuid);
1194 if (CPU_ISSET(cpuid, &idle_cpus_mask)) {
1195 forward_wakeups_delivered++;
1196 if (!cpu_idle_wakeup(cpuid))
1197 ipi_cpu(cpuid, IPI_AST);
1198 return;
1199 }
1200
1201 cpri = pcpu->pc_curthread->td_priority;
1202 if (pri >= cpri)
1203 return;
1204
1205 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
1206 #if !defined(FULL_PREEMPTION)
1207 if (pri <= PRI_MAX_ITHD)
1208 #endif /* ! FULL_PREEMPTION */
1209 {
1210 ipi_cpu(cpuid, IPI_PREEMPT);
1211 return;
1212 }
1213 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
1214
1215 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
1216 ipi_cpu(cpuid, IPI_AST);
1217 return;
1218 }
1219 #endif /* SMP */
1220
1221 #ifdef SMP
1222 static int
sched_pickcpu(struct thread * td)1223 sched_pickcpu(struct thread *td)
1224 {
1225 int best, cpu;
1226
1227 mtx_assert(&sched_lock, MA_OWNED);
1228
1229 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu))
1230 best = td->td_lastcpu;
1231 else
1232 best = NOCPU;
1233 CPU_FOREACH(cpu) {
1234 if (!THREAD_CAN_SCHED(td, cpu))
1235 continue;
1236
1237 if (best == NOCPU)
1238 best = cpu;
1239 else if (runq_length[cpu] < runq_length[best])
1240 best = cpu;
1241 }
1242 KASSERT(best != NOCPU, ("no valid CPUs"));
1243
1244 return (best);
1245 }
1246 #endif
1247
1248 void
sched_add(struct thread * td,int flags)1249 sched_add(struct thread *td, int flags)
1250 #ifdef SMP
1251 {
1252 cpuset_t tidlemsk;
1253 struct td_sched *ts;
1254 u_int cpu, cpuid;
1255 int forwarded = 0;
1256 int single_cpu = 0;
1257
1258 ts = td->td_sched;
1259 THREAD_LOCK_ASSERT(td, MA_OWNED);
1260 KASSERT((td->td_inhibitors == 0),
1261 ("sched_add: trying to run inhibited thread"));
1262 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1263 ("sched_add: bad thread state"));
1264 KASSERT(td->td_flags & TDF_INMEM,
1265 ("sched_add: thread swapped out"));
1266
1267 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1268 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1269 sched_tdname(curthread));
1270 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1271 KTR_ATTR_LINKED, sched_tdname(td));
1272 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1273 flags & SRQ_PREEMPTED);
1274
1275
1276 /*
1277 * Now that the thread is moving to the run-queue, set the lock
1278 * to the scheduler's lock.
1279 */
1280 if (td->td_lock != &sched_lock) {
1281 mtx_lock_spin(&sched_lock);
1282 thread_lock_set(td, &sched_lock);
1283 }
1284 TD_SET_RUNQ(td);
1285
1286 /*
1287 * If SMP is started and the thread is pinned or otherwise limited to
1288 * a specific set of CPUs, queue the thread to a per-CPU run queue.
1289 * Otherwise, queue the thread to the global run queue.
1290 *
1291 * If SMP has not yet been started we must use the global run queue
1292 * as per-CPU state may not be initialized yet and we may crash if we
1293 * try to access the per-CPU run queues.
1294 */
1295 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
1296 ts->ts_flags & TSF_AFFINITY)) {
1297 if (td->td_pinned != 0)
1298 cpu = td->td_lastcpu;
1299 else if (td->td_flags & TDF_BOUND) {
1300 /* Find CPU from bound runq. */
1301 KASSERT(SKE_RUNQ_PCPU(ts),
1302 ("sched_add: bound td_sched not on cpu runq"));
1303 cpu = ts->ts_runq - &runq_pcpu[0];
1304 } else
1305 /* Find a valid CPU for our cpuset */
1306 cpu = sched_pickcpu(td);
1307 ts->ts_runq = &runq_pcpu[cpu];
1308 single_cpu = 1;
1309 CTR3(KTR_RUNQ,
1310 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
1311 cpu);
1312 } else {
1313 CTR2(KTR_RUNQ,
1314 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
1315 td);
1316 cpu = NOCPU;
1317 ts->ts_runq = &runq;
1318 }
1319
1320 if ((td->td_flags & TDF_NOLOAD) == 0)
1321 sched_load_add();
1322 runq_add(ts->ts_runq, td, flags);
1323 if (cpu != NOCPU)
1324 runq_length[cpu]++;
1325
1326 cpuid = PCPU_GET(cpuid);
1327 if (single_cpu && cpu != cpuid) {
1328 kick_other_cpu(td->td_priority, cpu);
1329 } else {
1330 if (!single_cpu) {
1331 tidlemsk = idle_cpus_mask;
1332 CPU_NAND(&tidlemsk, &hlt_cpus_mask);
1333 CPU_CLR(cpuid, &tidlemsk);
1334
1335 if (!CPU_ISSET(cpuid, &idle_cpus_mask) &&
1336 ((flags & SRQ_INTR) == 0) &&
1337 !CPU_EMPTY(&tidlemsk))
1338 forwarded = forward_wakeup(cpu);
1339 }
1340
1341 if (!forwarded) {
1342 if (!maybe_preempt(td))
1343 maybe_resched(td);
1344 }
1345 }
1346 }
1347 #else /* SMP */
1348 {
1349 struct td_sched *ts;
1350
1351 ts = td->td_sched;
1352 THREAD_LOCK_ASSERT(td, MA_OWNED);
1353 KASSERT((td->td_inhibitors == 0),
1354 ("sched_add: trying to run inhibited thread"));
1355 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1356 ("sched_add: bad thread state"));
1357 KASSERT(td->td_flags & TDF_INMEM,
1358 ("sched_add: thread swapped out"));
1359 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
1360 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1361 sched_tdname(curthread));
1362 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
1363 KTR_ATTR_LINKED, sched_tdname(td));
1364 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
1365 flags & SRQ_PREEMPTED);
1366
1367 /*
1368 * Now that the thread is moving to the run-queue, set the lock
1369 * to the scheduler's lock.
1370 */
1371 if (td->td_lock != &sched_lock) {
1372 mtx_lock_spin(&sched_lock);
1373 thread_lock_set(td, &sched_lock);
1374 }
1375 TD_SET_RUNQ(td);
1376 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
1377 ts->ts_runq = &runq;
1378
1379 if ((td->td_flags & TDF_NOLOAD) == 0)
1380 sched_load_add();
1381 runq_add(ts->ts_runq, td, flags);
1382 if (!maybe_preempt(td))
1383 maybe_resched(td);
1384 }
1385 #endif /* SMP */
1386
1387 void
sched_rem(struct thread * td)1388 sched_rem(struct thread *td)
1389 {
1390 struct td_sched *ts;
1391
1392 ts = td->td_sched;
1393 KASSERT(td->td_flags & TDF_INMEM,
1394 ("sched_rem: thread swapped out"));
1395 KASSERT(TD_ON_RUNQ(td),
1396 ("sched_rem: thread not on run queue"));
1397 mtx_assert(&sched_lock, MA_OWNED);
1398 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
1399 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
1400 sched_tdname(curthread));
1401 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
1402
1403 if ((td->td_flags & TDF_NOLOAD) == 0)
1404 sched_load_rem();
1405 #ifdef SMP
1406 if (ts->ts_runq != &runq)
1407 runq_length[ts->ts_runq - runq_pcpu]--;
1408 #endif
1409 runq_remove(ts->ts_runq, td);
1410 TD_SET_CAN_RUN(td);
1411 }
1412
1413 /*
1414 * Select threads to run. Note that running threads still consume a
1415 * slot.
1416 */
1417 struct thread *
sched_choose(void)1418 sched_choose(void)
1419 {
1420 struct thread *td;
1421 struct runq *rq;
1422
1423 mtx_assert(&sched_lock, MA_OWNED);
1424 #ifdef SMP
1425 struct thread *tdcpu;
1426
1427 rq = &runq;
1428 td = runq_choose_fuzz(&runq, runq_fuzz);
1429 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
1430
1431 if (td == NULL ||
1432 (tdcpu != NULL &&
1433 tdcpu->td_priority < td->td_priority)) {
1434 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
1435 PCPU_GET(cpuid));
1436 td = tdcpu;
1437 rq = &runq_pcpu[PCPU_GET(cpuid)];
1438 } else {
1439 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
1440 }
1441
1442 #else
1443 rq = &runq;
1444 td = runq_choose(&runq);
1445 #endif
1446
1447 if (td) {
1448 #ifdef SMP
1449 if (td == tdcpu)
1450 runq_length[PCPU_GET(cpuid)]--;
1451 #endif
1452 runq_remove(rq, td);
1453 td->td_flags |= TDF_DIDRUN;
1454
1455 KASSERT(td->td_flags & TDF_INMEM,
1456 ("sched_choose: thread swapped out"));
1457 return (td);
1458 }
1459 return (PCPU_GET(idlethread));
1460 }
1461
1462 void
sched_preempt(struct thread * td)1463 sched_preempt(struct thread *td)
1464 {
1465
1466 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
1467 thread_lock(td);
1468 if (td->td_critnest > 1)
1469 td->td_owepreempt = 1;
1470 else
1471 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
1472 thread_unlock(td);
1473 }
1474
1475 void
sched_userret(struct thread * td)1476 sched_userret(struct thread *td)
1477 {
1478 /*
1479 * XXX we cheat slightly on the locking here to avoid locking in
1480 * the usual case. Setting td_priority here is essentially an
1481 * incomplete workaround for not setting it properly elsewhere.
1482 * Now that some interrupt handlers are threads, not setting it
1483 * properly elsewhere can clobber it in the window between setting
1484 * it here and returning to user mode, so don't waste time setting
1485 * it perfectly here.
1486 */
1487 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1488 ("thread with borrowed priority returning to userland"));
1489 if (td->td_priority != td->td_user_pri) {
1490 thread_lock(td);
1491 td->td_priority = td->td_user_pri;
1492 td->td_base_pri = td->td_user_pri;
1493 thread_unlock(td);
1494 }
1495 }
1496
1497 void
sched_bind(struct thread * td,int cpu)1498 sched_bind(struct thread *td, int cpu)
1499 {
1500 struct td_sched *ts;
1501
1502 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
1503 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
1504
1505 ts = td->td_sched;
1506
1507 td->td_flags |= TDF_BOUND;
1508 #ifdef SMP
1509 ts->ts_runq = &runq_pcpu[cpu];
1510 if (PCPU_GET(cpuid) == cpu)
1511 return;
1512
1513 mi_switch(SW_VOL, NULL);
1514 #endif
1515 }
1516
1517 void
sched_unbind(struct thread * td)1518 sched_unbind(struct thread* td)
1519 {
1520 THREAD_LOCK_ASSERT(td, MA_OWNED);
1521 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
1522 td->td_flags &= ~TDF_BOUND;
1523 }
1524
1525 int
sched_is_bound(struct thread * td)1526 sched_is_bound(struct thread *td)
1527 {
1528 THREAD_LOCK_ASSERT(td, MA_OWNED);
1529 return (td->td_flags & TDF_BOUND);
1530 }
1531
1532 void
sched_relinquish(struct thread * td)1533 sched_relinquish(struct thread *td)
1534 {
1535 thread_lock(td);
1536 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
1537 thread_unlock(td);
1538 }
1539
1540 int
sched_load(void)1541 sched_load(void)
1542 {
1543 return (sched_tdcnt);
1544 }
1545
1546 int
sched_sizeof_proc(void)1547 sched_sizeof_proc(void)
1548 {
1549 return (sizeof(struct proc));
1550 }
1551
1552 int
sched_sizeof_thread(void)1553 sched_sizeof_thread(void)
1554 {
1555 return (sizeof(struct thread) + sizeof(struct td_sched));
1556 }
1557
1558 fixpt_t
sched_pctcpu(struct thread * td)1559 sched_pctcpu(struct thread *td)
1560 {
1561 struct td_sched *ts;
1562
1563 THREAD_LOCK_ASSERT(td, MA_OWNED);
1564 ts = td->td_sched;
1565 return (ts->ts_pctcpu);
1566 }
1567
1568 #ifdef RACCT
1569 /*
1570 * Calculates the contribution to the thread cpu usage for the latest
1571 * (unfinished) second.
1572 */
1573 fixpt_t
sched_pctcpu_delta(struct thread * td)1574 sched_pctcpu_delta(struct thread *td)
1575 {
1576 struct td_sched *ts;
1577 fixpt_t delta;
1578 int realstathz;
1579
1580 THREAD_LOCK_ASSERT(td, MA_OWNED);
1581 ts = td->td_sched;
1582 delta = 0;
1583 realstathz = stathz ? stathz : hz;
1584 if (ts->ts_cpticks != 0) {
1585 #if (FSHIFT >= CCPU_SHIFT)
1586 delta = (realstathz == 100)
1587 ? ((fixpt_t) ts->ts_cpticks) <<
1588 (FSHIFT - CCPU_SHIFT) :
1589 100 * (((fixpt_t) ts->ts_cpticks)
1590 << (FSHIFT - CCPU_SHIFT)) / realstathz;
1591 #else
1592 delta = ((FSCALE - ccpu) *
1593 (ts->ts_cpticks *
1594 FSCALE / realstathz)) >> FSHIFT;
1595 #endif
1596 }
1597
1598 return (delta);
1599 }
1600 #endif
1601
1602 void
sched_tick(int cnt)1603 sched_tick(int cnt)
1604 {
1605 }
1606
1607 /*
1608 * The actual idle process.
1609 */
1610 void
sched_idletd(void * dummy)1611 sched_idletd(void *dummy)
1612 {
1613 struct pcpuidlestat *stat;
1614
1615 THREAD_NO_SLEEPING();
1616 stat = DPCPU_PTR(idlestat);
1617 for (;;) {
1618 mtx_assert(&Giant, MA_NOTOWNED);
1619
1620 while (sched_runnable() == 0) {
1621 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
1622 stat->idlecalls++;
1623 }
1624
1625 mtx_lock_spin(&sched_lock);
1626 mi_switch(SW_VOL | SWT_IDLE, NULL);
1627 mtx_unlock_spin(&sched_lock);
1628 }
1629 }
1630
1631 /*
1632 * A CPU is entering for the first time or a thread is exiting.
1633 */
1634 void
sched_throw(struct thread * td)1635 sched_throw(struct thread *td)
1636 {
1637 /*
1638 * Correct spinlock nesting. The idle thread context that we are
1639 * borrowing was created so that it would start out with a single
1640 * spin lock (sched_lock) held in fork_trampoline(). Since we've
1641 * explicitly acquired locks in this function, the nesting count
1642 * is now 2 rather than 1. Since we are nested, calling
1643 * spinlock_exit() will simply adjust the counts without allowing
1644 * spin lock using code to interrupt us.
1645 */
1646 if (td == NULL) {
1647 mtx_lock_spin(&sched_lock);
1648 spinlock_exit();
1649 PCPU_SET(switchtime, cpu_ticks());
1650 PCPU_SET(switchticks, ticks);
1651 } else {
1652 lock_profile_release_lock(&sched_lock.lock_object);
1653 MPASS(td->td_lock == &sched_lock);
1654 td->td_lastcpu = td->td_oncpu;
1655 td->td_oncpu = NOCPU;
1656 }
1657 mtx_assert(&sched_lock, MA_OWNED);
1658 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
1659 cpu_throw(td, choosethread()); /* doesn't return */
1660 }
1661
1662 void
sched_fork_exit(struct thread * td)1663 sched_fork_exit(struct thread *td)
1664 {
1665
1666 /*
1667 * Finish setting up thread glue so that it begins execution in a
1668 * non-nested critical section with sched_lock held but not recursed.
1669 */
1670 td->td_oncpu = PCPU_GET(cpuid);
1671 sched_lock.mtx_lock = (uintptr_t)td;
1672 lock_profile_obtain_lock_success(&sched_lock.lock_object,
1673 0, 0, __FILE__, __LINE__);
1674 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
1675
1676 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
1677 "prio:%d", td->td_priority);
1678 SDT_PROBE0(sched, , , on__cpu);
1679 }
1680
1681 char *
sched_tdname(struct thread * td)1682 sched_tdname(struct thread *td)
1683 {
1684 #ifdef KTR
1685 struct td_sched *ts;
1686
1687 ts = td->td_sched;
1688 if (ts->ts_name[0] == '\0')
1689 snprintf(ts->ts_name, sizeof(ts->ts_name),
1690 "%s tid %d", td->td_name, td->td_tid);
1691 return (ts->ts_name);
1692 #else
1693 return (td->td_name);
1694 #endif
1695 }
1696
1697 #ifdef KTR
1698 void
sched_clear_tdname(struct thread * td)1699 sched_clear_tdname(struct thread *td)
1700 {
1701 struct td_sched *ts;
1702
1703 ts = td->td_sched;
1704 ts->ts_name[0] = '\0';
1705 }
1706 #endif
1707
1708 void
sched_affinity(struct thread * td)1709 sched_affinity(struct thread *td)
1710 {
1711 #ifdef SMP
1712 struct td_sched *ts;
1713 int cpu;
1714
1715 THREAD_LOCK_ASSERT(td, MA_OWNED);
1716
1717 /*
1718 * Set the TSF_AFFINITY flag if there is at least one CPU this
1719 * thread can't run on.
1720 */
1721 ts = td->td_sched;
1722 ts->ts_flags &= ~TSF_AFFINITY;
1723 CPU_FOREACH(cpu) {
1724 if (!THREAD_CAN_SCHED(td, cpu)) {
1725 ts->ts_flags |= TSF_AFFINITY;
1726 break;
1727 }
1728 }
1729
1730 /*
1731 * If this thread can run on all CPUs, nothing else to do.
1732 */
1733 if (!(ts->ts_flags & TSF_AFFINITY))
1734 return;
1735
1736 /* Pinned threads and bound threads should be left alone. */
1737 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
1738 return;
1739
1740 switch (td->td_state) {
1741 case TDS_RUNQ:
1742 /*
1743 * If we are on a per-CPU runqueue that is in the set,
1744 * then nothing needs to be done.
1745 */
1746 if (ts->ts_runq != &runq &&
1747 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
1748 return;
1749
1750 /* Put this thread on a valid per-CPU runqueue. */
1751 sched_rem(td);
1752 sched_add(td, SRQ_BORING);
1753 break;
1754 case TDS_RUNNING:
1755 /*
1756 * See if our current CPU is in the set. If not, force a
1757 * context switch.
1758 */
1759 if (THREAD_CAN_SCHED(td, td->td_oncpu))
1760 return;
1761
1762 td->td_flags |= TDF_NEEDRESCHED;
1763 if (td != curthread)
1764 ipi_cpu(cpu, IPI_AST);
1765 break;
1766 default:
1767 break;
1768 }
1769 #endif
1770 }
1771