2 * Copyright (c) 2002-2003, 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 schedulable entity that can be given a context to run.
87 * A process may have several of these. Probably one per processor
88 * but posibly a few more. In this universe they are grouped
89 * with a KSEG that contains the priority and niceness
93 TAILQ_ENTRY(kse) ke_procq; /* (j/z) Run queue. */
94 int ke_flags; /* (j) KEF_* flags. */
95 struct thread *ke_thread; /* (*) Active associated thread. */
96 fixpt_t ke_pctcpu; /* (j) %cpu during p_swtime. */
97 char ke_rqindex; /* (j) Run queue index. */
99 KES_THREAD = 0x0, /* slaved to thread state */
101 } ke_state; /* (j) thread sched specific status. */
104 struct runq *ke_runq;
105 u_char ke_cpu; /* CPU that we have affinity for. */
106 /* The following variables are only used for pctcpu calculation */
107 int ke_ltick; /* Last tick that we were running on */
108 int ke_ftick; /* First tick that we were running on */
109 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
119 /* flags kept in ke_flags */
120 #define KEF_SCHED0 0x00001 /* For scheduler-specific use. */
121 #define KEF_SCHED1 0x00002 /* For scheduler-specific use. */
122 #define KEF_SCHED2 0x00004 /* For scheduler-specific use. */
123 #define KEF_SCHED3 0x00008 /* For scheduler-specific use. */
124 #define KEF_SCHED4 0x00010
125 #define KEF_SCHED5 0x00020
126 #define KEF_DIDRUN 0x02000 /* Thread actually ran. */
127 #define KEF_EXIT 0x04000 /* Thread is being killed. */
130 * These datastructures are allocated within their parent datastructure but
131 * are scheduler specific.
134 #define ke_assign ke_procq.tqe_next
136 #define KEF_ASSIGNED 0x0001 /* Thread is being migrated. */
137 #define KEF_BOUND 0x0002 /* Thread can not migrate. */
138 #define KEF_XFERABLE 0x0004 /* Thread was added as transferable. */
139 #define KEF_HOLD 0x0008 /* Thread is temporarily bound. */
140 #define KEF_REMOVED 0x0010 /* Thread was removed while ASSIGNED */
141 #define KEF_INTERNAL 0x0020
144 struct thread *skg_last_assigned; /* (j) Last thread assigned to */
145 /* the system scheduler */
146 int skg_slptime; /* Number of ticks we vol. slept */
147 int skg_runtime; /* Number of ticks we were running */
148 int skg_avail_opennings; /* (j) Num unfilled slots in group.*/
149 int skg_concurrency; /* (j) Num threads requested in group.*/
151 #define kg_last_assigned kg_sched->skg_last_assigned
152 #define kg_avail_opennings kg_sched->skg_avail_opennings
153 #define kg_concurrency kg_sched->skg_concurrency
154 #define kg_runtime kg_sched->skg_runtime
155 #define kg_slptime kg_sched->skg_slptime
157 #define SLOT_RELEASE(kg) \
159 kg->kg_avail_opennings++; \
160 CTR3(KTR_RUNQ, "kg %p(%d) Slot released (->%d)", \
162 kg->kg_concurrency, \
163 kg->kg_avail_opennings); \
164 /*KASSERT((kg->kg_avail_opennings <= kg->kg_concurrency), \
165 ("slots out of whack")); */ \
168 #define SLOT_USE(kg) \
170 kg->kg_avail_opennings--; \
171 CTR3(KTR_RUNQ, "kg %p(%d) Slot used (->%d)", \
173 kg->kg_concurrency, \
174 kg->kg_avail_opennings); \
175 /*KASSERT((kg->kg_avail_opennings >= 0), \
176 ("slots out of whack"));*/ \
179 static struct kse kse0;
180 static struct kg_sched kg_sched0;
183 * The priority is primarily determined by the interactivity score. Thus, we
184 * give lower(better) priorities to kse groups that use less CPU. The nice
185 * value is then directly added to this to allow nice to have some effect
188 * PRI_RANGE: Total priority range for timeshare threads.
189 * PRI_NRESV: Number of nice values.
190 * PRI_BASE: The start of the dynamic range.
192 #define SCHED_PRI_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
193 #define SCHED_PRI_NRESV ((PRIO_MAX - PRIO_MIN) + 1)
194 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
195 #define SCHED_PRI_BASE (PRI_MIN_TIMESHARE)
196 #define SCHED_PRI_INTERACT(score) \
197 ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
200 * These determine the interactivity of a process.
202 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
203 * before throttling back.
204 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
205 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
206 * INTERACT_THRESH: Threshhold for placement on the current runq.
208 #define SCHED_SLP_RUN_MAX ((hz * 5) << 10)
209 #define SCHED_SLP_RUN_FORK ((hz / 2) << 10)
210 #define SCHED_INTERACT_MAX (100)
211 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
212 #define SCHED_INTERACT_THRESH (30)
215 * These parameters and macros determine the size of the time slice that is
216 * granted to each thread.
218 * SLICE_MIN: Minimum time slice granted, in units of ticks.
219 * SLICE_MAX: Maximum time slice granted.
220 * SLICE_RANGE: Range of available time slices scaled by hz.
221 * SLICE_SCALE: The number slices granted per val in the range of [0, max].
222 * SLICE_NICE: Determine the amount of slice granted to a scaled nice.
223 * SLICE_NTHRESH: The nice cutoff point for slice assignment.
225 #define SCHED_SLICE_MIN (slice_min)
226 #define SCHED_SLICE_MAX (slice_max)
227 #define SCHED_SLICE_INTERACTIVE (slice_max)
228 #define SCHED_SLICE_NTHRESH (SCHED_PRI_NHALF - 1)
229 #define SCHED_SLICE_RANGE (SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
230 #define SCHED_SLICE_SCALE(val, max) (((val) * SCHED_SLICE_RANGE) / (max))
231 #define SCHED_SLICE_NICE(nice) \
232 (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
235 * This macro determines whether or not the thread belongs on the current or
238 #define SCHED_INTERACTIVE(kg) \
239 (sched_interact_score(kg) < SCHED_INTERACT_THRESH)
240 #define SCHED_CURR(kg, ke) \
241 ((ke->ke_thread->td_flags & TDF_BORROWING) || SCHED_INTERACTIVE(kg))
244 * Cpu percentage computation macros and defines.
246 * SCHED_CPU_TIME: Number of seconds to average the cpu usage across.
247 * SCHED_CPU_TICKS: Number of hz ticks to average the cpu usage across.
250 #define SCHED_CPU_TIME 10
251 #define SCHED_CPU_TICKS (hz * SCHED_CPU_TIME)
254 * kseq - per processor runqs and statistics.
257 struct runq ksq_idle; /* Queue of IDLE threads. */
258 struct runq ksq_timeshare[2]; /* Run queues for !IDLE. */
259 struct runq *ksq_next; /* Next timeshare queue. */
260 struct runq *ksq_curr; /* Current queue. */
261 int ksq_load_timeshare; /* Load for timeshare. */
262 int ksq_load; /* Aggregate load. */
263 short ksq_nice[SCHED_PRI_NRESV]; /* KSEs in each nice bin. */
264 short ksq_nicemin; /* Least nice. */
266 int ksq_transferable;
267 LIST_ENTRY(kseq) ksq_siblings; /* Next in kseq group. */
268 struct kseq_group *ksq_group; /* Our processor group. */
269 volatile struct kse *ksq_assigned; /* assigned by another CPU. */
271 int ksq_sysload; /* For loadavg, !ITHD load. */
277 * kseq groups are groups of processors which can cheaply share threads. When
278 * one processor in the group goes idle it will check the runqs of the other
279 * processors in its group prior to halting and waiting for an interrupt.
280 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
281 * In a numa environment we'd want an idle bitmap per group and a two tiered
285 int ksg_cpus; /* Count of CPUs in this kseq group. */
286 cpumask_t ksg_cpumask; /* Mask of cpus in this group. */
287 cpumask_t ksg_idlemask; /* Idle cpus in this group. */
288 cpumask_t ksg_mask; /* Bit mask for first cpu. */
289 int ksg_load; /* Total load of this group. */
290 int ksg_transferable; /* Transferable load of this group. */
291 LIST_HEAD(, kseq) ksg_members; /* Linked list of all members. */
296 * One kse queue per processor.
299 static cpumask_t kseq_idle;
300 static int ksg_maxid;
301 static struct kseq kseq_cpu[MAXCPU];
302 static struct kseq_group kseq_groups[MAXCPU];
304 static int gbal_tick;
305 static int balance_groups;
307 #define KSEQ_SELF() (&kseq_cpu[PCPU_GET(cpuid)])
308 #define KSEQ_CPU(x) (&kseq_cpu[(x)])
309 #define KSEQ_ID(x) ((x) - kseq_cpu)
310 #define KSEQ_GROUP(x) (&kseq_groups[(x)])
312 static struct kseq kseq_cpu;
314 #define KSEQ_SELF() (&kseq_cpu)
315 #define KSEQ_CPU(x) (&kseq_cpu)
318 static void slot_fill(struct ksegrp *kg);
319 static struct kse *sched_choose(void); /* XXX Should be thread * */
320 static void sched_slice(struct kse *ke);
321 static void sched_priority(struct ksegrp *kg);
322 static void sched_thread_priority(struct thread *td, u_char prio);
323 static int sched_interact_score(struct ksegrp *kg);
324 static void sched_interact_update(struct ksegrp *kg);
325 static void sched_interact_fork(struct ksegrp *kg);
326 static void sched_pctcpu_update(struct kse *ke);
328 /* Operations on per processor queues */
329 static struct kse * kseq_choose(struct kseq *kseq);
330 static void kseq_setup(struct kseq *kseq);
331 static void kseq_load_add(struct kseq *kseq, struct kse *ke);
332 static void kseq_load_rem(struct kseq *kseq, struct kse *ke);
333 static __inline void kseq_runq_add(struct kseq *kseq, struct kse *ke, int);
334 static __inline void kseq_runq_rem(struct kseq *kseq, struct kse *ke);
335 static void kseq_nice_add(struct kseq *kseq, int nice);
336 static void kseq_nice_rem(struct kseq *kseq, int nice);
337 void kseq_print(int cpu);
339 static int kseq_transfer(struct kseq *ksq, struct kse *ke, int class);
340 static struct kse *runq_steal(struct runq *rq);
341 static void sched_balance(void);
342 static void sched_balance_groups(void);
343 static void sched_balance_group(struct kseq_group *ksg);
344 static void sched_balance_pair(struct kseq *high, struct kseq *low);
345 static void kseq_move(struct kseq *from, int cpu);
346 static int kseq_idled(struct kseq *kseq);
347 static void kseq_notify(struct kse *ke, int cpu);
348 static void kseq_assign(struct kseq *);
349 static struct kse *kseq_steal(struct kseq *kseq, int stealidle);
350 #define KSE_CAN_MIGRATE(ke) \
351 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
360 kseq = KSEQ_CPU(cpu);
363 printf("\tload: %d\n", kseq->ksq_load);
364 printf("\tload TIMESHARE: %d\n", kseq->ksq_load_timeshare);
366 printf("\tload transferable: %d\n", kseq->ksq_transferable);
368 printf("\tnicemin:\t%d\n", kseq->ksq_nicemin);
369 printf("\tnice counts:\n");
370 for (i = 0; i < SCHED_PRI_NRESV; i++)
371 if (kseq->ksq_nice[i])
372 printf("\t\t%d = %d\n",
373 i - SCHED_PRI_NHALF, kseq->ksq_nice[i]);
377 kseq_runq_add(struct kseq *kseq, struct kse *ke, int flags)
380 if (KSE_CAN_MIGRATE(ke)) {
381 kseq->ksq_transferable++;
382 kseq->ksq_group->ksg_transferable++;
383 ke->ke_flags |= KEF_XFERABLE;
386 runq_add(ke->ke_runq, ke, flags);
390 kseq_runq_rem(struct kseq *kseq, struct kse *ke)
393 if (ke->ke_flags & KEF_XFERABLE) {
394 kseq->ksq_transferable--;
395 kseq->ksq_group->ksg_transferable--;
396 ke->ke_flags &= ~KEF_XFERABLE;
399 runq_remove(ke->ke_runq, ke);
403 kseq_load_add(struct kseq *kseq, struct kse *ke)
406 mtx_assert(&sched_lock, MA_OWNED);
407 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
408 if (class == PRI_TIMESHARE)
409 kseq->ksq_load_timeshare++;
411 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
412 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
414 kseq->ksq_group->ksg_load++;
418 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
419 kseq_nice_add(kseq, ke->ke_proc->p_nice);
423 kseq_load_rem(struct kseq *kseq, struct kse *ke)
426 mtx_assert(&sched_lock, MA_OWNED);
427 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
428 if (class == PRI_TIMESHARE)
429 kseq->ksq_load_timeshare--;
430 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
432 kseq->ksq_group->ksg_load--;
437 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
439 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
440 kseq_nice_rem(kseq, ke->ke_proc->p_nice);
444 kseq_nice_add(struct kseq *kseq, int nice)
446 mtx_assert(&sched_lock, MA_OWNED);
447 /* Normalize to zero. */
448 kseq->ksq_nice[nice + SCHED_PRI_NHALF]++;
449 if (nice < kseq->ksq_nicemin || kseq->ksq_load_timeshare == 1)
450 kseq->ksq_nicemin = nice;
454 kseq_nice_rem(struct kseq *kseq, int nice)
458 mtx_assert(&sched_lock, MA_OWNED);
459 /* Normalize to zero. */
460 n = nice + SCHED_PRI_NHALF;
462 KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count."));
465 * If this wasn't the smallest nice value or there are more in
466 * this bucket we can just return. Otherwise we have to recalculate
469 if (nice != kseq->ksq_nicemin ||
470 kseq->ksq_nice[n] != 0 ||
471 kseq->ksq_load_timeshare == 0)
474 for (; n < SCHED_PRI_NRESV; n++)
475 if (kseq->ksq_nice[n]) {
476 kseq->ksq_nicemin = n - SCHED_PRI_NHALF;
483 * sched_balance is a simple CPU load balancing algorithm. It operates by
484 * finding the least loaded and most loaded cpu and equalizing their load
485 * by migrating some processes.
487 * Dealing only with two CPUs at a time has two advantages. Firstly, most
488 * installations will only have 2 cpus. Secondly, load balancing too much at
489 * once can have an unpleasant effect on the system. The scheduler rarely has
490 * enough information to make perfect decisions. So this algorithm chooses
491 * algorithm simplicity and more gradual effects on load in larger systems.
493 * It could be improved by considering the priorities and slices assigned to
494 * each task prior to balancing them. There are many pathological cases with
495 * any approach and so the semi random algorithm below may work as well as any.
501 struct kseq_group *high;
502 struct kseq_group *low;
503 struct kseq_group *ksg;
507 bal_tick = ticks + (random() % (hz * 2));
508 if (smp_started == 0)
511 i = random() % (ksg_maxid + 1);
512 for (cnt = 0; cnt <= ksg_maxid; cnt++) {
515 * Find the CPU with the highest load that has some
516 * threads to transfer.
518 if ((high == NULL || ksg->ksg_load > high->ksg_load)
519 && ksg->ksg_transferable)
521 if (low == NULL || ksg->ksg_load < low->ksg_load)
526 if (low != NULL && high != NULL && high != low)
527 sched_balance_pair(LIST_FIRST(&high->ksg_members),
528 LIST_FIRST(&low->ksg_members));
532 sched_balance_groups(void)
536 gbal_tick = ticks + (random() % (hz * 2));
537 mtx_assert(&sched_lock, MA_OWNED);
539 for (i = 0; i <= ksg_maxid; i++)
540 sched_balance_group(KSEQ_GROUP(i));
544 sched_balance_group(struct kseq_group *ksg)
551 if (ksg->ksg_transferable == 0)
555 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
556 load = kseq->ksq_load;
557 if (high == NULL || load > high->ksq_load)
559 if (low == NULL || load < low->ksq_load)
562 if (high != NULL && low != NULL && high != low)
563 sched_balance_pair(high, low);
567 sched_balance_pair(struct kseq *high, struct kseq *low)
577 * If we're transfering within a group we have to use this specific
578 * kseq's transferable count, otherwise we can steal from other members
581 if (high->ksq_group == low->ksq_group) {
582 transferable = high->ksq_transferable;
583 high_load = high->ksq_load;
584 low_load = low->ksq_load;
586 transferable = high->ksq_group->ksg_transferable;
587 high_load = high->ksq_group->ksg_load;
588 low_load = low->ksq_group->ksg_load;
590 if (transferable == 0)
593 * Determine what the imbalance is and then adjust that to how many
594 * kses we actually have to give up (transferable).
596 diff = high_load - low_load;
600 move = min(move, transferable);
601 for (i = 0; i < move; i++)
602 kseq_move(high, KSEQ_ID(low));
607 kseq_move(struct kseq *from, int cpu)
615 ke = kseq_steal(kseq, 1);
617 struct kseq_group *ksg;
619 ksg = kseq->ksq_group;
620 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
621 if (kseq == from || kseq->ksq_transferable == 0)
623 ke = kseq_steal(kseq, 1);
627 panic("kseq_move: No KSEs available with a "
628 "transferable count of %d\n",
629 ksg->ksg_transferable);
633 ke->ke_state = KES_THREAD;
634 kseq_runq_rem(kseq, ke);
635 kseq_load_rem(kseq, ke);
636 kseq_notify(ke, cpu);
640 kseq_idled(struct kseq *kseq)
642 struct kseq_group *ksg;
646 ksg = kseq->ksq_group;
648 * If we're in a cpu group, try and steal kses from another cpu in
649 * the group before idling.
651 if (ksg->ksg_cpus > 1 && ksg->ksg_transferable) {
652 LIST_FOREACH(steal, &ksg->ksg_members, ksq_siblings) {
653 if (steal == kseq || steal->ksq_transferable == 0)
655 ke = kseq_steal(steal, 0);
658 ke->ke_state = KES_THREAD;
659 kseq_runq_rem(steal, ke);
660 kseq_load_rem(steal, ke);
661 ke->ke_cpu = PCPU_GET(cpuid);
662 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
663 sched_add(ke->ke_thread, SRQ_YIELDING);
668 * We only set the idled bit when all of the cpus in the group are
669 * idle. Otherwise we could get into a situation where a KSE bounces
670 * back and forth between two idle cores on seperate physical CPUs.
672 ksg->ksg_idlemask |= PCPU_GET(cpumask);
673 if (ksg->ksg_idlemask != ksg->ksg_cpumask)
675 atomic_set_int(&kseq_idle, ksg->ksg_mask);
680 kseq_assign(struct kseq *kseq)
686 *(volatile struct kse **)&ke = kseq->ksq_assigned;
687 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke, NULL));
688 for (; ke != NULL; ke = nke) {
690 kseq->ksq_group->ksg_load--;
692 ke->ke_flags &= ~KEF_ASSIGNED;
693 ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
694 sched_add(ke->ke_thread, SRQ_YIELDING);
699 kseq_notify(struct kse *ke, int cpu)
707 kseq = KSEQ_CPU(cpu);
709 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
710 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
711 (kseq_idle & kseq->ksq_group->ksg_mask))
712 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
713 kseq->ksq_group->ksg_load++;
716 ke->ke_flags |= KEF_ASSIGNED;
717 prio = ke->ke_thread->td_priority;
720 * Place a KSE on another cpu's queue and force a resched.
723 *(volatile struct kse **)&ke->ke_assign = kseq->ksq_assigned;
724 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke->ke_assign, ke));
726 * Without sched_lock we could lose a race where we set NEEDRESCHED
727 * on a thread that is switched out before the IPI is delivered. This
728 * would lead us to miss the resched. This will be a problem once
729 * sched_lock is pushed down.
731 pcpu = pcpu_find(cpu);
732 td = pcpu->pc_curthread;
733 if (ke->ke_thread->td_priority < td->td_priority ||
734 td == pcpu->pc_idlethread) {
735 td->td_flags |= TDF_NEEDRESCHED;
736 ipi_selected(1 << cpu, IPI_AST);
741 runq_steal(struct runq *rq)
749 mtx_assert(&sched_lock, MA_OWNED);
750 rqb = &rq->rq_status;
751 for (word = 0; word < RQB_LEN; word++) {
752 if (rqb->rqb_bits[word] == 0)
754 for (bit = 0; bit < RQB_BPW; bit++) {
755 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
757 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
758 TAILQ_FOREACH(ke, rqh, ke_procq) {
759 if (KSE_CAN_MIGRATE(ke))
768 kseq_steal(struct kseq *kseq, int stealidle)
773 * Steal from next first to try to get a non-interactive task that
774 * may not have run for a while.
776 if ((ke = runq_steal(kseq->ksq_next)) != NULL)
778 if ((ke = runq_steal(kseq->ksq_curr)) != NULL)
781 return (runq_steal(&kseq->ksq_idle));
786 kseq_transfer(struct kseq *kseq, struct kse *ke, int class)
788 struct kseq_group *nksg;
789 struct kseq_group *ksg;
794 if (smp_started == 0)
798 * If our load exceeds a certain threshold we should attempt to
799 * reassign this thread. The first candidate is the cpu that
800 * originally ran the thread. If it is idle, assign it there,
801 * otherwise, pick an idle cpu.
803 * The threshold at which we start to reassign kses has a large impact
804 * on the overall performance of the system. Tuned too high and
805 * some CPUs may idle. Too low and there will be excess migration
806 * and context switches.
808 old = KSEQ_CPU(ke->ke_cpu);
809 nksg = old->ksq_group;
810 ksg = kseq->ksq_group;
812 if (kseq_idle & nksg->ksg_mask) {
813 cpu = ffs(nksg->ksg_idlemask);
816 "kseq_transfer: %p found old cpu %X "
817 "in idlemask.", ke, cpu);
822 * Multiple cpus could find this bit simultaneously
823 * but the race shouldn't be terrible.
825 cpu = ffs(kseq_idle);
827 CTR2(KTR_SCHED, "kseq_transfer: %p found %X "
828 "in idlemask.", ke, cpu);
834 if (old->ksq_load < kseq->ksq_load) {
835 cpu = ke->ke_cpu + 1;
836 CTR2(KTR_SCHED, "kseq_transfer: %p old cpu %X "
837 "load less than ours.", ke, cpu);
841 * No new CPU was found, look for one with less load.
843 for (idx = 0; idx <= ksg_maxid; idx++) {
844 nksg = KSEQ_GROUP(idx);
845 if (nksg->ksg_load /*+ (nksg->ksg_cpus * 2)*/ < ksg->ksg_load) {
846 cpu = ffs(nksg->ksg_cpumask);
847 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X load less "
848 "than ours.", ke, cpu);
854 * If another cpu in this group has idled, assign a thread over
855 * to them after checking to see if there are idled groups.
857 if (ksg->ksg_idlemask) {
858 cpu = ffs(ksg->ksg_idlemask);
860 CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X idle in "
868 * Now that we've found an idle CPU, migrate the thread.
872 kseq_notify(ke, cpu);
880 * Pick the highest priority task we have and return it.
884 kseq_choose(struct kseq *kseq)
890 mtx_assert(&sched_lock, MA_OWNED);
894 ke = runq_choose(kseq->ksq_curr);
897 * We already swapped once and didn't get anywhere.
901 swap = kseq->ksq_curr;
902 kseq->ksq_curr = kseq->ksq_next;
903 kseq->ksq_next = swap;
907 * If we encounter a slice of 0 the kse is in a
908 * TIMESHARE kse group and its nice was too far out
909 * of the range that receives slices.
911 nice = ke->ke_proc->p_nice + (0 - kseq->ksq_nicemin);
912 if (ke->ke_slice == 0 || (nice > SCHED_SLICE_NTHRESH &&
913 ke->ke_proc->p_nice != 0)) {
914 runq_remove(ke->ke_runq, ke);
916 ke->ke_runq = kseq->ksq_next;
917 runq_add(ke->ke_runq, ke, 0);
923 return (runq_choose(&kseq->ksq_idle));
927 kseq_setup(struct kseq *kseq)
929 runq_init(&kseq->ksq_timeshare[0]);
930 runq_init(&kseq->ksq_timeshare[1]);
931 runq_init(&kseq->ksq_idle);
932 kseq->ksq_curr = &kseq->ksq_timeshare[0];
933 kseq->ksq_next = &kseq->ksq_timeshare[1];
935 kseq->ksq_load_timeshare = 0;
939 sched_setup(void *dummy)
945 slice_min = (hz/100); /* 10ms */
946 slice_max = (hz/7); /* ~140ms */
951 * Initialize the kseqs.
953 for (i = 0; i < MAXCPU; i++) {
957 ksq->ksq_assigned = NULL;
958 kseq_setup(&kseq_cpu[i]);
960 if (smp_topology == NULL) {
961 struct kseq_group *ksg;
965 for (cpus = 0, i = 0; i < MAXCPU; i++) {
968 ksq = &kseq_cpu[cpus];
969 ksg = &kseq_groups[cpus];
971 * Setup a kseq group with one member.
973 ksq->ksq_transferable = 0;
974 ksq->ksq_group = ksg;
976 ksg->ksg_idlemask = 0;
977 ksg->ksg_cpumask = ksg->ksg_mask = 1 << i;
979 ksg->ksg_transferable = 0;
980 LIST_INIT(&ksg->ksg_members);
981 LIST_INSERT_HEAD(&ksg->ksg_members, ksq, ksq_siblings);
984 ksg_maxid = cpus - 1;
986 struct kseq_group *ksg;
987 struct cpu_group *cg;
990 for (i = 0; i < smp_topology->ct_count; i++) {
991 cg = &smp_topology->ct_group[i];
992 ksg = &kseq_groups[i];
994 * Initialize the group.
996 ksg->ksg_idlemask = 0;
998 ksg->ksg_transferable = 0;
999 ksg->ksg_cpus = cg->cg_count;
1000 ksg->ksg_cpumask = cg->cg_mask;
1001 LIST_INIT(&ksg->ksg_members);
1003 * Find all of the group members and add them.
1005 for (j = 0; j < MAXCPU; j++) {
1006 if ((cg->cg_mask & (1 << j)) != 0) {
1007 if (ksg->ksg_mask == 0)
1008 ksg->ksg_mask = 1 << j;
1009 kseq_cpu[j].ksq_transferable = 0;
1010 kseq_cpu[j].ksq_group = ksg;
1011 LIST_INSERT_HEAD(&ksg->ksg_members,
1012 &kseq_cpu[j], ksq_siblings);
1015 if (ksg->ksg_cpus > 1)
1018 ksg_maxid = smp_topology->ct_count - 1;
1021 * Stagger the group and global load balancer so they do not
1022 * interfere with each other.
1024 bal_tick = ticks + hz;
1026 gbal_tick = ticks + (hz / 2);
1028 kseq_setup(KSEQ_SELF());
1030 mtx_lock_spin(&sched_lock);
1031 kseq_load_add(KSEQ_SELF(), &kse0);
1032 mtx_unlock_spin(&sched_lock);
1036 * Scale the scheduling priority according to the "interactivity" of this
1040 sched_priority(struct ksegrp *kg)
1044 if (kg->kg_pri_class != PRI_TIMESHARE)
1047 pri = SCHED_PRI_INTERACT(sched_interact_score(kg));
1048 pri += SCHED_PRI_BASE;
1049 pri += kg->kg_proc->p_nice;
1051 if (pri > PRI_MAX_TIMESHARE)
1052 pri = PRI_MAX_TIMESHARE;
1053 else if (pri < PRI_MIN_TIMESHARE)
1054 pri = PRI_MIN_TIMESHARE;
1056 kg->kg_user_pri = pri;
1062 * Calculate a time slice based on the properties of the kseg and the runq
1063 * that we're on. This is only for PRI_TIMESHARE ksegrps.
1066 sched_slice(struct kse *ke)
1072 kseq = KSEQ_CPU(ke->ke_cpu);
1074 if (ke->ke_thread->td_flags & TDF_BORROWING) {
1075 ke->ke_slice = SCHED_SLICE_MIN;
1081 * KSEs in interactive ksegs get a minimal slice so that we
1082 * quickly notice if it abuses its advantage.
1084 * KSEs in non-interactive ksegs are assigned a slice that is
1085 * based on the ksegs nice value relative to the least nice kseg
1086 * on the run queue for this cpu.
1088 * If the KSE is less nice than all others it gets the maximum
1089 * slice and other KSEs will adjust their slice relative to
1090 * this when they first expire.
1092 * There is 20 point window that starts relative to the least
1093 * nice kse on the run queue. Slice size is determined by
1094 * the kse distance from the last nice ksegrp.
1096 * If the kse is outside of the window it will get no slice
1097 * and will be reevaluated each time it is selected on the
1098 * run queue. The exception to this is nice 0 ksegs when
1099 * a nice -20 is running. They are always granted a minimum
1102 if (!SCHED_INTERACTIVE(kg)) {
1105 nice = kg->kg_proc->p_nice + (0 - kseq->ksq_nicemin);
1106 if (kseq->ksq_load_timeshare == 0 ||
1107 kg->kg_proc->p_nice < kseq->ksq_nicemin)
1108 ke->ke_slice = SCHED_SLICE_MAX;
1109 else if (nice <= SCHED_SLICE_NTHRESH)
1110 ke->ke_slice = SCHED_SLICE_NICE(nice);
1111 else if (kg->kg_proc->p_nice == 0)
1112 ke->ke_slice = SCHED_SLICE_MIN;
1116 ke->ke_slice = SCHED_SLICE_INTERACTIVE;
1122 * This routine enforces a maximum limit on the amount of scheduling history
1123 * kept. It is called after either the slptime or runtime is adjusted.
1124 * This routine will not operate correctly when slp or run times have been
1125 * adjusted to more than double their maximum.
1128 sched_interact_update(struct ksegrp *kg)
1132 sum = kg->kg_runtime + kg->kg_slptime;
1133 if (sum < SCHED_SLP_RUN_MAX)
1136 * If we have exceeded by more than 1/5th then the algorithm below
1137 * will not bring us back into range. Dividing by two here forces
1138 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1140 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1141 kg->kg_runtime /= 2;
1142 kg->kg_slptime /= 2;
1145 kg->kg_runtime = (kg->kg_runtime / 5) * 4;
1146 kg->kg_slptime = (kg->kg_slptime / 5) * 4;
1150 sched_interact_fork(struct ksegrp *kg)
1155 sum = kg->kg_runtime + kg->kg_slptime;
1156 if (sum > SCHED_SLP_RUN_FORK) {
1157 ratio = sum / SCHED_SLP_RUN_FORK;
1158 kg->kg_runtime /= ratio;
1159 kg->kg_slptime /= ratio;
1164 sched_interact_score(struct ksegrp *kg)
1168 if (kg->kg_runtime > kg->kg_slptime) {
1169 div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF);
1170 return (SCHED_INTERACT_HALF +
1171 (SCHED_INTERACT_HALF - (kg->kg_slptime / div)));
1172 } if (kg->kg_slptime > kg->kg_runtime) {
1173 div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF);
1174 return (kg->kg_runtime / div);
1178 * This can happen if slptime and runtime are 0.
1185 * Very early in the boot some setup of scheduler-specific
1186 * parts of proc0 and of soem scheduler resources needs to be done.
1194 * Set up the scheduler specific parts of proc0.
1196 proc0.p_sched = NULL; /* XXX */
1197 ksegrp0.kg_sched = &kg_sched0;
1198 thread0.td_sched = &kse0;
1199 kse0.ke_thread = &thread0;
1200 kse0.ke_state = KES_THREAD;
1201 kg_sched0.skg_concurrency = 1;
1202 kg_sched0.skg_avail_opennings = 0; /* we are already running */
1206 * This is only somewhat accurate since given many processes of the same
1207 * priority they will switch when their slices run out, which will be
1208 * at most SCHED_SLICE_MAX.
1211 sched_rr_interval(void)
1213 return (SCHED_SLICE_MAX);
1217 sched_pctcpu_update(struct kse *ke)
1220 * Adjust counters and watermark for pctcpu calc.
1222 if (ke->ke_ltick > ticks - SCHED_CPU_TICKS) {
1224 * Shift the tick count out so that the divide doesn't
1225 * round away our results.
1227 ke->ke_ticks <<= 10;
1228 ke->ke_ticks = (ke->ke_ticks / (ticks - ke->ke_ftick)) *
1230 ke->ke_ticks >>= 10;
1233 ke->ke_ltick = ticks;
1234 ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS;
1238 sched_thread_priority(struct thread *td, u_char prio)
1242 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1243 td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1244 curthread->td_proc->p_comm);
1246 mtx_assert(&sched_lock, MA_OWNED);
1247 if (td->td_priority == prio)
1249 if (TD_ON_RUNQ(td)) {
1251 * If the priority has been elevated due to priority
1252 * propagation, we may have to move ourselves to a new
1253 * queue. We still call adjustrunqueue below in case kse
1254 * needs to fix things up.
1256 if (prio < td->td_priority && ke->ke_runq != NULL &&
1257 (ke->ke_flags & KEF_ASSIGNED) == 0 &&
1258 ke->ke_runq != KSEQ_CPU(ke->ke_cpu)->ksq_curr) {
1259 runq_remove(ke->ke_runq, ke);
1260 ke->ke_runq = KSEQ_CPU(ke->ke_cpu)->ksq_curr;
1261 runq_add(ke->ke_runq, ke, 0);
1264 * Hold this kse on this cpu so that sched_prio() doesn't
1265 * cause excessive migration. We only want migration to
1266 * happen as the result of a wakeup.
1268 ke->ke_flags |= KEF_HOLD;
1269 adjustrunqueue(td, prio);
1270 ke->ke_flags &= ~KEF_HOLD;
1272 td->td_priority = prio;
1276 * Update a thread's priority when it is lent another thread's
1280 sched_lend_prio(struct thread *td, u_char prio)
1283 td->td_flags |= TDF_BORROWING;
1284 sched_thread_priority(td, prio);
1288 * Restore a thread's priority when priority propagation is
1289 * over. The prio argument is the minimum priority the thread
1290 * needs to have to satisfy other possible priority lending
1291 * requests. If the thread's regular priority is less
1292 * important than prio, the thread will keep a priority boost
1296 sched_unlend_prio(struct thread *td, u_char prio)
1300 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1301 td->td_base_pri <= PRI_MAX_TIMESHARE)
1302 base_pri = td->td_ksegrp->kg_user_pri;
1304 base_pri = td->td_base_pri;
1305 if (prio >= base_pri) {
1306 td->td_flags &= ~TDF_BORROWING;
1307 sched_thread_priority(td, base_pri);
1309 sched_lend_prio(td, prio);
1313 sched_prio(struct thread *td, u_char prio)
1317 /* First, update the base priority. */
1318 td->td_base_pri = prio;
1321 * If the thread is borrowing another thread's priority, don't
1322 * ever lower the priority.
1324 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1327 /* Change the real priority. */
1328 oldprio = td->td_priority;
1329 sched_thread_priority(td, prio);
1332 * If the thread is on a turnstile, then let the turnstile update
1335 if (TD_ON_LOCK(td) && oldprio != prio)
1336 turnstile_adjust(td, oldprio);
1340 sched_switch(struct thread *td, struct thread *newtd, int flags)
1345 mtx_assert(&sched_lock, MA_OWNED);
1350 td->td_lastcpu = td->td_oncpu;
1351 td->td_oncpu = NOCPU;
1352 td->td_flags &= ~TDF_NEEDRESCHED;
1353 td->td_owepreempt = 0;
1356 * If the KSE has been assigned it may be in the process of switching
1357 * to the new cpu. This is the case in sched_bind().
1359 if (td == PCPU_GET(idlethread)) {
1361 } else if ((ke->ke_flags & KEF_ASSIGNED) == 0) {
1362 /* We are ending our run so make our slot available again */
1363 SLOT_RELEASE(td->td_ksegrp);
1364 kseq_load_rem(ksq, ke);
1365 if (TD_IS_RUNNING(td)) {
1367 * Don't allow the thread to migrate
1368 * from a preemption.
1370 ke->ke_flags |= KEF_HOLD;
1371 setrunqueue(td, (flags & SW_PREEMPT) ?
1372 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1373 SRQ_OURSELF|SRQ_YIELDING);
1374 ke->ke_flags &= ~KEF_HOLD;
1375 } else if ((td->td_proc->p_flag & P_HADTHREADS) &&
1376 (newtd == NULL || newtd->td_ksegrp != td->td_ksegrp))
1378 * We will not be on the run queue.
1379 * So we must be sleeping or similar.
1380 * Don't use the slot if we will need it
1383 slot_fill(td->td_ksegrp);
1385 if (newtd != NULL) {
1387 * If we bring in a thread,
1388 * then account for it as if it had been added to the
1389 * run queue and then chosen.
1391 newtd->td_kse->ke_flags |= KEF_DIDRUN;
1392 newtd->td_kse->ke_runq = ksq->ksq_curr;
1393 SLOT_USE(newtd->td_ksegrp);
1394 TD_SET_RUNNING(newtd);
1395 kseq_load_add(KSEQ_SELF(), newtd->td_kse);
1397 newtd = choosethread();
1400 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1401 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1403 cpu_switch(td, newtd);
1405 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1406 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1410 sched_lock.mtx_lock = (uintptr_t)td;
1412 td->td_oncpu = PCPU_GET(cpuid);
1416 sched_nice(struct proc *p, int nice)
1423 PROC_LOCK_ASSERT(p, MA_OWNED);
1424 mtx_assert(&sched_lock, MA_OWNED);
1426 * We need to adjust the nice counts for running KSEs.
1428 FOREACH_KSEGRP_IN_PROC(p, kg) {
1429 if (kg->kg_pri_class == PRI_TIMESHARE) {
1430 FOREACH_THREAD_IN_GROUP(kg, td) {
1432 if (ke->ke_runq == NULL)
1434 kseq = KSEQ_CPU(ke->ke_cpu);
1435 kseq_nice_rem(kseq, p->p_nice);
1436 kseq_nice_add(kseq, nice);
1441 FOREACH_KSEGRP_IN_PROC(p, kg) {
1443 FOREACH_THREAD_IN_GROUP(kg, td)
1444 td->td_flags |= TDF_NEEDRESCHED;
1449 sched_sleep(struct thread *td)
1451 mtx_assert(&sched_lock, MA_OWNED);
1453 td->td_slptime = ticks;
1457 sched_wakeup(struct thread *td)
1459 mtx_assert(&sched_lock, MA_OWNED);
1462 * Let the kseg know how long we slept for. This is because process
1463 * interactivity behavior is modeled in the kseg.
1465 if (td->td_slptime) {
1470 hzticks = (ticks - td->td_slptime) << 10;
1471 if (hzticks >= SCHED_SLP_RUN_MAX) {
1472 kg->kg_slptime = SCHED_SLP_RUN_MAX;
1475 kg->kg_slptime += hzticks;
1476 sched_interact_update(kg);
1479 sched_slice(td->td_kse);
1482 setrunqueue(td, SRQ_BORING);
1486 * Penalize the parent for creating a new child and initialize the child's
1490 sched_fork(struct thread *td, struct thread *childtd)
1493 mtx_assert(&sched_lock, MA_OWNED);
1495 sched_fork_ksegrp(td, childtd->td_ksegrp);
1496 sched_fork_thread(td, childtd);
1500 sched_fork_ksegrp(struct thread *td, struct ksegrp *child)
1502 struct ksegrp *kg = td->td_ksegrp;
1503 mtx_assert(&sched_lock, MA_OWNED);
1505 child->kg_slptime = kg->kg_slptime;
1506 child->kg_runtime = kg->kg_runtime;
1507 child->kg_user_pri = kg->kg_user_pri;
1508 sched_interact_fork(child);
1509 kg->kg_runtime += tickincr << 10;
1510 sched_interact_update(kg);
1514 sched_fork_thread(struct thread *td, struct thread *child)
1519 sched_newthread(child);
1521 ke2 = child->td_kse;
1522 ke2->ke_slice = 1; /* Attempt to quickly learn interactivity. */
1523 ke2->ke_cpu = ke->ke_cpu;
1524 ke2->ke_runq = NULL;
1526 /* Grab our parents cpu estimation information. */
1527 ke2->ke_ticks = ke->ke_ticks;
1528 ke2->ke_ltick = ke->ke_ltick;
1529 ke2->ke_ftick = ke->ke_ftick;
1533 sched_class(struct ksegrp *kg, int class)
1541 mtx_assert(&sched_lock, MA_OWNED);
1542 if (kg->kg_pri_class == class)
1545 nclass = PRI_BASE(class);
1546 oclass = PRI_BASE(kg->kg_pri_class);
1547 FOREACH_THREAD_IN_GROUP(kg, td) {
1549 if ((ke->ke_state != KES_ONRUNQ &&
1550 ke->ke_state != KES_THREAD) || ke->ke_runq == NULL)
1552 kseq = KSEQ_CPU(ke->ke_cpu);
1556 * On SMP if we're on the RUNQ we must adjust the transferable
1557 * count because could be changing to or from an interrupt
1560 if (ke->ke_state == KES_ONRUNQ) {
1561 if (KSE_CAN_MIGRATE(ke)) {
1562 kseq->ksq_transferable--;
1563 kseq->ksq_group->ksg_transferable--;
1565 if (KSE_CAN_MIGRATE(ke)) {
1566 kseq->ksq_transferable++;
1567 kseq->ksq_group->ksg_transferable++;
1571 if (oclass == PRI_TIMESHARE) {
1572 kseq->ksq_load_timeshare--;
1573 kseq_nice_rem(kseq, kg->kg_proc->p_nice);
1575 if (nclass == PRI_TIMESHARE) {
1576 kseq->ksq_load_timeshare++;
1577 kseq_nice_add(kseq, kg->kg_proc->p_nice);
1581 kg->kg_pri_class = class;
1585 * Return some of the child's priority and interactivity to the parent.
1588 sched_exit(struct proc *p, struct thread *childtd)
1590 mtx_assert(&sched_lock, MA_OWNED);
1591 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), childtd);
1592 sched_exit_thread(NULL, childtd);
1596 sched_exit_ksegrp(struct ksegrp *kg, struct thread *td)
1598 /* kg->kg_slptime += td->td_ksegrp->kg_slptime; */
1599 kg->kg_runtime += td->td_ksegrp->kg_runtime;
1600 sched_interact_update(kg);
1604 sched_exit_thread(struct thread *td, struct thread *childtd)
1606 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1607 childtd, childtd->td_proc->p_comm, childtd->td_priority);
1608 kseq_load_rem(KSEQ_CPU(childtd->td_kse->ke_cpu), childtd->td_kse);
1612 sched_clock(struct thread *td)
1618 mtx_assert(&sched_lock, MA_OWNED);
1621 if (ticks >= bal_tick)
1623 if (ticks >= gbal_tick && balance_groups)
1624 sched_balance_groups();
1626 * We could have been assigned a non real-time thread without an
1629 if (kseq->ksq_assigned)
1630 kseq_assign(kseq); /* Potentially sets NEEDRESCHED */
1633 * sched_setup() apparently happens prior to stathz being set. We
1634 * need to resolve the timers earlier in the boot so we can avoid
1635 * calculating this here.
1637 if (realstathz == 0) {
1638 realstathz = stathz ? stathz : hz;
1639 tickincr = hz / realstathz;
1641 * XXX This does not work for values of stathz that are much
1651 /* Adjust ticks for pctcpu */
1653 ke->ke_ltick = ticks;
1655 /* Go up to one second beyond our max and then trim back down */
1656 if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick)
1657 sched_pctcpu_update(ke);
1659 if (td->td_flags & TDF_IDLETD)
1662 * We only do slicing code for TIMESHARE ksegrps.
1664 if (kg->kg_pri_class != PRI_TIMESHARE)
1667 * We used a tick charge it to the ksegrp so that we can compute our
1670 kg->kg_runtime += tickincr << 10;
1671 sched_interact_update(kg);
1674 * We used up one time slice.
1676 if (--ke->ke_slice > 0)
1679 * We're out of time, recompute priorities and requeue.
1681 kseq_load_rem(kseq, ke);
1684 if (SCHED_CURR(kg, ke))
1685 ke->ke_runq = kseq->ksq_curr;
1687 ke->ke_runq = kseq->ksq_next;
1688 kseq_load_add(kseq, ke);
1689 td->td_flags |= TDF_NEEDRESCHED;
1693 sched_runnable(void)
1702 if (kseq->ksq_assigned) {
1703 mtx_lock_spin(&sched_lock);
1705 mtx_unlock_spin(&sched_lock);
1708 if ((curthread->td_flags & TDF_IDLETD) != 0) {
1709 if (kseq->ksq_load > 0)
1712 if (kseq->ksq_load - 1 > 0)
1720 sched_userret(struct thread *td)
1724 KASSERT((td->td_flags & TDF_BORROWING) == 0,
1725 ("thread with borrowed priority returning to userland"));
1727 if (td->td_priority != kg->kg_user_pri) {
1728 mtx_lock_spin(&sched_lock);
1729 td->td_priority = kg->kg_user_pri;
1730 td->td_base_pri = kg->kg_user_pri;
1731 mtx_unlock_spin(&sched_lock);
1741 mtx_assert(&sched_lock, MA_OWNED);
1745 if (kseq->ksq_assigned)
1748 ke = kseq_choose(kseq);
1751 if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE)
1752 if (kseq_idled(kseq) == 0)
1755 kseq_runq_rem(kseq, ke);
1756 ke->ke_state = KES_THREAD;
1760 if (kseq_idled(kseq) == 0)
1767 sched_add(struct thread *td, int flags)
1776 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1777 td, td->td_proc->p_comm, td->td_priority, curthread,
1778 curthread->td_proc->p_comm);
1779 mtx_assert(&sched_lock, MA_OWNED);
1783 preemptive = !(flags & SRQ_YIELDING);
1784 class = PRI_BASE(kg->kg_pri_class);
1786 if ((ke->ke_flags & KEF_INTERNAL) == 0)
1787 SLOT_USE(td->td_ksegrp);
1788 ke->ke_flags &= ~KEF_INTERNAL;
1790 if (ke->ke_flags & KEF_ASSIGNED) {
1791 if (ke->ke_flags & KEF_REMOVED)
1792 ke->ke_flags &= ~KEF_REMOVED;
1795 canmigrate = KSE_CAN_MIGRATE(ke);
1797 KASSERT(ke->ke_state != KES_ONRUNQ,
1798 ("sched_add: kse %p (%s) already in run queue", ke,
1799 ke->ke_proc->p_comm));
1800 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
1801 ("sched_add: process swapped out"));
1802 KASSERT(ke->ke_runq == NULL,
1803 ("sched_add: KSE %p is still assigned to a run queue", ke));
1807 ke->ke_runq = kseq->ksq_curr;
1808 ke->ke_slice = SCHED_SLICE_MAX;
1810 ke->ke_cpu = PCPU_GET(cpuid);
1813 if (SCHED_CURR(kg, ke))
1814 ke->ke_runq = kseq->ksq_curr;
1816 ke->ke_runq = kseq->ksq_next;
1820 * This is for priority prop.
1822 if (ke->ke_thread->td_priority < PRI_MIN_IDLE)
1823 ke->ke_runq = kseq->ksq_curr;
1825 ke->ke_runq = &kseq->ksq_idle;
1826 ke->ke_slice = SCHED_SLICE_MIN;
1829 panic("Unknown pri class.");
1834 * Don't migrate running threads here. Force the long term balancer
1837 if (ke->ke_flags & KEF_HOLD) {
1838 ke->ke_flags &= ~KEF_HOLD;
1842 * If this thread is pinned or bound, notify the target cpu.
1844 if (!canmigrate && ke->ke_cpu != PCPU_GET(cpuid) ) {
1846 kseq_notify(ke, ke->ke_cpu);
1850 * If we had been idle, clear our bit in the group and potentially
1851 * the global bitmap. If not, see if we should transfer this thread.
1853 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1854 (kseq->ksq_group->ksg_idlemask & PCPU_GET(cpumask)) != 0) {
1856 * Check to see if our group is unidling, and if so, remove it
1857 * from the global idle mask.
1859 if (kseq->ksq_group->ksg_idlemask ==
1860 kseq->ksq_group->ksg_cpumask)
1861 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
1863 * Now remove ourselves from the group specific idle mask.
1865 kseq->ksq_group->ksg_idlemask &= ~PCPU_GET(cpumask);
1866 } else if (canmigrate && kseq->ksq_load > 1 && class != PRI_ITHD)
1867 if (kseq_transfer(kseq, ke, class))
1869 ke->ke_cpu = PCPU_GET(cpuid);
1871 if (td->td_priority < curthread->td_priority &&
1872 ke->ke_runq == kseq->ksq_curr)
1873 curthread->td_flags |= TDF_NEEDRESCHED;
1874 if (preemptive && maybe_preempt(td))
1876 ke->ke_state = KES_ONRUNQ;
1878 kseq_runq_add(kseq, ke, flags);
1879 kseq_load_add(kseq, ke);
1883 sched_rem(struct thread *td)
1888 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1889 td, td->td_proc->p_comm, td->td_priority, curthread,
1890 curthread->td_proc->p_comm);
1891 mtx_assert(&sched_lock, MA_OWNED);
1893 SLOT_RELEASE(td->td_ksegrp);
1894 if (ke->ke_flags & KEF_ASSIGNED) {
1895 ke->ke_flags |= KEF_REMOVED;
1898 KASSERT((ke->ke_state == KES_ONRUNQ),
1899 ("sched_rem: KSE not on run queue"));
1901 ke->ke_state = KES_THREAD;
1902 kseq = KSEQ_CPU(ke->ke_cpu);
1903 kseq_runq_rem(kseq, ke);
1904 kseq_load_rem(kseq, ke);
1908 sched_pctcpu(struct thread *td)
1918 mtx_lock_spin(&sched_lock);
1923 * Don't update more frequently than twice a second. Allowing
1924 * this causes the cpu usage to decay away too quickly due to
1927 if (ke->ke_ftick + SCHED_CPU_TICKS < ke->ke_ltick ||
1928 ke->ke_ltick < (ticks - (hz / 2)))
1929 sched_pctcpu_update(ke);
1930 /* How many rtick per second ? */
1931 rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1932 pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1935 ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick;
1936 mtx_unlock_spin(&sched_lock);
1942 sched_bind(struct thread *td, int cpu)
1946 mtx_assert(&sched_lock, MA_OWNED);
1948 ke->ke_flags |= KEF_BOUND;
1950 if (PCPU_GET(cpuid) == cpu)
1952 /* sched_rem without the runq_remove */
1953 ke->ke_state = KES_THREAD;
1954 kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1955 kseq_notify(ke, cpu);
1956 /* When we return from mi_switch we'll be on the correct cpu. */
1957 mi_switch(SW_VOL, NULL);
1962 sched_unbind(struct thread *td)
1964 mtx_assert(&sched_lock, MA_OWNED);
1965 td->td_kse->ke_flags &= ~KEF_BOUND;
1969 sched_is_bound(struct thread *td)
1971 mtx_assert(&sched_lock, MA_OWNED);
1972 return (td->td_kse->ke_flags & KEF_BOUND);
1983 for (i = 0; i <= ksg_maxid; i++)
1984 total += KSEQ_GROUP(i)->ksg_load;
1987 return (KSEQ_SELF()->ksq_sysload);
1992 sched_sizeof_ksegrp(void)
1994 return (sizeof(struct ksegrp) + sizeof(struct kg_sched));
1998 sched_sizeof_proc(void)
2000 return (sizeof(struct proc));
2004 sched_sizeof_thread(void)
2006 return (sizeof(struct thread) + sizeof(struct td_sched));
2008 #define KERN_SWITCH_INCLUDE 1
2009 #include "kern/kern_switch.c"