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_PREEMPTED 0x0040 /* Thread was preempted */
127 #define KEF_DIDRUN 0x02000 /* Thread actually ran. */
128 #define KEF_EXIT 0x04000 /* Thread is being killed. */
131 struct thread *skg_last_assigned; /* (j) Last thread assigned to */
132 /* the system scheduler */
133 int skg_slptime; /* Number of ticks we vol. slept */
134 int skg_runtime; /* Number of ticks we were running */
135 int skg_avail_opennings; /* (j) Num unfilled slots in group.*/
136 int skg_concurrency; /* (j) Num threads requested in group.*/
138 #define kg_last_assigned kg_sched->skg_last_assigned
139 #define kg_avail_opennings kg_sched->skg_avail_opennings
140 #define kg_concurrency kg_sched->skg_concurrency
141 #define kg_runtime kg_sched->skg_runtime
142 #define kg_slptime kg_sched->skg_slptime
144 #define SLOT_RELEASE(kg) (kg)->kg_avail_opennings++
145 #define SLOT_USE(kg) (kg)->kg_avail_opennings--
147 static struct kse kse0;
148 static struct kg_sched kg_sched0;
151 * The priority is primarily determined by the interactivity score. Thus, we
152 * give lower(better) priorities to kse groups that use less CPU. The nice
153 * value is then directly added to this to allow nice to have some effect
156 * PRI_RANGE: Total priority range for timeshare threads.
157 * PRI_NRESV: Number of nice values.
158 * PRI_BASE: The start of the dynamic range.
160 #define SCHED_PRI_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
161 #define SCHED_PRI_NRESV ((PRIO_MAX - PRIO_MIN) + 1)
162 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
163 #define SCHED_PRI_BASE (PRI_MIN_TIMESHARE)
164 #define SCHED_PRI_INTERACT(score) \
165 ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
168 * These determine the interactivity of a process.
170 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
171 * before throttling back.
172 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
173 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
174 * INTERACT_THRESH: Threshhold for placement on the current runq.
176 #define SCHED_SLP_RUN_MAX ((hz * 5) << 10)
177 #define SCHED_SLP_RUN_FORK ((hz / 2) << 10)
178 #define SCHED_INTERACT_MAX (100)
179 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
180 #define SCHED_INTERACT_THRESH (30)
183 * These parameters and macros determine the size of the time slice that is
184 * granted to each thread.
186 * SLICE_MIN: Minimum time slice granted, in units of ticks.
187 * SLICE_MAX: Maximum time slice granted.
188 * SLICE_RANGE: Range of available time slices scaled by hz.
189 * SLICE_SCALE: The number slices granted per val in the range of [0, max].
190 * SLICE_NICE: Determine the amount of slice granted to a scaled nice.
191 * SLICE_NTHRESH: The nice cutoff point for slice assignment.
193 #define SCHED_SLICE_MIN (slice_min)
194 #define SCHED_SLICE_MAX (slice_max)
195 #define SCHED_SLICE_INTERACTIVE (slice_max)
196 #define SCHED_SLICE_NTHRESH (SCHED_PRI_NHALF - 1)
197 #define SCHED_SLICE_RANGE (SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
198 #define SCHED_SLICE_SCALE(val, max) (((val) * SCHED_SLICE_RANGE) / (max))
199 #define SCHED_SLICE_NICE(nice) \
200 (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
203 * This macro determines whether or not the thread belongs on the current or
206 #define SCHED_INTERACTIVE(kg) \
207 (sched_interact_score(kg) < SCHED_INTERACT_THRESH)
208 #define SCHED_CURR(kg, ke) \
209 ((ke->ke_thread->td_flags & TDF_BORROWING) || \
210 (ke->ke_flags & KEF_PREEMPTED) || SCHED_INTERACTIVE(kg))
213 * Cpu percentage computation macros and defines.
215 * SCHED_CPU_TIME: Number of seconds to average the cpu usage across.
216 * SCHED_CPU_TICKS: Number of hz ticks to average the cpu usage across.
219 #define SCHED_CPU_TIME 10
220 #define SCHED_CPU_TICKS (hz * SCHED_CPU_TIME)
223 * kseq - per processor runqs and statistics.
226 struct runq ksq_idle; /* Queue of IDLE threads. */
227 struct runq ksq_timeshare[2]; /* Run queues for !IDLE. */
228 struct runq *ksq_next; /* Next timeshare queue. */
229 struct runq *ksq_curr; /* Current queue. */
230 int ksq_load_timeshare; /* Load for timeshare. */
231 int ksq_load; /* Aggregate load. */
232 short ksq_nice[SCHED_PRI_NRESV]; /* KSEs in each nice bin. */
233 short ksq_nicemin; /* Least nice. */
235 int ksq_transferable;
236 LIST_ENTRY(kseq) ksq_siblings; /* Next in kseq group. */
237 struct kseq_group *ksq_group; /* Our processor group. */
238 volatile struct kse *ksq_assigned; /* assigned by another CPU. */
240 int ksq_sysload; /* For loadavg, !ITHD load. */
246 * kseq groups are groups of processors which can cheaply share threads. When
247 * one processor in the group goes idle it will check the runqs of the other
248 * processors in its group prior to halting and waiting for an interrupt.
249 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
250 * In a numa environment we'd want an idle bitmap per group and a two tiered
254 int ksg_cpus; /* Count of CPUs in this kseq group. */
255 cpumask_t ksg_cpumask; /* Mask of cpus in this group. */
256 cpumask_t ksg_idlemask; /* Idle cpus in this group. */
257 cpumask_t ksg_mask; /* Bit mask for first cpu. */
258 int ksg_load; /* Total load of this group. */
259 int ksg_transferable; /* Transferable load of this group. */
260 LIST_HEAD(, kseq) ksg_members; /* Linked list of all members. */
265 * One kse queue per processor.
268 static cpumask_t kseq_idle;
269 static int ksg_maxid;
270 static struct kseq kseq_cpu[MAXCPU];
271 static struct kseq_group kseq_groups[MAXCPU];
273 static int gbal_tick;
274 static int balance_groups;
276 #define KSEQ_SELF() (&kseq_cpu[PCPU_GET(cpuid)])
277 #define KSEQ_CPU(x) (&kseq_cpu[(x)])
278 #define KSEQ_ID(x) ((x) - kseq_cpu)
279 #define KSEQ_GROUP(x) (&kseq_groups[(x)])
281 static struct kseq kseq_cpu;
283 #define KSEQ_SELF() (&kseq_cpu)
284 #define KSEQ_CPU(x) (&kseq_cpu)
287 static void slot_fill(struct ksegrp *);
288 static struct kse *sched_choose(void); /* XXX Should be thread * */
289 static void sched_slice(struct kse *);
290 static void sched_priority(struct ksegrp *);
291 static void sched_thread_priority(struct thread *, u_char);
292 static int sched_interact_score(struct ksegrp *);
293 static void sched_interact_update(struct ksegrp *);
294 static void sched_interact_fork(struct ksegrp *);
295 static void sched_pctcpu_update(struct kse *);
297 /* Operations on per processor queues */
298 static struct kse * kseq_choose(struct kseq *);
299 static void kseq_setup(struct kseq *);
300 static void kseq_load_add(struct kseq *, struct kse *);
301 static void kseq_load_rem(struct kseq *, struct kse *);
302 static __inline void kseq_runq_add(struct kseq *, struct kse *, int);
303 static __inline void kseq_runq_rem(struct kseq *, struct kse *);
304 static void kseq_nice_add(struct kseq *, int);
305 static void kseq_nice_rem(struct kseq *, int);
306 void kseq_print(int cpu);
308 static int kseq_transfer(struct kseq *, struct kse *, int);
309 static struct kse *runq_steal(struct runq *);
310 static void sched_balance(void);
311 static void sched_balance_groups(void);
312 static void sched_balance_group(struct kseq_group *);
313 static void sched_balance_pair(struct kseq *, struct kseq *);
314 static void kseq_move(struct kseq *, int);
315 static int kseq_idled(struct kseq *);
316 static void kseq_notify(struct kse *, int);
317 static void kseq_assign(struct kseq *);
318 static struct kse *kseq_steal(struct kseq *, int);
319 #define KSE_CAN_MIGRATE(ke) \
320 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
329 kseq = KSEQ_CPU(cpu);
332 printf("\tload: %d\n", kseq->ksq_load);
333 printf("\tload TIMESHARE: %d\n", kseq->ksq_load_timeshare);
335 printf("\tload transferable: %d\n", kseq->ksq_transferable);
337 printf("\tnicemin:\t%d\n", kseq->ksq_nicemin);
338 printf("\tnice counts:\n");
339 for (i = 0; i < SCHED_PRI_NRESV; i++)
340 if (kseq->ksq_nice[i])
341 printf("\t\t%d = %d\n",
342 i - SCHED_PRI_NHALF, kseq->ksq_nice[i]);
346 kseq_runq_add(struct kseq *kseq, struct kse *ke, int flags)
349 if (KSE_CAN_MIGRATE(ke)) {
350 kseq->ksq_transferable++;
351 kseq->ksq_group->ksg_transferable++;
352 ke->ke_flags |= KEF_XFERABLE;
355 if (ke->ke_flags & KEF_PREEMPTED)
356 flags |= SRQ_PREEMPTED;
357 runq_add(ke->ke_runq, ke, flags);
361 kseq_runq_rem(struct kseq *kseq, struct kse *ke)
364 if (ke->ke_flags & KEF_XFERABLE) {
365 kseq->ksq_transferable--;
366 kseq->ksq_group->ksg_transferable--;
367 ke->ke_flags &= ~KEF_XFERABLE;
370 runq_remove(ke->ke_runq, ke);
374 kseq_load_add(struct kseq *kseq, struct kse *ke)
377 mtx_assert(&sched_lock, MA_OWNED);
378 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
379 if (class == PRI_TIMESHARE)
380 kseq->ksq_load_timeshare++;
382 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
383 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
385 kseq->ksq_group->ksg_load++;
389 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
390 kseq_nice_add(kseq, ke->ke_proc->p_nice);
394 kseq_load_rem(struct kseq *kseq, struct kse *ke)
397 mtx_assert(&sched_lock, MA_OWNED);
398 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
399 if (class == PRI_TIMESHARE)
400 kseq->ksq_load_timeshare--;
401 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
403 kseq->ksq_group->ksg_load--;
408 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
410 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
411 kseq_nice_rem(kseq, ke->ke_proc->p_nice);
415 kseq_nice_add(struct kseq *kseq, int nice)
417 mtx_assert(&sched_lock, MA_OWNED);
418 /* Normalize to zero. */
419 kseq->ksq_nice[nice + SCHED_PRI_NHALF]++;
420 if (nice < kseq->ksq_nicemin || kseq->ksq_load_timeshare == 1)
421 kseq->ksq_nicemin = nice;
425 kseq_nice_rem(struct kseq *kseq, int nice)
429 mtx_assert(&sched_lock, MA_OWNED);
430 /* Normalize to zero. */
431 n = nice + SCHED_PRI_NHALF;
433 KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count."));
436 * If this wasn't the smallest nice value or there are more in
437 * this bucket we can just return. Otherwise we have to recalculate
440 if (nice != kseq->ksq_nicemin ||
441 kseq->ksq_nice[n] != 0 ||
442 kseq->ksq_load_timeshare == 0)
445 for (; n < SCHED_PRI_NRESV; n++)
446 if (kseq->ksq_nice[n]) {
447 kseq->ksq_nicemin = n - SCHED_PRI_NHALF;
454 * sched_balance is a simple CPU load balancing algorithm. It operates by
455 * finding the least loaded and most loaded cpu and equalizing their load
456 * by migrating some processes.
458 * Dealing only with two CPUs at a time has two advantages. Firstly, most
459 * installations will only have 2 cpus. Secondly, load balancing too much at
460 * once can have an unpleasant effect on the system. The scheduler rarely has
461 * enough information to make perfect decisions. So this algorithm chooses
462 * algorithm simplicity and more gradual effects on load in larger systems.
464 * It could be improved by considering the priorities and slices assigned to
465 * each task prior to balancing them. There are many pathological cases with
466 * any approach and so the semi random algorithm below may work as well as any.
472 struct kseq_group *high;
473 struct kseq_group *low;
474 struct kseq_group *ksg;
478 bal_tick = ticks + (random() % (hz * 2));
479 if (smp_started == 0)
482 i = random() % (ksg_maxid + 1);
483 for (cnt = 0; cnt <= ksg_maxid; cnt++) {
486 * Find the CPU with the highest load that has some
487 * threads to transfer.
489 if ((high == NULL || ksg->ksg_load > high->ksg_load)
490 && ksg->ksg_transferable)
492 if (low == NULL || ksg->ksg_load < low->ksg_load)
497 if (low != NULL && high != NULL && high != low)
498 sched_balance_pair(LIST_FIRST(&high->ksg_members),
499 LIST_FIRST(&low->ksg_members));
503 sched_balance_groups(void)
507 gbal_tick = ticks + (random() % (hz * 2));
508 mtx_assert(&sched_lock, MA_OWNED);
510 for (i = 0; i <= ksg_maxid; i++)
511 sched_balance_group(KSEQ_GROUP(i));
515 sched_balance_group(struct kseq_group *ksg)
522 if (ksg->ksg_transferable == 0)
526 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
527 load = kseq->ksq_load;
528 if (high == NULL || load > high->ksq_load)
530 if (low == NULL || load < low->ksq_load)
533 if (high != NULL && low != NULL && high != low)
534 sched_balance_pair(high, low);
538 sched_balance_pair(struct kseq *high, struct kseq *low)
548 * If we're transfering within a group we have to use this specific
549 * kseq's transferable count, otherwise we can steal from other members
552 if (high->ksq_group == low->ksq_group) {
553 transferable = high->ksq_transferable;
554 high_load = high->ksq_load;
555 low_load = low->ksq_load;
557 transferable = high->ksq_group->ksg_transferable;
558 high_load = high->ksq_group->ksg_load;
559 low_load = low->ksq_group->ksg_load;
561 if (transferable == 0)
564 * Determine what the imbalance is and then adjust that to how many
565 * kses we actually have to give up (transferable).
567 diff = high_load - low_load;
571 move = min(move, transferable);
572 for (i = 0; i < move; i++)
573 kseq_move(high, KSEQ_ID(low));
578 kseq_move(struct kseq *from, int cpu)
586 ke = kseq_steal(kseq, 1);
588 struct kseq_group *ksg;
590 ksg = kseq->ksq_group;
591 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
592 if (kseq == from || kseq->ksq_transferable == 0)
594 ke = kseq_steal(kseq, 1);
598 panic("kseq_move: No KSEs available with a "
599 "transferable count of %d\n",
600 ksg->ksg_transferable);
604 ke->ke_state = KES_THREAD;
605 kseq_runq_rem(kseq, ke);
606 kseq_load_rem(kseq, ke);
607 kseq_notify(ke, cpu);
611 kseq_idled(struct kseq *kseq)
613 struct kseq_group *ksg;
617 ksg = kseq->ksq_group;
619 * If we're in a cpu group, try and steal kses from another cpu in
620 * the group before idling.
622 if (ksg->ksg_cpus > 1 && ksg->ksg_transferable) {
623 LIST_FOREACH(steal, &ksg->ksg_members, ksq_siblings) {
624 if (steal == kseq || steal->ksq_transferable == 0)
626 ke = kseq_steal(steal, 0);
629 ke->ke_state = KES_THREAD;
630 kseq_runq_rem(steal, ke);
631 kseq_load_rem(steal, ke);
632 ke->ke_cpu = PCPU_GET(cpuid);
633 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
634 sched_add(ke->ke_thread, SRQ_YIELDING);
639 * We only set the idled bit when all of the cpus in the group are
640 * idle. Otherwise we could get into a situation where a KSE bounces
641 * back and forth between two idle cores on seperate physical CPUs.
643 ksg->ksg_idlemask |= PCPU_GET(cpumask);
644 if (ksg->ksg_idlemask != ksg->ksg_cpumask)
646 atomic_set_int(&kseq_idle, ksg->ksg_mask);
651 kseq_assign(struct kseq *kseq)
657 *(volatile struct kse **)&ke = kseq->ksq_assigned;
658 } while(!atomic_cmpset_ptr((volatile uintptr_t *)&kseq->ksq_assigned,
659 (uintptr_t)ke, (uintptr_t)NULL));
660 for (; ke != NULL; ke = nke) {
662 kseq->ksq_group->ksg_load--;
664 ke->ke_flags &= ~KEF_ASSIGNED;
665 if (ke->ke_flags & KEF_REMOVED) {
666 ke->ke_flags &= ~KEF_REMOVED;
669 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
670 sched_add(ke->ke_thread, SRQ_YIELDING);
675 kseq_notify(struct kse *ke, int cpu)
683 kseq = KSEQ_CPU(cpu);
685 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
686 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
687 (kseq_idle & kseq->ksq_group->ksg_mask))
688 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
689 kseq->ksq_group->ksg_load++;
692 ke->ke_flags |= KEF_ASSIGNED;
693 prio = ke->ke_thread->td_priority;
696 * Place a KSE on another cpu's queue and force a resched.
699 *(volatile struct kse **)&ke->ke_assign = kseq->ksq_assigned;
700 } while(!atomic_cmpset_ptr((volatile uintptr_t *)&kseq->ksq_assigned,
701 (uintptr_t)ke->ke_assign, (uintptr_t)ke));
703 * Without sched_lock we could lose a race where we set NEEDRESCHED
704 * on a thread that is switched out before the IPI is delivered. This
705 * would lead us to miss the resched. This will be a problem once
706 * sched_lock is pushed down.
708 pcpu = pcpu_find(cpu);
709 td = pcpu->pc_curthread;
710 if (ke->ke_thread->td_priority < td->td_priority ||
711 td == pcpu->pc_idlethread) {
712 td->td_flags |= TDF_NEEDRESCHED;
713 ipi_selected(1 << cpu, IPI_AST);
718 runq_steal(struct runq *rq)
726 mtx_assert(&sched_lock, MA_OWNED);
727 rqb = &rq->rq_status;
728 for (word = 0; word < RQB_LEN; word++) {
729 if (rqb->rqb_bits[word] == 0)
731 for (bit = 0; bit < RQB_BPW; bit++) {
732 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
734 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
735 TAILQ_FOREACH(ke, rqh, ke_procq) {
736 if (KSE_CAN_MIGRATE(ke))
745 kseq_steal(struct kseq *kseq, int stealidle)
750 * Steal from next first to try to get a non-interactive task that
751 * may not have run for a while.
753 if ((ke = runq_steal(kseq->ksq_next)) != NULL)
755 if ((ke = runq_steal(kseq->ksq_curr)) != NULL)
758 return (runq_steal(&kseq->ksq_idle));
763 kseq_transfer(struct kseq *kseq, struct kse *ke, int class)
765 struct kseq_group *nksg;
766 struct kseq_group *ksg;
771 if (smp_started == 0)
775 * If our load exceeds a certain threshold we should attempt to
776 * reassign this thread. The first candidate is the cpu that
777 * originally ran the thread. If it is idle, assign it there,
778 * otherwise, pick an idle cpu.
780 * The threshold at which we start to reassign kses has a large impact
781 * on the overall performance of the system. Tuned too high and
782 * some CPUs may idle. Too low and there will be excess migration
783 * and context switches.
785 old = KSEQ_CPU(ke->ke_cpu);
786 nksg = old->ksq_group;
787 ksg = kseq->ksq_group;
789 if (kseq_idle & nksg->ksg_mask) {
790 cpu = ffs(nksg->ksg_idlemask);
793 "kseq_transfer: %p found old cpu %X "
794 "in idlemask.", ke, cpu);
799 * Multiple cpus could find this bit simultaneously
800 * but the race shouldn't be terrible.
802 cpu = ffs(kseq_idle);
804 CTR2(KTR_SCHED, "kseq_transfer: %p found %X "
805 "in idlemask.", ke, cpu);
811 if (old->ksq_load < kseq->ksq_load) {
812 cpu = ke->ke_cpu + 1;
813 CTR2(KTR_SCHED, "kseq_transfer: %p old cpu %X "
814 "load less than ours.", ke, cpu);
818 * No new CPU was found, look for one with less load.
820 for (idx = 0; idx <= ksg_maxid; idx++) {
821 nksg = KSEQ_GROUP(idx);
822 if (nksg->ksg_load /*+ (nksg->ksg_cpus * 2)*/ < ksg->ksg_load) {
823 cpu = ffs(nksg->ksg_cpumask);
824 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X load less "
825 "than ours.", ke, cpu);
831 * If another cpu in this group has idled, assign a thread over
832 * to them after checking to see if there are idled groups.
834 if (ksg->ksg_idlemask) {
835 cpu = ffs(ksg->ksg_idlemask);
837 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X idle in "
845 * Now that we've found an idle CPU, migrate the thread.
849 kseq_notify(ke, cpu);
857 * Pick the highest priority task we have and return it.
861 kseq_choose(struct kseq *kseq)
867 mtx_assert(&sched_lock, MA_OWNED);
871 ke = runq_choose(kseq->ksq_curr);
874 * We already swapped once and didn't get anywhere.
878 swap = kseq->ksq_curr;
879 kseq->ksq_curr = kseq->ksq_next;
880 kseq->ksq_next = swap;
884 * If we encounter a slice of 0 the kse is in a
885 * TIMESHARE kse group and its nice was too far out
886 * of the range that receives slices.
888 nice = ke->ke_proc->p_nice + (0 - kseq->ksq_nicemin);
889 if (ke->ke_slice == 0 || (nice > SCHED_SLICE_NTHRESH &&
890 ke->ke_proc->p_nice != 0)) {
891 runq_remove(ke->ke_runq, ke);
893 ke->ke_runq = kseq->ksq_next;
894 runq_add(ke->ke_runq, ke, 0);
900 return (runq_choose(&kseq->ksq_idle));
904 kseq_setup(struct kseq *kseq)
906 runq_init(&kseq->ksq_timeshare[0]);
907 runq_init(&kseq->ksq_timeshare[1]);
908 runq_init(&kseq->ksq_idle);
909 kseq->ksq_curr = &kseq->ksq_timeshare[0];
910 kseq->ksq_next = &kseq->ksq_timeshare[1];
912 kseq->ksq_load_timeshare = 0;
916 sched_setup(void *dummy)
922 slice_min = (hz/100); /* 10ms */
923 slice_max = (hz/7); /* ~140ms */
928 * Initialize the kseqs.
930 for (i = 0; i < MAXCPU; i++) {
934 ksq->ksq_assigned = NULL;
935 kseq_setup(&kseq_cpu[i]);
937 if (smp_topology == NULL) {
938 struct kseq_group *ksg;
942 for (cpus = 0, i = 0; i < MAXCPU; i++) {
945 ksq = &kseq_cpu[cpus];
946 ksg = &kseq_groups[cpus];
948 * Setup a kseq group with one member.
950 ksq->ksq_transferable = 0;
951 ksq->ksq_group = ksg;
953 ksg->ksg_idlemask = 0;
954 ksg->ksg_cpumask = ksg->ksg_mask = 1 << i;
956 ksg->ksg_transferable = 0;
957 LIST_INIT(&ksg->ksg_members);
958 LIST_INSERT_HEAD(&ksg->ksg_members, ksq, ksq_siblings);
961 ksg_maxid = cpus - 1;
963 struct kseq_group *ksg;
964 struct cpu_group *cg;
967 for (i = 0; i < smp_topology->ct_count; i++) {
968 cg = &smp_topology->ct_group[i];
969 ksg = &kseq_groups[i];
971 * Initialize the group.
973 ksg->ksg_idlemask = 0;
975 ksg->ksg_transferable = 0;
976 ksg->ksg_cpus = cg->cg_count;
977 ksg->ksg_cpumask = cg->cg_mask;
978 LIST_INIT(&ksg->ksg_members);
980 * Find all of the group members and add them.
982 for (j = 0; j < MAXCPU; j++) {
983 if ((cg->cg_mask & (1 << j)) != 0) {
984 if (ksg->ksg_mask == 0)
985 ksg->ksg_mask = 1 << j;
986 kseq_cpu[j].ksq_transferable = 0;
987 kseq_cpu[j].ksq_group = ksg;
988 LIST_INSERT_HEAD(&ksg->ksg_members,
989 &kseq_cpu[j], ksq_siblings);
992 if (ksg->ksg_cpus > 1)
995 ksg_maxid = smp_topology->ct_count - 1;
998 * Stagger the group and global load balancer so they do not
999 * interfere with each other.
1001 bal_tick = ticks + hz;
1003 gbal_tick = ticks + (hz / 2);
1005 kseq_setup(KSEQ_SELF());
1007 mtx_lock_spin(&sched_lock);
1008 kseq_load_add(KSEQ_SELF(), &kse0);
1009 mtx_unlock_spin(&sched_lock);
1013 * Scale the scheduling priority according to the "interactivity" of this
1017 sched_priority(struct ksegrp *kg)
1021 if (kg->kg_pri_class != PRI_TIMESHARE)
1024 pri = SCHED_PRI_INTERACT(sched_interact_score(kg));
1025 pri += SCHED_PRI_BASE;
1026 pri += kg->kg_proc->p_nice;
1028 if (pri > PRI_MAX_TIMESHARE)
1029 pri = PRI_MAX_TIMESHARE;
1030 else if (pri < PRI_MIN_TIMESHARE)
1031 pri = PRI_MIN_TIMESHARE;
1033 kg->kg_user_pri = pri;
1039 * Calculate a time slice based on the properties of the kseg and the runq
1040 * that we're on. This is only for PRI_TIMESHARE ksegrps.
1043 sched_slice(struct kse *ke)
1049 kseq = KSEQ_CPU(ke->ke_cpu);
1051 if (ke->ke_thread->td_flags & TDF_BORROWING) {
1052 ke->ke_slice = SCHED_SLICE_MIN;
1058 * KSEs in interactive ksegs get a minimal slice so that we
1059 * quickly notice if it abuses its advantage.
1061 * KSEs in non-interactive ksegs are assigned a slice that is
1062 * based on the ksegs nice value relative to the least nice kseg
1063 * on the run queue for this cpu.
1065 * If the KSE is less nice than all others it gets the maximum
1066 * slice and other KSEs will adjust their slice relative to
1067 * this when they first expire.
1069 * There is 20 point window that starts relative to the least
1070 * nice kse on the run queue. Slice size is determined by
1071 * the kse distance from the last nice ksegrp.
1073 * If the kse is outside of the window it will get no slice
1074 * and will be reevaluated each time it is selected on the
1075 * run queue. The exception to this is nice 0 ksegs when
1076 * a nice -20 is running. They are always granted a minimum
1079 if (!SCHED_INTERACTIVE(kg)) {
1082 nice = kg->kg_proc->p_nice + (0 - kseq->ksq_nicemin);
1083 if (kseq->ksq_load_timeshare == 0 ||
1084 kg->kg_proc->p_nice < kseq->ksq_nicemin)
1085 ke->ke_slice = SCHED_SLICE_MAX;
1086 else if (nice <= SCHED_SLICE_NTHRESH)
1087 ke->ke_slice = SCHED_SLICE_NICE(nice);
1088 else if (kg->kg_proc->p_nice == 0)
1089 ke->ke_slice = SCHED_SLICE_MIN;
1093 ke->ke_slice = SCHED_SLICE_INTERACTIVE;
1099 * This routine enforces a maximum limit on the amount of scheduling history
1100 * kept. It is called after either the slptime or runtime is adjusted.
1101 * This routine will not operate correctly when slp or run times have been
1102 * adjusted to more than double their maximum.
1105 sched_interact_update(struct ksegrp *kg)
1109 sum = kg->kg_runtime + kg->kg_slptime;
1110 if (sum < SCHED_SLP_RUN_MAX)
1113 * If we have exceeded by more than 1/5th then the algorithm below
1114 * will not bring us back into range. Dividing by two here forces
1115 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1117 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1118 kg->kg_runtime /= 2;
1119 kg->kg_slptime /= 2;
1122 kg->kg_runtime = (kg->kg_runtime / 5) * 4;
1123 kg->kg_slptime = (kg->kg_slptime / 5) * 4;
1127 sched_interact_fork(struct ksegrp *kg)
1132 sum = kg->kg_runtime + kg->kg_slptime;
1133 if (sum > SCHED_SLP_RUN_FORK) {
1134 ratio = sum / SCHED_SLP_RUN_FORK;
1135 kg->kg_runtime /= ratio;
1136 kg->kg_slptime /= ratio;
1141 sched_interact_score(struct ksegrp *kg)
1145 if (kg->kg_runtime > kg->kg_slptime) {
1146 div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF);
1147 return (SCHED_INTERACT_HALF +
1148 (SCHED_INTERACT_HALF - (kg->kg_slptime / div)));
1149 } if (kg->kg_slptime > kg->kg_runtime) {
1150 div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF);
1151 return (kg->kg_runtime / div);
1155 * This can happen if slptime and runtime are 0.
1162 * Very early in the boot some setup of scheduler-specific
1163 * parts of proc0 and of soem scheduler resources needs to be done.
1171 * Set up the scheduler specific parts of proc0.
1173 proc0.p_sched = NULL; /* XXX */
1174 ksegrp0.kg_sched = &kg_sched0;
1175 thread0.td_sched = &kse0;
1176 kse0.ke_thread = &thread0;
1177 kse0.ke_state = KES_THREAD;
1178 kg_sched0.skg_concurrency = 1;
1179 kg_sched0.skg_avail_opennings = 0; /* we are already running */
1183 * This is only somewhat accurate since given many processes of the same
1184 * priority they will switch when their slices run out, which will be
1185 * at most SCHED_SLICE_MAX.
1188 sched_rr_interval(void)
1190 return (SCHED_SLICE_MAX);
1194 sched_pctcpu_update(struct kse *ke)
1197 * Adjust counters and watermark for pctcpu calc.
1199 if (ke->ke_ltick > ticks - SCHED_CPU_TICKS) {
1201 * Shift the tick count out so that the divide doesn't
1202 * round away our results.
1204 ke->ke_ticks <<= 10;
1205 ke->ke_ticks = (ke->ke_ticks / (ticks - ke->ke_ftick)) *
1207 ke->ke_ticks >>= 10;
1210 ke->ke_ltick = ticks;
1211 ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS;
1215 sched_thread_priority(struct thread *td, u_char prio)
1219 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1220 td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1221 curthread->td_proc->p_comm);
1223 mtx_assert(&sched_lock, MA_OWNED);
1224 if (td->td_priority == prio)
1226 if (TD_ON_RUNQ(td)) {
1228 * If the priority has been elevated due to priority
1229 * propagation, we may have to move ourselves to a new
1230 * queue. We still call adjustrunqueue below in case kse
1231 * needs to fix things up.
1233 if (prio < td->td_priority && ke->ke_runq != NULL &&
1234 (ke->ke_flags & KEF_ASSIGNED) == 0 &&
1235 ke->ke_runq != KSEQ_CPU(ke->ke_cpu)->ksq_curr) {
1236 runq_remove(ke->ke_runq, ke);
1237 ke->ke_runq = KSEQ_CPU(ke->ke_cpu)->ksq_curr;
1238 runq_add(ke->ke_runq, ke, 0);
1241 * Hold this kse on this cpu so that sched_prio() doesn't
1242 * cause excessive migration. We only want migration to
1243 * happen as the result of a wakeup.
1245 ke->ke_flags |= KEF_HOLD;
1246 adjustrunqueue(td, prio);
1247 ke->ke_flags &= ~KEF_HOLD;
1249 td->td_priority = prio;
1253 * Update a thread's priority when it is lent another thread's
1257 sched_lend_prio(struct thread *td, u_char prio)
1260 td->td_flags |= TDF_BORROWING;
1261 sched_thread_priority(td, prio);
1265 * Restore a thread's priority when priority propagation is
1266 * over. The prio argument is the minimum priority the thread
1267 * needs to have to satisfy other possible priority lending
1268 * requests. If the thread's regular priority is less
1269 * important than prio, the thread will keep a priority boost
1273 sched_unlend_prio(struct thread *td, u_char prio)
1277 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1278 td->td_base_pri <= PRI_MAX_TIMESHARE)
1279 base_pri = td->td_ksegrp->kg_user_pri;
1281 base_pri = td->td_base_pri;
1282 if (prio >= base_pri) {
1283 td->td_flags &= ~TDF_BORROWING;
1284 sched_thread_priority(td, base_pri);
1286 sched_lend_prio(td, prio);
1290 sched_prio(struct thread *td, u_char prio)
1294 /* First, update the base priority. */
1295 td->td_base_pri = prio;
1298 * If the thread is borrowing another thread's priority, don't
1299 * ever lower the priority.
1301 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1304 /* Change the real priority. */
1305 oldprio = td->td_priority;
1306 sched_thread_priority(td, prio);
1309 * If the thread is on a turnstile, then let the turnstile update
1312 if (TD_ON_LOCK(td) && oldprio != prio)
1313 turnstile_adjust(td, oldprio);
1317 sched_switch(struct thread *td, struct thread *newtd, int flags)
1322 mtx_assert(&sched_lock, MA_OWNED);
1327 td->td_lastcpu = td->td_oncpu;
1328 td->td_oncpu = NOCPU;
1329 td->td_flags &= ~TDF_NEEDRESCHED;
1330 td->td_owepreempt = 0;
1333 * If the KSE has been assigned it may be in the process of switching
1334 * to the new cpu. This is the case in sched_bind().
1336 if (td == PCPU_GET(idlethread)) {
1338 } else if ((ke->ke_flags & KEF_ASSIGNED) == 0) {
1339 /* We are ending our run so make our slot available again */
1340 SLOT_RELEASE(td->td_ksegrp);
1341 kseq_load_rem(ksq, ke);
1342 if (TD_IS_RUNNING(td)) {
1344 * Don't allow the thread to migrate
1345 * from a preemption.
1347 ke->ke_flags |= KEF_HOLD;
1348 setrunqueue(td, (flags & SW_PREEMPT) ?
1349 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1350 SRQ_OURSELF|SRQ_YIELDING);
1351 ke->ke_flags &= ~KEF_HOLD;
1352 } else if ((td->td_proc->p_flag & P_HADTHREADS) &&
1353 (newtd == NULL || newtd->td_ksegrp != td->td_ksegrp))
1355 * We will not be on the run queue.
1356 * So we must be sleeping or similar.
1357 * Don't use the slot if we will need it
1360 slot_fill(td->td_ksegrp);
1362 if (newtd != NULL) {
1364 * If we bring in a thread account for it as if it had been
1365 * added to the run queue and then chosen.
1367 newtd->td_kse->ke_flags |= KEF_DIDRUN;
1368 newtd->td_kse->ke_runq = ksq->ksq_curr;
1369 TD_SET_RUNNING(newtd);
1370 kseq_load_add(KSEQ_SELF(), newtd->td_kse);
1372 * XXX When we preempt, we've already consumed a slot because
1373 * we got here through sched_add(). However, newtd can come
1374 * from thread_switchout() which can't SLOT_USE() because
1375 * the SLOT code is scheduler dependent. We must use the
1376 * slot here otherwise.
1378 if ((flags & SW_PREEMPT) == 0)
1379 SLOT_USE(newtd->td_ksegrp);
1381 newtd = choosethread();
1384 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1385 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1387 cpu_switch(td, newtd);
1389 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1390 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1394 sched_lock.mtx_lock = (uintptr_t)td;
1396 td->td_oncpu = PCPU_GET(cpuid);
1400 sched_nice(struct proc *p, int nice)
1407 PROC_LOCK_ASSERT(p, MA_OWNED);
1408 mtx_assert(&sched_lock, MA_OWNED);
1410 * We need to adjust the nice counts for running KSEs.
1412 FOREACH_KSEGRP_IN_PROC(p, kg) {
1413 if (kg->kg_pri_class == PRI_TIMESHARE) {
1414 FOREACH_THREAD_IN_GROUP(kg, td) {
1416 if (ke->ke_runq == NULL)
1418 kseq = KSEQ_CPU(ke->ke_cpu);
1419 kseq_nice_rem(kseq, p->p_nice);
1420 kseq_nice_add(kseq, nice);
1425 FOREACH_KSEGRP_IN_PROC(p, kg) {
1427 FOREACH_THREAD_IN_GROUP(kg, td)
1428 td->td_flags |= TDF_NEEDRESCHED;
1433 sched_sleep(struct thread *td)
1435 mtx_assert(&sched_lock, MA_OWNED);
1437 td->td_slptime = ticks;
1441 sched_wakeup(struct thread *td)
1443 mtx_assert(&sched_lock, MA_OWNED);
1446 * Let the kseg know how long we slept for. This is because process
1447 * interactivity behavior is modeled in the kseg.
1449 if (td->td_slptime) {
1454 hzticks = (ticks - td->td_slptime) << 10;
1455 if (hzticks >= SCHED_SLP_RUN_MAX) {
1456 kg->kg_slptime = SCHED_SLP_RUN_MAX;
1459 kg->kg_slptime += hzticks;
1460 sched_interact_update(kg);
1463 sched_slice(td->td_kse);
1466 setrunqueue(td, SRQ_BORING);
1470 * Penalize the parent for creating a new child and initialize the child's
1474 sched_fork(struct thread *td, struct thread *childtd)
1477 mtx_assert(&sched_lock, MA_OWNED);
1479 sched_fork_ksegrp(td, childtd->td_ksegrp);
1480 sched_fork_thread(td, childtd);
1484 sched_fork_ksegrp(struct thread *td, struct ksegrp *child)
1486 struct ksegrp *kg = td->td_ksegrp;
1487 mtx_assert(&sched_lock, MA_OWNED);
1489 child->kg_slptime = kg->kg_slptime;
1490 child->kg_runtime = kg->kg_runtime;
1491 child->kg_user_pri = kg->kg_user_pri;
1492 sched_interact_fork(child);
1493 kg->kg_runtime += tickincr << 10;
1494 sched_interact_update(kg);
1498 sched_fork_thread(struct thread *td, struct thread *child)
1503 sched_newthread(child);
1505 ke2 = child->td_kse;
1506 ke2->ke_slice = 1; /* Attempt to quickly learn interactivity. */
1507 ke2->ke_cpu = ke->ke_cpu;
1508 ke2->ke_runq = NULL;
1510 /* Grab our parents cpu estimation information. */
1511 ke2->ke_ticks = ke->ke_ticks;
1512 ke2->ke_ltick = ke->ke_ltick;
1513 ke2->ke_ftick = ke->ke_ftick;
1517 sched_class(struct ksegrp *kg, int class)
1525 mtx_assert(&sched_lock, MA_OWNED);
1526 if (kg->kg_pri_class == class)
1529 nclass = PRI_BASE(class);
1530 oclass = PRI_BASE(kg->kg_pri_class);
1531 FOREACH_THREAD_IN_GROUP(kg, td) {
1533 if ((ke->ke_state != KES_ONRUNQ &&
1534 ke->ke_state != KES_THREAD) || ke->ke_runq == NULL)
1536 kseq = KSEQ_CPU(ke->ke_cpu);
1540 * On SMP if we're on the RUNQ we must adjust the transferable
1541 * count because could be changing to or from an interrupt
1544 if (ke->ke_state == KES_ONRUNQ) {
1545 if (KSE_CAN_MIGRATE(ke)) {
1546 kseq->ksq_transferable--;
1547 kseq->ksq_group->ksg_transferable--;
1549 if (KSE_CAN_MIGRATE(ke)) {
1550 kseq->ksq_transferable++;
1551 kseq->ksq_group->ksg_transferable++;
1555 if (oclass == PRI_TIMESHARE) {
1556 kseq->ksq_load_timeshare--;
1557 kseq_nice_rem(kseq, kg->kg_proc->p_nice);
1559 if (nclass == PRI_TIMESHARE) {
1560 kseq->ksq_load_timeshare++;
1561 kseq_nice_add(kseq, kg->kg_proc->p_nice);
1565 kg->kg_pri_class = class;
1569 * Return some of the child's priority and interactivity to the parent.
1572 sched_exit(struct proc *p, struct thread *childtd)
1574 mtx_assert(&sched_lock, MA_OWNED);
1575 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), childtd);
1576 sched_exit_thread(NULL, childtd);
1580 sched_exit_ksegrp(struct ksegrp *kg, struct thread *td)
1582 /* kg->kg_slptime += td->td_ksegrp->kg_slptime; */
1583 kg->kg_runtime += td->td_ksegrp->kg_runtime;
1584 sched_interact_update(kg);
1588 sched_exit_thread(struct thread *td, struct thread *childtd)
1590 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1591 childtd, childtd->td_proc->p_comm, childtd->td_priority);
1592 kseq_load_rem(KSEQ_CPU(childtd->td_kse->ke_cpu), childtd->td_kse);
1596 sched_clock(struct thread *td)
1602 mtx_assert(&sched_lock, MA_OWNED);
1605 if (ticks >= bal_tick)
1607 if (ticks >= gbal_tick && balance_groups)
1608 sched_balance_groups();
1610 * We could have been assigned a non real-time thread without an
1613 if (kseq->ksq_assigned)
1614 kseq_assign(kseq); /* Potentially sets NEEDRESCHED */
1617 * sched_setup() apparently happens prior to stathz being set. We
1618 * need to resolve the timers earlier in the boot so we can avoid
1619 * calculating this here.
1621 if (realstathz == 0) {
1622 realstathz = stathz ? stathz : hz;
1623 tickincr = hz / realstathz;
1625 * XXX This does not work for values of stathz that are much
1635 /* Adjust ticks for pctcpu */
1637 ke->ke_ltick = ticks;
1639 /* Go up to one second beyond our max and then trim back down */
1640 if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick)
1641 sched_pctcpu_update(ke);
1643 if (td->td_flags & TDF_IDLETD)
1646 * We only do slicing code for TIMESHARE ksegrps.
1648 if (kg->kg_pri_class != PRI_TIMESHARE)
1651 * We used a tick charge it to the ksegrp so that we can compute our
1654 kg->kg_runtime += tickincr << 10;
1655 sched_interact_update(kg);
1658 * We used up one time slice.
1660 if (--ke->ke_slice > 0)
1663 * We're out of time, recompute priorities and requeue.
1665 kseq_load_rem(kseq, ke);
1668 if (SCHED_CURR(kg, ke))
1669 ke->ke_runq = kseq->ksq_curr;
1671 ke->ke_runq = kseq->ksq_next;
1672 kseq_load_add(kseq, ke);
1673 td->td_flags |= TDF_NEEDRESCHED;
1677 sched_runnable(void)
1686 if (kseq->ksq_assigned) {
1687 mtx_lock_spin(&sched_lock);
1689 mtx_unlock_spin(&sched_lock);
1692 if ((curthread->td_flags & TDF_IDLETD) != 0) {
1693 if (kseq->ksq_load > 0)
1696 if (kseq->ksq_load - 1 > 0)
1704 sched_userret(struct thread *td)
1708 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1709 ("thread with borrowed priority returning to userland"));
1711 if (td->td_priority != kg->kg_user_pri) {
1712 mtx_lock_spin(&sched_lock);
1713 td->td_priority = kg->kg_user_pri;
1714 td->td_base_pri = kg->kg_user_pri;
1715 mtx_unlock_spin(&sched_lock);
1725 mtx_assert(&sched_lock, MA_OWNED);
1729 if (kseq->ksq_assigned)
1732 ke = kseq_choose(kseq);
1735 if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE)
1736 if (kseq_idled(kseq) == 0)
1739 kseq_runq_rem(kseq, ke);
1740 ke->ke_state = KES_THREAD;
1741 ke->ke_flags &= ~KEF_PREEMPTED;
1745 if (kseq_idled(kseq) == 0)
1752 sched_add(struct thread *td, int flags)
1761 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1762 td, td->td_proc->p_comm, td->td_priority, curthread,
1763 curthread->td_proc->p_comm);
1764 mtx_assert(&sched_lock, MA_OWNED);
1768 preemptive = !(flags & SRQ_YIELDING);
1769 class = PRI_BASE(kg->kg_pri_class);
1771 if ((ke->ke_flags & KEF_INTERNAL) == 0)
1772 SLOT_USE(td->td_ksegrp);
1773 ke->ke_flags &= ~KEF_INTERNAL;
1775 if (ke->ke_flags & KEF_ASSIGNED) {
1776 if (ke->ke_flags & KEF_REMOVED)
1777 ke->ke_flags &= ~KEF_REMOVED;
1780 canmigrate = KSE_CAN_MIGRATE(ke);
1782 KASSERT(ke->ke_state != KES_ONRUNQ,
1783 ("sched_add: kse %p (%s) already in run queue", ke,
1784 ke->ke_proc->p_comm));
1785 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
1786 ("sched_add: process swapped out"));
1787 KASSERT(ke->ke_runq == NULL,
1788 ("sched_add: KSE %p is still assigned to a run queue", ke));
1789 if (flags & SRQ_PREEMPTED)
1790 ke->ke_flags |= KEF_PREEMPTED;
1794 ke->ke_runq = kseq->ksq_curr;
1795 ke->ke_slice = SCHED_SLICE_MAX;
1797 ke->ke_cpu = PCPU_GET(cpuid);
1800 if (SCHED_CURR(kg, ke))
1801 ke->ke_runq = kseq->ksq_curr;
1803 ke->ke_runq = kseq->ksq_next;
1807 * This is for priority prop.
1809 if (ke->ke_thread->td_priority < PRI_MIN_IDLE)
1810 ke->ke_runq = kseq->ksq_curr;
1812 ke->ke_runq = &kseq->ksq_idle;
1813 ke->ke_slice = SCHED_SLICE_MIN;
1816 panic("Unknown pri class.");
1821 * Don't migrate running threads here. Force the long term balancer
1824 if (ke->ke_flags & KEF_HOLD) {
1825 ke->ke_flags &= ~KEF_HOLD;
1829 * If this thread is pinned or bound, notify the target cpu.
1831 if (!canmigrate && ke->ke_cpu != PCPU_GET(cpuid) ) {
1833 kseq_notify(ke, ke->ke_cpu);
1837 * If we had been idle, clear our bit in the group and potentially
1838 * the global bitmap. If not, see if we should transfer this thread.
1840 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1841 (kseq->ksq_group->ksg_idlemask & PCPU_GET(cpumask)) != 0) {
1843 * Check to see if our group is unidling, and if so, remove it
1844 * from the global idle mask.
1846 if (kseq->ksq_group->ksg_idlemask ==
1847 kseq->ksq_group->ksg_cpumask)
1848 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
1850 * Now remove ourselves from the group specific idle mask.
1852 kseq->ksq_group->ksg_idlemask &= ~PCPU_GET(cpumask);
1853 } else if (canmigrate && kseq->ksq_load > 1 && class != PRI_ITHD)
1854 if (kseq_transfer(kseq, ke, class))
1856 ke->ke_cpu = PCPU_GET(cpuid);
1858 if (td->td_priority < curthread->td_priority &&
1859 ke->ke_runq == kseq->ksq_curr)
1860 curthread->td_flags |= TDF_NEEDRESCHED;
1861 if (preemptive && maybe_preempt(td))
1863 ke->ke_state = KES_ONRUNQ;
1865 kseq_runq_add(kseq, ke, flags);
1866 kseq_load_add(kseq, ke);
1870 sched_rem(struct thread *td)
1875 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1876 td, td->td_proc->p_comm, td->td_priority, curthread,
1877 curthread->td_proc->p_comm);
1878 mtx_assert(&sched_lock, MA_OWNED);
1880 SLOT_RELEASE(td->td_ksegrp);
1881 ke->ke_flags &= ~KEF_PREEMPTED;
1882 if (ke->ke_flags & KEF_ASSIGNED) {
1883 ke->ke_flags |= KEF_REMOVED;
1886 KASSERT((ke->ke_state == KES_ONRUNQ),
1887 ("sched_rem: KSE not on run queue"));
1889 ke->ke_state = KES_THREAD;
1890 kseq = KSEQ_CPU(ke->ke_cpu);
1891 kseq_runq_rem(kseq, ke);
1892 kseq_load_rem(kseq, ke);
1896 sched_pctcpu(struct thread *td)
1906 mtx_lock_spin(&sched_lock);
1911 * Don't update more frequently than twice a second. Allowing
1912 * this causes the cpu usage to decay away too quickly due to
1915 if (ke->ke_ftick + SCHED_CPU_TICKS < ke->ke_ltick ||
1916 ke->ke_ltick < (ticks - (hz / 2)))
1917 sched_pctcpu_update(ke);
1918 /* How many rtick per second ? */
1919 rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1920 pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1923 ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick;
1924 mtx_unlock_spin(&sched_lock);
1930 sched_bind(struct thread *td, int cpu)
1934 mtx_assert(&sched_lock, MA_OWNED);
1936 ke->ke_flags |= KEF_BOUND;
1938 if (PCPU_GET(cpuid) == cpu)
1940 /* sched_rem without the runq_remove */
1941 ke->ke_state = KES_THREAD;
1942 kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1943 kseq_notify(ke, cpu);
1944 /* When we return from mi_switch we'll be on the correct cpu. */
1945 mi_switch(SW_VOL, NULL);
1950 sched_unbind(struct thread *td)
1952 mtx_assert(&sched_lock, MA_OWNED);
1953 td->td_kse->ke_flags &= ~KEF_BOUND;
1957 sched_is_bound(struct thread *td)
1959 mtx_assert(&sched_lock, MA_OWNED);
1960 return (td->td_kse->ke_flags & KEF_BOUND);
1971 for (i = 0; i <= ksg_maxid; i++)
1972 total += KSEQ_GROUP(i)->ksg_load;
1975 return (KSEQ_SELF()->ksq_sysload);
1980 sched_sizeof_ksegrp(void)
1982 return (sizeof(struct ksegrp) + sizeof(struct kg_sched));
1986 sched_sizeof_proc(void)
1988 return (sizeof(struct proc));
1992 sched_sizeof_thread(void)
1994 return (sizeof(struct thread) + sizeof(struct td_sched));
1996 #define KERN_SWITCH_INCLUDE 1
1997 #include "kern/kern_switch.c"