2 * Copyright (c) 1999 Peter Wemm <peter@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * $FreeBSD: src/sys/kern/kern_switch.c,v 1.3.2.1 2000/05/16 06:58:12 dillon Exp $
27 * $DragonFly: src/sys/kern/Attic/kern_switch.c,v 1.17 2004/02/12 06:57:48 dillon Exp $
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
34 #include <sys/queue.h>
36 #include <sys/rtprio.h>
37 #include <sys/thread2.h>
39 #include <sys/sysctl.h>
40 #include <machine/ipl.h>
41 #include <machine/cpu.h>
44 * debugging only YYY Remove me! define to schedule user processes only
45 * on the BSP. Interrupts can still be taken on the APs.
47 #undef ONLY_ONE_USER_CPU
50 * We have NQS (32) run queues per scheduling class. For the normal
51 * class, there are 128 priorities scaled onto these 32 queues. New
52 * processes are added to the last entry in each queue, and processes
53 * are selected for running by taking them from the head and maintaining
54 * a simple FIFO arrangement. Realtime and Idle priority processes have
55 * and explicit 0-31 priority which maps directly onto their class queue
56 * index. When a queue has something in it, the corresponding bit is
57 * set in the queuebits variable, allowing a single read to determine
58 * the state of all 32 queues and then a ffs() to find the first busy
61 static struct rq queues[NQS];
62 static struct rq rtqueues[NQS];
63 static struct rq idqueues[NQS];
64 static u_int32_t queuebits;
65 static u_int32_t rtqueuebits;
66 static u_int32_t idqueuebits;
67 static u_int32_t curprocmask = -1; /* currently running a user process */
68 static u_int32_t rdyprocmask; /* ready to accept a user process */
74 SYSCTL_INT(_debug, OID_AUTO, runqcount, CTLFLAG_RD, &runqcount, 0, "");
75 static int usched_steal;
76 SYSCTL_INT(_debug, OID_AUTO, usched_steal, CTLFLAG_RW,
77 &usched_steal, 0, "Passive Release was nonoptimal");
78 static int usched_optimal;
79 SYSCTL_INT(_debug, OID_AUTO, usched_optimal, CTLFLAG_RW,
80 &usched_optimal, 0, "Passive Release was nonoptimal");
82 static int remote_resched = 1;
83 static int remote_resched_nonaffinity;
84 static int remote_resched_affinity;
85 static int choose_affinity;
86 SYSCTL_INT(_debug, OID_AUTO, remote_resched, CTLFLAG_RW,
87 &remote_resched, 0, "Resched to another cpu");
88 SYSCTL_INT(_debug, OID_AUTO, remote_resched_nonaffinity, CTLFLAG_RD,
89 &remote_resched_nonaffinity, 0, "Number of remote rescheds");
90 SYSCTL_INT(_debug, OID_AUTO, remote_resched_affinity, CTLFLAG_RD,
91 &remote_resched_affinity, 0, "Number of remote rescheds");
92 SYSCTL_INT(_debug, OID_AUTO, choose_affinity, CTLFLAG_RD,
93 &choose_affinity, 0, "chooseproc() was smart");
96 #define USCHED_COUNTER(td) ((td->td_gd == mycpu) ? ++usched_optimal : ++usched_steal)
99 * Initialize the run queues at boot time.
106 for (i = 0; i < NQS; i++) {
107 TAILQ_INIT(&queues[i]);
108 TAILQ_INIT(&rtqueues[i]);
109 TAILQ_INIT(&idqueues[i]);
117 SYSINIT(runqueue, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, rqinit, NULL)
121 test_resched(struct proc *curp, struct proc *newp)
123 if (newp->p_priority / PPQ <= curp->p_priority / PPQ)
129 * chooseproc() is called when a cpu needs a user process to LWKT schedule.
130 * chooseproc() will select a user process and return it.
134 chooseproc(struct proc *chkp)
143 pri = bsfl(rtqueuebits);
145 which = &rtqueuebits;
146 } else if (queuebits) {
147 pri = bsfl(queuebits);
150 } else if (idqueuebits) {
151 pri = bsfl(idqueuebits);
153 which = &idqueuebits;
158 KASSERT(p, ("chooseproc: no proc on busy queue"));
161 * If the chosen process is not at a higher priority then chkp
162 * then return NULL without dequeueing a new process.
164 if (chkp && !test_resched(chkp, p))
169 * If the chosen process does not reside on this cpu spend a few
170 * cycles looking for a better candidate at the same priority level.
171 * This is a fallback check, setrunqueue() tries to wakeup the
172 * correct cpu and is our front-line affinity.
174 if (p->p_thread->td_gd != mycpu &&
175 (chkp = TAILQ_NEXT(p, p_procq)) != NULL
177 if (chkp->p_thread->td_gd == mycpu) {
184 TAILQ_REMOVE(q, p, p_procq);
187 *which &= ~(1 << pri);
188 KASSERT((p->p_flag & P_ONRUNQ) != 0, ("not on runq6!"));
189 p->p_flag &= ~P_ONRUNQ;
195 * called via an ipi message to reschedule on another cpu.
199 need_resched_remote(void *dummy)
207 * setrunqueue() 'wakes up' a 'user' process, which can mean several things.
209 * If P_CP_RELEASED is set the user process is under the control of the
210 * LWKT subsystem and we simply wake the thread up. This is ALWAYS the
211 * case when setrunqueue() is called from wakeup() and, in fact wakeup()
212 * asserts that P_CP_RELEASED is set.
214 * Note that acquire_curproc() already optimizes making the current process
215 * P_CURPROC, so setrunqueue() does not need to.
217 * If P_CP_RELEASED is not set we place the process on the run queue and we
218 * signal other cpus in the system that may need to be woken up to service
219 * the new 'user' process.
221 * If P_PASSIVE_ACQ is set setrunqueue() will not wakeup potential target
222 * cpus in an attempt to keep the process on the current cpu at least for
223 * a little while to take advantage of locality of reference (e.g. fork/exec
224 * or short fork/exit).
226 * CPU AFFINITY: cpu affinity is handled by attempting to either schedule
227 * or (user level) preempt on the same cpu that a process was previously
228 * scheduled to. If we cannot do this but we are at enough of a higher
229 * priority then the processes running on other cpus, we will allow the
230 * process to be stolen by another cpu.
232 * WARNING! a thread can be acquired by another cpu the moment it is put
233 * on the user scheduler's run queue AND we release the MP lock. Since we
234 * release the MP lock before switching out another cpu may begin stealing
235 * our current thread before we are completely switched out! The
236 * lwkt_acquire() function will stall until TDF_RUNNING is cleared on the
237 * thread before stealing it.
239 * The associated thread must NOT be scheduled.
240 * The process must be runnable.
241 * This must be called at splhigh().
244 setrunqueue(struct proc *p)
247 struct globaldata *gd;
256 KASSERT(p->p_stat == SRUN, ("setrunqueue: proc not SRUN"));
257 KASSERT((p->p_flag & (P_ONRUNQ|P_CURPROC)) == 0,
258 ("process %d already on runq! flag %08x", p->p_pid, p->p_flag));
259 KKASSERT((p->p_thread->td_flags & TDF_RUNQ) == 0);
262 * If we have been released from the userland scheduler we
263 * directly schedule its thread.
265 if (p->p_flag & P_CP_RELEASED) {
266 lwkt_schedule(p->p_thread);
272 * Check cpu affinity. The associated thread is stable at the
273 * moment. Note that we may be checking another cpu here so we
274 * have to be careful. Note that gd_upri only counts when the
275 * curprocmask bit is set for the cpu in question, and since it is
276 * only a hint we can modify it on another cpu's globaldata structure.
277 * We use it to prevent unnecessary IPIs (hence the - PPQ).
279 gd = p->p_thread->td_gd;
280 cpuid = gd->gd_cpuid;
282 if ((curprocmask & (1 << cpuid)) == 0) {
283 curprocmask |= 1 << cpuid;
284 p->p_flag |= P_CURPROC;
285 gd->gd_upri = p->p_priority;
286 USCHED_COUNTER(p->p_thread);
287 lwkt_schedule(p->p_thread);
288 /* CANNOT TOUCH PROC OR TD AFTER SCHEDULE CALL TO REMOTE CPU */
292 ++remote_resched_affinity;
298 * gd and cpuid may still 'hint' at another cpu. Even so we have
299 * to place this process on the userland scheduler's run queue for
300 * action by the target cpu.
303 p->p_flag |= P_ONRUNQ;
304 if (p->p_rtprio.type == RTP_PRIO_NORMAL) {
305 pri = (p->p_priority & PRIMASK) >> 2;
307 queuebits |= 1 << pri;
308 } else if (p->p_rtprio.type == RTP_PRIO_REALTIME ||
309 p->p_rtprio.type == RTP_PRIO_FIFO) {
310 pri = (u_int8_t)p->p_rtprio.prio;
312 rtqueuebits |= 1 << pri;
313 } else if (p->p_rtprio.type == RTP_PRIO_IDLE) {
314 pri = (u_int8_t)p->p_rtprio.prio;
316 idqueuebits |= 1 << pri;
318 panic("setrunqueue: invalid rtprio type");
321 p->p_rqindex = pri; /* remember the queue index */
322 TAILQ_INSERT_TAIL(q, p, p_procq);
326 * Either wakeup other cpus user thread scheduler or request
327 * preemption on other cpus (which will also wakeup a HLT).
329 * NOTE! gd and cpuid may still be our 'hint', not our current
336 * Check cpu affinity for user preemption (when the curprocmask bit
340 if (p->p_priority / PPQ < gd->gd_upri / PPQ) {
344 } else if (remote_resched) {
345 if (p->p_priority / PPQ < gd->gd_upri / PPQ) {
346 gd->gd_upri = p->p_priority;
347 lwkt_send_ipiq(gd, need_resched_remote, NULL);
349 ++remote_resched_affinity;
354 * No affinity, first schedule to any cpus that do not have a current
355 * process. If there is a free cpu we always schedule to it.
358 (mask = ~curprocmask & rdyprocmask & mycpu->gd_other_cpus) != 0 &&
359 (p->p_flag & P_PASSIVE_ACQ) == 0) {
361 printf("PROC %d nocpu to schedule it on\n", p->p_pid);
362 while (mask && count) {
364 KKASSERT((curprocmask & (1 << cpuid)) == 0);
365 rdyprocmask &= ~(1 << cpuid);
366 lwkt_schedule(&globaldata_find(cpuid)->gd_schedthread);
368 mask &= ~(1 << cpuid);
373 * If there are still runnable processes try to wakeup a random
374 * cpu that is running a much lower priority process in order to
375 * preempt on it. Note that gd_upri is only a hint, so we can
376 * overwrite it from the wrong cpu. If we can't find one, we
379 * We depress the priority check so multiple cpu bound programs
380 * do not bounce between cpus. Remember that the clock interrupt
381 * will also cause all cpus to reschedule.
383 if (count && remote_resched && ncpus > 1) {
386 if (++cpuid == ncpus)
388 } while (cpuid == mycpu->gd_cpuid);
391 gd = globaldata_find(cpuid);
393 if (p->p_priority / PPQ < gd->gd_upri / PPQ - 2) {
394 gd->gd_upri = p->p_priority;
395 lwkt_send_ipiq(gd, need_resched_remote, NULL);
396 ++remote_resched_nonaffinity;
400 if (p->p_priority / PPQ < gd->gd_upri / PPQ) {
408 * remrunqueue() removes a given process from the run queue that it is on,
409 * clearing the queue busy bit if it becomes empty. This function is called
410 * when a userland process is selected for LWKT scheduling. Note that
411 * LWKT scheduling is an abstraction of 'curproc'.. there could very well be
412 * several userland processes whos threads are scheduled or otherwise in
413 * a special state, and such processes are NOT on the userland scheduler's
416 * This must be called at splhigh().
419 remrunqueue(struct proc *p)
426 KASSERT((p->p_flag & P_ONRUNQ) != 0, ("not on runq4!"));
427 p->p_flag &= ~P_ONRUNQ;
429 KKASSERT(runqcount >= 0);
431 if (p->p_rtprio.type == RTP_PRIO_NORMAL) {
434 } else if (p->p_rtprio.type == RTP_PRIO_REALTIME ||
435 p->p_rtprio.type == RTP_PRIO_FIFO) {
437 which = &rtqueuebits;
438 } else if (p->p_rtprio.type == RTP_PRIO_IDLE) {
440 which = &idqueuebits;
442 panic("remrunqueue: invalid rtprio type");
444 TAILQ_REMOVE(q, p, p_procq);
445 if (TAILQ_EMPTY(q)) {
446 KASSERT((*which & (1 << pri)) != 0,
447 ("remrunqueue: remove from empty queue"));
448 *which &= ~(1 << pri);
454 * Release the P_CURPROC designation on the current process for this cpu
455 * and attempt to assign a new current process from the run queue.
457 * If we do not have or cannot get the MP lock we just wakeup the userland
458 * helper scheduler thread for this cpu.
460 * WARNING! The MP lock may be in an unsynchronized state due to the
461 * way get_mplock() works and the fact that this function may be called
462 * from a passive release during a lwkt_switch(). try_mplock() will deal
463 * with this for us but you should be aware that td_mpcount may not be
467 release_curproc(struct proc *p)
472 #ifdef ONLY_ONE_USER_CPU
473 KKASSERT(mycpu->gd_cpuid == 0 && p->p_thread->td_gd == mycpu);
477 cpuid = p->p_thread->td_gd->gd_cpuid;
478 if ((p->p_flag & P_CP_RELEASED) == 0) {
479 p->p_flag |= P_CP_RELEASED;
480 lwkt_setpri_self(TDPRI_KERN_USER);
482 if (p->p_flag & P_CURPROC) {
483 p->p_flag &= ~P_CURPROC;
484 curprocmask &= ~(1 << cpuid);
487 * Choose the next process to assign P_CURPROC to.
488 * Note that we cannot schedule gd_schedthread
489 * if runqcount is 0 without creating a scheduling
492 if ((np = chooseproc(NULL)) != NULL) {
493 curprocmask |= 1 << cpuid;
494 np->p_flag |= P_CURPROC;
495 mycpu->gd_upri = np->p_priority;
496 USCHED_COUNTER(np->p_thread);
497 lwkt_acquire(np->p_thread);
498 lwkt_schedule(np->p_thread);
499 } else if (runqcount && (rdyprocmask & (1 << cpuid))) {
500 rdyprocmask &= ~(1 << cpuid);
501 lwkt_schedule(&mycpu->gd_schedthread);
505 KKASSERT(0); /* MP LOCK ALWAYS HELD AT THE MOMENT */
506 if (runqcount && (rdyprocmask & (1 << cpuid))) {
507 rdyprocmask &= ~(1 << cpuid);
508 lwkt_schedule(&mycpu->gd_schedthread);
516 * Acquire the P_CURPROC designation on the CURRENT process only. This
517 * function is called prior to returning to userland. If the system
518 * call or trap did not block and if no reschedule was requested it is
519 * highly likely that the P_CURPROC flag is still set in the proc, and
520 * we do almost nothing here.
523 acquire_curproc(struct proc *p)
529 * Short cut, we've already acquired the designation or we never
530 * lost it in the first place. P_CP_RELEASED is cleared, meaning
531 * that the process is again under the control of the userland
532 * scheduler. We do not have to fiddle with the LWKT priority,
533 * the trap code (userret/userexit) will do that for us.
535 if ((p->p_flag & P_CURPROC) != 0) {
536 p->p_flag &= ~P_CP_RELEASED;
541 * Long cut. This pulls in a bit of the userland scheduler as
542 * an optimization. If our cpu has not scheduled a userland
543 * process we gladly fill the slot, otherwise we choose the best
544 * candidate from the run queue and compare it against ourselves,
545 * scheduling either us or him depending.
547 * If our cpu's slot isn't free we put ourselves on the userland
548 * run queue and switch away. We should have P_CURPROC when we
549 * come back. Note that a cpu change can occur when we come back.
551 * YYY don't need critical section, we hold giant and no interrupt
552 * will mess w/ this proc? Or will it? What about curprocmask?
554 #ifdef ONLY_ONE_USER_CPU
555 KKASSERT(mycpu->gd_cpuid == 0 && p->p_thread->td_gd == mycpu);
559 while ((p->p_flag & P_CURPROC) == 0) {
563 cpuid = p->p_thread->td_gd->gd_cpuid;
566 * (broken out from setrunqueue() as an optimization that
567 * allows us to avoid descheduling and rescheduling ourself)
569 * Interlock against the helper scheduler thread by setting
570 * curprocmask while we choose a new process. Check our
571 * process against the new process to shortcut setrunqueue()
572 * and remrunqueue() operations.
574 if ((curprocmask & (1 << cpuid)) == 0) {
575 curprocmask |= 1 << cpuid;
577 if ((np = chooseproc(p)) != NULL) {
578 KKASSERT((np->p_flag & P_CP_RELEASED) == 0);
579 np->p_flag |= P_CURPROC;
580 mycpu->gd_upri = np->p_priority;
581 USCHED_COUNTER(np->p_thread);
582 lwkt_acquire(np->p_thread);
583 lwkt_schedule(np->p_thread);
585 p->p_flag |= P_CURPROC;
589 lwkt_deschedule_self();
590 p->p_flag &= ~P_CP_RELEASED;
592 lwkt_switch(); /* CPU CAN CHANGE DUE TO SETRUNQUEUE() */
593 KASSERT((p->p_flag & (P_ONRUNQ|P_CURPROC|P_CP_RELEASED)) == P_CURPROC, ("unexpected p_flag %08x acquiring P_CURPROC\n", p->p_flag));
599 * Yield / synchronous reschedule. This is a bit tricky because the trap
600 * code might have set a lazy release on the switch function. Setting
601 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call
602 * switch, and that we will not be rescheduled to another cpu when we attempt
603 * to re-acquire P_CURPROC.
605 * We have to release P_CURPROC (by calling lwkt_switch(), and acquire it
606 * again to yield to another user process. Note that the release will
607 * ensure that we are running at a kernel LWKT priority, and this priority
608 * is not lowered through the reacquisition and rerelease sequence to ensure
609 * that we do not deadlock against a higher priority *user* process.
614 struct thread *td = curthread;
615 struct proc *p = td->td_proc;
618 p->p_flag |= P_PASSIVE_ACQ;
622 p->p_flag &= ~P_PASSIVE_ACQ;
630 * For SMP systems a user scheduler helper thread is created for each
631 * cpu and is used to allow one cpu to wakeup another for the purposes of
632 * scheduling userland threads from setrunqueue(). UP systems do not
633 * need the helper since there is only one cpu. We can't use the idle
634 * thread for this because we need to hold the MP lock. Additionally,
635 * doing things this way allows us to HLT idle cpus on MP systems.
641 sched_thread(void *dummy)
643 int cpuid = mycpu->gd_cpuid; /* doesn't change */
644 u_int32_t cpumask = 1 << cpuid; /* doesn't change */
646 #ifdef ONLY_ONE_USER_CPU
647 KKASSERT(cpuid == 0);
650 get_mplock(); /* hold the MP lock */
654 lwkt_deschedule_self(); /* interlock */
655 rdyprocmask |= cpumask;
657 if ((curprocmask & cpumask) == 0 && (np = chooseproc(NULL)) != NULL) {
658 curprocmask |= cpumask;
659 np->p_flag |= P_CURPROC;
660 mycpu->gd_upri = np->p_priority;
661 USCHED_COUNTER(np->p_thread);
662 lwkt_acquire(np->p_thread);
663 lwkt_schedule(np->p_thread);
671 sched_thread_init(void)
673 int cpuid = mycpu->gd_cpuid;
675 lwkt_create(sched_thread, NULL, NULL, &mycpu->gd_schedthread,
678 curprocmask &= ~(1 << cpuid); /* schedule user proc on cpu */
679 #ifdef ONLY_ONE_USER_CPU
681 curprocmask |= 1 << cpuid; /* DISABLE USER PROCS */
683 rdyprocmask |= 1 << cpuid;