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 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * SUCH DAMAGE.
37 */
38
39#include <sys/cdefs.h>
| 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 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * SUCH DAMAGE.
37 */
38
39#include <sys/cdefs.h>
|
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/ktr.h>
46#include <sys/lock.h>
47#include <sys/kthread.h>
48#include <sys/mutex.h>
49#include <sys/proc.h>
50#include <sys/resourcevar.h>
51#include <sys/sched.h>
52#include <sys/smp.h>
53#include <sys/sysctl.h>
54#include <sys/sx.h>
55
56#define KTR_4BSD 0x0
57
58/*
59 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
60 * the range 100-256 Hz (approximately).
61 */
62#define ESTCPULIM(e) \
63 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
64 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
65#ifdef SMP
66#define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
67#else
68#define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
69#endif
70#define NICE_WEIGHT 1 /* Priorities per nice level. */
71
72struct ke_sched {
73 int ske_cpticks; /* (j) Ticks of cpu time. */
74 struct runq *ske_runq; /* runq the kse is currently on */
75};
76#define ke_runq ke_sched->ske_runq
77#define KEF_BOUND KEF_SCHED1
78
79#define SKE_RUNQ_PCPU(ke) \
80 ((ke)->ke_runq != 0 && (ke)->ke_runq != &runq)
81
82/*
83 * KSE_CAN_MIGRATE macro returns true if the kse can migrate between
84 * cpus.
85 */
86#define KSE_CAN_MIGRATE(ke) \
87 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
88static struct ke_sched ke_sched;
89
90struct ke_sched *kse0_sched = &ke_sched;
91struct kg_sched *ksegrp0_sched = NULL;
92struct p_sched *proc0_sched = NULL;
93struct td_sched *thread0_sched = NULL;
94
95static int sched_tdcnt; /* Total runnable threads in the system. */
96static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
97#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
98
99static struct callout roundrobin_callout;
100
101static void setup_runqs(void);
102static void roundrobin(void *arg);
103static void schedcpu(void);
104static void schedcpu_thread(void);
105static void sched_setup(void *dummy);
106static void maybe_resched(struct thread *td);
107static void updatepri(struct ksegrp *kg);
108static void resetpriority(struct ksegrp *kg);
109
110static struct kproc_desc sched_kp = {
111 "schedcpu",
112 schedcpu_thread,
113 NULL
114};
115SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp)
116SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
117
118/*
119 * Global run queue.
120 */
121static struct runq runq;
122
123#ifdef SMP
124/*
125 * Per-CPU run queues
126 */
127static struct runq runq_pcpu[MAXCPU];
128#endif
129
130static void
131setup_runqs(void)
132{
133#ifdef SMP
134 int i;
135
136 for (i = 0; i < MAXCPU; ++i)
137 runq_init(&runq_pcpu[i]);
138#endif
139
140 runq_init(&runq);
141}
142
143static int
144sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
145{
146 int error, new_val;
147
148 new_val = sched_quantum * tick;
149 error = sysctl_handle_int(oidp, &new_val, 0, req);
150 if (error != 0 || req->newptr == NULL)
151 return (error);
152 if (new_val < tick)
153 return (EINVAL);
154 sched_quantum = new_val / tick;
155 hogticks = 2 * sched_quantum;
156 return (0);
157}
158
159SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
160 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
161 "Roundrobin scheduling quantum in microseconds");
162
163/*
164 * Arrange to reschedule if necessary, taking the priorities and
165 * schedulers into account.
166 */
167static void
168maybe_resched(struct thread *td)
169{
170
171 mtx_assert(&sched_lock, MA_OWNED);
172 if (td->td_priority < curthread->td_priority && curthread->td_kse)
173 curthread->td_flags |= TDF_NEEDRESCHED;
174}
175
176/*
177 * Force switch among equal priority processes every 100ms.
178 * We don't actually need to force a context switch of the current process.
179 * The act of firing the event triggers a context switch to softclock() and
180 * then switching back out again which is equivalent to a preemption, thus
181 * no further work is needed on the local CPU.
182 */
183/* ARGSUSED */
184static void
185roundrobin(void *arg)
186{
187
188#ifdef SMP
189 mtx_lock_spin(&sched_lock);
190 forward_roundrobin();
191 mtx_unlock_spin(&sched_lock);
192#endif
193
194 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
195}
196
197/*
198 * Constants for digital decay and forget:
199 * 90% of (kg_estcpu) usage in 5 * loadav time
200 * 95% of (ke_pctcpu) usage in 60 seconds (load insensitive)
201 * Note that, as ps(1) mentions, this can let percentages
202 * total over 100% (I've seen 137.9% for 3 processes).
203 *
204 * Note that schedclock() updates kg_estcpu and p_cpticks asynchronously.
205 *
206 * We wish to decay away 90% of kg_estcpu in (5 * loadavg) seconds.
207 * That is, the system wants to compute a value of decay such
208 * that the following for loop:
209 * for (i = 0; i < (5 * loadavg); i++)
210 * kg_estcpu *= decay;
211 * will compute
212 * kg_estcpu *= 0.1;
213 * for all values of loadavg:
214 *
215 * Mathematically this loop can be expressed by saying:
216 * decay ** (5 * loadavg) ~= .1
217 *
218 * The system computes decay as:
219 * decay = (2 * loadavg) / (2 * loadavg + 1)
220 *
221 * We wish to prove that the system's computation of decay
222 * will always fulfill the equation:
223 * decay ** (5 * loadavg) ~= .1
224 *
225 * If we compute b as:
226 * b = 2 * loadavg
227 * then
228 * decay = b / (b + 1)
229 *
230 * We now need to prove two things:
231 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
232 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
233 *
234 * Facts:
235 * For x close to zero, exp(x) =~ 1 + x, since
236 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
237 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
238 * For x close to zero, ln(1+x) =~ x, since
239 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
240 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
241 * ln(.1) =~ -2.30
242 *
243 * Proof of (1):
244 * Solve (factor)**(power) =~ .1 given power (5*loadav):
245 * solving for factor,
246 * ln(factor) =~ (-2.30/5*loadav), or
247 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
248 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
249 *
250 * Proof of (2):
251 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
252 * solving for power,
253 * power*ln(b/(b+1)) =~ -2.30, or
254 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
255 *
256 * Actual power values for the implemented algorithm are as follows:
257 * loadav: 1 2 3 4
258 * power: 5.68 10.32 14.94 19.55
259 */
260
261/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
262#define loadfactor(loadav) (2 * (loadav))
263#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
264
265/* decay 95% of `ke_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
266static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
267SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
268
269/*
270 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
271 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
272 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
273 *
274 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
275 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
276 *
277 * If you don't want to bother with the faster/more-accurate formula, you
278 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
279 * (more general) method of calculating the %age of CPU used by a process.
280 */
281#define CCPU_SHIFT 11
282
283/*
284 * Recompute process priorities, every hz ticks.
285 * MP-safe, called without the Giant mutex.
286 */
287/* ARGSUSED */
288static void
289schedcpu(void)
290{
291 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
292 struct thread *td;
293 struct proc *p;
294 struct kse *ke;
295 struct ksegrp *kg;
296 int awake, realstathz;
297
298 realstathz = stathz ? stathz : hz;
299 sx_slock(&allproc_lock);
300 FOREACH_PROC_IN_SYSTEM(p) {
301 /*
302 * Prevent state changes and protect run queue.
303 */
304 mtx_lock_spin(&sched_lock);
305 /*
306 * Increment time in/out of memory. We ignore overflow; with
307 * 16-bit int's (remember them?) overflow takes 45 days.
308 */
309 p->p_swtime++;
310 FOREACH_KSEGRP_IN_PROC(p, kg) {
311 awake = 0;
312 FOREACH_KSE_IN_GROUP(kg, ke) {
313 /*
314 * Increment sleep time (if sleeping). We
315 * ignore overflow, as above.
316 */
317 /*
318 * The kse slptimes are not touched in wakeup
319 * because the thread may not HAVE a KSE.
320 */
321 if (ke->ke_state == KES_ONRUNQ) {
322 awake = 1;
323 ke->ke_flags &= ~KEF_DIDRUN;
324 } else if ((ke->ke_state == KES_THREAD) &&
325 (TD_IS_RUNNING(ke->ke_thread))) {
326 awake = 1;
327 /* Do not clear KEF_DIDRUN */
328 } else if (ke->ke_flags & KEF_DIDRUN) {
329 awake = 1;
330 ke->ke_flags &= ~KEF_DIDRUN;
331 }
332
333 /*
334 * ke_pctcpu is only for ps and ttyinfo().
335 * Do it per kse, and add them up at the end?
336 * XXXKSE
337 */
338 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >>
339 FSHIFT;
340 /*
341 * If the kse has been idle the entire second,
342 * stop recalculating its priority until
343 * it wakes up.
344 */
345 if (ke->ke_sched->ske_cpticks == 0)
346 continue;
347#if (FSHIFT >= CCPU_SHIFT)
348 ke->ke_pctcpu += (realstathz == 100)
349 ? ((fixpt_t) ke->ke_sched->ske_cpticks) <<
350 (FSHIFT - CCPU_SHIFT) :
351 100 * (((fixpt_t) ke->ke_sched->ske_cpticks)
352 << (FSHIFT - CCPU_SHIFT)) / realstathz;
353#else
354 ke->ke_pctcpu += ((FSCALE - ccpu) *
355 (ke->ke_sched->ske_cpticks *
356 FSCALE / realstathz)) >> FSHIFT;
357#endif
358 ke->ke_sched->ske_cpticks = 0;
359 } /* end of kse loop */
360 /*
361 * If there are ANY running threads in this KSEGRP,
362 * then don't count it as sleeping.
363 */
364 if (awake) {
365 if (kg->kg_slptime > 1) {
366 /*
367 * In an ideal world, this should not
368 * happen, because whoever woke us
369 * up from the long sleep should have
370 * unwound the slptime and reset our
371 * priority before we run at the stale
372 * priority. Should KASSERT at some
373 * point when all the cases are fixed.
374 */
375 updatepri(kg);
376 }
377 kg->kg_slptime = 0;
378 } else
379 kg->kg_slptime++;
380 if (kg->kg_slptime > 1)
381 continue;
382 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu);
383 resetpriority(kg);
384 FOREACH_THREAD_IN_GROUP(kg, td) {
385 if (td->td_priority >= PUSER) {
386 sched_prio(td, kg->kg_user_pri);
387 }
388 }
389 } /* end of ksegrp loop */
390 mtx_unlock_spin(&sched_lock);
391 } /* end of process loop */
392 sx_sunlock(&allproc_lock);
393}
394
395/*
396 * Main loop for a kthread that executes schedcpu once a second.
397 */
398static void
399schedcpu_thread(void)
400{
401 int nowake;
402
403 for (;;) {
404 schedcpu();
405 tsleep(&nowake, curthread->td_priority, "-", hz);
406 }
407}
408
409/*
410 * Recalculate the priority of a process after it has slept for a while.
411 * For all load averages >= 1 and max kg_estcpu of 255, sleeping for at
412 * least six times the loadfactor will decay kg_estcpu to zero.
413 */
414static void
415updatepri(struct ksegrp *kg)
416{
417 register fixpt_t loadfac;
418 register unsigned int newcpu;
419
420 loadfac = loadfactor(averunnable.ldavg[0]);
421 if (kg->kg_slptime > 5 * loadfac)
422 kg->kg_estcpu = 0;
423 else {
424 newcpu = kg->kg_estcpu;
425 kg->kg_slptime--; /* was incremented in schedcpu() */
426 while (newcpu && --kg->kg_slptime)
427 newcpu = decay_cpu(loadfac, newcpu);
428 kg->kg_estcpu = newcpu;
429 }
430 resetpriority(kg);
431}
432
433/*
434 * Compute the priority of a process when running in user mode.
435 * Arrange to reschedule if the resulting priority is better
436 * than that of the current process.
437 */
438static void
439resetpriority(struct ksegrp *kg)
440{
441 register unsigned int newpriority;
442 struct thread *td;
443
444 if (kg->kg_pri_class == PRI_TIMESHARE) {
445 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT +
446 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN);
447 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
448 PRI_MAX_TIMESHARE);
449 kg->kg_user_pri = newpriority;
450 }
451 FOREACH_THREAD_IN_GROUP(kg, td) {
452 maybe_resched(td); /* XXXKSE silly */
453 }
454}
455
456/* ARGSUSED */
457static void
458sched_setup(void *dummy)
459{
460 setup_runqs();
461
462 if (sched_quantum == 0)
463 sched_quantum = SCHED_QUANTUM;
464 hogticks = 2 * sched_quantum;
465
466 callout_init(&roundrobin_callout, 0);
467
468 /* Kick off timeout driven events by calling first time. */
469 roundrobin(NULL);
470
471 /* Account for thread0. */
472 sched_tdcnt++;
473}
474
475/* External interfaces start here */
476int
477sched_runnable(void)
478{
479#ifdef SMP
480 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
481#else
482 return runq_check(&runq);
483#endif
484}
485
486int
487sched_rr_interval(void)
488{
489 if (sched_quantum == 0)
490 sched_quantum = SCHED_QUANTUM;
491 return (sched_quantum);
492}
493
494/*
495 * We adjust the priority of the current process. The priority of
496 * a process gets worse as it accumulates CPU time. The cpu usage
497 * estimator (kg_estcpu) is increased here. resetpriority() will
498 * compute a different priority each time kg_estcpu increases by
499 * INVERSE_ESTCPU_WEIGHT
500 * (until MAXPRI is reached). The cpu usage estimator ramps up
501 * quite quickly when the process is running (linearly), and decays
502 * away exponentially, at a rate which is proportionally slower when
503 * the system is busy. The basic principle is that the system will
504 * 90% forget that the process used a lot of CPU time in 5 * loadav
505 * seconds. This causes the system to favor processes which haven't
506 * run much recently, and to round-robin among other processes.
507 */
508void
509sched_clock(struct thread *td)
510{
511 struct ksegrp *kg;
512 struct kse *ke;
513
514 mtx_assert(&sched_lock, MA_OWNED);
515 kg = td->td_ksegrp;
516 ke = td->td_kse;
517
518 ke->ke_sched->ske_cpticks++;
519 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1);
520 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
521 resetpriority(kg);
522 if (td->td_priority >= PUSER)
523 td->td_priority = kg->kg_user_pri;
524 }
525}
526
527/*
528 * charge childs scheduling cpu usage to parent.
529 *
530 * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp.
531 * Charge it to the ksegrp that did the wait since process estcpu is sum of
532 * all ksegrps, this is strictly as expected. Assume that the child process
533 * aggregated all the estcpu into the 'built-in' ksegrp.
534 */
535void
536sched_exit(struct proc *p, struct proc *p1)
537{
538 sched_exit_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1));
539 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1));
540 sched_exit_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1));
541}
542
543void
544sched_exit_kse(struct kse *ke, struct kse *child)
545{
546}
547
548void
549sched_exit_ksegrp(struct ksegrp *kg, struct ksegrp *child)
550{
551
552 mtx_assert(&sched_lock, MA_OWNED);
553 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + child->kg_estcpu);
554}
555
556void
557sched_exit_thread(struct thread *td, struct thread *child)
558{
559 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
560 sched_tdcnt--;
561}
562
563void
564sched_fork(struct proc *p, struct proc *p1)
565{
566 sched_fork_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1));
567 sched_fork_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1));
568 sched_fork_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1));
569}
570
571void
572sched_fork_kse(struct kse *ke, struct kse *child)
573{
574 child->ke_sched->ske_cpticks = 0;
575}
576
577void
578sched_fork_ksegrp(struct ksegrp *kg, struct ksegrp *child)
579{
580 mtx_assert(&sched_lock, MA_OWNED);
581 child->kg_estcpu = kg->kg_estcpu;
582}
583
584void
585sched_fork_thread(struct thread *td, struct thread *child)
586{
587}
588
589void
590sched_nice(struct ksegrp *kg, int nice)
591{
592
593 PROC_LOCK_ASSERT(kg->kg_proc, MA_OWNED);
594 mtx_assert(&sched_lock, MA_OWNED);
595 kg->kg_nice = nice;
596 resetpriority(kg);
597}
598
599void
600sched_class(struct ksegrp *kg, int class)
601{
602 mtx_assert(&sched_lock, MA_OWNED);
603 kg->kg_pri_class = class;
604}
605
606/*
607 * Adjust the priority of a thread.
608 * This may include moving the thread within the KSEGRP,
609 * changing the assignment of a kse to the thread,
610 * and moving a KSE in the system run queue.
611 */
612void
613sched_prio(struct thread *td, u_char prio)
614{
615
616 mtx_assert(&sched_lock, MA_OWNED);
617 if (TD_ON_RUNQ(td)) {
618 adjustrunqueue(td, prio);
619 } else {
620 td->td_priority = prio;
621 }
622}
623
624void
| 41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/ktr.h>
46#include <sys/lock.h>
47#include <sys/kthread.h>
48#include <sys/mutex.h>
49#include <sys/proc.h>
50#include <sys/resourcevar.h>
51#include <sys/sched.h>
52#include <sys/smp.h>
53#include <sys/sysctl.h>
54#include <sys/sx.h>
55
56#define KTR_4BSD 0x0
57
58/*
59 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
60 * the range 100-256 Hz (approximately).
61 */
62#define ESTCPULIM(e) \
63 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
64 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
65#ifdef SMP
66#define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus)
67#else
68#define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */
69#endif
70#define NICE_WEIGHT 1 /* Priorities per nice level. */
71
72struct ke_sched {
73 int ske_cpticks; /* (j) Ticks of cpu time. */
74 struct runq *ske_runq; /* runq the kse is currently on */
75};
76#define ke_runq ke_sched->ske_runq
77#define KEF_BOUND KEF_SCHED1
78
79#define SKE_RUNQ_PCPU(ke) \
80 ((ke)->ke_runq != 0 && (ke)->ke_runq != &runq)
81
82/*
83 * KSE_CAN_MIGRATE macro returns true if the kse can migrate between
84 * cpus.
85 */
86#define KSE_CAN_MIGRATE(ke) \
87 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
88static struct ke_sched ke_sched;
89
90struct ke_sched *kse0_sched = &ke_sched;
91struct kg_sched *ksegrp0_sched = NULL;
92struct p_sched *proc0_sched = NULL;
93struct td_sched *thread0_sched = NULL;
94
95static int sched_tdcnt; /* Total runnable threads in the system. */
96static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */
97#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */
98
99static struct callout roundrobin_callout;
100
101static void setup_runqs(void);
102static void roundrobin(void *arg);
103static void schedcpu(void);
104static void schedcpu_thread(void);
105static void sched_setup(void *dummy);
106static void maybe_resched(struct thread *td);
107static void updatepri(struct ksegrp *kg);
108static void resetpriority(struct ksegrp *kg);
109
110static struct kproc_desc sched_kp = {
111 "schedcpu",
112 schedcpu_thread,
113 NULL
114};
115SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp)
116SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
117
118/*
119 * Global run queue.
120 */
121static struct runq runq;
122
123#ifdef SMP
124/*
125 * Per-CPU run queues
126 */
127static struct runq runq_pcpu[MAXCPU];
128#endif
129
130static void
131setup_runqs(void)
132{
133#ifdef SMP
134 int i;
135
136 for (i = 0; i < MAXCPU; ++i)
137 runq_init(&runq_pcpu[i]);
138#endif
139
140 runq_init(&runq);
141}
142
143static int
144sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
145{
146 int error, new_val;
147
148 new_val = sched_quantum * tick;
149 error = sysctl_handle_int(oidp, &new_val, 0, req);
150 if (error != 0 || req->newptr == NULL)
151 return (error);
152 if (new_val < tick)
153 return (EINVAL);
154 sched_quantum = new_val / tick;
155 hogticks = 2 * sched_quantum;
156 return (0);
157}
158
159SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW,
160 0, sizeof sched_quantum, sysctl_kern_quantum, "I",
161 "Roundrobin scheduling quantum in microseconds");
162
163/*
164 * Arrange to reschedule if necessary, taking the priorities and
165 * schedulers into account.
166 */
167static void
168maybe_resched(struct thread *td)
169{
170
171 mtx_assert(&sched_lock, MA_OWNED);
172 if (td->td_priority < curthread->td_priority && curthread->td_kse)
173 curthread->td_flags |= TDF_NEEDRESCHED;
174}
175
176/*
177 * Force switch among equal priority processes every 100ms.
178 * We don't actually need to force a context switch of the current process.
179 * The act of firing the event triggers a context switch to softclock() and
180 * then switching back out again which is equivalent to a preemption, thus
181 * no further work is needed on the local CPU.
182 */
183/* ARGSUSED */
184static void
185roundrobin(void *arg)
186{
187
188#ifdef SMP
189 mtx_lock_spin(&sched_lock);
190 forward_roundrobin();
191 mtx_unlock_spin(&sched_lock);
192#endif
193
194 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL);
195}
196
197/*
198 * Constants for digital decay and forget:
199 * 90% of (kg_estcpu) usage in 5 * loadav time
200 * 95% of (ke_pctcpu) usage in 60 seconds (load insensitive)
201 * Note that, as ps(1) mentions, this can let percentages
202 * total over 100% (I've seen 137.9% for 3 processes).
203 *
204 * Note that schedclock() updates kg_estcpu and p_cpticks asynchronously.
205 *
206 * We wish to decay away 90% of kg_estcpu in (5 * loadavg) seconds.
207 * That is, the system wants to compute a value of decay such
208 * that the following for loop:
209 * for (i = 0; i < (5 * loadavg); i++)
210 * kg_estcpu *= decay;
211 * will compute
212 * kg_estcpu *= 0.1;
213 * for all values of loadavg:
214 *
215 * Mathematically this loop can be expressed by saying:
216 * decay ** (5 * loadavg) ~= .1
217 *
218 * The system computes decay as:
219 * decay = (2 * loadavg) / (2 * loadavg + 1)
220 *
221 * We wish to prove that the system's computation of decay
222 * will always fulfill the equation:
223 * decay ** (5 * loadavg) ~= .1
224 *
225 * If we compute b as:
226 * b = 2 * loadavg
227 * then
228 * decay = b / (b + 1)
229 *
230 * We now need to prove two things:
231 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
232 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
233 *
234 * Facts:
235 * For x close to zero, exp(x) =~ 1 + x, since
236 * exp(x) = 0! + x**1/1! + x**2/2! + ... .
237 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
238 * For x close to zero, ln(1+x) =~ x, since
239 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
240 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
241 * ln(.1) =~ -2.30
242 *
243 * Proof of (1):
244 * Solve (factor)**(power) =~ .1 given power (5*loadav):
245 * solving for factor,
246 * ln(factor) =~ (-2.30/5*loadav), or
247 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
248 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
249 *
250 * Proof of (2):
251 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
252 * solving for power,
253 * power*ln(b/(b+1)) =~ -2.30, or
254 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
255 *
256 * Actual power values for the implemented algorithm are as follows:
257 * loadav: 1 2 3 4
258 * power: 5.68 10.32 14.94 19.55
259 */
260
261/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
262#define loadfactor(loadav) (2 * (loadav))
263#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
264
265/* decay 95% of `ke_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
266static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
267SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
268
269/*
270 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
271 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
272 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
273 *
274 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
275 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
276 *
277 * If you don't want to bother with the faster/more-accurate formula, you
278 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
279 * (more general) method of calculating the %age of CPU used by a process.
280 */
281#define CCPU_SHIFT 11
282
283/*
284 * Recompute process priorities, every hz ticks.
285 * MP-safe, called without the Giant mutex.
286 */
287/* ARGSUSED */
288static void
289schedcpu(void)
290{
291 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
292 struct thread *td;
293 struct proc *p;
294 struct kse *ke;
295 struct ksegrp *kg;
296 int awake, realstathz;
297
298 realstathz = stathz ? stathz : hz;
299 sx_slock(&allproc_lock);
300 FOREACH_PROC_IN_SYSTEM(p) {
301 /*
302 * Prevent state changes and protect run queue.
303 */
304 mtx_lock_spin(&sched_lock);
305 /*
306 * Increment time in/out of memory. We ignore overflow; with
307 * 16-bit int's (remember them?) overflow takes 45 days.
308 */
309 p->p_swtime++;
310 FOREACH_KSEGRP_IN_PROC(p, kg) {
311 awake = 0;
312 FOREACH_KSE_IN_GROUP(kg, ke) {
313 /*
314 * Increment sleep time (if sleeping). We
315 * ignore overflow, as above.
316 */
317 /*
318 * The kse slptimes are not touched in wakeup
319 * because the thread may not HAVE a KSE.
320 */
321 if (ke->ke_state == KES_ONRUNQ) {
322 awake = 1;
323 ke->ke_flags &= ~KEF_DIDRUN;
324 } else if ((ke->ke_state == KES_THREAD) &&
325 (TD_IS_RUNNING(ke->ke_thread))) {
326 awake = 1;
327 /* Do not clear KEF_DIDRUN */
328 } else if (ke->ke_flags & KEF_DIDRUN) {
329 awake = 1;
330 ke->ke_flags &= ~KEF_DIDRUN;
331 }
332
333 /*
334 * ke_pctcpu is only for ps and ttyinfo().
335 * Do it per kse, and add them up at the end?
336 * XXXKSE
337 */
338 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >>
339 FSHIFT;
340 /*
341 * If the kse has been idle the entire second,
342 * stop recalculating its priority until
343 * it wakes up.
344 */
345 if (ke->ke_sched->ske_cpticks == 0)
346 continue;
347#if (FSHIFT >= CCPU_SHIFT)
348 ke->ke_pctcpu += (realstathz == 100)
349 ? ((fixpt_t) ke->ke_sched->ske_cpticks) <<
350 (FSHIFT - CCPU_SHIFT) :
351 100 * (((fixpt_t) ke->ke_sched->ske_cpticks)
352 << (FSHIFT - CCPU_SHIFT)) / realstathz;
353#else
354 ke->ke_pctcpu += ((FSCALE - ccpu) *
355 (ke->ke_sched->ske_cpticks *
356 FSCALE / realstathz)) >> FSHIFT;
357#endif
358 ke->ke_sched->ske_cpticks = 0;
359 } /* end of kse loop */
360 /*
361 * If there are ANY running threads in this KSEGRP,
362 * then don't count it as sleeping.
363 */
364 if (awake) {
365 if (kg->kg_slptime > 1) {
366 /*
367 * In an ideal world, this should not
368 * happen, because whoever woke us
369 * up from the long sleep should have
370 * unwound the slptime and reset our
371 * priority before we run at the stale
372 * priority. Should KASSERT at some
373 * point when all the cases are fixed.
374 */
375 updatepri(kg);
376 }
377 kg->kg_slptime = 0;
378 } else
379 kg->kg_slptime++;
380 if (kg->kg_slptime > 1)
381 continue;
382 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu);
383 resetpriority(kg);
384 FOREACH_THREAD_IN_GROUP(kg, td) {
385 if (td->td_priority >= PUSER) {
386 sched_prio(td, kg->kg_user_pri);
387 }
388 }
389 } /* end of ksegrp loop */
390 mtx_unlock_spin(&sched_lock);
391 } /* end of process loop */
392 sx_sunlock(&allproc_lock);
393}
394
395/*
396 * Main loop for a kthread that executes schedcpu once a second.
397 */
398static void
399schedcpu_thread(void)
400{
401 int nowake;
402
403 for (;;) {
404 schedcpu();
405 tsleep(&nowake, curthread->td_priority, "-", hz);
406 }
407}
408
409/*
410 * Recalculate the priority of a process after it has slept for a while.
411 * For all load averages >= 1 and max kg_estcpu of 255, sleeping for at
412 * least six times the loadfactor will decay kg_estcpu to zero.
413 */
414static void
415updatepri(struct ksegrp *kg)
416{
417 register fixpt_t loadfac;
418 register unsigned int newcpu;
419
420 loadfac = loadfactor(averunnable.ldavg[0]);
421 if (kg->kg_slptime > 5 * loadfac)
422 kg->kg_estcpu = 0;
423 else {
424 newcpu = kg->kg_estcpu;
425 kg->kg_slptime--; /* was incremented in schedcpu() */
426 while (newcpu && --kg->kg_slptime)
427 newcpu = decay_cpu(loadfac, newcpu);
428 kg->kg_estcpu = newcpu;
429 }
430 resetpriority(kg);
431}
432
433/*
434 * Compute the priority of a process when running in user mode.
435 * Arrange to reschedule if the resulting priority is better
436 * than that of the current process.
437 */
438static void
439resetpriority(struct ksegrp *kg)
440{
441 register unsigned int newpriority;
442 struct thread *td;
443
444 if (kg->kg_pri_class == PRI_TIMESHARE) {
445 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT +
446 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN);
447 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
448 PRI_MAX_TIMESHARE);
449 kg->kg_user_pri = newpriority;
450 }
451 FOREACH_THREAD_IN_GROUP(kg, td) {
452 maybe_resched(td); /* XXXKSE silly */
453 }
454}
455
456/* ARGSUSED */
457static void
458sched_setup(void *dummy)
459{
460 setup_runqs();
461
462 if (sched_quantum == 0)
463 sched_quantum = SCHED_QUANTUM;
464 hogticks = 2 * sched_quantum;
465
466 callout_init(&roundrobin_callout, 0);
467
468 /* Kick off timeout driven events by calling first time. */
469 roundrobin(NULL);
470
471 /* Account for thread0. */
472 sched_tdcnt++;
473}
474
475/* External interfaces start here */
476int
477sched_runnable(void)
478{
479#ifdef SMP
480 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
481#else
482 return runq_check(&runq);
483#endif
484}
485
486int
487sched_rr_interval(void)
488{
489 if (sched_quantum == 0)
490 sched_quantum = SCHED_QUANTUM;
491 return (sched_quantum);
492}
493
494/*
495 * We adjust the priority of the current process. The priority of
496 * a process gets worse as it accumulates CPU time. The cpu usage
497 * estimator (kg_estcpu) is increased here. resetpriority() will
498 * compute a different priority each time kg_estcpu increases by
499 * INVERSE_ESTCPU_WEIGHT
500 * (until MAXPRI is reached). The cpu usage estimator ramps up
501 * quite quickly when the process is running (linearly), and decays
502 * away exponentially, at a rate which is proportionally slower when
503 * the system is busy. The basic principle is that the system will
504 * 90% forget that the process used a lot of CPU time in 5 * loadav
505 * seconds. This causes the system to favor processes which haven't
506 * run much recently, and to round-robin among other processes.
507 */
508void
509sched_clock(struct thread *td)
510{
511 struct ksegrp *kg;
512 struct kse *ke;
513
514 mtx_assert(&sched_lock, MA_OWNED);
515 kg = td->td_ksegrp;
516 ke = td->td_kse;
517
518 ke->ke_sched->ske_cpticks++;
519 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1);
520 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
521 resetpriority(kg);
522 if (td->td_priority >= PUSER)
523 td->td_priority = kg->kg_user_pri;
524 }
525}
526
527/*
528 * charge childs scheduling cpu usage to parent.
529 *
530 * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp.
531 * Charge it to the ksegrp that did the wait since process estcpu is sum of
532 * all ksegrps, this is strictly as expected. Assume that the child process
533 * aggregated all the estcpu into the 'built-in' ksegrp.
534 */
535void
536sched_exit(struct proc *p, struct proc *p1)
537{
538 sched_exit_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1));
539 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1));
540 sched_exit_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1));
541}
542
543void
544sched_exit_kse(struct kse *ke, struct kse *child)
545{
546}
547
548void
549sched_exit_ksegrp(struct ksegrp *kg, struct ksegrp *child)
550{
551
552 mtx_assert(&sched_lock, MA_OWNED);
553 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + child->kg_estcpu);
554}
555
556void
557sched_exit_thread(struct thread *td, struct thread *child)
558{
559 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
560 sched_tdcnt--;
561}
562
563void
564sched_fork(struct proc *p, struct proc *p1)
565{
566 sched_fork_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1));
567 sched_fork_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1));
568 sched_fork_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1));
569}
570
571void
572sched_fork_kse(struct kse *ke, struct kse *child)
573{
574 child->ke_sched->ske_cpticks = 0;
575}
576
577void
578sched_fork_ksegrp(struct ksegrp *kg, struct ksegrp *child)
579{
580 mtx_assert(&sched_lock, MA_OWNED);
581 child->kg_estcpu = kg->kg_estcpu;
582}
583
584void
585sched_fork_thread(struct thread *td, struct thread *child)
586{
587}
588
589void
590sched_nice(struct ksegrp *kg, int nice)
591{
592
593 PROC_LOCK_ASSERT(kg->kg_proc, MA_OWNED);
594 mtx_assert(&sched_lock, MA_OWNED);
595 kg->kg_nice = nice;
596 resetpriority(kg);
597}
598
599void
600sched_class(struct ksegrp *kg, int class)
601{
602 mtx_assert(&sched_lock, MA_OWNED);
603 kg->kg_pri_class = class;
604}
605
606/*
607 * Adjust the priority of a thread.
608 * This may include moving the thread within the KSEGRP,
609 * changing the assignment of a kse to the thread,
610 * and moving a KSE in the system run queue.
611 */
612void
613sched_prio(struct thread *td, u_char prio)
614{
615
616 mtx_assert(&sched_lock, MA_OWNED);
617 if (TD_ON_RUNQ(td)) {
618 adjustrunqueue(td, prio);
619 } else {
620 td->td_priority = prio;
621 }
622}
623
624void
|
631}
632
633void
634sched_switch(struct thread *td)
635{
636 struct thread *newtd;
637 struct kse *ke;
638 struct proc *p;
639
640 ke = td->td_kse;
641 p = td->td_proc;
642
643 mtx_assert(&sched_lock, MA_OWNED);
644 KASSERT((ke->ke_state == KES_THREAD), ("sched_switch: kse state?"));
645
646 if ((p->p_flag & P_NOLOAD) == 0)
647 sched_tdcnt--;
648 td->td_lastcpu = td->td_oncpu;
649 td->td_last_kse = ke;
650 td->td_flags &= ~TDF_NEEDRESCHED;
651 td->td_oncpu = NOCPU;
652 /*
653 * At the last moment, if this thread is still marked RUNNING,
654 * then put it back on the run queue as it has not been suspended
655 * or stopped or any thing else similar.
656 */
657 if (TD_IS_RUNNING(td)) {
658 /* Put us back on the run queue (kse and all). */
659 setrunqueue(td);
660 } else if (p->p_flag & P_SA) {
661 /*
662 * We will not be on the run queue. So we must be
663 * sleeping or similar. As it's available,
664 * someone else can use the KSE if they need it.
665 */
666 kse_reassign(ke);
667 }
668 newtd = choosethread();
669 if (td != newtd)
670 cpu_switch(td, newtd);
671 sched_lock.mtx_lock = (uintptr_t)td;
672 td->td_oncpu = PCPU_GET(cpuid);
673}
674
675void
676sched_wakeup(struct thread *td)
677{
678 struct ksegrp *kg;
679
680 mtx_assert(&sched_lock, MA_OWNED);
681 kg = td->td_ksegrp;
682 if (kg->kg_slptime > 1)
683 updatepri(kg);
684 kg->kg_slptime = 0;
685 setrunqueue(td);
686 maybe_resched(td);
687}
688
689void
690sched_add(struct thread *td)
691{
692 struct kse *ke;
693
694 ke = td->td_kse;
695 mtx_assert(&sched_lock, MA_OWNED);
696 KASSERT((ke->ke_thread != NULL), ("sched_add: No thread on KSE"));
697 KASSERT((ke->ke_thread->td_kse != NULL),
698 ("sched_add: No KSE on thread"));
699 KASSERT(ke->ke_state != KES_ONRUNQ,
700 ("sched_add: kse %p (%s) already in run queue", ke,
701 ke->ke_proc->p_comm));
702 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
703 ("sched_add: process swapped out"));
704 ke->ke_ksegrp->kg_runq_kses++;
705 ke->ke_state = KES_ONRUNQ;
706
707#ifdef SMP
708 if (KSE_CAN_MIGRATE(ke)) {
709 CTR1(KTR_4BSD, "adding kse:%p to gbl runq", ke);
710 ke->ke_runq = &runq;
711 } else {
712 CTR1(KTR_4BSD, "adding kse:%p to pcpu runq", ke);
713 if (!SKE_RUNQ_PCPU(ke))
714 ke->ke_runq = &runq_pcpu[PCPU_GET(cpuid)];
715 }
716#else
717 ke->ke_runq = &runq;
718#endif
719 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
720 sched_tdcnt++;
721 runq_add(ke->ke_runq, ke);
722}
723
724void
725sched_rem(struct thread *td)
726{
727 struct kse *ke;
728
729 ke = td->td_kse;
730 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
731 ("sched_rem: process swapped out"));
732 KASSERT((ke->ke_state == KES_ONRUNQ),
733 ("sched_rem: KSE not on run queue"));
734 mtx_assert(&sched_lock, MA_OWNED);
735
736 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
737 sched_tdcnt--;
738 runq_remove(ke->ke_sched->ske_runq, ke);
739
740 ke->ke_state = KES_THREAD;
741 ke->ke_ksegrp->kg_runq_kses--;
742}
743
744struct kse *
745sched_choose(void)
746{
747 struct kse *ke;
748 struct runq *rq;
749
750#ifdef SMP
751 struct kse *kecpu;
752
753 rq = &runq;
754 ke = runq_choose(&runq);
755 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
756
757 if (ke == NULL ||
758 (kecpu != NULL &&
759 kecpu->ke_thread->td_priority < ke->ke_thread->td_priority)) {
760 CTR2(KTR_4BSD, "choosing kse %p from pcpu runq %d", kecpu,
761 PCPU_GET(cpuid));
762 ke = kecpu;
763 rq = &runq_pcpu[PCPU_GET(cpuid)];
764 } else {
765 CTR1(KTR_4BSD, "choosing kse %p from main runq", ke);
766 }
767
768#else
769 rq = &runq;
770 ke = runq_choose(&runq);
771#endif
772
773 if (ke != NULL) {
774 runq_remove(rq, ke);
775 ke->ke_state = KES_THREAD;
776
777 KASSERT((ke->ke_thread != NULL),
778 ("sched_choose: No thread on KSE"));
779 KASSERT((ke->ke_thread->td_kse != NULL),
780 ("sched_choose: No KSE on thread"));
781 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
782 ("sched_choose: process swapped out"));
783 }
784 return (ke);
785}
786
787void
788sched_userret(struct thread *td)
789{
790 struct ksegrp *kg;
791 /*
792 * XXX we cheat slightly on the locking here to avoid locking in
793 * the usual case. Setting td_priority here is essentially an
794 * incomplete workaround for not setting it properly elsewhere.
795 * Now that some interrupt handlers are threads, not setting it
796 * properly elsewhere can clobber it in the window between setting
797 * it here and returning to user mode, so don't waste time setting
798 * it perfectly here.
799 */
800 kg = td->td_ksegrp;
801 if (td->td_priority != kg->kg_user_pri) {
802 mtx_lock_spin(&sched_lock);
803 td->td_priority = kg->kg_user_pri;
804 mtx_unlock_spin(&sched_lock);
805 }
806}
807
808void
809sched_bind(struct thread *td, int cpu)
810{
811 struct kse *ke;
812
813 mtx_assert(&sched_lock, MA_OWNED);
814 KASSERT(TD_IS_RUNNING(td),
815 ("sched_bind: cannot bind non-running thread"));
816
817 ke = td->td_kse;
818
819 ke->ke_flags |= KEF_BOUND;
820#ifdef SMP
821 ke->ke_runq = &runq_pcpu[cpu];
822 if (PCPU_GET(cpuid) == cpu)
823 return;
824
825 ke->ke_state = KES_THREAD;
826
827 mi_switch(SW_VOL);
828#endif
829}
830
831void
832sched_unbind(struct thread* td)
833{
834 mtx_assert(&sched_lock, MA_OWNED);
835 td->td_kse->ke_flags &= ~KEF_BOUND;
836}
837
838int
839sched_load(void)
840{
841 return (sched_tdcnt);
842}
843
844int
845sched_sizeof_kse(void)
846{
847 return (sizeof(struct kse) + sizeof(struct ke_sched));
848}
849int
850sched_sizeof_ksegrp(void)
851{
852 return (sizeof(struct ksegrp));
853}
854int
855sched_sizeof_proc(void)
856{
857 return (sizeof(struct proc));
858}
859int
860sched_sizeof_thread(void)
861{
862 return (sizeof(struct thread));
863}
864
865fixpt_t
866sched_pctcpu(struct thread *td)
867{
868 struct kse *ke;
869
870 ke = td->td_kse;
871 if (ke == NULL)
872 ke = td->td_last_kse;
873 if (ke)
874 return (ke->ke_pctcpu);
875
876 return (0);
877}
| 631}
632
633void
634sched_switch(struct thread *td)
635{
636 struct thread *newtd;
637 struct kse *ke;
638 struct proc *p;
639
640 ke = td->td_kse;
641 p = td->td_proc;
642
643 mtx_assert(&sched_lock, MA_OWNED);
644 KASSERT((ke->ke_state == KES_THREAD), ("sched_switch: kse state?"));
645
646 if ((p->p_flag & P_NOLOAD) == 0)
647 sched_tdcnt--;
648 td->td_lastcpu = td->td_oncpu;
649 td->td_last_kse = ke;
650 td->td_flags &= ~TDF_NEEDRESCHED;
651 td->td_oncpu = NOCPU;
652 /*
653 * At the last moment, if this thread is still marked RUNNING,
654 * then put it back on the run queue as it has not been suspended
655 * or stopped or any thing else similar.
656 */
657 if (TD_IS_RUNNING(td)) {
658 /* Put us back on the run queue (kse and all). */
659 setrunqueue(td);
660 } else if (p->p_flag & P_SA) {
661 /*
662 * We will not be on the run queue. So we must be
663 * sleeping or similar. As it's available,
664 * someone else can use the KSE if they need it.
665 */
666 kse_reassign(ke);
667 }
668 newtd = choosethread();
669 if (td != newtd)
670 cpu_switch(td, newtd);
671 sched_lock.mtx_lock = (uintptr_t)td;
672 td->td_oncpu = PCPU_GET(cpuid);
673}
674
675void
676sched_wakeup(struct thread *td)
677{
678 struct ksegrp *kg;
679
680 mtx_assert(&sched_lock, MA_OWNED);
681 kg = td->td_ksegrp;
682 if (kg->kg_slptime > 1)
683 updatepri(kg);
684 kg->kg_slptime = 0;
685 setrunqueue(td);
686 maybe_resched(td);
687}
688
689void
690sched_add(struct thread *td)
691{
692 struct kse *ke;
693
694 ke = td->td_kse;
695 mtx_assert(&sched_lock, MA_OWNED);
696 KASSERT((ke->ke_thread != NULL), ("sched_add: No thread on KSE"));
697 KASSERT((ke->ke_thread->td_kse != NULL),
698 ("sched_add: No KSE on thread"));
699 KASSERT(ke->ke_state != KES_ONRUNQ,
700 ("sched_add: kse %p (%s) already in run queue", ke,
701 ke->ke_proc->p_comm));
702 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
703 ("sched_add: process swapped out"));
704 ke->ke_ksegrp->kg_runq_kses++;
705 ke->ke_state = KES_ONRUNQ;
706
707#ifdef SMP
708 if (KSE_CAN_MIGRATE(ke)) {
709 CTR1(KTR_4BSD, "adding kse:%p to gbl runq", ke);
710 ke->ke_runq = &runq;
711 } else {
712 CTR1(KTR_4BSD, "adding kse:%p to pcpu runq", ke);
713 if (!SKE_RUNQ_PCPU(ke))
714 ke->ke_runq = &runq_pcpu[PCPU_GET(cpuid)];
715 }
716#else
717 ke->ke_runq = &runq;
718#endif
719 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
720 sched_tdcnt++;
721 runq_add(ke->ke_runq, ke);
722}
723
724void
725sched_rem(struct thread *td)
726{
727 struct kse *ke;
728
729 ke = td->td_kse;
730 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
731 ("sched_rem: process swapped out"));
732 KASSERT((ke->ke_state == KES_ONRUNQ),
733 ("sched_rem: KSE not on run queue"));
734 mtx_assert(&sched_lock, MA_OWNED);
735
736 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
737 sched_tdcnt--;
738 runq_remove(ke->ke_sched->ske_runq, ke);
739
740 ke->ke_state = KES_THREAD;
741 ke->ke_ksegrp->kg_runq_kses--;
742}
743
744struct kse *
745sched_choose(void)
746{
747 struct kse *ke;
748 struct runq *rq;
749
750#ifdef SMP
751 struct kse *kecpu;
752
753 rq = &runq;
754 ke = runq_choose(&runq);
755 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
756
757 if (ke == NULL ||
758 (kecpu != NULL &&
759 kecpu->ke_thread->td_priority < ke->ke_thread->td_priority)) {
760 CTR2(KTR_4BSD, "choosing kse %p from pcpu runq %d", kecpu,
761 PCPU_GET(cpuid));
762 ke = kecpu;
763 rq = &runq_pcpu[PCPU_GET(cpuid)];
764 } else {
765 CTR1(KTR_4BSD, "choosing kse %p from main runq", ke);
766 }
767
768#else
769 rq = &runq;
770 ke = runq_choose(&runq);
771#endif
772
773 if (ke != NULL) {
774 runq_remove(rq, ke);
775 ke->ke_state = KES_THREAD;
776
777 KASSERT((ke->ke_thread != NULL),
778 ("sched_choose: No thread on KSE"));
779 KASSERT((ke->ke_thread->td_kse != NULL),
780 ("sched_choose: No KSE on thread"));
781 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
782 ("sched_choose: process swapped out"));
783 }
784 return (ke);
785}
786
787void
788sched_userret(struct thread *td)
789{
790 struct ksegrp *kg;
791 /*
792 * XXX we cheat slightly on the locking here to avoid locking in
793 * the usual case. Setting td_priority here is essentially an
794 * incomplete workaround for not setting it properly elsewhere.
795 * Now that some interrupt handlers are threads, not setting it
796 * properly elsewhere can clobber it in the window between setting
797 * it here and returning to user mode, so don't waste time setting
798 * it perfectly here.
799 */
800 kg = td->td_ksegrp;
801 if (td->td_priority != kg->kg_user_pri) {
802 mtx_lock_spin(&sched_lock);
803 td->td_priority = kg->kg_user_pri;
804 mtx_unlock_spin(&sched_lock);
805 }
806}
807
808void
809sched_bind(struct thread *td, int cpu)
810{
811 struct kse *ke;
812
813 mtx_assert(&sched_lock, MA_OWNED);
814 KASSERT(TD_IS_RUNNING(td),
815 ("sched_bind: cannot bind non-running thread"));
816
817 ke = td->td_kse;
818
819 ke->ke_flags |= KEF_BOUND;
820#ifdef SMP
821 ke->ke_runq = &runq_pcpu[cpu];
822 if (PCPU_GET(cpuid) == cpu)
823 return;
824
825 ke->ke_state = KES_THREAD;
826
827 mi_switch(SW_VOL);
828#endif
829}
830
831void
832sched_unbind(struct thread* td)
833{
834 mtx_assert(&sched_lock, MA_OWNED);
835 td->td_kse->ke_flags &= ~KEF_BOUND;
836}
837
838int
839sched_load(void)
840{
841 return (sched_tdcnt);
842}
843
844int
845sched_sizeof_kse(void)
846{
847 return (sizeof(struct kse) + sizeof(struct ke_sched));
848}
849int
850sched_sizeof_ksegrp(void)
851{
852 return (sizeof(struct ksegrp));
853}
854int
855sched_sizeof_proc(void)
856{
857 return (sizeof(struct proc));
858}
859int
860sched_sizeof_thread(void)
861{
862 return (sizeof(struct thread));
863}
864
865fixpt_t
866sched_pctcpu(struct thread *td)
867{
868 struct kse *ke;
869
870 ke = td->td_kse;
871 if (ke == NULL)
872 ke = td->td_last_kse;
873 if (ke)
874 return (ke->ke_pctcpu);
875
876 return (0);
877}
|