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_hwpmc_hooks.h"
31 #include "opt_sched.h"
35 #include <sys/param.h>
36 #include <sys/systm.h>
38 #include <sys/kernel.h>
41 #include <sys/mutex.h>
43 #include <sys/resource.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sched.h>
48 #include <sys/sysctl.h>
49 #include <sys/sysproto.h>
50 #include <sys/turnstile.h>
51 #include <sys/vmmeter.h>
54 #include <sys/ktrace.h>
58 #include <sys/pmckern.h>
61 #include <machine/cpu.h>
62 #include <machine/smp.h>
64 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
65 /* XXX This is bogus compatability crap for ps */
66 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
67 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
69 static void sched_setup(void *dummy);
70 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
72 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
74 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0,
77 static int slice_min = 1;
78 SYSCTL_INT(_kern_sched, OID_AUTO, slice_min, CTLFLAG_RW, &slice_min, 0, "");
80 static int slice_max = 10;
81 SYSCTL_INT(_kern_sched, OID_AUTO, slice_max, CTLFLAG_RW, &slice_max, 0, "");
87 * The following datastructures are allocated within their parent structure
88 * but are scheduler specific.
91 * The schedulable entity that can be given a context to run. A process may
92 * have several of these.
95 TAILQ_ENTRY(kse) ke_procq; /* (j/z) Run queue. */
96 int ke_flags; /* (j) KEF_* flags. */
97 struct thread *ke_thread; /* (*) Active associated thread. */
98 fixpt_t ke_pctcpu; /* (j) %cpu during p_swtime. */
99 char ke_rqindex; /* (j) Run queue index. */
101 KES_THREAD = 0x0, /* slaved to thread state */
103 } ke_state; /* (j) thread sched specific status. */
106 struct runq *ke_runq;
107 u_char ke_cpu; /* CPU that we have affinity for. */
108 /* The following variables are only used for pctcpu calculation */
109 int ke_ltick; /* Last tick that we were running on */
110 int ke_ftick; /* First tick that we were running on */
111 int ke_ticks; /* Tick count */
114 #define td_kse td_sched
115 #define td_slptime td_kse->ke_slptime
116 #define ke_proc ke_thread->td_proc
117 #define ke_ksegrp ke_thread->td_ksegrp
118 #define ke_assign ke_procq.tqe_next
119 /* flags kept in ke_flags */
120 #define KEF_ASSIGNED 0x0001 /* Thread is being migrated. */
121 #define KEF_BOUND 0x0002 /* Thread can not migrate. */
122 #define KEF_XFERABLE 0x0004 /* Thread was added as transferable. */
123 #define KEF_HOLD 0x0008 /* Thread is temporarily bound. */
124 #define KEF_REMOVED 0x0010 /* Thread was removed while ASSIGNED */
125 #define KEF_INTERNAL 0x0020 /* Thread added due to migration. */
126 #define KEF_DIDRUN 0x02000 /* Thread actually ran. */
127 #define KEF_EXIT 0x04000 /* Thread is being killed. */
130 struct thread *skg_last_assigned; /* (j) Last thread assigned to */
131 /* the system scheduler */
132 int skg_slptime; /* Number of ticks we vol. slept */
133 int skg_runtime; /* Number of ticks we were running */
134 int skg_avail_opennings; /* (j) Num unfilled slots in group.*/
135 int skg_concurrency; /* (j) Num threads requested in group.*/
137 #define kg_last_assigned kg_sched->skg_last_assigned
138 #define kg_avail_opennings kg_sched->skg_avail_opennings
139 #define kg_concurrency kg_sched->skg_concurrency
140 #define kg_runtime kg_sched->skg_runtime
141 #define kg_slptime kg_sched->skg_slptime
143 #define SLOT_RELEASE(kg) (kg)->kg_avail_opennings++
144 #define SLOT_USE(kg) (kg)->kg_avail_opennings--
146 static struct kse kse0;
147 static struct kg_sched kg_sched0;
150 * The priority is primarily determined by the interactivity score. Thus, we
151 * give lower(better) priorities to kse groups that use less CPU. The nice
152 * value is then directly added to this to allow nice to have some effect
155 * PRI_RANGE: Total priority range for timeshare threads.
156 * PRI_NRESV: Number of nice values.
157 * PRI_BASE: The start of the dynamic range.
159 #define SCHED_PRI_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
160 #define SCHED_PRI_NRESV ((PRIO_MAX - PRIO_MIN) + 1)
161 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
162 #define SCHED_PRI_BASE (PRI_MIN_TIMESHARE)
163 #define SCHED_PRI_INTERACT(score) \
164 ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
167 * These determine the interactivity of a process.
169 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
170 * before throttling back.
171 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
172 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
173 * INTERACT_THRESH: Threshhold for placement on the current runq.
175 #define SCHED_SLP_RUN_MAX ((hz * 5) << 10)
176 #define SCHED_SLP_RUN_FORK ((hz / 2) << 10)
177 #define SCHED_INTERACT_MAX (100)
178 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
179 #define SCHED_INTERACT_THRESH (30)
182 * These parameters and macros determine the size of the time slice that is
183 * granted to each thread.
185 * SLICE_MIN: Minimum time slice granted, in units of ticks.
186 * SLICE_MAX: Maximum time slice granted.
187 * SLICE_RANGE: Range of available time slices scaled by hz.
188 * SLICE_SCALE: The number slices granted per val in the range of [0, max].
189 * SLICE_NICE: Determine the amount of slice granted to a scaled nice.
190 * SLICE_NTHRESH: The nice cutoff point for slice assignment.
192 #define SCHED_SLICE_MIN (slice_min)
193 #define SCHED_SLICE_MAX (slice_max)
194 #define SCHED_SLICE_INTERACTIVE (slice_max)
195 #define SCHED_SLICE_NTHRESH (SCHED_PRI_NHALF - 1)
196 #define SCHED_SLICE_RANGE (SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
197 #define SCHED_SLICE_SCALE(val, max) (((val) * SCHED_SLICE_RANGE) / (max))
198 #define SCHED_SLICE_NICE(nice) \
199 (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
202 * This macro determines whether or not the thread belongs on the current or
205 #define SCHED_INTERACTIVE(kg) \
206 (sched_interact_score(kg) < SCHED_INTERACT_THRESH)
207 #define SCHED_CURR(kg, ke) \
208 ((ke->ke_thread->td_flags & TDF_BORROWING) || SCHED_INTERACTIVE(kg))
211 * Cpu percentage computation macros and defines.
213 * SCHED_CPU_TIME: Number of seconds to average the cpu usage across.
214 * SCHED_CPU_TICKS: Number of hz ticks to average the cpu usage across.
217 #define SCHED_CPU_TIME 10
218 #define SCHED_CPU_TICKS (hz * SCHED_CPU_TIME)
221 * kseq - per processor runqs and statistics.
224 struct runq ksq_idle; /* Queue of IDLE threads. */
225 struct runq ksq_timeshare[2]; /* Run queues for !IDLE. */
226 struct runq *ksq_next; /* Next timeshare queue. */
227 struct runq *ksq_curr; /* Current queue. */
228 int ksq_load_timeshare; /* Load for timeshare. */
229 int ksq_load; /* Aggregate load. */
230 short ksq_nice[SCHED_PRI_NRESV]; /* KSEs in each nice bin. */
231 short ksq_nicemin; /* Least nice. */
233 int ksq_transferable;
234 LIST_ENTRY(kseq) ksq_siblings; /* Next in kseq group. */
235 struct kseq_group *ksq_group; /* Our processor group. */
236 volatile struct kse *ksq_assigned; /* assigned by another CPU. */
238 int ksq_sysload; /* For loadavg, !ITHD load. */
244 * kseq groups are groups of processors which can cheaply share threads. When
245 * one processor in the group goes idle it will check the runqs of the other
246 * processors in its group prior to halting and waiting for an interrupt.
247 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
248 * In a numa environment we'd want an idle bitmap per group and a two tiered
252 int ksg_cpus; /* Count of CPUs in this kseq group. */
253 cpumask_t ksg_cpumask; /* Mask of cpus in this group. */
254 cpumask_t ksg_idlemask; /* Idle cpus in this group. */
255 cpumask_t ksg_mask; /* Bit mask for first cpu. */
256 int ksg_load; /* Total load of this group. */
257 int ksg_transferable; /* Transferable load of this group. */
258 LIST_HEAD(, kseq) ksg_members; /* Linked list of all members. */
263 * One kse queue per processor.
266 static cpumask_t kseq_idle;
267 static int ksg_maxid;
268 static struct kseq kseq_cpu[MAXCPU];
269 static struct kseq_group kseq_groups[MAXCPU];
271 static int gbal_tick;
272 static int balance_groups;
274 #define KSEQ_SELF() (&kseq_cpu[PCPU_GET(cpuid)])
275 #define KSEQ_CPU(x) (&kseq_cpu[(x)])
276 #define KSEQ_ID(x) ((x) - kseq_cpu)
277 #define KSEQ_GROUP(x) (&kseq_groups[(x)])
279 static struct kseq kseq_cpu;
281 #define KSEQ_SELF() (&kseq_cpu)
282 #define KSEQ_CPU(x) (&kseq_cpu)
285 static void slot_fill(struct ksegrp *);
286 static struct kse *sched_choose(void); /* XXX Should be thread * */
287 static void sched_slice(struct kse *);
288 static void sched_priority(struct ksegrp *);
289 static void sched_thread_priority(struct thread *, u_char);
290 static int sched_interact_score(struct ksegrp *);
291 static void sched_interact_update(struct ksegrp *);
292 static void sched_interact_fork(struct ksegrp *);
293 static void sched_pctcpu_update(struct kse *);
295 /* Operations on per processor queues */
296 static struct kse * kseq_choose(struct kseq *);
297 static void kseq_setup(struct kseq *);
298 static void kseq_load_add(struct kseq *, struct kse *);
299 static void kseq_load_rem(struct kseq *, struct kse *);
300 static __inline void kseq_runq_add(struct kseq *, struct kse *, int);
301 static __inline void kseq_runq_rem(struct kseq *, struct kse *);
302 static void kseq_nice_add(struct kseq *, int);
303 static void kseq_nice_rem(struct kseq *, int);
304 void kseq_print(int cpu);
306 static int kseq_transfer(struct kseq *, struct kse *, int);
307 static struct kse *runq_steal(struct runq *);
308 static void sched_balance(void);
309 static void sched_balance_groups(void);
310 static void sched_balance_group(struct kseq_group *);
311 static void sched_balance_pair(struct kseq *, struct kseq *);
312 static void kseq_move(struct kseq *, int);
313 static int kseq_idled(struct kseq *);
314 static void kseq_notify(struct kse *, int);
315 static void kseq_assign(struct kseq *);
316 static struct kse *kseq_steal(struct kseq *, int);
317 #define KSE_CAN_MIGRATE(ke) \
318 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
327 kseq = KSEQ_CPU(cpu);
330 printf("\tload: %d\n", kseq->ksq_load);
331 printf("\tload TIMESHARE: %d\n", kseq->ksq_load_timeshare);
333 printf("\tload transferable: %d\n", kseq->ksq_transferable);
335 printf("\tnicemin:\t%d\n", kseq->ksq_nicemin);
336 printf("\tnice counts:\n");
337 for (i = 0; i < SCHED_PRI_NRESV; i++)
338 if (kseq->ksq_nice[i])
339 printf("\t\t%d = %d\n",
340 i - SCHED_PRI_NHALF, kseq->ksq_nice[i]);
344 kseq_runq_add(struct kseq *kseq, struct kse *ke, int flags)
347 if (KSE_CAN_MIGRATE(ke)) {
348 kseq->ksq_transferable++;
349 kseq->ksq_group->ksg_transferable++;
350 ke->ke_flags |= KEF_XFERABLE;
353 runq_add(ke->ke_runq, ke, flags);
357 kseq_runq_rem(struct kseq *kseq, struct kse *ke)
360 if (ke->ke_flags & KEF_XFERABLE) {
361 kseq->ksq_transferable--;
362 kseq->ksq_group->ksg_transferable--;
363 ke->ke_flags &= ~KEF_XFERABLE;
366 runq_remove(ke->ke_runq, ke);
370 kseq_load_add(struct kseq *kseq, struct kse *ke)
373 mtx_assert(&sched_lock, MA_OWNED);
374 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
375 if (class == PRI_TIMESHARE)
376 kseq->ksq_load_timeshare++;
378 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
379 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
381 kseq->ksq_group->ksg_load++;
385 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
386 kseq_nice_add(kseq, ke->ke_proc->p_nice);
390 kseq_load_rem(struct kseq *kseq, struct kse *ke)
393 mtx_assert(&sched_lock, MA_OWNED);
394 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
395 if (class == PRI_TIMESHARE)
396 kseq->ksq_load_timeshare--;
397 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
399 kseq->ksq_group->ksg_load--;
404 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
406 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
407 kseq_nice_rem(kseq, ke->ke_proc->p_nice);
411 kseq_nice_add(struct kseq *kseq, int nice)
413 mtx_assert(&sched_lock, MA_OWNED);
414 /* Normalize to zero. */
415 kseq->ksq_nice[nice + SCHED_PRI_NHALF]++;
416 if (nice < kseq->ksq_nicemin || kseq->ksq_load_timeshare == 1)
417 kseq->ksq_nicemin = nice;
421 kseq_nice_rem(struct kseq *kseq, int nice)
425 mtx_assert(&sched_lock, MA_OWNED);
426 /* Normalize to zero. */
427 n = nice + SCHED_PRI_NHALF;
429 KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count."));
432 * If this wasn't the smallest nice value or there are more in
433 * this bucket we can just return. Otherwise we have to recalculate
436 if (nice != kseq->ksq_nicemin ||
437 kseq->ksq_nice[n] != 0 ||
438 kseq->ksq_load_timeshare == 0)
441 for (; n < SCHED_PRI_NRESV; n++)
442 if (kseq->ksq_nice[n]) {
443 kseq->ksq_nicemin = n - SCHED_PRI_NHALF;
450 * sched_balance is a simple CPU load balancing algorithm. It operates by
451 * finding the least loaded and most loaded cpu and equalizing their load
452 * by migrating some processes.
454 * Dealing only with two CPUs at a time has two advantages. Firstly, most
455 * installations will only have 2 cpus. Secondly, load balancing too much at
456 * once can have an unpleasant effect on the system. The scheduler rarely has
457 * enough information to make perfect decisions. So this algorithm chooses
458 * algorithm simplicity and more gradual effects on load in larger systems.
460 * It could be improved by considering the priorities and slices assigned to
461 * each task prior to balancing them. There are many pathological cases with
462 * any approach and so the semi random algorithm below may work as well as any.
468 struct kseq_group *high;
469 struct kseq_group *low;
470 struct kseq_group *ksg;
474 bal_tick = ticks + (random() % (hz * 2));
475 if (smp_started == 0)
478 i = random() % (ksg_maxid + 1);
479 for (cnt = 0; cnt <= ksg_maxid; cnt++) {
482 * Find the CPU with the highest load that has some
483 * threads to transfer.
485 if ((high == NULL || ksg->ksg_load > high->ksg_load)
486 && ksg->ksg_transferable)
488 if (low == NULL || ksg->ksg_load < low->ksg_load)
493 if (low != NULL && high != NULL && high != low)
494 sched_balance_pair(LIST_FIRST(&high->ksg_members),
495 LIST_FIRST(&low->ksg_members));
499 sched_balance_groups(void)
503 gbal_tick = ticks + (random() % (hz * 2));
504 mtx_assert(&sched_lock, MA_OWNED);
506 for (i = 0; i <= ksg_maxid; i++)
507 sched_balance_group(KSEQ_GROUP(i));
511 sched_balance_group(struct kseq_group *ksg)
518 if (ksg->ksg_transferable == 0)
522 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
523 load = kseq->ksq_load;
524 if (high == NULL || load > high->ksq_load)
526 if (low == NULL || load < low->ksq_load)
529 if (high != NULL && low != NULL && high != low)
530 sched_balance_pair(high, low);
534 sched_balance_pair(struct kseq *high, struct kseq *low)
544 * If we're transfering within a group we have to use this specific
545 * kseq's transferable count, otherwise we can steal from other members
548 if (high->ksq_group == low->ksq_group) {
549 transferable = high->ksq_transferable;
550 high_load = high->ksq_load;
551 low_load = low->ksq_load;
553 transferable = high->ksq_group->ksg_transferable;
554 high_load = high->ksq_group->ksg_load;
555 low_load = low->ksq_group->ksg_load;
557 if (transferable == 0)
560 * Determine what the imbalance is and then adjust that to how many
561 * kses we actually have to give up (transferable).
563 diff = high_load - low_load;
567 move = min(move, transferable);
568 for (i = 0; i < move; i++)
569 kseq_move(high, KSEQ_ID(low));
574 kseq_move(struct kseq *from, int cpu)
582 ke = kseq_steal(kseq, 1);
584 struct kseq_group *ksg;
586 ksg = kseq->ksq_group;
587 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
588 if (kseq == from || kseq->ksq_transferable == 0)
590 ke = kseq_steal(kseq, 1);
594 panic("kseq_move: No KSEs available with a "
595 "transferable count of %d\n",
596 ksg->ksg_transferable);
600 ke->ke_state = KES_THREAD;
601 kseq_runq_rem(kseq, ke);
602 kseq_load_rem(kseq, ke);
603 kseq_notify(ke, cpu);
607 kseq_idled(struct kseq *kseq)
609 struct kseq_group *ksg;
613 ksg = kseq->ksq_group;
615 * If we're in a cpu group, try and steal kses from another cpu in
616 * the group before idling.
618 if (ksg->ksg_cpus > 1 && ksg->ksg_transferable) {
619 LIST_FOREACH(steal, &ksg->ksg_members, ksq_siblings) {
620 if (steal == kseq || steal->ksq_transferable == 0)
622 ke = kseq_steal(steal, 0);
625 ke->ke_state = KES_THREAD;
626 kseq_runq_rem(steal, ke);
627 kseq_load_rem(steal, ke);
628 ke->ke_cpu = PCPU_GET(cpuid);
629 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
630 sched_add(ke->ke_thread, SRQ_YIELDING);
635 * We only set the idled bit when all of the cpus in the group are
636 * idle. Otherwise we could get into a situation where a KSE bounces
637 * back and forth between two idle cores on seperate physical CPUs.
639 ksg->ksg_idlemask |= PCPU_GET(cpumask);
640 if (ksg->ksg_idlemask != ksg->ksg_cpumask)
642 atomic_set_int(&kseq_idle, ksg->ksg_mask);
647 kseq_assign(struct kseq *kseq)
653 *(volatile struct kse **)&ke = kseq->ksq_assigned;
654 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke, NULL));
655 for (; ke != NULL; ke = nke) {
657 kseq->ksq_group->ksg_load--;
659 ke->ke_flags &= ~KEF_ASSIGNED;
660 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
661 sched_add(ke->ke_thread, SRQ_YIELDING);
666 kseq_notify(struct kse *ke, int cpu)
674 kseq = KSEQ_CPU(cpu);
676 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
677 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
678 (kseq_idle & kseq->ksq_group->ksg_mask))
679 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
680 kseq->ksq_group->ksg_load++;
683 ke->ke_flags |= KEF_ASSIGNED;
684 prio = ke->ke_thread->td_priority;
687 * Place a KSE on another cpu's queue and force a resched.
690 *(volatile struct kse **)&ke->ke_assign = kseq->ksq_assigned;
691 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke->ke_assign, ke));
693 * Without sched_lock we could lose a race where we set NEEDRESCHED
694 * on a thread that is switched out before the IPI is delivered. This
695 * would lead us to miss the resched. This will be a problem once
696 * sched_lock is pushed down.
698 pcpu = pcpu_find(cpu);
699 td = pcpu->pc_curthread;
700 if (ke->ke_thread->td_priority < td->td_priority ||
701 td == pcpu->pc_idlethread) {
702 td->td_flags |= TDF_NEEDRESCHED;
703 ipi_selected(1 << cpu, IPI_AST);
708 runq_steal(struct runq *rq)
716 mtx_assert(&sched_lock, MA_OWNED);
717 rqb = &rq->rq_status;
718 for (word = 0; word < RQB_LEN; word++) {
719 if (rqb->rqb_bits[word] == 0)
721 for (bit = 0; bit < RQB_BPW; bit++) {
722 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
724 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
725 TAILQ_FOREACH(ke, rqh, ke_procq) {
726 if (KSE_CAN_MIGRATE(ke))
735 kseq_steal(struct kseq *kseq, int stealidle)
740 * Steal from next first to try to get a non-interactive task that
741 * may not have run for a while.
743 if ((ke = runq_steal(kseq->ksq_next)) != NULL)
745 if ((ke = runq_steal(kseq->ksq_curr)) != NULL)
748 return (runq_steal(&kseq->ksq_idle));
753 kseq_transfer(struct kseq *kseq, struct kse *ke, int class)
755 struct kseq_group *nksg;
756 struct kseq_group *ksg;
761 if (smp_started == 0)
765 * If our load exceeds a certain threshold we should attempt to
766 * reassign this thread. The first candidate is the cpu that
767 * originally ran the thread. If it is idle, assign it there,
768 * otherwise, pick an idle cpu.
770 * The threshold at which we start to reassign kses has a large impact
771 * on the overall performance of the system. Tuned too high and
772 * some CPUs may idle. Too low and there will be excess migration
773 * and context switches.
775 old = KSEQ_CPU(ke->ke_cpu);
776 nksg = old->ksq_group;
777 ksg = kseq->ksq_group;
779 if (kseq_idle & nksg->ksg_mask) {
780 cpu = ffs(nksg->ksg_idlemask);
783 "kseq_transfer: %p found old cpu %X "
784 "in idlemask.", ke, cpu);
789 * Multiple cpus could find this bit simultaneously
790 * but the race shouldn't be terrible.
792 cpu = ffs(kseq_idle);
794 CTR2(KTR_SCHED, "kseq_transfer: %p found %X "
795 "in idlemask.", ke, cpu);
801 if (old->ksq_load < kseq->ksq_load) {
802 cpu = ke->ke_cpu + 1;
803 CTR2(KTR_SCHED, "kseq_transfer: %p old cpu %X "
804 "load less than ours.", ke, cpu);
808 * No new CPU was found, look for one with less load.
810 for (idx = 0; idx <= ksg_maxid; idx++) {
811 nksg = KSEQ_GROUP(idx);
812 if (nksg->ksg_load /*+ (nksg->ksg_cpus * 2)*/ < ksg->ksg_load) {
813 cpu = ffs(nksg->ksg_cpumask);
814 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X load less "
815 "than ours.", ke, cpu);
821 * If another cpu in this group has idled, assign a thread over
822 * to them after checking to see if there are idled groups.
824 if (ksg->ksg_idlemask) {
825 cpu = ffs(ksg->ksg_idlemask);
827 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X idle in "
835 * Now that we've found an idle CPU, migrate the thread.
839 kseq_notify(ke, cpu);
847 * Pick the highest priority task we have and return it.
851 kseq_choose(struct kseq *kseq)
857 mtx_assert(&sched_lock, MA_OWNED);
861 ke = runq_choose(kseq->ksq_curr);
864 * We already swapped once and didn't get anywhere.
868 swap = kseq->ksq_curr;
869 kseq->ksq_curr = kseq->ksq_next;
870 kseq->ksq_next = swap;
874 * If we encounter a slice of 0 the kse is in a
875 * TIMESHARE kse group and its nice was too far out
876 * of the range that receives slices.
878 nice = ke->ke_proc->p_nice + (0 - kseq->ksq_nicemin);
879 if (ke->ke_slice == 0 || (nice > SCHED_SLICE_NTHRESH &&
880 ke->ke_proc->p_nice != 0)) {
881 runq_remove(ke->ke_runq, ke);
883 ke->ke_runq = kseq->ksq_next;
884 runq_add(ke->ke_runq, ke, 0);
890 return (runq_choose(&kseq->ksq_idle));
894 kseq_setup(struct kseq *kseq)
896 runq_init(&kseq->ksq_timeshare[0]);
897 runq_init(&kseq->ksq_timeshare[1]);
898 runq_init(&kseq->ksq_idle);
899 kseq->ksq_curr = &kseq->ksq_timeshare[0];
900 kseq->ksq_next = &kseq->ksq_timeshare[1];
902 kseq->ksq_load_timeshare = 0;
906 sched_setup(void *dummy)
912 slice_min = (hz/100); /* 10ms */
913 slice_max = (hz/7); /* ~140ms */
918 * Initialize the kseqs.
920 for (i = 0; i < MAXCPU; i++) {
924 ksq->ksq_assigned = NULL;
925 kseq_setup(&kseq_cpu[i]);
927 if (smp_topology == NULL) {
928 struct kseq_group *ksg;
932 for (cpus = 0, i = 0; i < MAXCPU; i++) {
935 ksq = &kseq_cpu[cpus];
936 ksg = &kseq_groups[cpus];
938 * Setup a kseq group with one member.
940 ksq->ksq_transferable = 0;
941 ksq->ksq_group = ksg;
943 ksg->ksg_idlemask = 0;
944 ksg->ksg_cpumask = ksg->ksg_mask = 1 << i;
946 ksg->ksg_transferable = 0;
947 LIST_INIT(&ksg->ksg_members);
948 LIST_INSERT_HEAD(&ksg->ksg_members, ksq, ksq_siblings);
951 ksg_maxid = cpus - 1;
953 struct kseq_group *ksg;
954 struct cpu_group *cg;
957 for (i = 0; i < smp_topology->ct_count; i++) {
958 cg = &smp_topology->ct_group[i];
959 ksg = &kseq_groups[i];
961 * Initialize the group.
963 ksg->ksg_idlemask = 0;
965 ksg->ksg_transferable = 0;
966 ksg->ksg_cpus = cg->cg_count;
967 ksg->ksg_cpumask = cg->cg_mask;
968 LIST_INIT(&ksg->ksg_members);
970 * Find all of the group members and add them.
972 for (j = 0; j < MAXCPU; j++) {
973 if ((cg->cg_mask & (1 << j)) != 0) {
974 if (ksg->ksg_mask == 0)
975 ksg->ksg_mask = 1 << j;
976 kseq_cpu[j].ksq_transferable = 0;
977 kseq_cpu[j].ksq_group = ksg;
978 LIST_INSERT_HEAD(&ksg->ksg_members,
979 &kseq_cpu[j], ksq_siblings);
982 if (ksg->ksg_cpus > 1)
985 ksg_maxid = smp_topology->ct_count - 1;
988 * Stagger the group and global load balancer so they do not
989 * interfere with each other.
991 bal_tick = ticks + hz;
993 gbal_tick = ticks + (hz / 2);
995 kseq_setup(KSEQ_SELF());
997 mtx_lock_spin(&sched_lock);
998 kseq_load_add(KSEQ_SELF(), &kse0);
999 mtx_unlock_spin(&sched_lock);
1003 * Scale the scheduling priority according to the "interactivity" of this
1007 sched_priority(struct ksegrp *kg)
1011 if (kg->kg_pri_class != PRI_TIMESHARE)
1014 pri = SCHED_PRI_INTERACT(sched_interact_score(kg));
1015 pri += SCHED_PRI_BASE;
1016 pri += kg->kg_proc->p_nice;
1018 if (pri > PRI_MAX_TIMESHARE)
1019 pri = PRI_MAX_TIMESHARE;
1020 else if (pri < PRI_MIN_TIMESHARE)
1021 pri = PRI_MIN_TIMESHARE;
1023 kg->kg_user_pri = pri;
1029 * Calculate a time slice based on the properties of the kseg and the runq
1030 * that we're on. This is only for PRI_TIMESHARE ksegrps.
1033 sched_slice(struct kse *ke)
1039 kseq = KSEQ_CPU(ke->ke_cpu);
1041 if (ke->ke_thread->td_flags & TDF_BORROWING) {
1042 ke->ke_slice = SCHED_SLICE_MIN;
1048 * KSEs in interactive ksegs get a minimal slice so that we
1049 * quickly notice if it abuses its advantage.
1051 * KSEs in non-interactive ksegs are assigned a slice that is
1052 * based on the ksegs nice value relative to the least nice kseg
1053 * on the run queue for this cpu.
1055 * If the KSE is less nice than all others it gets the maximum
1056 * slice and other KSEs will adjust their slice relative to
1057 * this when they first expire.
1059 * There is 20 point window that starts relative to the least
1060 * nice kse on the run queue. Slice size is determined by
1061 * the kse distance from the last nice ksegrp.
1063 * If the kse is outside of the window it will get no slice
1064 * and will be reevaluated each time it is selected on the
1065 * run queue. The exception to this is nice 0 ksegs when
1066 * a nice -20 is running. They are always granted a minimum
1069 if (!SCHED_INTERACTIVE(kg)) {
1072 nice = kg->kg_proc->p_nice + (0 - kseq->ksq_nicemin);
1073 if (kseq->ksq_load_timeshare == 0 ||
1074 kg->kg_proc->p_nice < kseq->ksq_nicemin)
1075 ke->ke_slice = SCHED_SLICE_MAX;
1076 else if (nice <= SCHED_SLICE_NTHRESH)
1077 ke->ke_slice = SCHED_SLICE_NICE(nice);
1078 else if (kg->kg_proc->p_nice == 0)
1079 ke->ke_slice = SCHED_SLICE_MIN;
1083 ke->ke_slice = SCHED_SLICE_INTERACTIVE;
1089 * This routine enforces a maximum limit on the amount of scheduling history
1090 * kept. It is called after either the slptime or runtime is adjusted.
1091 * This routine will not operate correctly when slp or run times have been
1092 * adjusted to more than double their maximum.
1095 sched_interact_update(struct ksegrp *kg)
1099 sum = kg->kg_runtime + kg->kg_slptime;
1100 if (sum < SCHED_SLP_RUN_MAX)
1103 * If we have exceeded by more than 1/5th then the algorithm below
1104 * will not bring us back into range. Dividing by two here forces
1105 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1107 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1108 kg->kg_runtime /= 2;
1109 kg->kg_slptime /= 2;
1112 kg->kg_runtime = (kg->kg_runtime / 5) * 4;
1113 kg->kg_slptime = (kg->kg_slptime / 5) * 4;
1117 sched_interact_fork(struct ksegrp *kg)
1122 sum = kg->kg_runtime + kg->kg_slptime;
1123 if (sum > SCHED_SLP_RUN_FORK) {
1124 ratio = sum / SCHED_SLP_RUN_FORK;
1125 kg->kg_runtime /= ratio;
1126 kg->kg_slptime /= ratio;
1131 sched_interact_score(struct ksegrp *kg)
1135 if (kg->kg_runtime > kg->kg_slptime) {
1136 div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF);
1137 return (SCHED_INTERACT_HALF +
1138 (SCHED_INTERACT_HALF - (kg->kg_slptime / div)));
1139 } if (kg->kg_slptime > kg->kg_runtime) {
1140 div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF);
1141 return (kg->kg_runtime / div);
1145 * This can happen if slptime and runtime are 0.
1152 * Very early in the boot some setup of scheduler-specific
1153 * parts of proc0 and of soem scheduler resources needs to be done.
1161 * Set up the scheduler specific parts of proc0.
1163 proc0.p_sched = NULL; /* XXX */
1164 ksegrp0.kg_sched = &kg_sched0;
1165 thread0.td_sched = &kse0;
1166 kse0.ke_thread = &thread0;
1167 kse0.ke_state = KES_THREAD;
1168 kg_sched0.skg_concurrency = 1;
1169 kg_sched0.skg_avail_opennings = 0; /* we are already running */
1173 * This is only somewhat accurate since given many processes of the same
1174 * priority they will switch when their slices run out, which will be
1175 * at most SCHED_SLICE_MAX.
1178 sched_rr_interval(void)
1180 return (SCHED_SLICE_MAX);
1184 sched_pctcpu_update(struct kse *ke)
1187 * Adjust counters and watermark for pctcpu calc.
1189 if (ke->ke_ltick > ticks - SCHED_CPU_TICKS) {
1191 * Shift the tick count out so that the divide doesn't
1192 * round away our results.
1194 ke->ke_ticks <<= 10;
1195 ke->ke_ticks = (ke->ke_ticks / (ticks - ke->ke_ftick)) *
1197 ke->ke_ticks >>= 10;
1200 ke->ke_ltick = ticks;
1201 ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS;
1205 sched_thread_priority(struct thread *td, u_char prio)
1209 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1210 td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1211 curthread->td_proc->p_comm);
1213 mtx_assert(&sched_lock, MA_OWNED);
1214 if (td->td_priority == prio)
1216 if (TD_ON_RUNQ(td)) {
1218 * If the priority has been elevated due to priority
1219 * propagation, we may have to move ourselves to a new
1220 * queue. We still call adjustrunqueue below in case kse
1221 * needs to fix things up.
1223 if (prio < td->td_priority && ke->ke_runq != NULL &&
1224 (ke->ke_flags & KEF_ASSIGNED) == 0 &&
1225 ke->ke_runq != KSEQ_CPU(ke->ke_cpu)->ksq_curr) {
1226 runq_remove(ke->ke_runq, ke);
1227 ke->ke_runq = KSEQ_CPU(ke->ke_cpu)->ksq_curr;
1228 runq_add(ke->ke_runq, ke, 0);
1231 * Hold this kse on this cpu so that sched_prio() doesn't
1232 * cause excessive migration. We only want migration to
1233 * happen as the result of a wakeup.
1235 ke->ke_flags |= KEF_HOLD;
1236 adjustrunqueue(td, prio);
1237 ke->ke_flags &= ~KEF_HOLD;
1239 td->td_priority = prio;
1243 * Update a thread's priority when it is lent another thread's
1247 sched_lend_prio(struct thread *td, u_char prio)
1250 td->td_flags |= TDF_BORROWING;
1251 sched_thread_priority(td, prio);
1255 * Restore a thread's priority when priority propagation is
1256 * over. The prio argument is the minimum priority the thread
1257 * needs to have to satisfy other possible priority lending
1258 * requests. If the thread's regular priority is less
1259 * important than prio, the thread will keep a priority boost
1263 sched_unlend_prio(struct thread *td, u_char prio)
1267 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1268 td->td_base_pri <= PRI_MAX_TIMESHARE)
1269 base_pri = td->td_ksegrp->kg_user_pri;
1271 base_pri = td->td_base_pri;
1272 if (prio >= base_pri) {
1273 td->td_flags &= ~TDF_BORROWING;
1274 sched_thread_priority(td, base_pri);
1276 sched_lend_prio(td, prio);
1280 sched_prio(struct thread *td, u_char prio)
1284 /* First, update the base priority. */
1285 td->td_base_pri = prio;
1288 * If the thread is borrowing another thread's priority, don't
1289 * ever lower the priority.
1291 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1294 /* Change the real priority. */
1295 oldprio = td->td_priority;
1296 sched_thread_priority(td, prio);
1299 * If the thread is on a turnstile, then let the turnstile update
1302 if (TD_ON_LOCK(td) && oldprio != prio)
1303 turnstile_adjust(td, oldprio);
1307 sched_switch(struct thread *td, struct thread *newtd, int flags)
1312 mtx_assert(&sched_lock, MA_OWNED);
1317 td->td_lastcpu = td->td_oncpu;
1318 td->td_oncpu = NOCPU;
1319 td->td_flags &= ~TDF_NEEDRESCHED;
1320 td->td_owepreempt = 0;
1323 * If the KSE has been assigned it may be in the process of switching
1324 * to the new cpu. This is the case in sched_bind().
1326 if (td == PCPU_GET(idlethread)) {
1328 } else if ((ke->ke_flags & KEF_ASSIGNED) == 0) {
1329 /* We are ending our run so make our slot available again */
1330 SLOT_RELEASE(td->td_ksegrp);
1331 kseq_load_rem(ksq, ke);
1332 if (TD_IS_RUNNING(td)) {
1334 * Don't allow the thread to migrate
1335 * from a preemption.
1337 ke->ke_flags |= KEF_HOLD;
1338 setrunqueue(td, (flags & SW_PREEMPT) ?
1339 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1340 SRQ_OURSELF|SRQ_YIELDING);
1341 ke->ke_flags &= ~KEF_HOLD;
1342 } else if ((td->td_proc->p_flag & P_HADTHREADS) &&
1343 (newtd == NULL || newtd->td_ksegrp != td->td_ksegrp))
1345 * We will not be on the run queue.
1346 * So we must be sleeping or similar.
1347 * Don't use the slot if we will need it
1350 slot_fill(td->td_ksegrp);
1352 if (newtd != NULL) {
1354 * If we bring in a thread account for it as if it had been
1355 * added to the 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 * XXX When we preempt, we've already consumed a slot because
1363 * we got here through sched_add(). However, newtd can come
1364 * from thread_switchout() which can't SLOT_USE() because
1365 * the SLOT code is scheduler dependent. We must use the
1366 * slot here otherwise.
1368 if ((flags & SW_PREEMPT) == 0)
1369 SLOT_USE(newtd->td_ksegrp);
1371 newtd = choosethread();
1374 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1375 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1377 cpu_switch(td, newtd);
1379 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1380 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1384 sched_lock.mtx_lock = (uintptr_t)td;
1386 td->td_oncpu = PCPU_GET(cpuid);
1390 sched_nice(struct proc *p, int nice)
1397 PROC_LOCK_ASSERT(p, MA_OWNED);
1398 mtx_assert(&sched_lock, MA_OWNED);
1400 * We need to adjust the nice counts for running KSEs.
1402 FOREACH_KSEGRP_IN_PROC(p, kg) {
1403 if (kg->kg_pri_class == PRI_TIMESHARE) {
1404 FOREACH_THREAD_IN_GROUP(kg, td) {
1406 if (ke->ke_runq == NULL)
1408 kseq = KSEQ_CPU(ke->ke_cpu);
1409 kseq_nice_rem(kseq, p->p_nice);
1410 kseq_nice_add(kseq, nice);
1415 FOREACH_KSEGRP_IN_PROC(p, kg) {
1417 FOREACH_THREAD_IN_GROUP(kg, td)
1418 td->td_flags |= TDF_NEEDRESCHED;
1423 sched_sleep(struct thread *td)
1425 mtx_assert(&sched_lock, MA_OWNED);
1427 td->td_slptime = ticks;
1431 sched_wakeup(struct thread *td)
1433 mtx_assert(&sched_lock, MA_OWNED);
1436 * Let the kseg know how long we slept for. This is because process
1437 * interactivity behavior is modeled in the kseg.
1439 if (td->td_slptime) {
1444 hzticks = (ticks - td->td_slptime) << 10;
1445 if (hzticks >= SCHED_SLP_RUN_MAX) {
1446 kg->kg_slptime = SCHED_SLP_RUN_MAX;
1449 kg->kg_slptime += hzticks;
1450 sched_interact_update(kg);
1453 sched_slice(td->td_kse);
1456 setrunqueue(td, SRQ_BORING);
1460 * Penalize the parent for creating a new child and initialize the child's
1464 sched_fork(struct thread *td, struct thread *childtd)
1467 mtx_assert(&sched_lock, MA_OWNED);
1469 sched_fork_ksegrp(td, childtd->td_ksegrp);
1470 sched_fork_thread(td, childtd);
1474 sched_fork_ksegrp(struct thread *td, struct ksegrp *child)
1476 struct ksegrp *kg = td->td_ksegrp;
1477 mtx_assert(&sched_lock, MA_OWNED);
1479 child->kg_slptime = kg->kg_slptime;
1480 child->kg_runtime = kg->kg_runtime;
1481 child->kg_user_pri = kg->kg_user_pri;
1482 sched_interact_fork(child);
1483 kg->kg_runtime += tickincr << 10;
1484 sched_interact_update(kg);
1488 sched_fork_thread(struct thread *td, struct thread *child)
1493 sched_newthread(child);
1495 ke2 = child->td_kse;
1496 ke2->ke_slice = 1; /* Attempt to quickly learn interactivity. */
1497 ke2->ke_cpu = ke->ke_cpu;
1498 ke2->ke_runq = NULL;
1500 /* Grab our parents cpu estimation information. */
1501 ke2->ke_ticks = ke->ke_ticks;
1502 ke2->ke_ltick = ke->ke_ltick;
1503 ke2->ke_ftick = ke->ke_ftick;
1507 sched_class(struct ksegrp *kg, int class)
1515 mtx_assert(&sched_lock, MA_OWNED);
1516 if (kg->kg_pri_class == class)
1519 nclass = PRI_BASE(class);
1520 oclass = PRI_BASE(kg->kg_pri_class);
1521 FOREACH_THREAD_IN_GROUP(kg, td) {
1523 if ((ke->ke_state != KES_ONRUNQ &&
1524 ke->ke_state != KES_THREAD) || ke->ke_runq == NULL)
1526 kseq = KSEQ_CPU(ke->ke_cpu);
1530 * On SMP if we're on the RUNQ we must adjust the transferable
1531 * count because could be changing to or from an interrupt
1534 if (ke->ke_state == KES_ONRUNQ) {
1535 if (KSE_CAN_MIGRATE(ke)) {
1536 kseq->ksq_transferable--;
1537 kseq->ksq_group->ksg_transferable--;
1539 if (KSE_CAN_MIGRATE(ke)) {
1540 kseq->ksq_transferable++;
1541 kseq->ksq_group->ksg_transferable++;
1545 if (oclass == PRI_TIMESHARE) {
1546 kseq->ksq_load_timeshare--;
1547 kseq_nice_rem(kseq, kg->kg_proc->p_nice);
1549 if (nclass == PRI_TIMESHARE) {
1550 kseq->ksq_load_timeshare++;
1551 kseq_nice_add(kseq, kg->kg_proc->p_nice);
1555 kg->kg_pri_class = class;
1559 * Return some of the child's priority and interactivity to the parent.
1562 sched_exit(struct proc *p, struct thread *childtd)
1564 mtx_assert(&sched_lock, MA_OWNED);
1565 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), childtd);
1566 sched_exit_thread(NULL, childtd);
1570 sched_exit_ksegrp(struct ksegrp *kg, struct thread *td)
1572 /* kg->kg_slptime += td->td_ksegrp->kg_slptime; */
1573 kg->kg_runtime += td->td_ksegrp->kg_runtime;
1574 sched_interact_update(kg);
1578 sched_exit_thread(struct thread *td, struct thread *childtd)
1580 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1581 childtd, childtd->td_proc->p_comm, childtd->td_priority);
1582 kseq_load_rem(KSEQ_CPU(childtd->td_kse->ke_cpu), childtd->td_kse);
1586 sched_clock(struct thread *td)
1592 mtx_assert(&sched_lock, MA_OWNED);
1595 if (ticks >= bal_tick)
1597 if (ticks >= gbal_tick && balance_groups)
1598 sched_balance_groups();
1600 * We could have been assigned a non real-time thread without an
1603 if (kseq->ksq_assigned)
1604 kseq_assign(kseq); /* Potentially sets NEEDRESCHED */
1607 * sched_setup() apparently happens prior to stathz being set. We
1608 * need to resolve the timers earlier in the boot so we can avoid
1609 * calculating this here.
1611 if (realstathz == 0) {
1612 realstathz = stathz ? stathz : hz;
1613 tickincr = hz / realstathz;
1615 * XXX This does not work for values of stathz that are much
1625 /* Adjust ticks for pctcpu */
1627 ke->ke_ltick = ticks;
1629 /* Go up to one second beyond our max and then trim back down */
1630 if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick)
1631 sched_pctcpu_update(ke);
1633 if (td->td_flags & TDF_IDLETD)
1636 * We only do slicing code for TIMESHARE ksegrps.
1638 if (kg->kg_pri_class != PRI_TIMESHARE)
1641 * We used a tick charge it to the ksegrp so that we can compute our
1644 kg->kg_runtime += tickincr << 10;
1645 sched_interact_update(kg);
1648 * We used up one time slice.
1650 if (--ke->ke_slice > 0)
1653 * We're out of time, recompute priorities and requeue.
1655 kseq_load_rem(kseq, ke);
1658 if (SCHED_CURR(kg, ke))
1659 ke->ke_runq = kseq->ksq_curr;
1661 ke->ke_runq = kseq->ksq_next;
1662 kseq_load_add(kseq, ke);
1663 td->td_flags |= TDF_NEEDRESCHED;
1667 sched_runnable(void)
1676 if (kseq->ksq_assigned) {
1677 mtx_lock_spin(&sched_lock);
1679 mtx_unlock_spin(&sched_lock);
1682 if ((curthread->td_flags & TDF_IDLETD) != 0) {
1683 if (kseq->ksq_load > 0)
1686 if (kseq->ksq_load - 1 > 0)
1694 sched_userret(struct thread *td)
1698 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1699 ("thread with borrowed priority returning to userland"));
1701 if (td->td_priority != kg->kg_user_pri) {
1702 mtx_lock_spin(&sched_lock);
1703 td->td_priority = kg->kg_user_pri;
1704 td->td_base_pri = kg->kg_user_pri;
1705 mtx_unlock_spin(&sched_lock);
1715 mtx_assert(&sched_lock, MA_OWNED);
1719 if (kseq->ksq_assigned)
1722 ke = kseq_choose(kseq);
1725 if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE)
1726 if (kseq_idled(kseq) == 0)
1729 kseq_runq_rem(kseq, ke);
1730 ke->ke_state = KES_THREAD;
1734 if (kseq_idled(kseq) == 0)
1741 sched_add(struct thread *td, int flags)
1750 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1751 td, td->td_proc->p_comm, td->td_priority, curthread,
1752 curthread->td_proc->p_comm);
1753 mtx_assert(&sched_lock, MA_OWNED);
1757 preemptive = !(flags & SRQ_YIELDING);
1758 class = PRI_BASE(kg->kg_pri_class);
1760 if ((ke->ke_flags & KEF_INTERNAL) == 0)
1761 SLOT_USE(td->td_ksegrp);
1762 ke->ke_flags &= ~KEF_INTERNAL;
1764 if (ke->ke_flags & KEF_ASSIGNED) {
1765 if (ke->ke_flags & KEF_REMOVED)
1766 ke->ke_flags &= ~KEF_REMOVED;
1769 canmigrate = KSE_CAN_MIGRATE(ke);
1771 KASSERT(ke->ke_state != KES_ONRUNQ,
1772 ("sched_add: kse %p (%s) already in run queue", ke,
1773 ke->ke_proc->p_comm));
1774 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
1775 ("sched_add: process swapped out"));
1776 KASSERT(ke->ke_runq == NULL,
1777 ("sched_add: KSE %p is still assigned to a run queue", ke));
1781 ke->ke_runq = kseq->ksq_curr;
1782 ke->ke_slice = SCHED_SLICE_MAX;
1784 ke->ke_cpu = PCPU_GET(cpuid);
1787 if (SCHED_CURR(kg, ke))
1788 ke->ke_runq = kseq->ksq_curr;
1790 ke->ke_runq = kseq->ksq_next;
1794 * This is for priority prop.
1796 if (ke->ke_thread->td_priority < PRI_MIN_IDLE)
1797 ke->ke_runq = kseq->ksq_curr;
1799 ke->ke_runq = &kseq->ksq_idle;
1800 ke->ke_slice = SCHED_SLICE_MIN;
1803 panic("Unknown pri class.");
1808 * Don't migrate running threads here. Force the long term balancer
1811 if (ke->ke_flags & KEF_HOLD) {
1812 ke->ke_flags &= ~KEF_HOLD;
1816 * If this thread is pinned or bound, notify the target cpu.
1818 if (!canmigrate && ke->ke_cpu != PCPU_GET(cpuid) ) {
1820 kseq_notify(ke, ke->ke_cpu);
1824 * If we had been idle, clear our bit in the group and potentially
1825 * the global bitmap. If not, see if we should transfer this thread.
1827 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1828 (kseq->ksq_group->ksg_idlemask & PCPU_GET(cpumask)) != 0) {
1830 * Check to see if our group is unidling, and if so, remove it
1831 * from the global idle mask.
1833 if (kseq->ksq_group->ksg_idlemask ==
1834 kseq->ksq_group->ksg_cpumask)
1835 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
1837 * Now remove ourselves from the group specific idle mask.
1839 kseq->ksq_group->ksg_idlemask &= ~PCPU_GET(cpumask);
1840 } else if (canmigrate && kseq->ksq_load > 1 && class != PRI_ITHD)
1841 if (kseq_transfer(kseq, ke, class))
1843 ke->ke_cpu = PCPU_GET(cpuid);
1845 if (td->td_priority < curthread->td_priority &&
1846 ke->ke_runq == kseq->ksq_curr)
1847 curthread->td_flags |= TDF_NEEDRESCHED;
1848 if (preemptive && maybe_preempt(td))
1850 ke->ke_state = KES_ONRUNQ;
1852 kseq_runq_add(kseq, ke, flags);
1853 kseq_load_add(kseq, ke);
1857 sched_rem(struct thread *td)
1862 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1863 td, td->td_proc->p_comm, td->td_priority, curthread,
1864 curthread->td_proc->p_comm);
1865 mtx_assert(&sched_lock, MA_OWNED);
1867 SLOT_RELEASE(td->td_ksegrp);
1868 if (ke->ke_flags & KEF_ASSIGNED) {
1869 ke->ke_flags |= KEF_REMOVED;
1872 KASSERT((ke->ke_state == KES_ONRUNQ),
1873 ("sched_rem: KSE not on run queue"));
1875 ke->ke_state = KES_THREAD;
1876 kseq = KSEQ_CPU(ke->ke_cpu);
1877 kseq_runq_rem(kseq, ke);
1878 kseq_load_rem(kseq, ke);
1882 sched_pctcpu(struct thread *td)
1892 mtx_lock_spin(&sched_lock);
1897 * Don't update more frequently than twice a second. Allowing
1898 * this causes the cpu usage to decay away too quickly due to
1901 if (ke->ke_ftick + SCHED_CPU_TICKS < ke->ke_ltick ||
1902 ke->ke_ltick < (ticks - (hz / 2)))
1903 sched_pctcpu_update(ke);
1904 /* How many rtick per second ? */
1905 rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1906 pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1909 ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick;
1910 mtx_unlock_spin(&sched_lock);
1916 sched_bind(struct thread *td, int cpu)
1920 mtx_assert(&sched_lock, MA_OWNED);
1922 ke->ke_flags |= KEF_BOUND;
1924 if (PCPU_GET(cpuid) == cpu)
1926 /* sched_rem without the runq_remove */
1927 ke->ke_state = KES_THREAD;
1928 kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1929 kseq_notify(ke, cpu);
1930 /* When we return from mi_switch we'll be on the correct cpu. */
1931 mi_switch(SW_VOL, NULL);
1936 sched_unbind(struct thread *td)
1938 mtx_assert(&sched_lock, MA_OWNED);
1939 td->td_kse->ke_flags &= ~KEF_BOUND;
1943 sched_is_bound(struct thread *td)
1945 mtx_assert(&sched_lock, MA_OWNED);
1946 return (td->td_kse->ke_flags & KEF_BOUND);
1957 for (i = 0; i <= ksg_maxid; i++)
1958 total += KSEQ_GROUP(i)->ksg_load;
1961 return (KSEQ_SELF()->ksq_sysload);
1966 sched_sizeof_ksegrp(void)
1968 return (sizeof(struct ksegrp) + sizeof(struct kg_sched));
1972 sched_sizeof_proc(void)
1974 return (sizeof(struct proc));
1978 sched_sizeof_thread(void)
1980 return (sizeof(struct thread) + sizeof(struct td_sched));
1982 #define KERN_SWITCH_INCLUDE 1
1983 #include "kern/kern_switch.c"