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
27 #include <sys/param.h>
28 #include <sys/systm.h>
29 #include <sys/kernel.h>
31 #include <sys/queue.h>
33 #include <sys/rtprio.h>
35 #include <sys/sysctl.h>
36 #include <sys/resourcevar.h>
37 #include <sys/spinlock.h>
38 #include <machine/cpu.h>
39 #include <machine/smp.h>
41 #include <sys/thread2.h>
42 #include <sys/spinlock2.h>
43 #include <sys/mplock2.h>
46 * Priorities. Note that with 32 run queues per scheduler each queue
47 * represents four priority levels.
51 #define PRIMASK (MAXPRI - 1)
52 #define PRIBASE_REALTIME 0
53 #define PRIBASE_NORMAL MAXPRI
54 #define PRIBASE_IDLE (MAXPRI * 2)
55 #define PRIBASE_THREAD (MAXPRI * 3)
56 #define PRIBASE_NULL (MAXPRI * 4)
58 #define NQS 32 /* 32 run queues. */
59 #define PPQ (MAXPRI / NQS) /* priorities per queue */
60 #define PPQMASK (PPQ - 1)
63 * NICEPPQ - number of nice units per priority queue
65 * ESTCPUPPQ - number of estcpu units per priority queue
66 * ESTCPUMAX - number of estcpu units
70 #define ESTCPUMAX (ESTCPUPPQ * NQS)
71 #define BATCHMAX (ESTCPUFREQ * 30)
72 #define PRIO_RANGE (PRIO_MAX - PRIO_MIN + 1)
74 #define ESTCPULIM(v) min((v), ESTCPUMAX)
78 #define lwp_priority lwp_usdata.bsd4.priority
79 #define lwp_rqindex lwp_usdata.bsd4.rqindex
80 #define lwp_estcpu lwp_usdata.bsd4.estcpu
81 #define lwp_batch lwp_usdata.bsd4.batch
82 #define lwp_rqtype lwp_usdata.bsd4.rqtype
84 static void bsd4_acquire_curproc(struct lwp *lp);
85 static void bsd4_release_curproc(struct lwp *lp);
86 static void bsd4_select_curproc(globaldata_t gd);
87 static void bsd4_setrunqueue(struct lwp *lp);
88 static void bsd4_schedulerclock(struct lwp *lp, sysclock_t period,
90 static void bsd4_recalculate_estcpu(struct lwp *lp);
91 static void bsd4_resetpriority(struct lwp *lp);
92 static void bsd4_forking(struct lwp *plp, struct lwp *lp);
93 static void bsd4_exiting(struct lwp *lp, struct proc *);
94 static void bsd4_yield(struct lwp *lp);
97 static void need_user_resched_remote(void *dummy);
99 static struct lwp *chooseproc_locked(struct lwp *chklp);
100 static void bsd4_remrunqueue_locked(struct lwp *lp);
101 static void bsd4_setrunqueue_locked(struct lwp *lp);
103 struct usched usched_bsd4 = {
105 "bsd4", "Original DragonFly Scheduler",
106 NULL, /* default registration */
107 NULL, /* default deregistration */
108 bsd4_acquire_curproc,
109 bsd4_release_curproc,
112 bsd4_recalculate_estcpu,
116 NULL, /* setcpumask not supported */
120 struct usched_bsd4_pcpu {
121 struct thread helper_thread;
124 struct lwp *uschedcp;
127 typedef struct usched_bsd4_pcpu *bsd4_pcpu_t;
130 * We have NQS (32) run queues per scheduling class. For the normal
131 * class, there are 128 priorities scaled onto these 32 queues. New
132 * processes are added to the last entry in each queue, and processes
133 * are selected for running by taking them from the head and maintaining
134 * a simple FIFO arrangement. Realtime and Idle priority processes have
135 * and explicit 0-31 priority which maps directly onto their class queue
136 * index. When a queue has something in it, the corresponding bit is
137 * set in the queuebits variable, allowing a single read to determine
138 * the state of all 32 queues and then a ffs() to find the first busy
141 static struct rq bsd4_queues[NQS];
142 static struct rq bsd4_rtqueues[NQS];
143 static struct rq bsd4_idqueues[NQS];
144 static u_int32_t bsd4_queuebits;
145 static u_int32_t bsd4_rtqueuebits;
146 static u_int32_t bsd4_idqueuebits;
147 static cpumask_t bsd4_curprocmask = -1; /* currently running a user process */
148 static cpumask_t bsd4_rdyprocmask; /* ready to accept a user process */
149 static int bsd4_runqcount;
151 static volatile int bsd4_scancpu;
153 static struct spinlock bsd4_spin;
154 static struct usched_bsd4_pcpu bsd4_pcpu[MAXCPU];
156 SYSCTL_INT(_debug, OID_AUTO, bsd4_runqcount, CTLFLAG_RD, &bsd4_runqcount, 0,
157 "Number of run queues");
159 static int usched_nonoptimal;
160 SYSCTL_INT(_debug, OID_AUTO, usched_nonoptimal, CTLFLAG_RW,
161 &usched_nonoptimal, 0, "acquire_curproc() was not optimal");
162 static int usched_optimal;
163 SYSCTL_INT(_debug, OID_AUTO, usched_optimal, CTLFLAG_RW,
164 &usched_optimal, 0, "acquire_curproc() was optimal");
166 static int usched_debug = -1;
167 SYSCTL_INT(_debug, OID_AUTO, scdebug, CTLFLAG_RW, &usched_debug, 0,
168 "Print debug information for this pid");
170 static int remote_resched_nonaffinity;
171 static int remote_resched_affinity;
172 static int choose_affinity;
173 SYSCTL_INT(_debug, OID_AUTO, remote_resched_nonaffinity, CTLFLAG_RD,
174 &remote_resched_nonaffinity, 0, "Number of remote rescheds");
175 SYSCTL_INT(_debug, OID_AUTO, remote_resched_affinity, CTLFLAG_RD,
176 &remote_resched_affinity, 0, "Number of remote rescheds");
177 SYSCTL_INT(_debug, OID_AUTO, choose_affinity, CTLFLAG_RD,
178 &choose_affinity, 0, "chooseproc() was smart");
181 static int usched_bsd4_rrinterval = (ESTCPUFREQ + 9) / 10;
182 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_rrinterval, CTLFLAG_RW,
183 &usched_bsd4_rrinterval, 0, "");
184 static int usched_bsd4_decay = 8;
185 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_decay, CTLFLAG_RW,
186 &usched_bsd4_decay, 0, "Extra decay when not running");
187 static int usched_bsd4_batch_time = 10;
188 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_batch_time, CTLFLAG_RW,
189 &usched_bsd4_batch_time, 0, "Minimum batch counter value");
192 * Initialize the run queues at boot time.
199 spin_init(&bsd4_spin);
200 for (i = 0; i < NQS; i++) {
201 TAILQ_INIT(&bsd4_queues[i]);
202 TAILQ_INIT(&bsd4_rtqueues[i]);
203 TAILQ_INIT(&bsd4_idqueues[i]);
205 atomic_clear_cpumask(&bsd4_curprocmask, 1);
207 SYSINIT(runqueue, SI_BOOT2_USCHED, SI_ORDER_FIRST, rqinit, NULL)
210 * BSD4_ACQUIRE_CURPROC
212 * This function is called when the kernel intends to return to userland.
213 * It is responsible for making the thread the current designated userland
214 * thread for this cpu, blocking if necessary.
216 * The kernel has already depressed our LWKT priority so we must not switch
217 * until we have either assigned or disposed of the thread.
219 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
220 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will
221 * occur, this function is called only under very controlled circumstances.
226 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
250 dd = &bsd4_pcpu[gd->gd_cpuid];
254 * Process any pending events and higher priority threads.
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 cpu's lwp designation
281 * away simply by replacing it. The other thread
282 * will stall when it tries to return to userland.
285 dd->upri = lp->lwp_priority;
287 lwkt_deschedule(olp->lwp_thread);
288 bsd4_setrunqueue(olp);
292 * We cannot become the current lwp, place the lp
293 * on the bsd4 run-queue and deschedule ourselves.
295 * When we are reactivated we will have another
298 lwkt_deschedule(lp->lwp_thread);
299 bsd4_setrunqueue(lp);
302 * Reload after a switch or setrunqueue/switch possibly
303 * moved us to another cpu.
306 dd = &bsd4_pcpu[gd->gd_cpuid];
308 } while (dd->uschedcp != lp);
311 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
315 * BSD4_RELEASE_CURPROC
317 * This routine detaches the current thread from the userland scheduler,
318 * usually because the thread needs to run or block in the kernel (at
319 * kernel priority) for a while.
321 * This routine is also responsible for selecting a new thread to
322 * make the current thread.
324 * NOTE: This implementation differs from the dummy example in that
325 * bsd4_select_curproc() is able to select the current process, whereas
326 * dummy_select_curproc() is not able to select the current process.
327 * This means we have to NULL out uschedcp.
329 * Additionally, note that we may already be on a run queue if releasing
330 * via the lwkt_switch() in bsd4_setrunqueue().
335 bsd4_release_curproc(struct lwp *lp)
337 globaldata_t gd = mycpu;
338 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
340 if (dd->uschedcp == lp) {
342 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
343 dd->uschedcp = NULL; /* don't let lp be selected */
344 dd->upri = PRIBASE_NULL;
345 atomic_clear_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
346 bsd4_select_curproc(gd);
352 * BSD4_SELECT_CURPROC
354 * Select a new current process for this cpu and clear any pending user
355 * reschedule request. The cpu currently has no current process.
357 * This routine is also responsible for equal-priority round-robining,
358 * typically triggered from bsd4_schedulerclock(). In our dummy example
359 * all the 'user' threads are LWKT scheduled all at once and we just
360 * call lwkt_switch().
362 * The calling process is not on the queue and cannot be selected.
368 bsd4_select_curproc(globaldata_t gd)
370 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
372 int cpuid = gd->gd_cpuid;
376 spin_lock(&bsd4_spin);
377 if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) {
378 atomic_set_cpumask(&bsd4_curprocmask, CPUMASK(cpuid));
379 dd->upri = nlp->lwp_priority;
381 spin_unlock(&bsd4_spin);
383 lwkt_acquire(nlp->lwp_thread);
385 lwkt_schedule(nlp->lwp_thread);
387 spin_unlock(&bsd4_spin);
390 } else if (bsd4_runqcount && (bsd4_rdyprocmask & CPUMASK(cpuid))) {
391 atomic_clear_cpumask(&bsd4_rdyprocmask, CPUMASK(cpuid));
392 spin_unlock(&bsd4_spin);
393 lwkt_schedule(&dd->helper_thread);
395 spin_unlock(&bsd4_spin);
404 * Place the specified lwp on the user scheduler's run queue. This routine
405 * must be called with the thread descheduled. The lwp must be runnable.
407 * The thread may be the current thread as a special case.
412 bsd4_setrunqueue(struct lwp *lp)
423 * First validate the process state relative to the current cpu.
424 * We don't need the spinlock for this, just a critical section.
425 * We are in control of the process.
428 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
429 KASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0,
430 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
431 lp->lwp_tid, lp->lwp_proc->p_flag, lp->lwp_flag));
432 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
435 * Note: gd and dd are relative to the target thread's last cpu,
436 * NOT our current cpu.
438 gd = lp->lwp_thread->td_gd;
439 dd = &bsd4_pcpu[gd->gd_cpuid];
442 * This process is not supposed to be scheduled anywhere or assigned
443 * as the current process anywhere. Assert the condition.
445 KKASSERT(dd->uschedcp != lp);
449 * If we are not SMP we do not have a scheduler helper to kick
450 * and must directly activate the process if none are scheduled.
452 * This is really only an issue when bootstrapping init since
453 * the caller in all other cases will be a user process, and
454 * even if released (dd->uschedcp == NULL), that process will
455 * kickstart the scheduler when it returns to user mode from
458 if (dd->uschedcp == NULL) {
459 atomic_set_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
461 dd->upri = lp->lwp_priority;
462 lwkt_schedule(lp->lwp_thread);
470 * XXX fixme. Could be part of a remrunqueue/setrunqueue
471 * operation when the priority is recalculated, so TDF_MIGRATING
472 * may already be set.
474 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
475 lwkt_giveaway(lp->lwp_thread);
479 * We lose control of lp the moment we release the spinlock after
480 * having placed lp on the queue. i.e. another cpu could pick it
481 * up and it could exit, or its priority could be further adjusted,
482 * or something like that.
484 spin_lock(&bsd4_spin);
485 bsd4_setrunqueue_locked(lp);
489 * Kick the scheduler helper on one of the other cpu's
490 * and request a reschedule if appropriate.
492 * NOTE: We check all cpus whos rdyprocmask is set. First we
493 * look for cpus without designated lps, then we look for
494 * cpus with designated lps with a worse priority than our
498 cpuid = (bsd4_scancpu & 0xFFFF) % ncpus;
499 mask = ~bsd4_curprocmask & bsd4_rdyprocmask & lp->lwp_cpumask &
500 smp_active_mask & usched_global_cpumask;
503 tmpmask = ~(CPUMASK(cpuid) - 1);
505 cpuid = BSFCPUMASK(mask & tmpmask);
507 cpuid = BSFCPUMASK(mask);
508 gd = globaldata_find(cpuid);
509 dd = &bsd4_pcpu[cpuid];
511 if ((dd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
513 mask &= ~CPUMASK(cpuid);
517 * Then cpus which might have a currently running lp
519 mask = bsd4_curprocmask & bsd4_rdyprocmask &
520 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
523 tmpmask = ~(CPUMASK(cpuid) - 1);
525 cpuid = BSFCPUMASK(mask & tmpmask);
527 cpuid = BSFCPUMASK(mask);
528 gd = globaldata_find(cpuid);
529 dd = &bsd4_pcpu[cpuid];
531 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
533 mask &= ~CPUMASK(cpuid);
537 * If we cannot find a suitable cpu we reload from bsd4_scancpu
538 * and round-robin. Other cpus will pickup as they release their
539 * current lwps or become ready.
541 * Avoid a degenerate system lockup case if usched_global_cpumask
542 * is set to 0 or otherwise does not cover lwp_cpumask.
544 * We only kick the target helper thread in this case, we do not
545 * set the user resched flag because
547 cpuid = (bsd4_scancpu & 0xFFFF) % ncpus;
548 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0) {
551 gd = globaldata_find(cpuid);
552 dd = &bsd4_pcpu[cpuid];
555 spin_unlock(&bsd4_spin);
556 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
557 if (dd->uschedcp == NULL) {
558 lwkt_schedule(&dd->helper_thread);
564 atomic_clear_cpumask(&bsd4_rdyprocmask, CPUMASK(cpuid));
565 spin_unlock(&bsd4_spin);
566 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
567 lwkt_send_ipiq(gd, need_user_resched_remote, NULL);
569 lwkt_schedule(&dd->helper_thread);
573 * Request a reschedule if appropriate.
575 spin_unlock(&bsd4_spin);
576 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
584 * This routine is called from a systimer IPI. It MUST be MP-safe and
585 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
592 bsd4_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
594 globaldata_t gd = mycpu;
595 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
598 * Do we need to round-robin? We round-robin 10 times a second.
599 * This should only occur for cpu-bound batch processes.
601 if (++dd->rrcount >= usched_bsd4_rrinterval) {
607 * Adjust estcpu upward using a real time equivalent calculation.
609 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
612 * Spinlocks also hold a critical section so there should not be
615 KKASSERT(gd->gd_spinlocks_wr == 0);
617 bsd4_resetpriority(lp);
620 * if we can't call bsd4_resetpriority for some reason we must call
621 * need user_resched().
628 * Called from acquire and from kern_synch's one-second timer (one of the
629 * callout helper threads) with a critical section held.
631 * Decay p_estcpu based on the number of ticks we haven't been running
632 * and our p_nice. As the load increases each process observes a larger
633 * number of idle ticks (because other processes are running in them).
634 * This observation leads to a larger correction which tends to make the
635 * system more 'batchy'.
637 * Note that no recalculation occurs for a process which sleeps and wakes
638 * up in the same tick. That is, a system doing thousands of context
639 * switches per second will still only do serious estcpu calculations
640 * ESTCPUFREQ times per second.
646 bsd4_recalculate_estcpu(struct lwp *lp)
648 globaldata_t gd = mycpu;
655 * We have to subtract periodic to get the last schedclock
656 * timeout time, otherwise we would get the upcoming timeout.
657 * Keep in mind that a process can migrate between cpus and
658 * while the scheduler clock should be very close, boundary
659 * conditions could lead to a small negative delta.
661 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
663 if (lp->lwp_slptime > 1) {
665 * Too much time has passed, do a coarse correction.
667 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
668 bsd4_resetpriority(lp);
669 lp->lwp_cpbase = cpbase;
671 lp->lwp_batch -= ESTCPUFREQ;
672 if (lp->lwp_batch < 0)
674 } else if (lp->lwp_cpbase != cpbase) {
676 * Adjust estcpu if we are in a different tick. Don't waste
677 * time if we are in the same tick.
679 * First calculate the number of ticks in the measurement
680 * interval. The ttlticks calculation can wind up 0 due to
681 * a bug in the handling of lwp_slptime (as yet not found),
682 * so make sure we do not get a divide by 0 panic.
684 ttlticks = (cpbase - lp->lwp_cpbase) /
685 gd->gd_schedclock.periodic;
688 lp->lwp_cpbase = cpbase;
692 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
695 * Calculate the percentage of one cpu used factoring in ncpus
696 * and the load and adjust estcpu. Handle degenerate cases
697 * by adding 1 to bsd4_runqcount.
699 * estcpu is scaled by ESTCPUMAX.
701 * bsd4_runqcount is the excess number of user processes
702 * that cannot be immediately scheduled to cpus. We want
703 * to count these as running to avoid range compression
704 * in the base calculation (which is the actual percentage
707 estcpu = (lp->lwp_cpticks * ESTCPUMAX) *
708 (bsd4_runqcount + ncpus) / (ncpus * ttlticks);
711 * If estcpu is > 50% we become more batch-like
712 * If estcpu is <= 50% we become less batch-like
714 * It takes 30 cpu seconds to traverse the entire range.
716 if (estcpu > ESTCPUMAX / 2) {
717 lp->lwp_batch += ttlticks;
718 if (lp->lwp_batch > BATCHMAX)
719 lp->lwp_batch = BATCHMAX;
721 lp->lwp_batch -= ttlticks;
722 if (lp->lwp_batch < 0)
726 if (usched_debug == lp->lwp_proc->p_pid) {
727 kprintf("pid %d lwp %p estcpu %3d %3d bat %d cp %d/%d",
728 lp->lwp_proc->p_pid, lp,
729 estcpu, lp->lwp_estcpu,
731 lp->lwp_cpticks, ttlticks);
735 * Adjust lp->lwp_esetcpu. The decay factor determines how
736 * quickly lwp_estcpu collapses to its realtime calculation.
737 * A slower collapse gives us a more accurate number but
738 * can cause a cpu hog to eat too much cpu before the
739 * scheduler decides to downgrade it.
741 * NOTE: p_nice is accounted for in bsd4_resetpriority(),
742 * and not here, but we must still ensure that a
743 * cpu-bound nice -20 process does not completely
744 * override a cpu-bound nice +20 process.
746 * NOTE: We must use ESTCPULIM() here to deal with any
749 decay_factor = usched_bsd4_decay;
750 if (decay_factor < 1)
752 if (decay_factor > 1024)
755 lp->lwp_estcpu = ESTCPULIM(
756 (lp->lwp_estcpu * decay_factor + estcpu) /
759 if (usched_debug == lp->lwp_proc->p_pid)
760 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
761 bsd4_resetpriority(lp);
762 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
768 * Compute the priority of a process when running in user mode.
769 * Arrange to reschedule if the resulting priority is better
770 * than that of the current process.
772 * This routine may be called with any process.
774 * This routine is called by fork1() for initial setup with the process
775 * of the run queue, and also may be called normally with the process on or
781 bsd4_resetpriority(struct lwp *lp)
791 * Calculate the new priority and queue type
794 spin_lock(&bsd4_spin);
796 newrqtype = lp->lwp_rtprio.type;
799 case RTP_PRIO_REALTIME:
801 newpriority = PRIBASE_REALTIME +
802 (lp->lwp_rtprio.prio & PRIMASK);
804 case RTP_PRIO_NORMAL:
806 * Detune estcpu based on batchiness. lwp_batch ranges
807 * from 0 to BATCHMAX. Limit estcpu for the sake of
808 * the priority calculation to between 50% and 100%.
810 estcpu = lp->lwp_estcpu * (lp->lwp_batch + BATCHMAX) /
814 * p_nice piece Adds (0-40) * 2 0-80
815 * estcpu Adds 16384 * 4 / 512 0-128
817 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
818 newpriority += estcpu * PPQ / ESTCPUPPQ;
819 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
820 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
821 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
824 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
826 case RTP_PRIO_THREAD:
827 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
830 panic("Bad RTP_PRIO %d", newrqtype);
835 * The newpriority incorporates the queue type so do a simple masked
836 * check to determine if the process has moved to another queue. If
837 * it has, and it is currently on a run queue, then move it.
839 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
840 lp->lwp_priority = newpriority;
841 if (lp->lwp_flag & LWP_ONRUNQ) {
842 bsd4_remrunqueue_locked(lp);
843 lp->lwp_rqtype = newrqtype;
844 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
845 bsd4_setrunqueue_locked(lp);
848 lp->lwp_rqtype = newrqtype;
849 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
852 reschedcpu = lp->lwp_thread->td_gd->gd_cpuid;
854 lp->lwp_priority = newpriority;
860 * Determine if we need to reschedule the target cpu. This only
861 * occurs if the LWP is already on a scheduler queue, which means
862 * that idle cpu notification has already occured. At most we
863 * need only issue a need_user_resched() on the appropriate cpu.
865 * The LWP may be owned by a CPU different from the current one,
866 * in which case dd->uschedcp may be modified without an MP lock
867 * or a spinlock held. The worst that happens is that the code
868 * below causes a spurious need_user_resched() on the target CPU
869 * and dd->pri to be wrong for a short period of time, both of
870 * which are harmless.
872 * If checkpri is 0 we are adjusting the priority of the current
873 * process, possibly higher (less desireable), so ignore the upri
874 * check which will fail in that case.
876 if (reschedcpu >= 0) {
877 dd = &bsd4_pcpu[reschedcpu];
878 if ((bsd4_rdyprocmask & CPUMASK(reschedcpu)) &&
880 (dd->upri & ~PRIMASK) > (lp->lwp_priority & ~PRIMASK))) {
882 if (reschedcpu == mycpu->gd_cpuid) {
883 spin_unlock(&bsd4_spin);
886 spin_unlock(&bsd4_spin);
887 atomic_clear_cpumask(&bsd4_rdyprocmask,
888 CPUMASK(reschedcpu));
889 lwkt_send_ipiq(lp->lwp_thread->td_gd,
890 need_user_resched_remote, NULL);
893 spin_unlock(&bsd4_spin);
897 spin_unlock(&bsd4_spin);
900 spin_unlock(&bsd4_spin);
910 bsd4_yield(struct lwp *lp)
913 /* FUTURE (or something similar) */
914 switch(lp->lwp_rqtype) {
915 case RTP_PRIO_NORMAL:
916 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
926 * Called from fork1() when a new child process is being created.
928 * Give the child process an initial estcpu that is more batch then
929 * its parent and dock the parent for the fork (but do not
930 * reschedule the parent). This comprises the main part of our batch
931 * detection heuristic for both parallel forking and sequential execs.
933 * XXX lwp should be "spawning" instead of "forking"
938 bsd4_forking(struct lwp *plp, struct lwp *lp)
941 * Put the child 4 queue slots (out of 32) higher than the parent
942 * (less desireable than the parent).
944 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
947 * The batch status of children always starts out centerline
948 * and will inch-up or inch-down as appropriate. It takes roughly
949 * ~15 seconds of >50% cpu to hit the limit.
951 lp->lwp_batch = BATCHMAX / 2;
954 * Dock the parent a cost for the fork, protecting us from fork
955 * bombs. If the parent is forking quickly make the child more
958 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
962 * Called when a parent waits for a child.
967 bsd4_exiting(struct lwp *lp, struct proc *child_proc)
972 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
973 * it selects a user process and returns it. If chklp is non-NULL and chklp
974 * has a better or equal priority then the process that would otherwise be
975 * chosen, NULL is returned.
977 * Until we fix the RUNQ code the chklp test has to be strict or we may
978 * bounce between processes trying to acquire the current process designation.
980 * MPSAFE - must be called with bsd4_spin exclusive held. The spinlock is
981 * left intact through the entire routine.
985 chooseproc_locked(struct lwp *chklp)
989 u_int32_t *which, *which2;
996 rtqbits = bsd4_rtqueuebits;
997 tsqbits = bsd4_queuebits;
998 idqbits = bsd4_idqueuebits;
999 cpumask = mycpu->gd_cpumask;
1005 pri = bsfl(rtqbits);
1006 q = &bsd4_rtqueues[pri];
1007 which = &bsd4_rtqueuebits;
1009 } else if (tsqbits) {
1010 pri = bsfl(tsqbits);
1011 q = &bsd4_queues[pri];
1012 which = &bsd4_queuebits;
1014 } else if (idqbits) {
1015 pri = bsfl(idqbits);
1016 q = &bsd4_idqueues[pri];
1017 which = &bsd4_idqueuebits;
1022 lp = TAILQ_FIRST(q);
1023 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1026 while ((lp->lwp_cpumask & cpumask) == 0) {
1027 lp = TAILQ_NEXT(lp, lwp_procq);
1029 *which2 &= ~(1 << pri);
1036 * If the passed lwp <chklp> is reasonably close to the selected
1037 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1039 * Note that we must error on the side of <chklp> to avoid bouncing
1040 * between threads in the acquire code.
1043 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1049 * If the chosen lwp does not reside on this cpu spend a few
1050 * cycles looking for a better candidate at the same priority level.
1051 * This is a fallback check, setrunqueue() tries to wakeup the
1052 * correct cpu and is our front-line affinity.
1054 if (lp->lwp_thread->td_gd != mycpu &&
1055 (chklp = TAILQ_NEXT(lp, lwp_procq)) != NULL
1057 if (chklp->lwp_thread->td_gd == mycpu) {
1064 TAILQ_REMOVE(q, lp, lwp_procq);
1067 *which &= ~(1 << pri);
1068 KASSERT((lp->lwp_flag & LWP_ONRUNQ) != 0, ("not on runq6!"));
1069 lp->lwp_flag &= ~LWP_ONRUNQ;
1077 need_user_resched_remote(void *dummy)
1079 globaldata_t gd = mycpu;
1080 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
1082 need_user_resched();
1083 lwkt_schedule(&dd->helper_thread);
1089 * bsd4_remrunqueue_locked() removes a given process from the run queue
1090 * that it is on, clearing the queue busy bit if it becomes empty.
1092 * Note that user process scheduler is different from the LWKT schedule.
1093 * The user process scheduler only manages user processes but it uses LWKT
1094 * underneath, and a user process operating in the kernel will often be
1095 * 'released' from our management.
1097 * MPSAFE - bsd4_spin must be held exclusively on call
1100 bsd4_remrunqueue_locked(struct lwp *lp)
1106 KKASSERT(lp->lwp_flag & LWP_ONRUNQ);
1107 lp->lwp_flag &= ~LWP_ONRUNQ;
1109 KKASSERT(bsd4_runqcount >= 0);
1111 pri = lp->lwp_rqindex;
1112 switch(lp->lwp_rqtype) {
1113 case RTP_PRIO_NORMAL:
1114 q = &bsd4_queues[pri];
1115 which = &bsd4_queuebits;
1117 case RTP_PRIO_REALTIME:
1119 q = &bsd4_rtqueues[pri];
1120 which = &bsd4_rtqueuebits;
1123 q = &bsd4_idqueues[pri];
1124 which = &bsd4_idqueuebits;
1127 panic("remrunqueue: invalid rtprio type");
1130 TAILQ_REMOVE(q, lp, lwp_procq);
1131 if (TAILQ_EMPTY(q)) {
1132 KASSERT((*which & (1 << pri)) != 0,
1133 ("remrunqueue: remove from empty queue"));
1134 *which &= ~(1 << pri);
1139 * bsd4_setrunqueue_locked()
1141 * Add a process whos rqtype and rqindex had previously been calculated
1142 * onto the appropriate run queue. Determine if the addition requires
1143 * a reschedule on a cpu and return the cpuid or -1.
1145 * NOTE: Lower priorities are better priorities.
1147 * MPSAFE - bsd4_spin must be held exclusively on call
1150 bsd4_setrunqueue_locked(struct lwp *lp)
1156 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
1157 lp->lwp_flag |= LWP_ONRUNQ;
1160 pri = lp->lwp_rqindex;
1162 switch(lp->lwp_rqtype) {
1163 case RTP_PRIO_NORMAL:
1164 q = &bsd4_queues[pri];
1165 which = &bsd4_queuebits;
1167 case RTP_PRIO_REALTIME:
1169 q = &bsd4_rtqueues[pri];
1170 which = &bsd4_rtqueuebits;
1173 q = &bsd4_idqueues[pri];
1174 which = &bsd4_idqueuebits;
1177 panic("remrunqueue: invalid rtprio type");
1182 * Add to the correct queue and set the appropriate bit. If no
1183 * lower priority (i.e. better) processes are in the queue then
1184 * we want a reschedule, calculate the best cpu for the job.
1186 * Always run reschedules on the LWPs original cpu.
1188 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1195 * For SMP systems a user scheduler helper thread is created for each
1196 * cpu and is used to allow one cpu to wakeup another for the purposes of
1197 * scheduling userland threads from setrunqueue().
1199 * UP systems do not need the helper since there is only one cpu.
1201 * We can't use the idle thread for this because we might block.
1202 * Additionally, doing things this way allows us to HLT idle cpus
1208 sched_thread(void *dummy)
1222 cpuid = gd->gd_cpuid; /* doesn't change */
1223 mask = gd->gd_cpumask; /* doesn't change */
1224 dd = &bsd4_pcpu[cpuid];
1227 * Since we are woken up only when no user processes are scheduled
1228 * on a cpu, we can run at an ultra low priority.
1230 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1234 * We use the LWKT deschedule-interlock trick to avoid racing
1235 * bsd4_rdyprocmask. This means we cannot block through to the
1236 * manual lwkt_switch() call we make below.
1239 lwkt_deschedule_self(gd->gd_curthread);
1240 spin_lock(&bsd4_spin);
1241 atomic_set_cpumask(&bsd4_rdyprocmask, mask);
1243 clear_user_resched(); /* This satisfied the reschedule request */
1244 dd->rrcount = 0; /* Reset the round-robin counter */
1246 if ((bsd4_curprocmask & mask) == 0) {
1248 * No thread is currently scheduled.
1250 KKASSERT(dd->uschedcp == NULL);
1251 if ((nlp = chooseproc_locked(NULL)) != NULL) {
1252 atomic_set_cpumask(&bsd4_curprocmask, mask);
1253 dd->upri = nlp->lwp_priority;
1255 spin_unlock(&bsd4_spin);
1257 lwkt_acquire(nlp->lwp_thread);
1259 lwkt_schedule(nlp->lwp_thread);
1261 spin_unlock(&bsd4_spin);
1263 } else if (bsd4_runqcount) {
1264 if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) {
1265 dd->upri = nlp->lwp_priority;
1267 spin_unlock(&bsd4_spin);
1269 lwkt_acquire(nlp->lwp_thread);
1271 lwkt_schedule(nlp->lwp_thread);
1274 * CHAINING CONDITION TRAIN
1276 * We could not deal with the scheduler wakeup
1277 * request on this cpu, locate a ready scheduler
1278 * with no current lp assignment and chain to it.
1280 * This ensures that a wakeup race which fails due
1281 * to priority test does not leave other unscheduled
1282 * cpus idle when the runqueue is not empty.
1284 tmpmask = ~bsd4_curprocmask & bsd4_rdyprocmask &
1287 tmpid = BSFCPUMASK(tmpmask);
1288 tmpdd = &bsd4_pcpu[tmpid];
1289 atomic_clear_cpumask(&bsd4_rdyprocmask,
1291 spin_unlock(&bsd4_spin);
1292 lwkt_schedule(&tmpdd->helper_thread);
1294 spin_unlock(&bsd4_spin);
1299 * The runq is empty.
1301 spin_unlock(&bsd4_spin);
1305 * We're descheduled unless someone scheduled us. Switch away.
1306 * Exiting the critical section will cause splz() to be called
1307 * for us if interrupts and such are pending.
1315 * Setup our scheduler helpers. Note that curprocmask bit 0 has already
1316 * been cleared by rqinit() and we should not mess with it further.
1319 sched_thread_cpu_init(void)
1324 kprintf("start scheduler helpers on cpus:");
1326 for (i = 0; i < ncpus; ++i) {
1327 bsd4_pcpu_t dd = &bsd4_pcpu[i];
1328 cpumask_t mask = CPUMASK(i);
1330 if ((mask & smp_active_mask) == 0)
1336 lwkt_create(sched_thread, NULL, NULL, &dd->helper_thread,
1337 TDF_STOPREQ, i, "usched %d", i);
1340 * Allow user scheduling on the target cpu. cpu #0 has already
1341 * been enabled in rqinit().
1344 atomic_clear_cpumask(&bsd4_curprocmask, mask);
1345 atomic_set_cpumask(&bsd4_rdyprocmask, mask);
1346 dd->upri = PRIBASE_NULL;
1351 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
1352 sched_thread_cpu_init, NULL)