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];
260 * Become the currently scheduled user thread for this cpu
261 * if we can do so trivially.
263 * We can steal another thread's current thread designation
264 * on this cpu since if we are running that other thread
265 * must not be, so we can safely deschedule it.
267 if (dd->uschedcp == lp) {
269 * We are already the current lwp (hot path).
271 dd->upri = lp->lwp_priority;
272 } else if (dd->uschedcp == NULL) {
274 * We can trivially become the current lwp.
276 atomic_set_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
278 dd->upri = lp->lwp_priority;
279 } else if (dd->upri > lp->lwp_priority) {
281 * We can steal the current lwp designation from the
282 * olp that was previously assigned to this cpu.
286 dd->upri = lp->lwp_priority;
287 lwkt_deschedule(olp->lwp_thread);
288 bsd4_setrunqueue(olp);
291 * We cannot become the current lwp, place the lp
292 * on the bsd4 run-queue and deschedule ourselves.
294 lwkt_deschedule(lp->lwp_thread);
295 bsd4_setrunqueue(lp);
300 * Because we are in a critical section interrupts may wind
301 * up pending and prevent an interrupt thread from being
302 * scheduled, we have to run splz() unconditionally to
303 * ensure that these folks are properly scheduled so we can
304 * then test the LWKT thread reschedule flag.
306 * Other threads at our current user priority have already
307 * put in their bids, but we must run any kernel threads
308 * at higher priorities, and we could lose our bid to
309 * another thread trying to return to user mode in the
312 * If we lose our bid we will be descheduled and put on
313 * the run queue. When we are reactivated we will have
317 if (lwkt_resched_wanted())
319 } while (dd->uschedcp != lp);
322 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
326 * BSD4_RELEASE_CURPROC
328 * This routine detaches the current thread from the userland scheduler,
329 * usually because the thread needs to run or block in the kernel (at
330 * kernel priority) for a while.
332 * This routine is also responsible for selecting a new thread to
333 * make the current thread.
335 * NOTE: This implementation differs from the dummy example in that
336 * bsd4_select_curproc() is able to select the current process, whereas
337 * dummy_select_curproc() is not able to select the current process.
338 * This means we have to NULL out uschedcp.
340 * Additionally, note that we may already be on a run queue if releasing
341 * via the lwkt_switch() in bsd4_setrunqueue().
346 bsd4_release_curproc(struct lwp *lp)
348 globaldata_t gd = mycpu;
349 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
351 if (dd->uschedcp == lp) {
353 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
354 dd->uschedcp = NULL; /* don't let lp be selected */
355 dd->upri = PRIBASE_NULL;
356 atomic_clear_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
357 bsd4_select_curproc(gd);
363 * BSD4_SELECT_CURPROC
365 * Select a new current process for this cpu and clear any pending user
366 * reschedule request. The cpu currently has no current process.
368 * This routine is also responsible for equal-priority round-robining,
369 * typically triggered from bsd4_schedulerclock(). In our dummy example
370 * all the 'user' threads are LWKT scheduled all at once and we just
371 * call lwkt_switch().
373 * The calling process is not on the queue and cannot be selected.
379 bsd4_select_curproc(globaldata_t gd)
381 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
383 int cpuid = gd->gd_cpuid;
387 spin_lock(&bsd4_spin);
388 if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) {
389 atomic_set_cpumask(&bsd4_curprocmask, CPUMASK(cpuid));
390 dd->upri = nlp->lwp_priority;
392 spin_unlock(&bsd4_spin);
394 lwkt_acquire(nlp->lwp_thread);
396 lwkt_schedule(nlp->lwp_thread);
398 spin_unlock(&bsd4_spin);
401 } else if (bsd4_runqcount && (bsd4_rdyprocmask & CPUMASK(cpuid))) {
402 atomic_clear_cpumask(&bsd4_rdyprocmask, CPUMASK(cpuid));
403 spin_unlock(&bsd4_spin);
404 lwkt_schedule(&dd->helper_thread);
406 spin_unlock(&bsd4_spin);
415 * Place the specified lwp on the user scheduler's run queue. This routine
416 * must be called with the thread descheduled. The lwp must be runnable.
418 * The thread may be the current thread as a special case.
423 bsd4_setrunqueue(struct lwp *lp)
434 * First validate the process state relative to the current cpu.
435 * We don't need the spinlock for this, just a critical section.
436 * We are in control of the process.
439 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
440 KASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0,
441 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
442 lp->lwp_tid, lp->lwp_proc->p_flag, lp->lwp_flag));
443 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
446 * Note: gd and dd are relative to the target thread's last cpu,
447 * NOT our current cpu.
449 gd = lp->lwp_thread->td_gd;
450 dd = &bsd4_pcpu[gd->gd_cpuid];
453 * This process is not supposed to be scheduled anywhere or assigned
454 * as the current process anywhere. Assert the condition.
456 KKASSERT(dd->uschedcp != lp);
460 * If we are not SMP we do not have a scheduler helper to kick
461 * and must directly activate the process if none are scheduled.
463 * This is really only an issue when bootstrapping init since
464 * the caller in all other cases will be a user process, and
465 * even if released (dd->uschedcp == NULL), that process will
466 * kickstart the scheduler when it returns to user mode from
469 if (dd->uschedcp == NULL) {
470 atomic_set_cpumask(&bsd4_curprocmask, gd->gd_cpumask);
472 dd->upri = lp->lwp_priority;
473 lwkt_schedule(lp->lwp_thread);
481 * XXX fixme. Could be part of a remrunqueue/setrunqueue
482 * operation when the priority is recalculated, so TDF_MIGRATING
483 * may already be set.
485 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
486 lwkt_giveaway(lp->lwp_thread);
490 * We lose control of lp the moment we release the spinlock after
491 * having placed lp on the queue. i.e. another cpu could pick it
492 * up and it could exit, or its priority could be further adjusted,
493 * or something like that.
495 spin_lock(&bsd4_spin);
496 bsd4_setrunqueue_locked(lp);
500 * Kick the scheduler helper on one of the other cpu's
501 * and request a reschedule if appropriate.
503 * NOTE: We check all cpus whos rdyprocmask is set. First we
504 * look for cpus without designated lps, then we look for
505 * cpus with designated lps with a worse priority than our
509 cpuid = (bsd4_scancpu & 0xFFFF) % ncpus;
510 mask = ~bsd4_curprocmask & bsd4_rdyprocmask & lp->lwp_cpumask &
511 smp_active_mask & usched_global_cpumask;
514 tmpmask = ~(CPUMASK(cpuid) - 1);
516 cpuid = BSFCPUMASK(mask & tmpmask);
518 cpuid = BSFCPUMASK(mask);
519 gd = globaldata_find(cpuid);
520 dd = &bsd4_pcpu[cpuid];
522 if ((dd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
524 mask &= ~CPUMASK(cpuid);
528 * Then cpus which might have a currently running lp
530 mask = bsd4_curprocmask & bsd4_rdyprocmask &
531 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
534 tmpmask = ~(CPUMASK(cpuid) - 1);
536 cpuid = BSFCPUMASK(mask & tmpmask);
538 cpuid = BSFCPUMASK(mask);
539 gd = globaldata_find(cpuid);
540 dd = &bsd4_pcpu[cpuid];
542 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
544 mask &= ~CPUMASK(cpuid);
548 * If we cannot find a suitable cpu we reload from bsd4_scancpu
549 * and round-robin. Other cpus will pickup as they release their
550 * current lwps or become ready.
552 * Avoid a degenerate system lockup case if usched_global_cpumask
553 * is set to 0 or otherwise does not cover lwp_cpumask.
555 * We only kick the target helper thread in this case, we do not
556 * set the user resched flag because
558 cpuid = (bsd4_scancpu & 0xFFFF) % ncpus;
559 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0) {
562 gd = globaldata_find(cpuid);
563 dd = &bsd4_pcpu[cpuid];
566 spin_unlock(&bsd4_spin);
567 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
568 if (dd->uschedcp == NULL) {
569 lwkt_schedule(&dd->helper_thread);
575 atomic_clear_cpumask(&bsd4_rdyprocmask, CPUMASK(cpuid));
576 spin_unlock(&bsd4_spin);
577 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
578 lwkt_send_ipiq(gd, need_user_resched_remote, NULL);
580 lwkt_schedule(&dd->helper_thread);
584 * Request a reschedule if appropriate.
586 spin_unlock(&bsd4_spin);
587 if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
595 * This routine is called from a systimer IPI. It MUST be MP-safe and
596 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
603 bsd4_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
605 globaldata_t gd = mycpu;
606 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
609 * Do we need to round-robin? We round-robin 10 times a second.
610 * This should only occur for cpu-bound batch processes.
612 if (++dd->rrcount >= usched_bsd4_rrinterval) {
618 * Adjust estcpu upward using a real time equivalent calculation.
620 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
623 * Spinlocks also hold a critical section so there should not be
626 KKASSERT(gd->gd_spinlocks_wr == 0);
628 bsd4_resetpriority(lp);
631 * if we can't call bsd4_resetpriority for some reason we must call
632 * need user_resched().
639 * Called from acquire and from kern_synch's one-second timer (one of the
640 * callout helper threads) with a critical section held.
642 * Decay p_estcpu based on the number of ticks we haven't been running
643 * and our p_nice. As the load increases each process observes a larger
644 * number of idle ticks (because other processes are running in them).
645 * This observation leads to a larger correction which tends to make the
646 * system more 'batchy'.
648 * Note that no recalculation occurs for a process which sleeps and wakes
649 * up in the same tick. That is, a system doing thousands of context
650 * switches per second will still only do serious estcpu calculations
651 * ESTCPUFREQ times per second.
657 bsd4_recalculate_estcpu(struct lwp *lp)
659 globaldata_t gd = mycpu;
666 * We have to subtract periodic to get the last schedclock
667 * timeout time, otherwise we would get the upcoming timeout.
668 * Keep in mind that a process can migrate between cpus and
669 * while the scheduler clock should be very close, boundary
670 * conditions could lead to a small negative delta.
672 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
674 if (lp->lwp_slptime > 1) {
676 * Too much time has passed, do a coarse correction.
678 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
679 bsd4_resetpriority(lp);
680 lp->lwp_cpbase = cpbase;
682 lp->lwp_batch -= ESTCPUFREQ;
683 if (lp->lwp_batch < 0)
685 } else if (lp->lwp_cpbase != cpbase) {
687 * Adjust estcpu if we are in a different tick. Don't waste
688 * time if we are in the same tick.
690 * First calculate the number of ticks in the measurement
691 * interval. The ttlticks calculation can wind up 0 due to
692 * a bug in the handling of lwp_slptime (as yet not found),
693 * so make sure we do not get a divide by 0 panic.
695 ttlticks = (cpbase - lp->lwp_cpbase) /
696 gd->gd_schedclock.periodic;
699 lp->lwp_cpbase = cpbase;
703 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
706 * Calculate the percentage of one cpu used factoring in ncpus
707 * and the load and adjust estcpu. Handle degenerate cases
708 * by adding 1 to bsd4_runqcount.
710 * estcpu is scaled by ESTCPUMAX.
712 * bsd4_runqcount is the excess number of user processes
713 * that cannot be immediately scheduled to cpus. We want
714 * to count these as running to avoid range compression
715 * in the base calculation (which is the actual percentage
718 estcpu = (lp->lwp_cpticks * ESTCPUMAX) *
719 (bsd4_runqcount + ncpus) / (ncpus * ttlticks);
722 * If estcpu is > 50% we become more batch-like
723 * If estcpu is <= 50% we become less batch-like
725 * It takes 30 cpu seconds to traverse the entire range.
727 if (estcpu > ESTCPUMAX / 2) {
728 lp->lwp_batch += ttlticks;
729 if (lp->lwp_batch > BATCHMAX)
730 lp->lwp_batch = BATCHMAX;
732 lp->lwp_batch -= ttlticks;
733 if (lp->lwp_batch < 0)
737 if (usched_debug == lp->lwp_proc->p_pid) {
738 kprintf("pid %d lwp %p estcpu %3d %3d bat %d cp %d/%d",
739 lp->lwp_proc->p_pid, lp,
740 estcpu, lp->lwp_estcpu,
742 lp->lwp_cpticks, ttlticks);
746 * Adjust lp->lwp_esetcpu. The decay factor determines how
747 * quickly lwp_estcpu collapses to its realtime calculation.
748 * A slower collapse gives us a more accurate number but
749 * can cause a cpu hog to eat too much cpu before the
750 * scheduler decides to downgrade it.
752 * NOTE: p_nice is accounted for in bsd4_resetpriority(),
753 * and not here, but we must still ensure that a
754 * cpu-bound nice -20 process does not completely
755 * override a cpu-bound nice +20 process.
757 * NOTE: We must use ESTCPULIM() here to deal with any
760 decay_factor = usched_bsd4_decay;
761 if (decay_factor < 1)
763 if (decay_factor > 1024)
766 lp->lwp_estcpu = ESTCPULIM(
767 (lp->lwp_estcpu * decay_factor + estcpu) /
770 if (usched_debug == lp->lwp_proc->p_pid)
771 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
772 bsd4_resetpriority(lp);
773 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
779 * Compute the priority of a process when running in user mode.
780 * Arrange to reschedule if the resulting priority is better
781 * than that of the current process.
783 * This routine may be called with any process.
785 * This routine is called by fork1() for initial setup with the process
786 * of the run queue, and also may be called normally with the process on or
792 bsd4_resetpriority(struct lwp *lp)
802 * Calculate the new priority and queue type
805 spin_lock(&bsd4_spin);
807 newrqtype = lp->lwp_rtprio.type;
810 case RTP_PRIO_REALTIME:
812 newpriority = PRIBASE_REALTIME +
813 (lp->lwp_rtprio.prio & PRIMASK);
815 case RTP_PRIO_NORMAL:
817 * Detune estcpu based on batchiness. lwp_batch ranges
818 * from 0 to BATCHMAX. Limit estcpu for the sake of
819 * the priority calculation to between 50% and 100%.
821 estcpu = lp->lwp_estcpu * (lp->lwp_batch + BATCHMAX) /
825 * p_nice piece Adds (0-40) * 2 0-80
826 * estcpu Adds 16384 * 4 / 512 0-128
828 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
829 newpriority += estcpu * PPQ / ESTCPUPPQ;
830 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
831 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
832 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
835 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
837 case RTP_PRIO_THREAD:
838 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
841 panic("Bad RTP_PRIO %d", newrqtype);
846 * The newpriority incorporates the queue type so do a simple masked
847 * check to determine if the process has moved to another queue. If
848 * it has, and it is currently on a run queue, then move it.
850 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
851 lp->lwp_priority = newpriority;
852 if (lp->lwp_flag & LWP_ONRUNQ) {
853 bsd4_remrunqueue_locked(lp);
854 lp->lwp_rqtype = newrqtype;
855 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
856 bsd4_setrunqueue_locked(lp);
859 lp->lwp_rqtype = newrqtype;
860 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
863 reschedcpu = lp->lwp_thread->td_gd->gd_cpuid;
865 lp->lwp_priority = newpriority;
871 * Determine if we need to reschedule the target cpu. This only
872 * occurs if the LWP is already on a scheduler queue, which means
873 * that idle cpu notification has already occured. At most we
874 * need only issue a need_user_resched() on the appropriate cpu.
876 * The LWP may be owned by a CPU different from the current one,
877 * in which case dd->uschedcp may be modified without an MP lock
878 * or a spinlock held. The worst that happens is that the code
879 * below causes a spurious need_user_resched() on the target CPU
880 * and dd->pri to be wrong for a short period of time, both of
881 * which are harmless.
883 * If checkpri is 0 we are adjusting the priority of the current
884 * process, possibly higher (less desireable), so ignore the upri
885 * check which will fail in that case.
887 if (reschedcpu >= 0) {
888 dd = &bsd4_pcpu[reschedcpu];
889 if ((bsd4_rdyprocmask & CPUMASK(reschedcpu)) &&
891 (dd->upri & ~PRIMASK) > (lp->lwp_priority & ~PRIMASK))) {
893 if (reschedcpu == mycpu->gd_cpuid) {
894 spin_unlock(&bsd4_spin);
897 spin_unlock(&bsd4_spin);
898 atomic_clear_cpumask(&bsd4_rdyprocmask,
899 CPUMASK(reschedcpu));
900 lwkt_send_ipiq(lp->lwp_thread->td_gd,
901 need_user_resched_remote, NULL);
904 spin_unlock(&bsd4_spin);
908 spin_unlock(&bsd4_spin);
911 spin_unlock(&bsd4_spin);
921 bsd4_yield(struct lwp *lp)
924 /* FUTURE (or something similar) */
925 switch(lp->lwp_rqtype) {
926 case RTP_PRIO_NORMAL:
927 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
937 * Called from fork1() when a new child process is being created.
939 * Give the child process an initial estcpu that is more batch then
940 * its parent and dock the parent for the fork (but do not
941 * reschedule the parent). This comprises the main part of our batch
942 * detection heuristic for both parallel forking and sequential execs.
944 * XXX lwp should be "spawning" instead of "forking"
949 bsd4_forking(struct lwp *plp, struct lwp *lp)
952 * Put the child 4 queue slots (out of 32) higher than the parent
953 * (less desireable than the parent).
955 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
958 * The batch status of children always starts out centerline
959 * and will inch-up or inch-down as appropriate. It takes roughly
960 * ~15 seconds of >50% cpu to hit the limit.
962 lp->lwp_batch = BATCHMAX / 2;
965 * Dock the parent a cost for the fork, protecting us from fork
966 * bombs. If the parent is forking quickly make the child more
969 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
973 * Called when a parent waits for a child.
978 bsd4_exiting(struct lwp *lp, struct proc *child_proc)
983 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
984 * it selects a user process and returns it. If chklp is non-NULL and chklp
985 * has a better or equal priority then the process that would otherwise be
986 * chosen, NULL is returned.
988 * Until we fix the RUNQ code the chklp test has to be strict or we may
989 * bounce between processes trying to acquire the current process designation.
991 * MPSAFE - must be called with bsd4_spin exclusive held. The spinlock is
992 * left intact through the entire routine.
996 chooseproc_locked(struct lwp *chklp)
1000 u_int32_t *which, *which2;
1007 rtqbits = bsd4_rtqueuebits;
1008 tsqbits = bsd4_queuebits;
1009 idqbits = bsd4_idqueuebits;
1010 cpumask = mycpu->gd_cpumask;
1016 pri = bsfl(rtqbits);
1017 q = &bsd4_rtqueues[pri];
1018 which = &bsd4_rtqueuebits;
1020 } else if (tsqbits) {
1021 pri = bsfl(tsqbits);
1022 q = &bsd4_queues[pri];
1023 which = &bsd4_queuebits;
1025 } else if (idqbits) {
1026 pri = bsfl(idqbits);
1027 q = &bsd4_idqueues[pri];
1028 which = &bsd4_idqueuebits;
1033 lp = TAILQ_FIRST(q);
1034 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1037 while ((lp->lwp_cpumask & cpumask) == 0) {
1038 lp = TAILQ_NEXT(lp, lwp_procq);
1040 *which2 &= ~(1 << pri);
1047 * If the passed lwp <chklp> is reasonably close to the selected
1048 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1050 * Note that we must error on the side of <chklp> to avoid bouncing
1051 * between threads in the acquire code.
1054 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1060 * If the chosen lwp does not reside on this cpu spend a few
1061 * cycles looking for a better candidate at the same priority level.
1062 * This is a fallback check, setrunqueue() tries to wakeup the
1063 * correct cpu and is our front-line affinity.
1065 if (lp->lwp_thread->td_gd != mycpu &&
1066 (chklp = TAILQ_NEXT(lp, lwp_procq)) != NULL
1068 if (chklp->lwp_thread->td_gd == mycpu) {
1075 TAILQ_REMOVE(q, lp, lwp_procq);
1078 *which &= ~(1 << pri);
1079 KASSERT((lp->lwp_flag & LWP_ONRUNQ) != 0, ("not on runq6!"));
1080 lp->lwp_flag &= ~LWP_ONRUNQ;
1088 need_user_resched_remote(void *dummy)
1090 globaldata_t gd = mycpu;
1091 bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
1093 need_user_resched();
1094 lwkt_schedule(&dd->helper_thread);
1100 * bsd4_remrunqueue_locked() removes a given process from the run queue
1101 * that it is on, clearing the queue busy bit if it becomes empty.
1103 * Note that user process scheduler is different from the LWKT schedule.
1104 * The user process scheduler only manages user processes but it uses LWKT
1105 * underneath, and a user process operating in the kernel will often be
1106 * 'released' from our management.
1108 * MPSAFE - bsd4_spin must be held exclusively on call
1111 bsd4_remrunqueue_locked(struct lwp *lp)
1117 KKASSERT(lp->lwp_flag & LWP_ONRUNQ);
1118 lp->lwp_flag &= ~LWP_ONRUNQ;
1120 KKASSERT(bsd4_runqcount >= 0);
1122 pri = lp->lwp_rqindex;
1123 switch(lp->lwp_rqtype) {
1124 case RTP_PRIO_NORMAL:
1125 q = &bsd4_queues[pri];
1126 which = &bsd4_queuebits;
1128 case RTP_PRIO_REALTIME:
1130 q = &bsd4_rtqueues[pri];
1131 which = &bsd4_rtqueuebits;
1134 q = &bsd4_idqueues[pri];
1135 which = &bsd4_idqueuebits;
1138 panic("remrunqueue: invalid rtprio type");
1141 TAILQ_REMOVE(q, lp, lwp_procq);
1142 if (TAILQ_EMPTY(q)) {
1143 KASSERT((*which & (1 << pri)) != 0,
1144 ("remrunqueue: remove from empty queue"));
1145 *which &= ~(1 << pri);
1150 * bsd4_setrunqueue_locked()
1152 * Add a process whos rqtype and rqindex had previously been calculated
1153 * onto the appropriate run queue. Determine if the addition requires
1154 * a reschedule on a cpu and return the cpuid or -1.
1156 * NOTE: Lower priorities are better priorities.
1158 * MPSAFE - bsd4_spin must be held exclusively on call
1161 bsd4_setrunqueue_locked(struct lwp *lp)
1167 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
1168 lp->lwp_flag |= LWP_ONRUNQ;
1171 pri = lp->lwp_rqindex;
1173 switch(lp->lwp_rqtype) {
1174 case RTP_PRIO_NORMAL:
1175 q = &bsd4_queues[pri];
1176 which = &bsd4_queuebits;
1178 case RTP_PRIO_REALTIME:
1180 q = &bsd4_rtqueues[pri];
1181 which = &bsd4_rtqueuebits;
1184 q = &bsd4_idqueues[pri];
1185 which = &bsd4_idqueuebits;
1188 panic("remrunqueue: invalid rtprio type");
1193 * Add to the correct queue and set the appropriate bit. If no
1194 * lower priority (i.e. better) processes are in the queue then
1195 * we want a reschedule, calculate the best cpu for the job.
1197 * Always run reschedules on the LWPs original cpu.
1199 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1206 * For SMP systems a user scheduler helper thread is created for each
1207 * cpu and is used to allow one cpu to wakeup another for the purposes of
1208 * scheduling userland threads from setrunqueue().
1210 * UP systems do not need the helper since there is only one cpu.
1212 * We can't use the idle thread for this because we might block.
1213 * Additionally, doing things this way allows us to HLT idle cpus
1219 sched_thread(void *dummy)
1232 cpuid = gd->gd_cpuid; /* doesn't change */
1233 mask = gd->gd_cpumask; /* doesn't change */
1234 dd = &bsd4_pcpu[cpuid];
1237 * Since we are woken up only when no user processes are scheduled
1238 * on a cpu, we can run at an ultra low priority.
1240 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1244 * We use the LWKT deschedule-interlock trick to avoid racing
1245 * bsd4_rdyprocmask. This means we cannot block through to the
1246 * manual lwkt_switch() call we make below.
1249 lwkt_deschedule_self(gd->gd_curthread);
1250 spin_lock(&bsd4_spin);
1251 atomic_set_cpumask(&bsd4_rdyprocmask, mask);
1253 clear_user_resched(); /* This satisfied the reschedule request */
1254 dd->rrcount = 0; /* Reset the round-robin counter */
1256 if ((bsd4_curprocmask & mask) == 0) {
1258 * No thread is currently scheduled.
1260 KKASSERT(dd->uschedcp == NULL);
1261 if ((nlp = chooseproc_locked(NULL)) != NULL) {
1262 atomic_set_cpumask(&bsd4_curprocmask, mask);
1263 dd->upri = nlp->lwp_priority;
1265 spin_unlock(&bsd4_spin);
1267 lwkt_acquire(nlp->lwp_thread);
1269 lwkt_schedule(nlp->lwp_thread);
1271 spin_unlock(&bsd4_spin);
1273 } else if (bsd4_runqcount) {
1274 if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) {
1275 dd->upri = nlp->lwp_priority;
1277 spin_unlock(&bsd4_spin);
1279 lwkt_acquire(nlp->lwp_thread);
1281 lwkt_schedule(nlp->lwp_thread);
1284 * CHAINING CONDITION TRAIN
1286 * We could not deal with the scheduler wakeup
1287 * request on this cpu, locate a ready scheduler
1288 * with no current lp assignment and chain to it.
1290 * This ensures that a wakeup race which fails due
1291 * to priority test does not leave other unscheduled
1292 * cpus idle when the runqueue is not empty.
1294 tmpmask = ~bsd4_curprocmask & bsd4_rdyprocmask &
1297 tmpid = BSFCPUMASK(tmpmask);
1298 gd = globaldata_find(cpuid);
1299 dd = &bsd4_pcpu[cpuid];
1300 atomic_clear_cpumask(&bsd4_rdyprocmask,
1302 spin_unlock(&bsd4_spin);
1303 lwkt_schedule(&dd->helper_thread);
1305 spin_unlock(&bsd4_spin);
1310 * The runq is empty.
1312 spin_unlock(&bsd4_spin);
1320 * Setup our scheduler helpers. Note that curprocmask bit 0 has already
1321 * been cleared by rqinit() and we should not mess with it further.
1324 sched_thread_cpu_init(void)
1329 kprintf("start scheduler helpers on cpus:");
1331 for (i = 0; i < ncpus; ++i) {
1332 bsd4_pcpu_t dd = &bsd4_pcpu[i];
1333 cpumask_t mask = CPUMASK(i);
1335 if ((mask & smp_active_mask) == 0)
1341 lwkt_create(sched_thread, NULL, NULL, &dd->helper_thread,
1342 TDF_STOPREQ, i, "usched %d", i);
1345 * Allow user scheduling on the target cpu. cpu #0 has already
1346 * been enabled in rqinit().
1349 atomic_clear_cpumask(&bsd4_curprocmask, mask);
1350 atomic_set_cpumask(&bsd4_rdyprocmask, mask);
1351 dd->upri = PRIBASE_NULL;
1356 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
1357 sched_thread_cpu_init, NULL)