2 * Copyright (c) 2002-2005, Jeffrey Roberson <jeff@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 unmodified, this list of conditions, and the following
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
30 #include <opt_sched.h>
34 #include <sys/param.h>
35 #include <sys/systm.h>
37 #include <sys/kernel.h>
40 #include <sys/mutex.h>
42 #include <sys/resource.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sched.h>
47 #include <sys/sysctl.h>
48 #include <sys/sysproto.h>
49 #include <sys/turnstile.h>
50 #include <sys/vmmeter.h>
53 #include <sys/ktrace.h>
57 #include <sys/pmckern.h>
60 #include <machine/cpu.h>
61 #include <machine/smp.h>
63 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
64 /* XXX This is bogus compatability crap for ps */
65 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
66 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
68 static void sched_setup(void *dummy);
69 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
71 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
73 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0,
76 static int slice_min = 1;
77 SYSCTL_INT(_kern_sched, OID_AUTO, slice_min, CTLFLAG_RW, &slice_min, 0, "");
79 static int slice_max = 10;
80 SYSCTL_INT(_kern_sched, OID_AUTO, slice_max, CTLFLAG_RW, &slice_max, 0, "");
86 * The following datastructures are allocated within their parent structure
87 * but are scheduler specific.
90 * The schedulable entity that can be given a context to run. A process may
91 * have several of these.
94 TAILQ_ENTRY(kse) ke_procq; /* (j/z) Run queue. */
95 int ke_flags; /* (j) KEF_* flags. */
96 struct thread *ke_thread; /* (*) Active associated thread. */
97 fixpt_t ke_pctcpu; /* (j) %cpu during p_swtime. */
98 char ke_rqindex; /* (j) Run queue index. */
100 KES_THREAD = 0x0, /* slaved to thread state */
102 } ke_state; /* (j) thread sched specific status. */
105 struct runq *ke_runq;
106 u_char ke_cpu; /* CPU that we have affinity for. */
107 /* The following variables are only used for pctcpu calculation */
108 int ke_ltick; /* Last tick that we were running on */
109 int ke_ftick; /* First tick that we were running on */
110 int ke_ticks; /* Tick count */
113 #define td_kse td_sched
114 #define td_slptime td_kse->ke_slptime
115 #define ke_proc ke_thread->td_proc
116 #define ke_ksegrp ke_thread->td_ksegrp
117 #define ke_assign ke_procq.tqe_next
118 /* flags kept in ke_flags */
119 #define KEF_ASSIGNED 0x0001 /* Thread is being migrated. */
120 #define KEF_BOUND 0x0002 /* Thread can not migrate. */
121 #define KEF_XFERABLE 0x0004 /* Thread was added as transferable. */
122 #define KEF_HOLD 0x0008 /* Thread is temporarily bound. */
123 #define KEF_REMOVED 0x0010 /* Thread was removed while ASSIGNED */
124 #define KEF_INTERNAL 0x0020 /* Thread added due to migration. */
125 #define KEF_DIDRUN 0x02000 /* Thread actually ran. */
126 #define KEF_EXIT 0x04000 /* Thread is being killed. */
129 struct thread *skg_last_assigned; /* (j) Last thread assigned to */
130 /* the system scheduler */
131 int skg_slptime; /* Number of ticks we vol. slept */
132 int skg_runtime; /* Number of ticks we were running */
133 int skg_avail_opennings; /* (j) Num unfilled slots in group.*/
134 int skg_concurrency; /* (j) Num threads requested in group.*/
136 #define kg_last_assigned kg_sched->skg_last_assigned
137 #define kg_avail_opennings kg_sched->skg_avail_opennings
138 #define kg_concurrency kg_sched->skg_concurrency
139 #define kg_runtime kg_sched->skg_runtime
140 #define kg_slptime kg_sched->skg_slptime
142 #define SLOT_RELEASE(kg) (kg)->kg_avail_opennings++
143 #define SLOT_USE(kg) (kg)->kg_avail_opennings--
145 static struct kse kse0;
146 static struct kg_sched kg_sched0;
149 * The priority is primarily determined by the interactivity score. Thus, we
150 * give lower(better) priorities to kse groups that use less CPU. The nice
151 * value is then directly added to this to allow nice to have some effect
154 * PRI_RANGE: Total priority range for timeshare threads.
155 * PRI_NRESV: Number of nice values.
156 * PRI_BASE: The start of the dynamic range.
158 #define SCHED_PRI_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
159 #define SCHED_PRI_NRESV ((PRIO_MAX - PRIO_MIN) + 1)
160 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
161 #define SCHED_PRI_BASE (PRI_MIN_TIMESHARE)
162 #define SCHED_PRI_INTERACT(score) \
163 ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
166 * These determine the interactivity of a process.
168 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
169 * before throttling back.
170 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
171 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
172 * INTERACT_THRESH: Threshhold for placement on the current runq.
174 #define SCHED_SLP_RUN_MAX ((hz * 5) << 10)
175 #define SCHED_SLP_RUN_FORK ((hz / 2) << 10)
176 #define SCHED_INTERACT_MAX (100)
177 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
178 #define SCHED_INTERACT_THRESH (30)
181 * These parameters and macros determine the size of the time slice that is
182 * granted to each thread.
184 * SLICE_MIN: Minimum time slice granted, in units of ticks.
185 * SLICE_MAX: Maximum time slice granted.
186 * SLICE_RANGE: Range of available time slices scaled by hz.
187 * SLICE_SCALE: The number slices granted per val in the range of [0, max].
188 * SLICE_NICE: Determine the amount of slice granted to a scaled nice.
189 * SLICE_NTHRESH: The nice cutoff point for slice assignment.
191 #define SCHED_SLICE_MIN (slice_min)
192 #define SCHED_SLICE_MAX (slice_max)
193 #define SCHED_SLICE_INTERACTIVE (slice_max)
194 #define SCHED_SLICE_NTHRESH (SCHED_PRI_NHALF - 1)
195 #define SCHED_SLICE_RANGE (SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
196 #define SCHED_SLICE_SCALE(val, max) (((val) * SCHED_SLICE_RANGE) / (max))
197 #define SCHED_SLICE_NICE(nice) \
198 (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
201 * This macro determines whether or not the thread belongs on the current or
204 #define SCHED_INTERACTIVE(kg) \
205 (sched_interact_score(kg) < SCHED_INTERACT_THRESH)
206 #define SCHED_CURR(kg, ke) \
207 ((ke->ke_thread->td_flags & TDF_BORROWING) || SCHED_INTERACTIVE(kg))
210 * Cpu percentage computation macros and defines.
212 * SCHED_CPU_TIME: Number of seconds to average the cpu usage across.
213 * SCHED_CPU_TICKS: Number of hz ticks to average the cpu usage across.
216 #define SCHED_CPU_TIME 10
217 #define SCHED_CPU_TICKS (hz * SCHED_CPU_TIME)
220 * kseq - per processor runqs and statistics.
223 struct runq ksq_idle; /* Queue of IDLE threads. */
224 struct runq ksq_timeshare[2]; /* Run queues for !IDLE. */
225 struct runq *ksq_next; /* Next timeshare queue. */
226 struct runq *ksq_curr; /* Current queue. */
227 int ksq_load_timeshare; /* Load for timeshare. */
228 int ksq_load; /* Aggregate load. */
229 short ksq_nice[SCHED_PRI_NRESV]; /* KSEs in each nice bin. */
230 short ksq_nicemin; /* Least nice. */
232 int ksq_transferable;
233 LIST_ENTRY(kseq) ksq_siblings; /* Next in kseq group. */
234 struct kseq_group *ksq_group; /* Our processor group. */
235 volatile struct kse *ksq_assigned; /* assigned by another CPU. */
237 int ksq_sysload; /* For loadavg, !ITHD load. */
243 * kseq groups are groups of processors which can cheaply share threads. When
244 * one processor in the group goes idle it will check the runqs of the other
245 * processors in its group prior to halting and waiting for an interrupt.
246 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
247 * In a numa environment we'd want an idle bitmap per group and a two tiered
251 int ksg_cpus; /* Count of CPUs in this kseq group. */
252 cpumask_t ksg_cpumask; /* Mask of cpus in this group. */
253 cpumask_t ksg_idlemask; /* Idle cpus in this group. */
254 cpumask_t ksg_mask; /* Bit mask for first cpu. */
255 int ksg_load; /* Total load of this group. */
256 int ksg_transferable; /* Transferable load of this group. */
257 LIST_HEAD(, kseq) ksg_members; /* Linked list of all members. */
262 * One kse queue per processor.
265 static cpumask_t kseq_idle;
266 static int ksg_maxid;
267 static struct kseq kseq_cpu[MAXCPU];
268 static struct kseq_group kseq_groups[MAXCPU];
270 static int gbal_tick;
271 static int balance_groups;
273 #define KSEQ_SELF() (&kseq_cpu[PCPU_GET(cpuid)])
274 #define KSEQ_CPU(x) (&kseq_cpu[(x)])
275 #define KSEQ_ID(x) ((x) - kseq_cpu)
276 #define KSEQ_GROUP(x) (&kseq_groups[(x)])
278 static struct kseq kseq_cpu;
280 #define KSEQ_SELF() (&kseq_cpu)
281 #define KSEQ_CPU(x) (&kseq_cpu)
284 static void slot_fill(struct ksegrp *);
285 static struct kse *sched_choose(void); /* XXX Should be thread * */
286 static void sched_slice(struct kse *);
287 static void sched_priority(struct ksegrp *);
288 static void sched_thread_priority(struct thread *, u_char);
289 static int sched_interact_score(struct ksegrp *);
290 static void sched_interact_update(struct ksegrp *);
291 static void sched_interact_fork(struct ksegrp *);
292 static void sched_pctcpu_update(struct kse *);
294 /* Operations on per processor queues */
295 static struct kse * kseq_choose(struct kseq *);
296 static void kseq_setup(struct kseq *);
297 static void kseq_load_add(struct kseq *, struct kse *);
298 static void kseq_load_rem(struct kseq *, struct kse *);
299 static __inline void kseq_runq_add(struct kseq *, struct kse *, int);
300 static __inline void kseq_runq_rem(struct kseq *, struct kse *);
301 static void kseq_nice_add(struct kseq *, int);
302 static void kseq_nice_rem(struct kseq *, int);
303 void kseq_print(int cpu);
305 static int kseq_transfer(struct kseq *, struct kse *, int);
306 static struct kse *runq_steal(struct runq *);
307 static void sched_balance(void);
308 static void sched_balance_groups(void);
309 static void sched_balance_group(struct kseq_group *);
310 static void sched_balance_pair(struct kseq *, struct kseq *);
311 static void kseq_move(struct kseq *, int);
312 static int kseq_idled(struct kseq *);
313 static void kseq_notify(struct kse *, int);
314 static void kseq_assign(struct kseq *);
315 static struct kse *kseq_steal(struct kseq *, int);
316 #define KSE_CAN_MIGRATE(ke) \
317 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
326 kseq = KSEQ_CPU(cpu);
329 printf("\tload: %d\n", kseq->ksq_load);
330 printf("\tload TIMESHARE: %d\n", kseq->ksq_load_timeshare);
332 printf("\tload transferable: %d\n", kseq->ksq_transferable);
334 printf("\tnicemin:\t%d\n", kseq->ksq_nicemin);
335 printf("\tnice counts:\n");
336 for (i = 0; i < SCHED_PRI_NRESV; i++)
337 if (kseq->ksq_nice[i])
338 printf("\t\t%d = %d\n",
339 i - SCHED_PRI_NHALF, kseq->ksq_nice[i]);
343 kseq_runq_add(struct kseq *kseq, struct kse *ke, int flags)
346 if (KSE_CAN_MIGRATE(ke)) {
347 kseq->ksq_transferable++;
348 kseq->ksq_group->ksg_transferable++;
349 ke->ke_flags |= KEF_XFERABLE;
352 runq_add(ke->ke_runq, ke, flags);
356 kseq_runq_rem(struct kseq *kseq, struct kse *ke)
359 if (ke->ke_flags & KEF_XFERABLE) {
360 kseq->ksq_transferable--;
361 kseq->ksq_group->ksg_transferable--;
362 ke->ke_flags &= ~KEF_XFERABLE;
365 runq_remove(ke->ke_runq, ke);
369 kseq_load_add(struct kseq *kseq, struct kse *ke)
372 mtx_assert(&sched_lock, MA_OWNED);
373 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
374 if (class == PRI_TIMESHARE)
375 kseq->ksq_load_timeshare++;
377 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
378 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
380 kseq->ksq_group->ksg_load++;
384 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
385 kseq_nice_add(kseq, ke->ke_proc->p_nice);
389 kseq_load_rem(struct kseq *kseq, struct kse *ke)
392 mtx_assert(&sched_lock, MA_OWNED);
393 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
394 if (class == PRI_TIMESHARE)
395 kseq->ksq_load_timeshare--;
396 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
398 kseq->ksq_group->ksg_load--;
403 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
405 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
406 kseq_nice_rem(kseq, ke->ke_proc->p_nice);
410 kseq_nice_add(struct kseq *kseq, int nice)
412 mtx_assert(&sched_lock, MA_OWNED);
413 /* Normalize to zero. */
414 kseq->ksq_nice[nice + SCHED_PRI_NHALF]++;
415 if (nice < kseq->ksq_nicemin || kseq->ksq_load_timeshare == 1)
416 kseq->ksq_nicemin = nice;
420 kseq_nice_rem(struct kseq *kseq, int nice)
424 mtx_assert(&sched_lock, MA_OWNED);
425 /* Normalize to zero. */
426 n = nice + SCHED_PRI_NHALF;
428 KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count."));
431 * If this wasn't the smallest nice value or there are more in
432 * this bucket we can just return. Otherwise we have to recalculate
435 if (nice != kseq->ksq_nicemin ||
436 kseq->ksq_nice[n] != 0 ||
437 kseq->ksq_load_timeshare == 0)
440 for (; n < SCHED_PRI_NRESV; n++)
441 if (kseq->ksq_nice[n]) {
442 kseq->ksq_nicemin = n - SCHED_PRI_NHALF;
449 * sched_balance is a simple CPU load balancing algorithm. It operates by
450 * finding the least loaded and most loaded cpu and equalizing their load
451 * by migrating some processes.
453 * Dealing only with two CPUs at a time has two advantages. Firstly, most
454 * installations will only have 2 cpus. Secondly, load balancing too much at
455 * once can have an unpleasant effect on the system. The scheduler rarely has
456 * enough information to make perfect decisions. So this algorithm chooses
457 * algorithm simplicity and more gradual effects on load in larger systems.
459 * It could be improved by considering the priorities and slices assigned to
460 * each task prior to balancing them. There are many pathological cases with
461 * any approach and so the semi random algorithm below may work as well as any.
467 struct kseq_group *high;
468 struct kseq_group *low;
469 struct kseq_group *ksg;
473 bal_tick = ticks + (random() % (hz * 2));
474 if (smp_started == 0)
477 i = random() % (ksg_maxid + 1);
478 for (cnt = 0; cnt <= ksg_maxid; cnt++) {
481 * Find the CPU with the highest load that has some
482 * threads to transfer.
484 if ((high == NULL || ksg->ksg_load > high->ksg_load)
485 && ksg->ksg_transferable)
487 if (low == NULL || ksg->ksg_load < low->ksg_load)
492 if (low != NULL && high != NULL && high != low)
493 sched_balance_pair(LIST_FIRST(&high->ksg_members),
494 LIST_FIRST(&low->ksg_members));
498 sched_balance_groups(void)
502 gbal_tick = ticks + (random() % (hz * 2));
503 mtx_assert(&sched_lock, MA_OWNED);
505 for (i = 0; i <= ksg_maxid; i++)
506 sched_balance_group(KSEQ_GROUP(i));
510 sched_balance_group(struct kseq_group *ksg)
517 if (ksg->ksg_transferable == 0)
521 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
522 load = kseq->ksq_load;
523 if (high == NULL || load > high->ksq_load)
525 if (low == NULL || load < low->ksq_load)
528 if (high != NULL && low != NULL && high != low)
529 sched_balance_pair(high, low);
533 sched_balance_pair(struct kseq *high, struct kseq *low)
543 * If we're transfering within a group we have to use this specific
544 * kseq's transferable count, otherwise we can steal from other members
547 if (high->ksq_group == low->ksq_group) {
548 transferable = high->ksq_transferable;
549 high_load = high->ksq_load;
550 low_load = low->ksq_load;
552 transferable = high->ksq_group->ksg_transferable;
553 high_load = high->ksq_group->ksg_load;
554 low_load = low->ksq_group->ksg_load;
556 if (transferable == 0)
559 * Determine what the imbalance is and then adjust that to how many
560 * kses we actually have to give up (transferable).
562 diff = high_load - low_load;
566 move = min(move, transferable);
567 for (i = 0; i < move; i++)
568 kseq_move(high, KSEQ_ID(low));
573 kseq_move(struct kseq *from, int cpu)
581 ke = kseq_steal(kseq, 1);
583 struct kseq_group *ksg;
585 ksg = kseq->ksq_group;
586 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
587 if (kseq == from || kseq->ksq_transferable == 0)
589 ke = kseq_steal(kseq, 1);
593 panic("kseq_move: No KSEs available with a "
594 "transferable count of %d\n",
595 ksg->ksg_transferable);
599 ke->ke_state = KES_THREAD;
600 kseq_runq_rem(kseq, ke);
601 kseq_load_rem(kseq, ke);
602 kseq_notify(ke, cpu);
606 kseq_idled(struct kseq *kseq)
608 struct kseq_group *ksg;
612 ksg = kseq->ksq_group;
614 * If we're in a cpu group, try and steal kses from another cpu in
615 * the group before idling.
617 if (ksg->ksg_cpus > 1 && ksg->ksg_transferable) {
618 LIST_FOREACH(steal, &ksg->ksg_members, ksq_siblings) {
619 if (steal == kseq || steal->ksq_transferable == 0)
621 ke = kseq_steal(steal, 0);
624 ke->ke_state = KES_THREAD;
625 kseq_runq_rem(steal, ke);
626 kseq_load_rem(steal, ke);
627 ke->ke_cpu = PCPU_GET(cpuid);
628 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
629 sched_add(ke->ke_thread, SRQ_YIELDING);
634 * We only set the idled bit when all of the cpus in the group are
635 * idle. Otherwise we could get into a situation where a KSE bounces
636 * back and forth between two idle cores on seperate physical CPUs.
638 ksg->ksg_idlemask |= PCPU_GET(cpumask);
639 if (ksg->ksg_idlemask != ksg->ksg_cpumask)
641 atomic_set_int(&kseq_idle, ksg->ksg_mask);
646 kseq_assign(struct kseq *kseq)
652 *(volatile struct kse **)&ke = kseq->ksq_assigned;
653 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke, NULL));
654 for (; ke != NULL; ke = nke) {
656 kseq->ksq_group->ksg_load--;
658 ke->ke_flags &= ~KEF_ASSIGNED;
659 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
660 sched_add(ke->ke_thread, SRQ_YIELDING);
665 kseq_notify(struct kse *ke, int cpu)
673 kseq = KSEQ_CPU(cpu);
675 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
676 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
677 (kseq_idle & kseq->ksq_group->ksg_mask))
678 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
679 kseq->ksq_group->ksg_load++;
682 ke->ke_flags |= KEF_ASSIGNED;
683 prio = ke->ke_thread->td_priority;
686 * Place a KSE on another cpu's queue and force a resched.
689 *(volatile struct kse **)&ke->ke_assign = kseq->ksq_assigned;
690 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke->ke_assign, ke));
692 * Without sched_lock we could lose a race where we set NEEDRESCHED
693 * on a thread that is switched out before the IPI is delivered. This
694 * would lead us to miss the resched. This will be a problem once
695 * sched_lock is pushed down.
697 pcpu = pcpu_find(cpu);
698 td = pcpu->pc_curthread;
699 if (ke->ke_thread->td_priority < td->td_priority ||
700 td == pcpu->pc_idlethread) {
701 td->td_flags |= TDF_NEEDRESCHED;
702 ipi_selected(1 << cpu, IPI_AST);
707 runq_steal(struct runq *rq)
715 mtx_assert(&sched_lock, MA_OWNED);
716 rqb = &rq->rq_status;
717 for (word = 0; word < RQB_LEN; word++) {
718 if (rqb->rqb_bits[word] == 0)
720 for (bit = 0; bit < RQB_BPW; bit++) {
721 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
723 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
724 TAILQ_FOREACH(ke, rqh, ke_procq) {
725 if (KSE_CAN_MIGRATE(ke))
734 kseq_steal(struct kseq *kseq, int stealidle)
739 * Steal from next first to try to get a non-interactive task that
740 * may not have run for a while.
742 if ((ke = runq_steal(kseq->ksq_next)) != NULL)
744 if ((ke = runq_steal(kseq->ksq_curr)) != NULL)
747 return (runq_steal(&kseq->ksq_idle));
752 kseq_transfer(struct kseq *kseq, struct kse *ke, int class)
754 struct kseq_group *nksg;
755 struct kseq_group *ksg;
760 if (smp_started == 0)
764 * If our load exceeds a certain threshold we should attempt to
765 * reassign this thread. The first candidate is the cpu that
766 * originally ran the thread. If it is idle, assign it there,
767 * otherwise, pick an idle cpu.
769 * The threshold at which we start to reassign kses has a large impact
770 * on the overall performance of the system. Tuned too high and
771 * some CPUs may idle. Too low and there will be excess migration
772 * and context switches.
774 old = KSEQ_CPU(ke->ke_cpu);
775 nksg = old->ksq_group;
776 ksg = kseq->ksq_group;
778 if (kseq_idle & nksg->ksg_mask) {
779 cpu = ffs(nksg->ksg_idlemask);
782 "kseq_transfer: %p found old cpu %X "
783 "in idlemask.", ke, cpu);
788 * Multiple cpus could find this bit simultaneously
789 * but the race shouldn't be terrible.
791 cpu = ffs(kseq_idle);
793 CTR2(KTR_SCHED, "kseq_transfer: %p found %X "
794 "in idlemask.", ke, cpu);
800 if (old->ksq_load < kseq->ksq_load) {
801 cpu = ke->ke_cpu + 1;
802 CTR2(KTR_SCHED, "kseq_transfer: %p old cpu %X "
803 "load less than ours.", ke, cpu);
807 * No new CPU was found, look for one with less load.
809 for (idx = 0; idx <= ksg_maxid; idx++) {
810 nksg = KSEQ_GROUP(idx);
811 if (nksg->ksg_load /*+ (nksg->ksg_cpus * 2)*/ < ksg->ksg_load) {
812 cpu = ffs(nksg->ksg_cpumask);
813 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X load less "
814 "than ours.", ke, cpu);
820 * If another cpu in this group has idled, assign a thread over
821 * to them after checking to see if there are idled groups.
823 if (ksg->ksg_idlemask) {
824 cpu = ffs(ksg->ksg_idlemask);
826 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X idle in "
834 * Now that we've found an idle CPU, migrate the thread.
838 kseq_notify(ke, cpu);
846 * Pick the highest priority task we have and return it.
850 kseq_choose(struct kseq *kseq)
856 mtx_assert(&sched_lock, MA_OWNED);
860 ke = runq_choose(kseq->ksq_curr);
863 * We already swapped once and didn't get anywhere.
867 swap = kseq->ksq_curr;
868 kseq->ksq_curr = kseq->ksq_next;
869 kseq->ksq_next = swap;
873 * If we encounter a slice of 0 the kse is in a
874 * TIMESHARE kse group and its nice was too far out
875 * of the range that receives slices.
877 nice = ke->ke_proc->p_nice + (0 - kseq->ksq_nicemin);
878 if (ke->ke_slice == 0 || (nice > SCHED_SLICE_NTHRESH &&
879 ke->ke_proc->p_nice != 0)) {
880 runq_remove(ke->ke_runq, ke);
882 ke->ke_runq = kseq->ksq_next;
883 runq_add(ke->ke_runq, ke, 0);
889 return (runq_choose(&kseq->ksq_idle));
893 kseq_setup(struct kseq *kseq)
895 runq_init(&kseq->ksq_timeshare[0]);
896 runq_init(&kseq->ksq_timeshare[1]);
897 runq_init(&kseq->ksq_idle);
898 kseq->ksq_curr = &kseq->ksq_timeshare[0];
899 kseq->ksq_next = &kseq->ksq_timeshare[1];
901 kseq->ksq_load_timeshare = 0;
905 sched_setup(void *dummy)
911 slice_min = (hz/100); /* 10ms */
912 slice_max = (hz/7); /* ~140ms */
917 * Initialize the kseqs.
919 for (i = 0; i < MAXCPU; i++) {
923 ksq->ksq_assigned = NULL;
924 kseq_setup(&kseq_cpu[i]);
926 if (smp_topology == NULL) {
927 struct kseq_group *ksg;
931 for (cpus = 0, i = 0; i < MAXCPU; i++) {
934 ksq = &kseq_cpu[cpus];
935 ksg = &kseq_groups[cpus];
937 * Setup a kseq group with one member.
939 ksq->ksq_transferable = 0;
940 ksq->ksq_group = ksg;
942 ksg->ksg_idlemask = 0;
943 ksg->ksg_cpumask = ksg->ksg_mask = 1 << i;
945 ksg->ksg_transferable = 0;
946 LIST_INIT(&ksg->ksg_members);
947 LIST_INSERT_HEAD(&ksg->ksg_members, ksq, ksq_siblings);
950 ksg_maxid = cpus - 1;
952 struct kseq_group *ksg;
953 struct cpu_group *cg;
956 for (i = 0; i < smp_topology->ct_count; i++) {
957 cg = &smp_topology->ct_group[i];
958 ksg = &kseq_groups[i];
960 * Initialize the group.
962 ksg->ksg_idlemask = 0;
964 ksg->ksg_transferable = 0;
965 ksg->ksg_cpus = cg->cg_count;
966 ksg->ksg_cpumask = cg->cg_mask;
967 LIST_INIT(&ksg->ksg_members);
969 * Find all of the group members and add them.
971 for (j = 0; j < MAXCPU; j++) {
972 if ((cg->cg_mask & (1 << j)) != 0) {
973 if (ksg->ksg_mask == 0)
974 ksg->ksg_mask = 1 << j;
975 kseq_cpu[j].ksq_transferable = 0;
976 kseq_cpu[j].ksq_group = ksg;
977 LIST_INSERT_HEAD(&ksg->ksg_members,
978 &kseq_cpu[j], ksq_siblings);
981 if (ksg->ksg_cpus > 1)
984 ksg_maxid = smp_topology->ct_count - 1;
987 * Stagger the group and global load balancer so they do not
988 * interfere with each other.
990 bal_tick = ticks + hz;
992 gbal_tick = ticks + (hz / 2);
994 kseq_setup(KSEQ_SELF());
996 mtx_lock_spin(&sched_lock);
997 kseq_load_add(KSEQ_SELF(), &kse0);
998 mtx_unlock_spin(&sched_lock);
1002 * Scale the scheduling priority according to the "interactivity" of this
1006 sched_priority(struct ksegrp *kg)
1010 if (kg->kg_pri_class != PRI_TIMESHARE)
1013 pri = SCHED_PRI_INTERACT(sched_interact_score(kg));
1014 pri += SCHED_PRI_BASE;
1015 pri += kg->kg_proc->p_nice;
1017 if (pri > PRI_MAX_TIMESHARE)
1018 pri = PRI_MAX_TIMESHARE;
1019 else if (pri < PRI_MIN_TIMESHARE)
1020 pri = PRI_MIN_TIMESHARE;
1022 kg->kg_user_pri = pri;
1028 * Calculate a time slice based on the properties of the kseg and the runq
1029 * that we're on. This is only for PRI_TIMESHARE ksegrps.
1032 sched_slice(struct kse *ke)
1038 kseq = KSEQ_CPU(ke->ke_cpu);
1040 if (ke->ke_thread->td_flags & TDF_BORROWING) {
1041 ke->ke_slice = SCHED_SLICE_MIN;
1047 * KSEs in interactive ksegs get a minimal slice so that we
1048 * quickly notice if it abuses its advantage.
1050 * KSEs in non-interactive ksegs are assigned a slice that is
1051 * based on the ksegs nice value relative to the least nice kseg
1052 * on the run queue for this cpu.
1054 * If the KSE is less nice than all others it gets the maximum
1055 * slice and other KSEs will adjust their slice relative to
1056 * this when they first expire.
1058 * There is 20 point window that starts relative to the least
1059 * nice kse on the run queue. Slice size is determined by
1060 * the kse distance from the last nice ksegrp.
1062 * If the kse is outside of the window it will get no slice
1063 * and will be reevaluated each time it is selected on the
1064 * run queue. The exception to this is nice 0 ksegs when
1065 * a nice -20 is running. They are always granted a minimum
1068 if (!SCHED_INTERACTIVE(kg)) {
1071 nice = kg->kg_proc->p_nice + (0 - kseq->ksq_nicemin);
1072 if (kseq->ksq_load_timeshare == 0 ||
1073 kg->kg_proc->p_nice < kseq->ksq_nicemin)
1074 ke->ke_slice = SCHED_SLICE_MAX;
1075 else if (nice <= SCHED_SLICE_NTHRESH)
1076 ke->ke_slice = SCHED_SLICE_NICE(nice);
1077 else if (kg->kg_proc->p_nice == 0)
1078 ke->ke_slice = SCHED_SLICE_MIN;
1082 ke->ke_slice = SCHED_SLICE_INTERACTIVE;
1088 * This routine enforces a maximum limit on the amount of scheduling history
1089 * kept. It is called after either the slptime or runtime is adjusted.
1090 * This routine will not operate correctly when slp or run times have been
1091 * adjusted to more than double their maximum.
1094 sched_interact_update(struct ksegrp *kg)
1098 sum = kg->kg_runtime + kg->kg_slptime;
1099 if (sum < SCHED_SLP_RUN_MAX)
1102 * If we have exceeded by more than 1/5th then the algorithm below
1103 * will not bring us back into range. Dividing by two here forces
1104 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1106 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1107 kg->kg_runtime /= 2;
1108 kg->kg_slptime /= 2;
1111 kg->kg_runtime = (kg->kg_runtime / 5) * 4;
1112 kg->kg_slptime = (kg->kg_slptime / 5) * 4;
1116 sched_interact_fork(struct ksegrp *kg)
1121 sum = kg->kg_runtime + kg->kg_slptime;
1122 if (sum > SCHED_SLP_RUN_FORK) {
1123 ratio = sum / SCHED_SLP_RUN_FORK;
1124 kg->kg_runtime /= ratio;
1125 kg->kg_slptime /= ratio;
1130 sched_interact_score(struct ksegrp *kg)
1134 if (kg->kg_runtime > kg->kg_slptime) {
1135 div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF);
1136 return (SCHED_INTERACT_HALF +
1137 (SCHED_INTERACT_HALF - (kg->kg_slptime / div)));
1138 } if (kg->kg_slptime > kg->kg_runtime) {
1139 div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF);
1140 return (kg->kg_runtime / div);
1144 * This can happen if slptime and runtime are 0.
1151 * Very early in the boot some setup of scheduler-specific
1152 * parts of proc0 and of soem scheduler resources needs to be done.
1160 * Set up the scheduler specific parts of proc0.
1162 proc0.p_sched = NULL; /* XXX */
1163 ksegrp0.kg_sched = &kg_sched0;
1164 thread0.td_sched = &kse0;
1165 kse0.ke_thread = &thread0;
1166 kse0.ke_state = KES_THREAD;
1167 kg_sched0.skg_concurrency = 1;
1168 kg_sched0.skg_avail_opennings = 0; /* we are already running */
1172 * This is only somewhat accurate since given many processes of the same
1173 * priority they will switch when their slices run out, which will be
1174 * at most SCHED_SLICE_MAX.
1177 sched_rr_interval(void)
1179 return (SCHED_SLICE_MAX);
1183 sched_pctcpu_update(struct kse *ke)
1186 * Adjust counters and watermark for pctcpu calc.
1188 if (ke->ke_ltick > ticks - SCHED_CPU_TICKS) {
1190 * Shift the tick count out so that the divide doesn't
1191 * round away our results.
1193 ke->ke_ticks <<= 10;
1194 ke->ke_ticks = (ke->ke_ticks / (ticks - ke->ke_ftick)) *
1196 ke->ke_ticks >>= 10;
1199 ke->ke_ltick = ticks;
1200 ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS;
1204 sched_thread_priority(struct thread *td, u_char prio)
1208 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1209 td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1210 curthread->td_proc->p_comm);
1212 mtx_assert(&sched_lock, MA_OWNED);
1213 if (td->td_priority == prio)
1215 if (TD_ON_RUNQ(td)) {
1217 * If the priority has been elevated due to priority
1218 * propagation, we may have to move ourselves to a new
1219 * queue. We still call adjustrunqueue below in case kse
1220 * needs to fix things up.
1222 if (prio < td->td_priority && ke->ke_runq != NULL &&
1223 (ke->ke_flags & KEF_ASSIGNED) == 0 &&
1224 ke->ke_runq != KSEQ_CPU(ke->ke_cpu)->ksq_curr) {
1225 runq_remove(ke->ke_runq, ke);
1226 ke->ke_runq = KSEQ_CPU(ke->ke_cpu)->ksq_curr;
1227 runq_add(ke->ke_runq, ke, 0);
1230 * Hold this kse on this cpu so that sched_prio() doesn't
1231 * cause excessive migration. We only want migration to
1232 * happen as the result of a wakeup.
1234 ke->ke_flags |= KEF_HOLD;
1235 adjustrunqueue(td, prio);
1236 ke->ke_flags &= ~KEF_HOLD;
1238 td->td_priority = prio;
1242 * Update a thread's priority when it is lent another thread's
1246 sched_lend_prio(struct thread *td, u_char prio)
1249 td->td_flags |= TDF_BORROWING;
1250 sched_thread_priority(td, prio);
1254 * Restore a thread's priority when priority propagation is
1255 * over. The prio argument is the minimum priority the thread
1256 * needs to have to satisfy other possible priority lending
1257 * requests. If the thread's regular priority is less
1258 * important than prio, the thread will keep a priority boost
1262 sched_unlend_prio(struct thread *td, u_char prio)
1266 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1267 td->td_base_pri <= PRI_MAX_TIMESHARE)
1268 base_pri = td->td_ksegrp->kg_user_pri;
1270 base_pri = td->td_base_pri;
1271 if (prio >= base_pri) {
1272 td->td_flags &= ~TDF_BORROWING;
1273 sched_thread_priority(td, base_pri);
1275 sched_lend_prio(td, prio);
1279 sched_prio(struct thread *td, u_char prio)
1283 /* First, update the base priority. */
1284 td->td_base_pri = prio;
1287 * If the thread is borrowing another thread's priority, don't
1288 * ever lower the priority.
1290 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1293 /* Change the real priority. */
1294 oldprio = td->td_priority;
1295 sched_thread_priority(td, prio);
1298 * If the thread is on a turnstile, then let the turnstile update
1301 if (TD_ON_LOCK(td) && oldprio != prio)
1302 turnstile_adjust(td, oldprio);
1306 sched_switch(struct thread *td, struct thread *newtd, int flags)
1311 mtx_assert(&sched_lock, MA_OWNED);
1316 td->td_lastcpu = td->td_oncpu;
1317 td->td_oncpu = NOCPU;
1318 td->td_flags &= ~TDF_NEEDRESCHED;
1319 td->td_owepreempt = 0;
1322 * If the KSE has been assigned it may be in the process of switching
1323 * to the new cpu. This is the case in sched_bind().
1325 if (td == PCPU_GET(idlethread)) {
1327 } else if ((ke->ke_flags & KEF_ASSIGNED) == 0) {
1328 /* We are ending our run so make our slot available again */
1329 SLOT_RELEASE(td->td_ksegrp);
1330 kseq_load_rem(ksq, ke);
1331 if (TD_IS_RUNNING(td)) {
1333 * Don't allow the thread to migrate
1334 * from a preemption.
1336 ke->ke_flags |= KEF_HOLD;
1337 setrunqueue(td, (flags & SW_PREEMPT) ?
1338 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1339 SRQ_OURSELF|SRQ_YIELDING);
1340 ke->ke_flags &= ~KEF_HOLD;
1341 } else if ((td->td_proc->p_flag & P_HADTHREADS) &&
1342 (newtd == NULL || newtd->td_ksegrp != td->td_ksegrp))
1344 * We will not be on the run queue.
1345 * So we must be sleeping or similar.
1346 * Don't use the slot if we will need it
1349 slot_fill(td->td_ksegrp);
1351 if (newtd != NULL) {
1353 * If we bring in a thread,
1354 * then account for it as if it had been added to the
1355 * run queue and then chosen.
1357 newtd->td_kse->ke_flags |= KEF_DIDRUN;
1358 newtd->td_kse->ke_runq = ksq->ksq_curr;
1359 TD_SET_RUNNING(newtd);
1360 kseq_load_add(KSEQ_SELF(), newtd->td_kse);
1362 newtd = choosethread();
1365 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1366 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1368 cpu_switch(td, newtd);
1370 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1371 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1375 sched_lock.mtx_lock = (uintptr_t)td;
1377 td->td_oncpu = PCPU_GET(cpuid);
1381 sched_nice(struct proc *p, int nice)
1388 PROC_LOCK_ASSERT(p, MA_OWNED);
1389 mtx_assert(&sched_lock, MA_OWNED);
1391 * We need to adjust the nice counts for running KSEs.
1393 FOREACH_KSEGRP_IN_PROC(p, kg) {
1394 if (kg->kg_pri_class == PRI_TIMESHARE) {
1395 FOREACH_THREAD_IN_GROUP(kg, td) {
1397 if (ke->ke_runq == NULL)
1399 kseq = KSEQ_CPU(ke->ke_cpu);
1400 kseq_nice_rem(kseq, p->p_nice);
1401 kseq_nice_add(kseq, nice);
1406 FOREACH_KSEGRP_IN_PROC(p, kg) {
1408 FOREACH_THREAD_IN_GROUP(kg, td)
1409 td->td_flags |= TDF_NEEDRESCHED;
1414 sched_sleep(struct thread *td)
1416 mtx_assert(&sched_lock, MA_OWNED);
1418 td->td_slptime = ticks;
1422 sched_wakeup(struct thread *td)
1424 mtx_assert(&sched_lock, MA_OWNED);
1427 * Let the kseg know how long we slept for. This is because process
1428 * interactivity behavior is modeled in the kseg.
1430 if (td->td_slptime) {
1435 hzticks = (ticks - td->td_slptime) << 10;
1436 if (hzticks >= SCHED_SLP_RUN_MAX) {
1437 kg->kg_slptime = SCHED_SLP_RUN_MAX;
1440 kg->kg_slptime += hzticks;
1441 sched_interact_update(kg);
1444 sched_slice(td->td_kse);
1447 setrunqueue(td, SRQ_BORING);
1451 * Penalize the parent for creating a new child and initialize the child's
1455 sched_fork(struct thread *td, struct thread *childtd)
1458 mtx_assert(&sched_lock, MA_OWNED);
1460 sched_fork_ksegrp(td, childtd->td_ksegrp);
1461 sched_fork_thread(td, childtd);
1465 sched_fork_ksegrp(struct thread *td, struct ksegrp *child)
1467 struct ksegrp *kg = td->td_ksegrp;
1468 mtx_assert(&sched_lock, MA_OWNED);
1470 child->kg_slptime = kg->kg_slptime;
1471 child->kg_runtime = kg->kg_runtime;
1472 child->kg_user_pri = kg->kg_user_pri;
1473 sched_interact_fork(child);
1474 kg->kg_runtime += tickincr << 10;
1475 sched_interact_update(kg);
1479 sched_fork_thread(struct thread *td, struct thread *child)
1484 sched_newthread(child);
1486 ke2 = child->td_kse;
1487 ke2->ke_slice = 1; /* Attempt to quickly learn interactivity. */
1488 ke2->ke_cpu = ke->ke_cpu;
1489 ke2->ke_runq = NULL;
1491 /* Grab our parents cpu estimation information. */
1492 ke2->ke_ticks = ke->ke_ticks;
1493 ke2->ke_ltick = ke->ke_ltick;
1494 ke2->ke_ftick = ke->ke_ftick;
1498 sched_class(struct ksegrp *kg, int class)
1506 mtx_assert(&sched_lock, MA_OWNED);
1507 if (kg->kg_pri_class == class)
1510 nclass = PRI_BASE(class);
1511 oclass = PRI_BASE(kg->kg_pri_class);
1512 FOREACH_THREAD_IN_GROUP(kg, td) {
1514 if ((ke->ke_state != KES_ONRUNQ &&
1515 ke->ke_state != KES_THREAD) || ke->ke_runq == NULL)
1517 kseq = KSEQ_CPU(ke->ke_cpu);
1521 * On SMP if we're on the RUNQ we must adjust the transferable
1522 * count because could be changing to or from an interrupt
1525 if (ke->ke_state == KES_ONRUNQ) {
1526 if (KSE_CAN_MIGRATE(ke)) {
1527 kseq->ksq_transferable--;
1528 kseq->ksq_group->ksg_transferable--;
1530 if (KSE_CAN_MIGRATE(ke)) {
1531 kseq->ksq_transferable++;
1532 kseq->ksq_group->ksg_transferable++;
1536 if (oclass == PRI_TIMESHARE) {
1537 kseq->ksq_load_timeshare--;
1538 kseq_nice_rem(kseq, kg->kg_proc->p_nice);
1540 if (nclass == PRI_TIMESHARE) {
1541 kseq->ksq_load_timeshare++;
1542 kseq_nice_add(kseq, kg->kg_proc->p_nice);
1546 kg->kg_pri_class = class;
1550 * Return some of the child's priority and interactivity to the parent.
1553 sched_exit(struct proc *p, struct thread *childtd)
1555 mtx_assert(&sched_lock, MA_OWNED);
1556 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), childtd);
1557 sched_exit_thread(NULL, childtd);
1561 sched_exit_ksegrp(struct ksegrp *kg, struct thread *td)
1563 /* kg->kg_slptime += td->td_ksegrp->kg_slptime; */
1564 kg->kg_runtime += td->td_ksegrp->kg_runtime;
1565 sched_interact_update(kg);
1569 sched_exit_thread(struct thread *td, struct thread *childtd)
1571 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1572 childtd, childtd->td_proc->p_comm, childtd->td_priority);
1573 kseq_load_rem(KSEQ_CPU(childtd->td_kse->ke_cpu), childtd->td_kse);
1577 sched_clock(struct thread *td)
1583 mtx_assert(&sched_lock, MA_OWNED);
1586 if (ticks >= bal_tick)
1588 if (ticks >= gbal_tick && balance_groups)
1589 sched_balance_groups();
1591 * We could have been assigned a non real-time thread without an
1594 if (kseq->ksq_assigned)
1595 kseq_assign(kseq); /* Potentially sets NEEDRESCHED */
1598 * sched_setup() apparently happens prior to stathz being set. We
1599 * need to resolve the timers earlier in the boot so we can avoid
1600 * calculating this here.
1602 if (realstathz == 0) {
1603 realstathz = stathz ? stathz : hz;
1604 tickincr = hz / realstathz;
1606 * XXX This does not work for values of stathz that are much
1616 /* Adjust ticks for pctcpu */
1618 ke->ke_ltick = ticks;
1620 /* Go up to one second beyond our max and then trim back down */
1621 if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick)
1622 sched_pctcpu_update(ke);
1624 if (td->td_flags & TDF_IDLETD)
1627 * We only do slicing code for TIMESHARE ksegrps.
1629 if (kg->kg_pri_class != PRI_TIMESHARE)
1632 * We used a tick charge it to the ksegrp so that we can compute our
1635 kg->kg_runtime += tickincr << 10;
1636 sched_interact_update(kg);
1639 * We used up one time slice.
1641 if (--ke->ke_slice > 0)
1644 * We're out of time, recompute priorities and requeue.
1646 kseq_load_rem(kseq, ke);
1649 if (SCHED_CURR(kg, ke))
1650 ke->ke_runq = kseq->ksq_curr;
1652 ke->ke_runq = kseq->ksq_next;
1653 kseq_load_add(kseq, ke);
1654 td->td_flags |= TDF_NEEDRESCHED;
1658 sched_runnable(void)
1667 if (kseq->ksq_assigned) {
1668 mtx_lock_spin(&sched_lock);
1670 mtx_unlock_spin(&sched_lock);
1673 if ((curthread->td_flags & TDF_IDLETD) != 0) {
1674 if (kseq->ksq_load > 0)
1677 if (kseq->ksq_load - 1 > 0)
1685 sched_userret(struct thread *td)
1689 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1690 ("thread with borrowed priority returning to userland"));
1692 if (td->td_priority != kg->kg_user_pri) {
1693 mtx_lock_spin(&sched_lock);
1694 td->td_priority = kg->kg_user_pri;
1695 td->td_base_pri = kg->kg_user_pri;
1696 mtx_unlock_spin(&sched_lock);
1706 mtx_assert(&sched_lock, MA_OWNED);
1710 if (kseq->ksq_assigned)
1713 ke = kseq_choose(kseq);
1716 if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE)
1717 if (kseq_idled(kseq) == 0)
1720 kseq_runq_rem(kseq, ke);
1721 ke->ke_state = KES_THREAD;
1725 if (kseq_idled(kseq) == 0)
1732 sched_add(struct thread *td, int flags)
1741 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1742 td, td->td_proc->p_comm, td->td_priority, curthread,
1743 curthread->td_proc->p_comm);
1744 mtx_assert(&sched_lock, MA_OWNED);
1748 preemptive = !(flags & SRQ_YIELDING);
1749 class = PRI_BASE(kg->kg_pri_class);
1751 if ((ke->ke_flags & KEF_INTERNAL) == 0)
1752 SLOT_USE(td->td_ksegrp);
1753 ke->ke_flags &= ~KEF_INTERNAL;
1755 if (ke->ke_flags & KEF_ASSIGNED) {
1756 if (ke->ke_flags & KEF_REMOVED)
1757 ke->ke_flags &= ~KEF_REMOVED;
1760 canmigrate = KSE_CAN_MIGRATE(ke);
1762 KASSERT(ke->ke_state != KES_ONRUNQ,
1763 ("sched_add: kse %p (%s) already in run queue", ke,
1764 ke->ke_proc->p_comm));
1765 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
1766 ("sched_add: process swapped out"));
1767 KASSERT(ke->ke_runq == NULL,
1768 ("sched_add: KSE %p is still assigned to a run queue", ke));
1772 ke->ke_runq = kseq->ksq_curr;
1773 ke->ke_slice = SCHED_SLICE_MAX;
1775 ke->ke_cpu = PCPU_GET(cpuid);
1778 if (SCHED_CURR(kg, ke))
1779 ke->ke_runq = kseq->ksq_curr;
1781 ke->ke_runq = kseq->ksq_next;
1785 * This is for priority prop.
1787 if (ke->ke_thread->td_priority < PRI_MIN_IDLE)
1788 ke->ke_runq = kseq->ksq_curr;
1790 ke->ke_runq = &kseq->ksq_idle;
1791 ke->ke_slice = SCHED_SLICE_MIN;
1794 panic("Unknown pri class.");
1799 * Don't migrate running threads here. Force the long term balancer
1802 if (ke->ke_flags & KEF_HOLD) {
1803 ke->ke_flags &= ~KEF_HOLD;
1807 * If this thread is pinned or bound, notify the target cpu.
1809 if (!canmigrate && ke->ke_cpu != PCPU_GET(cpuid) ) {
1811 kseq_notify(ke, ke->ke_cpu);
1815 * If we had been idle, clear our bit in the group and potentially
1816 * the global bitmap. If not, see if we should transfer this thread.
1818 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1819 (kseq->ksq_group->ksg_idlemask & PCPU_GET(cpumask)) != 0) {
1821 * Check to see if our group is unidling, and if so, remove it
1822 * from the global idle mask.
1824 if (kseq->ksq_group->ksg_idlemask ==
1825 kseq->ksq_group->ksg_cpumask)
1826 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
1828 * Now remove ourselves from the group specific idle mask.
1830 kseq->ksq_group->ksg_idlemask &= ~PCPU_GET(cpumask);
1831 } else if (canmigrate && kseq->ksq_load > 1 && class != PRI_ITHD)
1832 if (kseq_transfer(kseq, ke, class))
1834 ke->ke_cpu = PCPU_GET(cpuid);
1836 if (td->td_priority < curthread->td_priority &&
1837 ke->ke_runq == kseq->ksq_curr)
1838 curthread->td_flags |= TDF_NEEDRESCHED;
1839 if (preemptive && maybe_preempt(td))
1841 ke->ke_state = KES_ONRUNQ;
1843 kseq_runq_add(kseq, ke, flags);
1844 kseq_load_add(kseq, ke);
1848 sched_rem(struct thread *td)
1853 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1854 td, td->td_proc->p_comm, td->td_priority, curthread,
1855 curthread->td_proc->p_comm);
1856 mtx_assert(&sched_lock, MA_OWNED);
1858 SLOT_RELEASE(td->td_ksegrp);
1859 if (ke->ke_flags & KEF_ASSIGNED) {
1860 ke->ke_flags |= KEF_REMOVED;
1863 KASSERT((ke->ke_state == KES_ONRUNQ),
1864 ("sched_rem: KSE not on run queue"));
1866 ke->ke_state = KES_THREAD;
1867 kseq = KSEQ_CPU(ke->ke_cpu);
1868 kseq_runq_rem(kseq, ke);
1869 kseq_load_rem(kseq, ke);
1873 sched_pctcpu(struct thread *td)
1883 mtx_lock_spin(&sched_lock);
1888 * Don't update more frequently than twice a second. Allowing
1889 * this causes the cpu usage to decay away too quickly due to
1892 if (ke->ke_ftick + SCHED_CPU_TICKS < ke->ke_ltick ||
1893 ke->ke_ltick < (ticks - (hz / 2)))
1894 sched_pctcpu_update(ke);
1895 /* How many rtick per second ? */
1896 rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1897 pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1900 ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick;
1901 mtx_unlock_spin(&sched_lock);
1907 sched_bind(struct thread *td, int cpu)
1911 mtx_assert(&sched_lock, MA_OWNED);
1913 ke->ke_flags |= KEF_BOUND;
1915 if (PCPU_GET(cpuid) == cpu)
1917 /* sched_rem without the runq_remove */
1918 ke->ke_state = KES_THREAD;
1919 kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1920 kseq_notify(ke, cpu);
1921 /* When we return from mi_switch we'll be on the correct cpu. */
1922 mi_switch(SW_VOL, NULL);
1927 sched_unbind(struct thread *td)
1929 mtx_assert(&sched_lock, MA_OWNED);
1930 td->td_kse->ke_flags &= ~KEF_BOUND;
1934 sched_is_bound(struct thread *td)
1936 mtx_assert(&sched_lock, MA_OWNED);
1937 return (td->td_kse->ke_flags & KEF_BOUND);
1948 for (i = 0; i <= ksg_maxid; i++)
1949 total += KSEQ_GROUP(i)->ksg_load;
1952 return (KSEQ_SELF()->ksq_sysload);
1957 sched_sizeof_ksegrp(void)
1959 return (sizeof(struct ksegrp) + sizeof(struct kg_sched));
1963 sched_sizeof_proc(void)
1965 return (sizeof(struct proc));
1969 sched_sizeof_thread(void)
1971 return (sizeof(struct thread) + sizeof(struct td_sched));
1973 #define KERN_SWITCH_INCLUDE 1
1974 #include "kern/kern_switch.c"