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 * $DragonFly: src/sys/kern/usched_bsd4.c,v 1.26 2008/11/01 23:31:19 dillon Exp $
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/kernel.h>
33 #include <sys/queue.h>
35 #include <sys/rtprio.h>
37 #include <sys/sysctl.h>
38 #include <sys/resourcevar.h>
39 #include <sys/spinlock.h>
40 #include <machine/cpu.h>
41 #include <machine/smp.h>
43 #include <sys/thread2.h>
44 #include <sys/spinlock2.h>
45 #include <sys/mplock2.h>
48 * Priorities. Note that with 32 run queues per scheduler each queue
49 * represents four priority levels.
53 #define PRIMASK (MAXPRI - 1)
54 #define PRIBASE_REALTIME 0
55 #define PRIBASE_NORMAL MAXPRI
56 #define PRIBASE_IDLE (MAXPRI * 2)
57 #define PRIBASE_THREAD (MAXPRI * 3)
58 #define PRIBASE_NULL (MAXPRI * 4)
60 #define NQS 32 /* 32 run queues. */
61 #define PPQ (MAXPRI / NQS) /* priorities per queue */
62 #define PPQMASK (PPQ - 1)
65 * NICEPPQ - number of nice units per priority queue
67 * ESTCPUPPQ - number of estcpu units per priority queue
68 * ESTCPUMAX - number of estcpu units
72 #define ESTCPUMAX (ESTCPUPPQ * NQS)
73 #define BATCHMAX (ESTCPUFREQ * 30)
74 #define PRIO_RANGE (PRIO_MAX - PRIO_MIN + 1)
76 #define ESTCPULIM(v) min((v), ESTCPUMAX)
80 #define lwp_priority lwp_usdata.bsd4.priority
81 #define lwp_rqindex lwp_usdata.bsd4.rqindex
82 #define lwp_estcpu lwp_usdata.bsd4.estcpu
83 #define lwp_batch lwp_usdata.bsd4.batch
84 #define lwp_rqtype lwp_usdata.bsd4.rqtype
86 static void bsd4_acquire_curproc(struct lwp *lp);
87 static void bsd4_release_curproc(struct lwp *lp);
88 static void bsd4_select_curproc(globaldata_t gd);
89 static void bsd4_setrunqueue(struct lwp *lp);
90 static void bsd4_schedulerclock(struct lwp *lp, sysclock_t period,
92 static void bsd4_recalculate_estcpu(struct lwp *lp);
93 static void bsd4_resetpriority(struct lwp *lp);
94 static void bsd4_forking(struct lwp *plp, struct lwp *lp);
95 static void bsd4_exiting(struct lwp *lp, struct proc *);
96 static void bsd4_yield(struct lwp *lp);
99 static void need_user_resched_remote(void *dummy);
101 static struct lwp *chooseproc_locked(struct lwp *chklp);
102 static void bsd4_remrunqueue_locked(struct lwp *lp);
103 static void bsd4_setrunqueue_locked(struct lwp *lp);
105 struct usched usched_bsd4 = {
107 "bsd4", "Original DragonFly Scheduler",
108 NULL, /* default registration */
109 NULL, /* default deregistration */
110 bsd4_acquire_curproc,
111 bsd4_release_curproc,
114 bsd4_recalculate_estcpu,
118 NULL, /* setcpumask not supported */
122 struct usched_bsd4_pcpu {
123 struct thread helper_thread;
126 struct lwp *uschedcp;
129 typedef struct usched_bsd4_pcpu *bsd4_pcpu_t;
132 * We have NQS (32) run queues per scheduling class. For the normal
133 * class, there are 128 priorities scaled onto these 32 queues. New
134 * processes are added to the last entry in each queue, and processes
135 * are selected for running by taking them from the head and maintaining
136 * a simple FIFO arrangement. Realtime and Idle priority processes have
137 * and explicit 0-31 priority which maps directly onto their class queue
138 * index. When a queue has something in it, the corresponding bit is
139 * set in the queuebits variable, allowing a single read to determine
140 * the state of all 32 queues and then a ffs() to find the first busy
143 static struct rq bsd4_queues[NQS];
144 static struct rq bsd4_rtqueues[NQS];
145 static struct rq bsd4_idqueues[NQS];
146 static u_int32_t bsd4_queuebits;
147 static u_int32_t bsd4_rtqueuebits;
148 static u_int32_t bsd4_idqueuebits;
149 static cpumask_t bsd4_curprocmask = -1; /* currently running a user process */
150 static cpumask_t bsd4_rdyprocmask; /* ready to accept a user process */
151 static int bsd4_runqcount;
153 static volatile int bsd4_scancpu;
155 static struct spinlock bsd4_spin;
156 static struct usched_bsd4_pcpu bsd4_pcpu[MAXCPU];
158 SYSCTL_INT(_debug, OID_AUTO, bsd4_runqcount, CTLFLAG_RD, &bsd4_runqcount, 0,
159 "Number of run queues");
161 static int usched_nonoptimal;
162 SYSCTL_INT(_debug, OID_AUTO, usched_nonoptimal, CTLFLAG_RW,
163 &usched_nonoptimal, 0, "acquire_curproc() was not optimal");
164 static int usched_optimal;
165 SYSCTL_INT(_debug, OID_AUTO, usched_optimal, CTLFLAG_RW,
166 &usched_optimal, 0, "acquire_curproc() was optimal");
168 static int usched_debug = -1;
169 SYSCTL_INT(_debug, OID_AUTO, scdebug, CTLFLAG_RW, &usched_debug, 0,
170 "Print debug information for this pid");
172 static int remote_resched_nonaffinity;
173 static int remote_resched_affinity;
174 static int choose_affinity;
175 SYSCTL_INT(_debug, OID_AUTO, remote_resched_nonaffinity, CTLFLAG_RD,
176 &remote_resched_nonaffinity, 0, "Number of remote rescheds");
177 SYSCTL_INT(_debug, OID_AUTO, remote_resched_affinity, CTLFLAG_RD,
178 &remote_resched_affinity, 0, "Number of remote rescheds");
179 SYSCTL_INT(_debug, OID_AUTO, choose_affinity, CTLFLAG_RD,
180 &choose_affinity, 0, "chooseproc() was smart");
183 static int usched_bsd4_rrinterval = (ESTCPUFREQ + 9) / 10;
184 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_rrinterval, CTLFLAG_RW,
185 &usched_bsd4_rrinterval, 0, "");
186 static int usched_bsd4_decay = 8;
187 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_decay, CTLFLAG_RW,
188 &usched_bsd4_decay, 0, "Extra decay when not running");
189 static int usched_bsd4_batch_time = 10;
190 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_batch_time, CTLFLAG_RW,
191 &usched_bsd4_batch_time, 0, "Minimum batch counter value");
194 * Initialize the run queues at boot time.
201 spin_init(&bsd4_spin);
202 for (i = 0; i < NQS; i++) {
203 TAILQ_INIT(&bsd4_queues[i]);
204 TAILQ_INIT(&bsd4_rtqueues[i]);
205 TAILQ_INIT(&bsd4_idqueues[i]);
207 atomic_clear_cpumask(&bsd4_curprocmask, 1);
209 SYSINIT(runqueue, SI_BOOT2_USCHED, SI_ORDER_FIRST, rqinit, NULL)
212 * BSD4_ACQUIRE_CURPROC
214 * This function is called when the kernel intends to return to userland.
215 * It is responsible for making the thread the current designated userland
216 * thread for this cpu, blocking if necessary.
218 * The kernel has already depressed our LWKT priority so we must not switch
219 * until we have either assigned or disposed of the thread.
221 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
222 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will
223 * occur, this function is called only under very controlled circumstances.
228 bsd4_acquire_curproc(struct lwp *lp)
235 bsd4_recalculate_estcpu(lp);
238 * If a reschedule was requested give another thread the
241 if (user_resched_wanted()) {
242 clear_user_resched();
243 bsd4_release_curproc(lp);
247 * Loop until we are the current user thread
251 * Reload after a switch or setrunqueue/switch possibly
252 * moved us to another cpu.
254 /*clear_lwkt_resched();*/
256 dd = &bsd4_pcpu[gd->gd_cpuid];
259 * Become the currently scheduled user thread for this cpu
260 * if we can do so trivially.
262 * We can steal another thread's current thread designation
263 * on this cpu since if we are running that other thread
264 * must not be, so we can safely deschedule it.
266 if (dd->uschedcp == lp) {
268 * We are already the current lwp (hot path).
270 dd->upri = lp->lwp_priority;
271 } else if (dd->uschedcp == NULL) {
273 * We can trivially become the current lwp.
275 atomic_set_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
277 dd->upri = lp->lwp_priority;
278 } else if (dd->upri > lp->lwp_priority) {
280 * We can steal the current lwp designation from the
281 * olp that was previously assigned to this cpu.
285 dd->upri = lp->lwp_priority;
286 lwkt_deschedule(olp->lwp_thread);
287 bsd4_setrunqueue(olp);
290 * We cannot become the current lwp, place the lp
291 * on the bsd4 run-queue and deschedule ourselves.
293 lwkt_deschedule(lp->lwp_thread);
294 bsd4_setrunqueue(lp);
299 * Other threads at our current user priority have already
300 * put in their bids, but we must run any kernel threads
301 * at higher priorities, and we could lose our bid to
302 * another thread trying to return to user mode in the
305 * If we lose our bid we will be descheduled and put on
306 * the run queue. When we are reactivated we will have
309 if (lwkt_resched_wanted() ||
310 lp->lwp_thread->td_fairq_accum < 0) {
313 } while (dd->uschedcp != lp);
316 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
320 * BSD4_RELEASE_CURPROC
322 * This routine detaches the current thread from the userland scheduler,
323 * usually because the thread needs to run or block in the kernel (at
324 * kernel priority) for a while.
326 * This routine is also responsible for selecting a new thread to
327 * make the current thread.
329 * NOTE: This implementation differs from the dummy example in that
330 * bsd4_select_curproc() is able to select the current process, whereas
331 * dummy_select_curproc() is not able to select the current process.
332 * This means we have to NULL out uschedcp.
334 * Additionally, note that we may already be on a run queue if releasing
335 * via the lwkt_switch() in bsd4_setrunqueue().
340 bsd4_release_curproc(struct lwp *lp)
342 globaldata_t gd = mycpu;
343 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
345 if (dd->uschedcp == lp) {
347 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
348 dd->uschedcp = NULL; /* don't let lp be selected */
349 dd->upri = PRIBASE_NULL;
350 atomic_clear_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
351 bsd4_select_curproc(gd);
357 * BSD4_SELECT_CURPROC
359 * Select a new current process for this cpu and clear any pending user
360 * reschedule request. The cpu currently has no current process.
362 * This routine is also responsible for equal-priority round-robining,
363 * typically triggered from bsd4_schedulerclock(). In our dummy example
364 * all the 'user' threads are LWKT scheduled all at once and we just
365 * call lwkt_switch().
367 * The calling process is not on the queue and cannot be selected.
373 bsd4_select_curproc(globaldata_t gd)
375 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
377 int cpuid = gd->gd_cpuid;
381 spin_lock(&bsd4_spin);
382 if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) {
383 atomic_set_cpumask(&bsd4_curprocmask, CPUMASK(cpuid));
384 dd->upri = nlp->lwp_priority;
386 spin_unlock(&bsd4_spin);
388 lwkt_acquire(nlp->lwp_thread);
390 lwkt_schedule(nlp->lwp_thread);
392 spin_unlock(&bsd4_spin);
395 } else if (bsd4_runqcount && (bsd4_rdyprocmask & CPUMASK(cpuid))) {
396 atomic_clear_cpumask(&bsd4_rdyprocmask, CPUMASK(cpuid));
397 spin_unlock(&bsd4_spin);
398 lwkt_schedule(&dd->helper_thread);
400 spin_unlock(&bsd4_spin);
409 * Place the specified lwp on the user scheduler's run queue. This routine
410 * must be called with the thread descheduled. The lwp must be runnable.
412 * The thread may be the current thread as a special case.
417 bsd4_setrunqueue(struct lwp *lp)
428 * First validate the process state relative to the current cpu.
429 * We don't need the spinlock for this, just a critical section.
430 * We are in control of the process.
433 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
434 KASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0,
435 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
436 lp->lwp_tid, lp->lwp_proc->p_flag, lp->lwp_flag));
437 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
440 * Note: gd and dd are relative to the target thread's last cpu,
441 * NOT our current cpu.
443 gd = lp->lwp_thread->td_gd;
444 dd = &bsd4_pcpu[gd->gd_cpuid];
447 * This process is not supposed to be scheduled anywhere or assigned
448 * as the current process anywhere. Assert the condition.
450 KKASSERT(dd->uschedcp != lp);
454 * If we are not SMP we do not have a scheduler helper to kick
455 * and must directly activate the process if none are scheduled.
457 * This is really only an issue when bootstrapping init since
458 * the caller in all other cases will be a user process, and
459 * even if released (dd->uschedcp == NULL), that process will
460 * kickstart the scheduler when it returns to user mode from
463 if (dd->uschedcp == NULL) {
464 atomic_set_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
466 dd->upri = lp->lwp_priority;
467 lwkt_schedule(lp->lwp_thread);
475 * XXX fixme. Could be part of a remrunqueue/setrunqueue
476 * operation when the priority is recalculated, so TDF_MIGRATING
477 * may already be set.
479 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
480 lwkt_giveaway(lp->lwp_thread);
484 * We lose control of lp the moment we release the spinlock after
485 * having placed lp on the queue. i.e. another cpu could pick it
486 * up and it could exit, or its priority could be further adjusted,
487 * or something like that.
489 spin_lock(&bsd4_spin);
490 bsd4_setrunqueue_locked(lp);
494 * Kick the scheduler helper on one of the other cpu's
495 * and request a reschedule if appropriate.
497 * NOTE: We check all cpus whos rdyprocmask is set. First we
498 * look for cpus without designated lps, then we look for
499 * cpus with designated lps with a worse priority than our
503 cpuid = (bsd4_scancpu & 0xFFFF) % ncpus;
504 mask = ~bsd4_curprocmask & bsd4_rdyprocmask & lp->lwp_cpumask &
505 smp_active_mask & usched_global_cpumask;
508 tmpmask = ~(CPUMASK(cpuid) - 1);
510 cpuid = BSFCPUMASK(mask & tmpmask);
512 cpuid = BSFCPUMASK(mask);
513 gd = globaldata_find(cpuid);
514 dd = &bsd4_pcpu[cpuid];
516 if ((dd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
518 mask &= ~CPUMASK(cpuid);
522 * Then cpus which might have a currently running lp
524 mask = bsd4_curprocmask & bsd4_rdyprocmask &
525 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
528 tmpmask = ~(CPUMASK(cpuid) - 1);
530 cpuid = BSFCPUMASK(mask & tmpmask);
532 cpuid = BSFCPUMASK(mask);
533 gd = globaldata_find(cpuid);
534 dd = &bsd4_pcpu[cpuid];
536 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
538 mask &= ~CPUMASK(cpuid);
542 * If we cannot find a suitable cpu we reload from bsd4_scancpu
543 * and round-robin. Other cpus will pickup as they release their
544 * current lwps or become ready.
546 * Avoid a degenerate system lockup case if usched_global_cpumask
547 * is set to 0 or otherwise does not cover lwp_cpumask.
549 * We only kick the target helper thread in this case, we do not
550 * set the user resched flag because
552 cpuid = (bsd4_scancpu & 0xFFFF) % ncpus;
553 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0) {
556 gd = globaldata_find(cpuid);
557 dd = &bsd4_pcpu[cpuid];
560 spin_unlock(&bsd4_spin);
561 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
562 if (dd->uschedcp == NULL) {
563 lwkt_schedule(&dd->helper_thread);
569 atomic_clear_cpumask(&bsd4_rdyprocmask, CPUMASK(cpuid));
570 spin_unlock(&bsd4_spin);
571 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
572 lwkt_send_ipiq(gd, need_user_resched_remote, NULL);
574 lwkt_schedule(&dd->helper_thread);
578 * Request a reschedule if appropriate.
580 spin_unlock(&bsd4_spin);
581 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
589 * This routine is called from a systimer IPI. It MUST be MP-safe and
590 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
597 bsd4_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
599 globaldata_t gd = mycpu;
600 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
603 * Do we need to round-robin? We round-robin 10 times a second.
604 * This should only occur for cpu-bound batch processes.
606 if (++dd->rrcount >= usched_bsd4_rrinterval) {
612 * Adjust estcpu upward using a real time equivalent calculation.
614 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
617 * Spinlocks also hold a critical section so there should not be
620 KKASSERT(gd->gd_spinlocks_wr == 0);
622 bsd4_resetpriority(lp);
625 * if we can't call bsd4_resetpriority for some reason we must call
626 * need user_resched().
633 * Called from acquire and from kern_synch's one-second timer (one of the
634 * callout helper threads) with a critical section held.
636 * Decay p_estcpu based on the number of ticks we haven't been running
637 * and our p_nice. As the load increases each process observes a larger
638 * number of idle ticks (because other processes are running in them).
639 * This observation leads to a larger correction which tends to make the
640 * system more 'batchy'.
642 * Note that no recalculation occurs for a process which sleeps and wakes
643 * up in the same tick. That is, a system doing thousands of context
644 * switches per second will still only do serious estcpu calculations
645 * ESTCPUFREQ times per second.
651 bsd4_recalculate_estcpu(struct lwp *lp)
653 globaldata_t gd = mycpu;
660 * We have to subtract periodic to get the last schedclock
661 * timeout time, otherwise we would get the upcoming timeout.
662 * Keep in mind that a process can migrate between cpus and
663 * while the scheduler clock should be very close, boundary
664 * conditions could lead to a small negative delta.
666 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
668 if (lp->lwp_slptime > 1) {
670 * Too much time has passed, do a coarse correction.
672 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
673 bsd4_resetpriority(lp);
674 lp->lwp_cpbase = cpbase;
676 lp->lwp_batch -= ESTCPUFREQ;
677 if (lp->lwp_batch < 0)
679 } else if (lp->lwp_cpbase != cpbase) {
681 * Adjust estcpu if we are in a different tick. Don't waste
682 * time if we are in the same tick.
684 * First calculate the number of ticks in the measurement
685 * interval. The ttlticks calculation can wind up 0 due to
686 * a bug in the handling of lwp_slptime (as yet not found),
687 * so make sure we do not get a divide by 0 panic.
689 ttlticks = (cpbase - lp->lwp_cpbase) /
690 gd->gd_schedclock.periodic;
693 lp->lwp_cpbase = cpbase;
697 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
700 * Calculate the percentage of one cpu used factoring in ncpus
701 * and the load and adjust estcpu. Handle degenerate cases
702 * by adding 1 to bsd4_runqcount.
704 * estcpu is scaled by ESTCPUMAX.
706 * bsd4_runqcount is the excess number of user processes
707 * that cannot be immediately scheduled to cpus. We want
708 * to count these as running to avoid range compression
709 * in the base calculation (which is the actual percentage
712 estcpu = (lp->lwp_cpticks * ESTCPUMAX) *
713 (bsd4_runqcount + ncpus) / (ncpus * ttlticks);
716 * If estcpu is > 50% we become more batch-like
717 * If estcpu is <= 50% we become less batch-like
719 * It takes 30 cpu seconds to traverse the entire range.
721 if (estcpu > ESTCPUMAX / 2) {
722 lp->lwp_batch += ttlticks;
723 if (lp->lwp_batch > BATCHMAX)
724 lp->lwp_batch = BATCHMAX;
726 lp->lwp_batch -= ttlticks;
727 if (lp->lwp_batch < 0)
731 if (usched_debug == lp->lwp_proc->p_pid) {
732 kprintf("pid %d lwp %p estcpu %3d %3d bat %d cp %d/%d",
733 lp->lwp_proc->p_pid, lp,
734 estcpu, lp->lwp_estcpu,
736 lp->lwp_cpticks, ttlticks);
740 * Adjust lp->lwp_esetcpu. The decay factor determines how
741 * quickly lwp_estcpu collapses to its realtime calculation.
742 * A slower collapse gives us a more accurate number but
743 * can cause a cpu hog to eat too much cpu before the
744 * scheduler decides to downgrade it.
746 * NOTE: p_nice is accounted for in bsd4_resetpriority(),
747 * and not here, but we must still ensure that a
748 * cpu-bound nice -20 process does not completely
749 * override a cpu-bound nice +20 process.
751 * NOTE: We must use ESTCPULIM() here to deal with any
754 decay_factor = usched_bsd4_decay;
755 if (decay_factor < 1)
757 if (decay_factor > 1024)
760 lp->lwp_estcpu = ESTCPULIM(
761 (lp->lwp_estcpu * decay_factor + estcpu) /
764 if (usched_debug == lp->lwp_proc->p_pid)
765 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
766 bsd4_resetpriority(lp);
767 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
773 * Compute the priority of a process when running in user mode.
774 * Arrange to reschedule if the resulting priority is better
775 * than that of the current process.
777 * This routine may be called with any process.
779 * This routine is called by fork1() for initial setup with the process
780 * of the run queue, and also may be called normally with the process on or
786 bsd4_resetpriority(struct lwp *lp)
796 * Calculate the new priority and queue type
799 spin_lock(&bsd4_spin);
801 newrqtype = lp->lwp_rtprio.type;
804 case RTP_PRIO_REALTIME:
806 newpriority = PRIBASE_REALTIME +
807 (lp->lwp_rtprio.prio & PRIMASK);
809 case RTP_PRIO_NORMAL:
811 * Detune estcpu based on batchiness. lwp_batch ranges
812 * from 0 to BATCHMAX. Limit estcpu for the sake of
813 * the priority calculation to between 50% and 100%.
815 estcpu = lp->lwp_estcpu * (lp->lwp_batch + BATCHMAX) /
819 * p_nice piece Adds (0-40) * 2 0-80
820 * estcpu Adds 16384 * 4 / 512 0-128
822 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
823 newpriority += estcpu * PPQ / ESTCPUPPQ;
824 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
825 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
826 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
829 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
831 case RTP_PRIO_THREAD:
832 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
835 panic("Bad RTP_PRIO %d", newrqtype);
840 * The newpriority incorporates the queue type so do a simple masked
841 * check to determine if the process has moved to another queue. If
842 * it has, and it is currently on a run queue, then move it.
844 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
845 lp->lwp_priority = newpriority;
846 if (lp->lwp_flag & LWP_ONRUNQ) {
847 bsd4_remrunqueue_locked(lp);
848 lp->lwp_rqtype = newrqtype;
849 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
850 bsd4_setrunqueue_locked(lp);
853 lp->lwp_rqtype = newrqtype;
854 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
857 reschedcpu = lp->lwp_thread->td_gd->gd_cpuid;
859 lp->lwp_priority = newpriority;
865 * Determine if we need to reschedule the target cpu. This only
866 * occurs if the LWP is already on a scheduler queue, which means
867 * that idle cpu notification has already occured. At most we
868 * need only issue a need_user_resched() on the appropriate cpu.
870 * The LWP may be owned by a CPU different from the current one,
871 * in which case dd->uschedcp may be modified without an MP lock
872 * or a spinlock held. The worst that happens is that the code
873 * below causes a spurious need_user_resched() on the target CPU
874 * and dd->pri to be wrong for a short period of time, both of
875 * which are harmless.
877 * If checkpri is 0 we are adjusting the priority of the current
878 * process, possibly higher (less desireable), so ignore the upri
879 * check which will fail in that case.
881 if (reschedcpu >= 0) {
882 dd = &bsd4_pcpu[reschedcpu];
883 if ((bsd4_rdyprocmask & CPUMASK(reschedcpu)) &&
885 (dd->upri & ~PRIMASK) > (lp->lwp_priority & ~PRIMASK))) {
887 if (reschedcpu == mycpu->gd_cpuid) {
888 spin_unlock(&bsd4_spin);
891 spin_unlock(&bsd4_spin);
892 atomic_clear_cpumask(&bsd4_rdyprocmask,
893 CPUMASK(reschedcpu));
894 lwkt_send_ipiq(lp->lwp_thread->td_gd,
895 need_user_resched_remote, NULL);
898 spin_unlock(&bsd4_spin);
902 spin_unlock(&bsd4_spin);
905 spin_unlock(&bsd4_spin);
915 bsd4_yield(struct lwp *lp)
918 /* FUTURE (or something similar) */
919 switch(lp->lwp_rqtype) {
920 case RTP_PRIO_NORMAL:
921 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
931 * Called from fork1() when a new child process is being created.
933 * Give the child process an initial estcpu that is more batch then
934 * its parent and dock the parent for the fork (but do not
935 * reschedule the parent). This comprises the main part of our batch
936 * detection heuristic for both parallel forking and sequential execs.
938 * XXX lwp should be "spawning" instead of "forking"
943 bsd4_forking(struct lwp *plp, struct lwp *lp)
946 * Put the child 4 queue slots (out of 32) higher than the parent
947 * (less desireable than the parent).
949 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
952 * The batch status of children always starts out centerline
953 * and will inch-up or inch-down as appropriate. It takes roughly
954 * ~15 seconds of >50% cpu to hit the limit.
956 lp->lwp_batch = BATCHMAX / 2;
959 * Dock the parent a cost for the fork, protecting us from fork
960 * bombs. If the parent is forking quickly make the child more
963 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
967 * Called when a parent waits for a child.
972 bsd4_exiting(struct lwp *lp, struct proc *child_proc)
977 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
978 * it selects a user process and returns it. If chklp is non-NULL and chklp
979 * has a better or equal priority then the process that would otherwise be
980 * chosen, NULL is returned.
982 * Until we fix the RUNQ code the chklp test has to be strict or we may
983 * bounce between processes trying to acquire the current process designation.
985 * MPSAFE - must be called with bsd4_spin exclusive held. The spinlock is
986 * left intact through the entire routine.
990 chooseproc_locked(struct lwp *chklp)
994 u_int32_t *which, *which2;
1001 rtqbits = bsd4_rtqueuebits;
1002 tsqbits = bsd4_queuebits;
1003 idqbits = bsd4_idqueuebits;
1004 cpumask = mycpu->gd_cpumask;
1010 pri = bsfl(rtqbits);
1011 q = &bsd4_rtqueues[pri];
1012 which = &bsd4_rtqueuebits;
1014 } else if (tsqbits) {
1015 pri = bsfl(tsqbits);
1016 q = &bsd4_queues[pri];
1017 which = &bsd4_queuebits;
1019 } else if (idqbits) {
1020 pri = bsfl(idqbits);
1021 q = &bsd4_idqueues[pri];
1022 which = &bsd4_idqueuebits;
1027 lp = TAILQ_FIRST(q);
1028 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1031 while ((lp->lwp_cpumask & cpumask) == 0) {
1032 lp = TAILQ_NEXT(lp, lwp_procq);
1034 *which2 &= ~(1 << pri);
1041 * If the passed lwp <chklp> is reasonably close to the selected
1042 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1044 * Note that we must error on the side of <chklp> to avoid bouncing
1045 * between threads in the acquire code.
1048 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1054 * If the chosen lwp does not reside on this cpu spend a few
1055 * cycles looking for a better candidate at the same priority level.
1056 * This is a fallback check, setrunqueue() tries to wakeup the
1057 * correct cpu and is our front-line affinity.
1059 if (lp->lwp_thread->td_gd != mycpu &&
1060 (chklp = TAILQ_NEXT(lp, lwp_procq)) != NULL
1062 if (chklp->lwp_thread->td_gd == mycpu) {
1069 TAILQ_REMOVE(q, lp, lwp_procq);
1072 *which &= ~(1 << pri);
1073 KASSERT((lp->lwp_flag & LWP_ONRUNQ) != 0, ("not on runq6!"));
1074 lp->lwp_flag &= ~LWP_ONRUNQ;
1082 need_user_resched_remote(void *dummy)
1084 globaldata_t gd = mycpu;
1085 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
1087 need_user_resched();
1088 lwkt_schedule(&dd->helper_thread);
1094 * bsd4_remrunqueue_locked() removes a given process from the run queue
1095 * that it is on, clearing the queue busy bit if it becomes empty.
1097 * Note that user process scheduler is different from the LWKT schedule.
1098 * The user process scheduler only manages user processes but it uses LWKT
1099 * underneath, and a user process operating in the kernel will often be
1100 * 'released' from our management.
1102 * MPSAFE - bsd4_spin must be held exclusively on call
1105 bsd4_remrunqueue_locked(struct lwp *lp)
1111 KKASSERT(lp->lwp_flag & LWP_ONRUNQ);
1112 lp->lwp_flag &= ~LWP_ONRUNQ;
1114 KKASSERT(bsd4_runqcount >= 0);
1116 pri = lp->lwp_rqindex;
1117 switch(lp->lwp_rqtype) {
1118 case RTP_PRIO_NORMAL:
1119 q = &bsd4_queues[pri];
1120 which = &bsd4_queuebits;
1122 case RTP_PRIO_REALTIME:
1124 q = &bsd4_rtqueues[pri];
1125 which = &bsd4_rtqueuebits;
1128 q = &bsd4_idqueues[pri];
1129 which = &bsd4_idqueuebits;
1132 panic("remrunqueue: invalid rtprio type");
1135 TAILQ_REMOVE(q, lp, lwp_procq);
1136 if (TAILQ_EMPTY(q)) {
1137 KASSERT((*which & (1 << pri)) != 0,
1138 ("remrunqueue: remove from empty queue"));
1139 *which &= ~(1 << pri);
1144 * bsd4_setrunqueue_locked()
1146 * Add a process whos rqtype and rqindex had previously been calculated
1147 * onto the appropriate run queue. Determine if the addition requires
1148 * a reschedule on a cpu and return the cpuid or -1.
1150 * NOTE: Lower priorities are better priorities.
1152 * MPSAFE - bsd4_spin must be held exclusively on call
1155 bsd4_setrunqueue_locked(struct lwp *lp)
1161 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
1162 lp->lwp_flag |= LWP_ONRUNQ;
1165 pri = lp->lwp_rqindex;
1167 switch(lp->lwp_rqtype) {
1168 case RTP_PRIO_NORMAL:
1169 q = &bsd4_queues[pri];
1170 which = &bsd4_queuebits;
1172 case RTP_PRIO_REALTIME:
1174 q = &bsd4_rtqueues[pri];
1175 which = &bsd4_rtqueuebits;
1178 q = &bsd4_idqueues[pri];
1179 which = &bsd4_idqueuebits;
1182 panic("remrunqueue: invalid rtprio type");
1187 * Add to the correct queue and set the appropriate bit. If no
1188 * lower priority (i.e. better) processes are in the queue then
1189 * we want a reschedule, calculate the best cpu for the job.
1191 * Always run reschedules on the LWPs original cpu.
1193 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1200 * For SMP systems a user scheduler helper thread is created for each
1201 * cpu and is used to allow one cpu to wakeup another for the purposes of
1202 * scheduling userland threads from setrunqueue().
1204 * UP systems do not need the helper since there is only one cpu.
1206 * We can't use the idle thread for this because we might block.
1207 * Additionally, doing things this way allows us to HLT idle cpus
1213 sched_thread(void *dummy)
1226 cpuid = gd->gd_cpuid; /* doesn't change */
1227 mask = gd->gd_cpumask; /* doesn't change */
1228 dd = &bsd4_pcpu[cpuid];
1231 * Since we are woken up only when no user processes are scheduled
1232 * on a cpu, we can run at an ultra low priority.
1234 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1238 * We use the LWKT deschedule-interlock trick to avoid racing
1239 * bsd4_rdyprocmask. This means we cannot block through to the
1240 * manual lwkt_switch() call we make below.
1243 lwkt_deschedule_self(gd->gd_curthread);
1244 spin_lock(&bsd4_spin);
1245 atomic_set_cpumask(&bsd4_rdyprocmask, mask);
1247 clear_user_resched(); /* This satisfied the reschedule request */
1248 dd->rrcount = 0; /* Reset the round-robin counter */
1250 if ((bsd4_curprocmask & mask) == 0) {
1252 * No thread is currently scheduled.
1254 KKASSERT(dd->uschedcp == NULL);
1255 if ((nlp = chooseproc_locked(NULL)) != NULL) {
1256 atomic_set_cpumask(&bsd4_curprocmask, mask);
1257 dd->upri = nlp->lwp_priority;
1259 spin_unlock(&bsd4_spin);
1261 lwkt_acquire(nlp->lwp_thread);
1263 lwkt_schedule(nlp->lwp_thread);
1265 spin_unlock(&bsd4_spin);
1267 } else if (bsd4_runqcount) {
1268 if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) {
1269 dd->upri = nlp->lwp_priority;
1271 spin_unlock(&bsd4_spin);
1273 lwkt_acquire(nlp->lwp_thread);
1275 lwkt_schedule(nlp->lwp_thread);
1278 * CHAINING CONDITION TRAIN
1280 * We could not deal with the scheduler wakeup
1281 * request on this cpu, locate a ready scheduler
1282 * with no current lp assignment and chain to it.
1284 * This ensures that a wakeup race which fails due
1285 * to priority test does not leave other unscheduled
1286 * cpus idle when the runqueue is not empty.
1288 tmpmask = ~bsd4_curprocmask & bsd4_rdyprocmask &
1291 tmpid = BSFCPUMASK(tmpmask);
1292 gd = globaldata_find(cpuid);
1293 dd = &bsd4_pcpu[cpuid];
1294 atomic_clear_cpumask(&bsd4_rdyprocmask,
1296 spin_unlock(&bsd4_spin);
1297 lwkt_schedule(&dd->helper_thread);
1299 spin_unlock(&bsd4_spin);
1304 * The runq is empty.
1306 spin_unlock(&bsd4_spin);
1314 * Setup our scheduler helpers. Note that curprocmask bit 0 has already
1315 * been cleared by rqinit() and we should not mess with it further.
1318 sched_thread_cpu_init(void)
1323 kprintf("start scheduler helpers on cpus:");
1325 for (i = 0; i < ncpus; ++i) {
1326 bsd4_pcpu_t dd = &bsd4_pcpu[i];
1327 cpumask_t mask = CPUMASK(i);
1329 if ((mask & smp_active_mask) == 0)
1335 lwkt_create(sched_thread, NULL, NULL, &dd->helper_thread,
1336 TDF_STOPREQ, i, "usched %d", i);
1339 * Allow user scheduling on the target cpu. cpu #0 has already
1340 * been enabled in rqinit().
1343 atomic_clear_cpumask(&bsd4_curprocmask, mask);
1344 atomic_set_cpumask(&bsd4_rdyprocmask, mask);
1345 dd->upri = PRIBASE_NULL;
1350 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
1351 sched_thread_cpu_init, NULL)