2 * Copyright (c) 2002-2007, 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.
28 * This file implements the ULE scheduler. ULE supports independent CPU
29 * run queues and fine grain locking. It has superior interactive
30 * performance under load even on uni-processor systems.
33 * ULE is the last three letters in schedule. It owes its name to a
34 * generic user created for a scheduling system by Paul Mikesell at
35 * Isilon Systems and a general lack of creativity on the part of the author.
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_kdtrace.h"
43 #include "opt_sched.h"
45 #include <sys/param.h>
46 #include <sys/systm.h>
48 #include <sys/kernel.h>
51 #include <sys/mutex.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
68 #include <sys/pmckern.h>
72 #include <sys/dtrace_bsd.h>
73 int dtrace_vtime_active;
74 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
80 #if defined(__powerpc__) && defined(BOOKE_E500)
81 #error "This architecture is not currently compatible with ULE"
86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
87 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
88 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
91 * Thread scheduler specific section. All fields are protected
95 struct runq *ts_runq; /* Run-queue we're queued on. */
96 short ts_flags; /* TSF_* flags. */
97 u_char ts_cpu; /* CPU that we have affinity for. */
98 int ts_rltick; /* Real last tick, for affinity. */
99 int ts_slice; /* Ticks of slice remaining. */
100 u_int ts_slptime; /* Number of ticks we vol. slept */
101 u_int ts_runtime; /* Number of ticks we were running */
102 int ts_ltick; /* Last tick that we were running on */
103 int ts_ftick; /* First tick that we were running on */
104 int ts_ticks; /* Tick count */
106 char ts_name[TS_NAME_LEN];
109 /* flags kept in ts_flags */
110 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
111 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
113 static struct td_sched td_sched0;
115 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
116 #define THREAD_CAN_SCHED(td, cpu) \
117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
120 * Priority ranges used for interactive and non-interactive timeshare
121 * threads. The timeshare priorities are split up into four ranges.
122 * The first range handles interactive threads. The last three ranges
123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124 * ranges supporting nice values.
126 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
130 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
131 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
136 * Cpu percentage computation macros and defines.
138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
145 #define SCHED_TICK_SECS 10
146 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
147 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
148 #define SCHED_TICK_SHIFT 10
149 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
153 * These macros determine priorities for non-interactive threads. They are
154 * assigned a priority based on their recent cpu utilization as expressed
155 * by the ratio of ticks to the tick total. NHALF priorities at the start
156 * and end of the MIN to MAX timeshare range are only reachable with negative
157 * or positive nice respectively.
159 * PRI_RANGE: Priority range for utilization dependent priorities.
160 * PRI_NRESV: Number of nice values.
161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
162 * PRI_NICE: Determines the part of the priority inherited from nice.
164 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
165 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
166 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define SCHED_PRI_TICKS(ts) \
170 (SCHED_TICK_HZ((ts)) / \
171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define SCHED_PRI_NICE(nice) (nice)
175 * These determine the interactivity of a process. Interactivity differs from
176 * cpu utilization in that it expresses the voluntary time slept vs time ran
177 * while cpu utilization includes all time not running. This more accurately
178 * models the intent of the thread.
180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
181 * before throttling back.
182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
183 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
184 * INTERACT_THRESH: Threshold for placement on the current runq.
186 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
187 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
188 #define SCHED_INTERACT_MAX (100)
189 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
190 #define SCHED_INTERACT_THRESH (30)
193 * tickincr: Converts a stathz tick into a hz domain scaled by
194 * the shift factor. Without the shift the error rate
195 * due to rounding would be unacceptably high.
196 * realstathz: stathz is sometimes 0 and run off of hz.
197 * sched_slice: Runtime of each thread before rescheduling.
198 * preempt_thresh: Priority threshold for preemption and remote IPIs.
200 static int sched_interact = SCHED_INTERACT_THRESH;
201 static int realstathz;
203 static int sched_slice = 1;
205 #ifdef FULL_PREEMPTION
206 static int preempt_thresh = PRI_MAX_IDLE;
208 static int preempt_thresh = PRI_MIN_KERN;
211 static int preempt_thresh = 0;
213 static int static_boost = PRI_MIN_BATCH;
214 static int sched_idlespins = 10000;
215 static int sched_idlespinthresh = -1;
218 * tdq - per processor runqs and statistics. All fields are protected by the
219 * tdq_lock. The load and lowpri may be accessed without to avoid excess
220 * locking in sched_pickcpu();
223 /* Ordered to improve efficiency of cpu_search() and switch(). */
224 struct mtx tdq_lock; /* run queue lock. */
225 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
226 volatile int tdq_load; /* Aggregate load. */
227 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
228 int tdq_sysload; /* For loadavg, !ITHD load. */
229 int tdq_transferable; /* Transferable thread count. */
230 short tdq_switchcnt; /* Switches this tick. */
231 short tdq_oldswitchcnt; /* Switches last tick. */
232 u_char tdq_lowpri; /* Lowest priority thread. */
233 u_char tdq_ipipending; /* IPI pending. */
234 u_char tdq_idx; /* Current insert index. */
235 u_char tdq_ridx; /* Current removal index. */
236 struct runq tdq_realtime; /* real-time run queue. */
237 struct runq tdq_timeshare; /* timeshare run queue. */
238 struct runq tdq_idle; /* Queue of IDLE threads. */
239 char tdq_name[TDQ_NAME_LEN];
241 char tdq_loadname[TDQ_LOADNAME_LEN];
245 /* Idle thread states and config. */
246 #define TDQ_RUNNING 1
250 struct cpu_group *cpu_top; /* CPU topology */
252 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
253 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
258 static int rebalance = 1;
259 static int balance_interval = 128; /* Default set in sched_initticks(). */
261 static int steal_idle = 1;
262 static int steal_thresh = 2;
265 * One thread queue per processor.
267 static struct tdq tdq_cpu[MAXCPU];
268 static struct tdq *balance_tdq;
269 static int balance_ticks;
270 static DPCPU_DEFINE(uint32_t, randomval);
272 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
273 #define TDQ_CPU(x) (&tdq_cpu[(x)])
274 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
276 static struct tdq tdq_cpu;
278 #define TDQ_ID(x) (0)
279 #define TDQ_SELF() (&tdq_cpu)
280 #define TDQ_CPU(x) (&tdq_cpu)
283 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
284 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
285 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
286 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
287 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
289 static void sched_priority(struct thread *);
290 static void sched_thread_priority(struct thread *, u_char);
291 static int sched_interact_score(struct thread *);
292 static void sched_interact_update(struct thread *);
293 static void sched_interact_fork(struct thread *);
294 static void sched_pctcpu_update(struct td_sched *, int);
296 /* Operations on per processor queues */
297 static struct thread *tdq_choose(struct tdq *);
298 static void tdq_setup(struct tdq *);
299 static void tdq_load_add(struct tdq *, struct thread *);
300 static void tdq_load_rem(struct tdq *, struct thread *);
301 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
302 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
303 static inline int sched_shouldpreempt(int, int, int);
304 void tdq_print(int cpu);
305 static void runq_print(struct runq *rq);
306 static void tdq_add(struct tdq *, struct thread *, int);
308 static int tdq_move(struct tdq *, struct tdq *);
309 static int tdq_idled(struct tdq *);
310 static void tdq_notify(struct tdq *, struct thread *);
311 static struct thread *tdq_steal(struct tdq *, int);
312 static struct thread *runq_steal(struct runq *, int);
313 static int sched_pickcpu(struct thread *, int);
314 static void sched_balance(void);
315 static int sched_balance_pair(struct tdq *, struct tdq *);
316 static inline struct tdq *sched_setcpu(struct thread *, int, int);
317 static inline void thread_unblock_switch(struct thread *, struct mtx *);
318 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
319 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
320 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
321 struct cpu_group *cg, int indent);
324 static void sched_setup(void *dummy);
325 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
327 static void sched_initticks(void *dummy);
328 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
331 SDT_PROVIDER_DEFINE(sched);
333 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
334 "struct proc *", "uint8_t");
335 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
336 "struct proc *", "void *");
337 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
338 "struct proc *", "void *", "int");
339 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
340 "struct proc *", "uint8_t", "struct thread *");
341 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
342 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
344 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
345 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
346 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
350 * Print the threads waiting on a run-queue.
353 runq_print(struct runq *rq)
361 for (i = 0; i < RQB_LEN; i++) {
362 printf("\t\trunq bits %d 0x%zx\n",
363 i, rq->rq_status.rqb_bits[i]);
364 for (j = 0; j < RQB_BPW; j++)
365 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
366 pri = j + (i << RQB_L2BPW);
367 rqh = &rq->rq_queues[pri];
368 TAILQ_FOREACH(td, rqh, td_runq) {
369 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
370 td, td->td_name, td->td_priority,
371 td->td_rqindex, pri);
378 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
387 printf("tdq %d:\n", TDQ_ID(tdq));
388 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
389 printf("\tLock name: %s\n", tdq->tdq_name);
390 printf("\tload: %d\n", tdq->tdq_load);
391 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
392 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
393 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
394 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
395 printf("\tload transferable: %d\n", tdq->tdq_transferable);
396 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
397 printf("\trealtime runq:\n");
398 runq_print(&tdq->tdq_realtime);
399 printf("\ttimeshare runq:\n");
400 runq_print(&tdq->tdq_timeshare);
401 printf("\tidle runq:\n");
402 runq_print(&tdq->tdq_idle);
406 sched_shouldpreempt(int pri, int cpri, int remote)
409 * If the new priority is not better than the current priority there is
415 * Always preempt idle.
417 if (cpri >= PRI_MIN_IDLE)
420 * If preemption is disabled don't preempt others.
422 if (preempt_thresh == 0)
425 * Preempt if we exceed the threshold.
427 if (pri <= preempt_thresh)
430 * If we're interactive or better and there is non-interactive
431 * or worse running preempt only remote processors.
433 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
439 * Add a thread to the actual run-queue. Keeps transferable counts up to
440 * date with what is actually on the run-queue. Selects the correct
441 * queue position for timeshare threads.
444 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
449 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
450 THREAD_LOCK_ASSERT(td, MA_OWNED);
452 pri = td->td_priority;
455 if (THREAD_CAN_MIGRATE(td)) {
456 tdq->tdq_transferable++;
457 ts->ts_flags |= TSF_XFERABLE;
459 if (pri < PRI_MIN_BATCH) {
460 ts->ts_runq = &tdq->tdq_realtime;
461 } else if (pri <= PRI_MAX_BATCH) {
462 ts->ts_runq = &tdq->tdq_timeshare;
463 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
464 ("Invalid priority %d on timeshare runq", pri));
466 * This queue contains only priorities between MIN and MAX
467 * realtime. Use the whole queue to represent these values.
469 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
470 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
471 pri = (pri + tdq->tdq_idx) % RQ_NQS;
473 * This effectively shortens the queue by one so we
474 * can have a one slot difference between idx and
475 * ridx while we wait for threads to drain.
477 if (tdq->tdq_ridx != tdq->tdq_idx &&
478 pri == tdq->tdq_ridx)
479 pri = (unsigned char)(pri - 1) % RQ_NQS;
482 runq_add_pri(ts->ts_runq, td, pri, flags);
485 ts->ts_runq = &tdq->tdq_idle;
486 runq_add(ts->ts_runq, td, flags);
490 * Remove a thread from a run-queue. This typically happens when a thread
491 * is selected to run. Running threads are not on the queue and the
492 * transferable count does not reflect them.
495 tdq_runq_rem(struct tdq *tdq, struct thread *td)
500 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
501 KASSERT(ts->ts_runq != NULL,
502 ("tdq_runq_remove: thread %p null ts_runq", td));
503 if (ts->ts_flags & TSF_XFERABLE) {
504 tdq->tdq_transferable--;
505 ts->ts_flags &= ~TSF_XFERABLE;
507 if (ts->ts_runq == &tdq->tdq_timeshare) {
508 if (tdq->tdq_idx != tdq->tdq_ridx)
509 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
511 runq_remove_idx(ts->ts_runq, td, NULL);
513 runq_remove(ts->ts_runq, td);
517 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
518 * for this thread to the referenced thread queue.
521 tdq_load_add(struct tdq *tdq, struct thread *td)
524 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
525 THREAD_LOCK_ASSERT(td, MA_OWNED);
528 if ((td->td_flags & TDF_NOLOAD) == 0)
530 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
531 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
535 * Remove the load from a thread that is transitioning to a sleep state or
539 tdq_load_rem(struct tdq *tdq, struct thread *td)
542 THREAD_LOCK_ASSERT(td, MA_OWNED);
543 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
544 KASSERT(tdq->tdq_load != 0,
545 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
548 if ((td->td_flags & TDF_NOLOAD) == 0)
550 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
551 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
555 * Set lowpri to its exact value by searching the run-queue and
556 * evaluating curthread. curthread may be passed as an optimization.
559 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
563 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
565 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
566 td = tdq_choose(tdq);
567 if (td == NULL || td->td_priority > ctd->td_priority)
568 tdq->tdq_lowpri = ctd->td_priority;
570 tdq->tdq_lowpri = td->td_priority;
577 int cs_pri; /* Min priority for low. */
578 int cs_limit; /* Max load for low, min load for high. */
583 #define CPU_SEARCH_LOWEST 0x1
584 #define CPU_SEARCH_HIGHEST 0x2
585 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
587 #define CPUSET_FOREACH(cpu, mask) \
588 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
589 if (CPU_ISSET(cpu, &mask))
591 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
592 struct cpu_search *high, const int match);
593 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
594 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
595 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
596 struct cpu_search *high);
599 * Search the tree of cpu_groups for the lowest or highest loaded cpu
600 * according to the match argument. This routine actually compares the
601 * load on all paths through the tree and finds the least loaded cpu on
602 * the least loaded path, which may differ from the least loaded cpu in
603 * the system. This balances work among caches and busses.
605 * This inline is instantiated in three forms below using constants for the
606 * match argument. It is reduced to the minimum set for each case. It is
607 * also recursive to the depth of the tree.
610 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
611 struct cpu_search *high, const int match)
613 struct cpu_search lgroup;
614 struct cpu_search hgroup;
616 struct cpu_group *child;
618 int cpu, i, hload, lload, load, total, rnd, *rndptr;
621 cpumask = cg->cg_mask;
622 if (match & CPU_SEARCH_LOWEST) {
626 if (match & CPU_SEARCH_HIGHEST) {
631 /* Iterate through the child CPU groups and then remaining CPUs. */
632 for (i = cg->cg_children, cpu = mp_maxid; i >= 0; ) {
634 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
640 child = &cg->cg_child[i - 1];
642 if (match & CPU_SEARCH_LOWEST)
644 if (match & CPU_SEARCH_HIGHEST)
646 if (child) { /* Handle child CPU group. */
647 CPU_NAND(&cpumask, &child->cg_mask);
649 case CPU_SEARCH_LOWEST:
650 load = cpu_search_lowest(child, &lgroup);
652 case CPU_SEARCH_HIGHEST:
653 load = cpu_search_highest(child, &hgroup);
655 case CPU_SEARCH_BOTH:
656 load = cpu_search_both(child, &lgroup, &hgroup);
659 } else { /* Handle child CPU. */
661 load = tdq->tdq_load * 256;
662 rndptr = DPCPU_PTR(randomval);
663 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
664 if (match & CPU_SEARCH_LOWEST) {
665 if (cpu == low->cs_prefer)
667 /* If that CPU is allowed and get data. */
668 if (tdq->tdq_lowpri > lgroup.cs_pri &&
669 tdq->tdq_load <= lgroup.cs_limit &&
670 CPU_ISSET(cpu, &lgroup.cs_mask)) {
672 lgroup.cs_load = load - rnd;
675 if (match & CPU_SEARCH_HIGHEST)
676 if (tdq->tdq_load >= hgroup.cs_limit &&
677 tdq->tdq_transferable &&
678 CPU_ISSET(cpu, &hgroup.cs_mask)) {
680 hgroup.cs_load = load - rnd;
685 /* We have info about child item. Compare it. */
686 if (match & CPU_SEARCH_LOWEST) {
687 if (lgroup.cs_cpu >= 0 &&
689 (load == lload && lgroup.cs_load < low->cs_load))) {
691 low->cs_cpu = lgroup.cs_cpu;
692 low->cs_load = lgroup.cs_load;
695 if (match & CPU_SEARCH_HIGHEST)
696 if (hgroup.cs_cpu >= 0 &&
698 (load == hload && hgroup.cs_load > high->cs_load))) {
700 high->cs_cpu = hgroup.cs_cpu;
701 high->cs_load = hgroup.cs_load;
705 if (i == 0 && CPU_EMPTY(&cpumask))
714 * cpu_search instantiations must pass constants to maintain the inline
718 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
720 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
724 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
726 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
730 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
731 struct cpu_search *high)
733 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
737 * Find the cpu with the least load via the least loaded path that has a
738 * lowpri greater than pri pri. A pri of -1 indicates any priority is
742 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
745 struct cpu_search low;
748 low.cs_prefer = prefer;
751 low.cs_limit = maxload;
752 cpu_search_lowest(cg, &low);
757 * Find the cpu with the highest load via the highest loaded path.
760 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
762 struct cpu_search high;
766 high.cs_limit = minload;
767 cpu_search_highest(cg, &high);
772 * Simultaneously find the highest and lowest loaded cpu reachable via
776 sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
778 struct cpu_search high;
779 struct cpu_search low;
784 low.cs_limit = INT_MAX;
789 cpu_search_both(cg, &low, &high);
790 *lowcpu = low.cs_cpu;
791 *highcpu = high.cs_cpu;
796 sched_balance_group(struct cpu_group *cg)
798 cpuset_t hmask, lmask;
799 int high, low, anylow;
803 high = sched_highest(cg, hmask, 1);
804 /* Stop if there is no more CPU with transferrable threads. */
807 CPU_CLR(high, &hmask);
808 CPU_COPY(&hmask, &lmask);
809 /* Stop if there is no more CPU left for low. */
810 if (CPU_EMPTY(&lmask))
814 low = sched_lowest(cg, lmask, -1,
815 TDQ_CPU(high)->tdq_load - 1, high);
816 /* Stop if we looked well and found no less loaded CPU. */
817 if (anylow && low == -1)
819 /* Go to next high if we found no less loaded CPU. */
822 /* Transfer thread from high to low. */
823 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
824 /* CPU that got thread can no longer be a donor. */
825 CPU_CLR(low, &hmask);
828 * If failed, then there is no threads on high
829 * that can run on this low. Drop low from low
830 * mask and look for different one.
832 CPU_CLR(low, &lmask);
845 * Select a random time between .5 * balance_interval and
846 * 1.5 * balance_interval.
848 balance_ticks = max(balance_interval / 2, 1);
849 balance_ticks += random() % balance_interval;
850 if (smp_started == 0 || rebalance == 0)
854 sched_balance_group(cpu_top);
859 * Lock two thread queues using their address to maintain lock order.
862 tdq_lock_pair(struct tdq *one, struct tdq *two)
866 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
869 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
874 * Unlock two thread queues. Order is not important here.
877 tdq_unlock_pair(struct tdq *one, struct tdq *two)
884 * Transfer load between two imbalanced thread queues.
887 sched_balance_pair(struct tdq *high, struct tdq *low)
892 tdq_lock_pair(high, low);
895 * Determine what the imbalance is and then adjust that to how many
896 * threads we actually have to give up (transferable).
898 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
899 (moved = tdq_move(high, low)) > 0) {
901 * In case the target isn't the current cpu IPI it to force a
902 * reschedule with the new workload.
906 if (cpu != PCPU_GET(cpuid))
907 ipi_cpu(cpu, IPI_PREEMPT);
910 tdq_unlock_pair(high, low);
915 * Move a thread from one thread queue to another.
918 tdq_move(struct tdq *from, struct tdq *to)
925 TDQ_LOCK_ASSERT(from, MA_OWNED);
926 TDQ_LOCK_ASSERT(to, MA_OWNED);
930 td = tdq_steal(tdq, cpu);
935 * Although the run queue is locked the thread may be blocked. Lock
936 * it to clear this and acquire the run-queue lock.
939 /* Drop recursive lock on from acquired via thread_lock(). */
943 td->td_lock = TDQ_LOCKPTR(to);
944 tdq_add(to, td, SRQ_YIELDING);
949 * This tdq has idled. Try to steal a thread from another cpu and switch
953 tdq_idled(struct tdq *tdq)
955 struct cpu_group *cg;
961 if (smp_started == 0 || steal_idle == 0)
964 CPU_CLR(PCPU_GET(cpuid), &mask);
965 /* We don't want to be preempted while we're iterating. */
967 for (cg = tdq->tdq_cg; cg != NULL; ) {
968 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
969 thresh = steal_thresh;
972 cpu = sched_highest(cg, mask, thresh);
977 steal = TDQ_CPU(cpu);
979 tdq_lock_pair(tdq, steal);
980 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
981 tdq_unlock_pair(tdq, steal);
985 * If a thread was added while interrupts were disabled don't
986 * steal one here. If we fail to acquire one due to affinity
987 * restrictions loop again with this cpu removed from the
990 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
991 tdq_unlock_pair(tdq, steal);
996 mi_switch(SW_VOL | SWT_IDLE, NULL);
997 thread_unlock(curthread);
1006 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1009 tdq_notify(struct tdq *tdq, struct thread *td)
1015 if (tdq->tdq_ipipending)
1017 cpu = td->td_sched->ts_cpu;
1018 pri = td->td_priority;
1019 ctd = pcpu_find(cpu)->pc_curthread;
1020 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1022 if (TD_IS_IDLETHREAD(ctd)) {
1024 * If the MD code has an idle wakeup routine try that before
1025 * falling back to IPI.
1027 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1030 tdq->tdq_ipipending = 1;
1031 ipi_cpu(cpu, IPI_PREEMPT);
1035 * Steals load from a timeshare queue. Honors the rotating queue head
1038 static struct thread *
1039 runq_steal_from(struct runq *rq, int cpu, u_char start)
1043 struct thread *td, *first;
1048 rqb = &rq->rq_status;
1049 bit = start & (RQB_BPW -1);
1053 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1054 if (rqb->rqb_bits[i] == 0)
1057 for (pri = bit; pri < RQB_BPW; pri++)
1058 if (rqb->rqb_bits[i] & (1ul << pri))
1063 pri = RQB_FFS(rqb->rqb_bits[i]);
1064 pri += (i << RQB_L2BPW);
1065 rqh = &rq->rq_queues[pri];
1066 TAILQ_FOREACH(td, rqh, td_runq) {
1067 if (first && THREAD_CAN_MIGRATE(td) &&
1068 THREAD_CAN_SCHED(td, cpu))
1078 if (first && THREAD_CAN_MIGRATE(first) &&
1079 THREAD_CAN_SCHED(first, cpu))
1085 * Steals load from a standard linear queue.
1087 static struct thread *
1088 runq_steal(struct runq *rq, int cpu)
1096 rqb = &rq->rq_status;
1097 for (word = 0; word < RQB_LEN; word++) {
1098 if (rqb->rqb_bits[word] == 0)
1100 for (bit = 0; bit < RQB_BPW; bit++) {
1101 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1103 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1104 TAILQ_FOREACH(td, rqh, td_runq)
1105 if (THREAD_CAN_MIGRATE(td) &&
1106 THREAD_CAN_SCHED(td, cpu))
1114 * Attempt to steal a thread in priority order from a thread queue.
1116 static struct thread *
1117 tdq_steal(struct tdq *tdq, int cpu)
1121 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1122 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1124 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1125 cpu, tdq->tdq_ridx)) != NULL)
1127 return (runq_steal(&tdq->tdq_idle, cpu));
1131 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1132 * current lock and returns with the assigned queue locked.
1134 static inline struct tdq *
1135 sched_setcpu(struct thread *td, int cpu, int flags)
1140 THREAD_LOCK_ASSERT(td, MA_OWNED);
1142 td->td_sched->ts_cpu = cpu;
1144 * If the lock matches just return the queue.
1146 if (td->td_lock == TDQ_LOCKPTR(tdq))
1150 * If the thread isn't running its lockptr is a
1151 * turnstile or a sleepqueue. We can just lock_set without
1154 if (TD_CAN_RUN(td)) {
1156 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1161 * The hard case, migration, we need to block the thread first to
1162 * prevent order reversals with other cpus locks.
1165 thread_lock_block(td);
1167 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1172 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1173 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1174 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1175 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1176 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1177 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1180 sched_pickcpu(struct thread *td, int flags)
1182 struct cpu_group *cg, *ccg;
1183 struct td_sched *ts;
1188 self = PCPU_GET(cpuid);
1190 if (smp_started == 0)
1193 * Don't migrate a running thread from sched_switch().
1195 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1196 return (ts->ts_cpu);
1198 * Prefer to run interrupt threads on the processors that generate
1201 pri = td->td_priority;
1202 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1203 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1204 SCHED_STAT_INC(pickcpu_intrbind);
1206 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1207 SCHED_STAT_INC(pickcpu_affinity);
1208 return (ts->ts_cpu);
1212 * If the thread can run on the last cpu and the affinity has not
1213 * expired or it is idle run it there.
1215 tdq = TDQ_CPU(ts->ts_cpu);
1217 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1218 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1219 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1220 if (cg->cg_flags & CG_FLAG_THREAD) {
1221 CPUSET_FOREACH(cpu, cg->cg_mask) {
1222 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1227 if (cpu > mp_maxid) {
1228 SCHED_STAT_INC(pickcpu_idle_affinity);
1229 return (ts->ts_cpu);
1233 * Search for the last level cache CPU group in the tree.
1234 * Skip caches with expired affinity time and SMT groups.
1235 * Affinity to higher level caches will be handled less aggressively.
1237 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1238 if (cg->cg_flags & CG_FLAG_THREAD)
1240 if (!SCHED_AFFINITY(ts, cg->cg_level))
1247 /* Search the group for the less loaded idle CPU we can run now. */
1248 mask = td->td_cpuset->cs_mask;
1249 if (cg != NULL && cg != cpu_top &&
1250 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1251 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1252 INT_MAX, ts->ts_cpu);
1253 /* Search globally for the less loaded CPU we can run now. */
1255 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1256 /* Search globally for the less loaded CPU. */
1258 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1259 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1261 * Compare the lowest loaded cpu to current cpu.
1263 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1264 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1265 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1266 SCHED_STAT_INC(pickcpu_local);
1269 SCHED_STAT_INC(pickcpu_lowest);
1270 if (cpu != ts->ts_cpu)
1271 SCHED_STAT_INC(pickcpu_migration);
1277 * Pick the highest priority task we have and return it.
1279 static struct thread *
1280 tdq_choose(struct tdq *tdq)
1284 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1285 td = runq_choose(&tdq->tdq_realtime);
1288 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1290 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1291 ("tdq_choose: Invalid priority on timeshare queue %d",
1295 td = runq_choose(&tdq->tdq_idle);
1297 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1298 ("tdq_choose: Invalid priority on idle queue %d",
1307 * Initialize a thread queue.
1310 tdq_setup(struct tdq *tdq)
1314 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1315 runq_init(&tdq->tdq_realtime);
1316 runq_init(&tdq->tdq_timeshare);
1317 runq_init(&tdq->tdq_idle);
1318 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1319 "sched lock %d", (int)TDQ_ID(tdq));
1320 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1321 MTX_SPIN | MTX_RECURSE);
1323 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1324 "CPU %d load", (int)TDQ_ID(tdq));
1330 sched_setup_smp(void)
1335 cpu_top = smp_topo();
1339 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1340 if (tdq->tdq_cg == NULL)
1341 panic("Can't find cpu group for %d\n", i);
1343 balance_tdq = TDQ_SELF();
1349 * Setup the thread queues and initialize the topology based on MD
1353 sched_setup(void *dummy)
1364 * To avoid divide-by-zero, we set realstathz a dummy value
1365 * in case which sched_clock() called before sched_initticks().
1368 sched_slice = (realstathz/10); /* ~100ms */
1369 tickincr = 1 << SCHED_TICK_SHIFT;
1371 /* Add thread0's load since it's running. */
1373 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1374 tdq_load_add(tdq, &thread0);
1375 tdq->tdq_lowpri = thread0.td_priority;
1380 * This routine determines the tickincr after stathz and hz are setup.
1384 sched_initticks(void *dummy)
1388 realstathz = stathz ? stathz : hz;
1389 sched_slice = (realstathz/10); /* ~100ms */
1392 * tickincr is shifted out by 10 to avoid rounding errors due to
1393 * hz not being evenly divisible by stathz on all platforms.
1395 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1397 * This does not work for values of stathz that are more than
1398 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1405 * Set the default balance interval now that we know
1406 * what realstathz is.
1408 balance_interval = realstathz;
1410 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1411 * This prevents excess thrashing on large machines and excess idle
1412 * on smaller machines.
1414 steal_thresh = min(fls(mp_ncpus) - 1, 3);
1415 affinity = SCHED_AFFINITY_DEFAULT;
1417 if (sched_idlespinthresh < 0)
1418 sched_idlespinthresh = max(16, 2 * hz / realstathz);
1423 * This is the core of the interactivity algorithm. Determines a score based
1424 * on past behavior. It is the ratio of sleep time to run time scaled to
1425 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1426 * differs from the cpu usage because it does not account for time spent
1427 * waiting on a run-queue. Would be prettier if we had floating point.
1430 sched_interact_score(struct thread *td)
1432 struct td_sched *ts;
1437 * The score is only needed if this is likely to be an interactive
1438 * task. Don't go through the expense of computing it if there's
1441 if (sched_interact <= SCHED_INTERACT_HALF &&
1442 ts->ts_runtime >= ts->ts_slptime)
1443 return (SCHED_INTERACT_HALF);
1445 if (ts->ts_runtime > ts->ts_slptime) {
1446 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1447 return (SCHED_INTERACT_HALF +
1448 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1450 if (ts->ts_slptime > ts->ts_runtime) {
1451 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1452 return (ts->ts_runtime / div);
1454 /* runtime == slptime */
1456 return (SCHED_INTERACT_HALF);
1459 * This can happen if slptime and runtime are 0.
1466 * Scale the scheduling priority according to the "interactivity" of this
1470 sched_priority(struct thread *td)
1475 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1478 * If the score is interactive we place the thread in the realtime
1479 * queue with a priority that is less than kernel and interrupt
1480 * priorities. These threads are not subject to nice restrictions.
1482 * Scores greater than this are placed on the normal timeshare queue
1483 * where the priority is partially decided by the most recent cpu
1484 * utilization and the rest is decided by nice value.
1486 * The nice value of the process has a linear effect on the calculated
1487 * score. Negative nice values make it easier for a thread to be
1488 * considered interactive.
1490 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1491 if (score < sched_interact) {
1492 pri = PRI_MIN_INTERACT;
1493 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1494 sched_interact) * score;
1495 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1496 ("sched_priority: invalid interactive priority %d score %d",
1499 pri = SCHED_PRI_MIN;
1500 if (td->td_sched->ts_ticks)
1501 pri += min(SCHED_PRI_TICKS(td->td_sched),
1503 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1504 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1505 ("sched_priority: invalid priority %d: nice %d, "
1506 "ticks %d ftick %d ltick %d tick pri %d",
1507 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1508 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1509 SCHED_PRI_TICKS(td->td_sched)));
1511 sched_user_prio(td, pri);
1517 * This routine enforces a maximum limit on the amount of scheduling history
1518 * kept. It is called after either the slptime or runtime is adjusted. This
1519 * function is ugly due to integer math.
1522 sched_interact_update(struct thread *td)
1524 struct td_sched *ts;
1528 sum = ts->ts_runtime + ts->ts_slptime;
1529 if (sum < SCHED_SLP_RUN_MAX)
1532 * This only happens from two places:
1533 * 1) We have added an unusual amount of run time from fork_exit.
1534 * 2) We have added an unusual amount of sleep time from sched_sleep().
1536 if (sum > SCHED_SLP_RUN_MAX * 2) {
1537 if (ts->ts_runtime > ts->ts_slptime) {
1538 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1541 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1547 * If we have exceeded by more than 1/5th then the algorithm below
1548 * will not bring us back into range. Dividing by two here forces
1549 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1551 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1552 ts->ts_runtime /= 2;
1553 ts->ts_slptime /= 2;
1556 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1557 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1561 * Scale back the interactivity history when a child thread is created. The
1562 * history is inherited from the parent but the thread may behave totally
1563 * differently. For example, a shell spawning a compiler process. We want
1564 * to learn that the compiler is behaving badly very quickly.
1567 sched_interact_fork(struct thread *td)
1572 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1573 if (sum > SCHED_SLP_RUN_FORK) {
1574 ratio = sum / SCHED_SLP_RUN_FORK;
1575 td->td_sched->ts_runtime /= ratio;
1576 td->td_sched->ts_slptime /= ratio;
1581 * Called from proc0_init() to setup the scheduler fields.
1588 * Set up the scheduler specific parts of proc0.
1590 proc0.p_sched = NULL; /* XXX */
1591 thread0.td_sched = &td_sched0;
1592 td_sched0.ts_ltick = ticks;
1593 td_sched0.ts_ftick = ticks;
1594 td_sched0.ts_slice = sched_slice;
1598 * This is only somewhat accurate since given many processes of the same
1599 * priority they will switch when their slices run out, which will be
1600 * at most sched_slice stathz ticks.
1603 sched_rr_interval(void)
1606 /* Convert sched_slice to hz */
1607 return (hz/(realstathz/sched_slice));
1611 * Update the percent cpu tracking information when it is requested or
1612 * the total history exceeds the maximum. We keep a sliding history of
1613 * tick counts that slowly decays. This is less precise than the 4BSD
1614 * mechanism since it happens with less regular and frequent events.
1617 sched_pctcpu_update(struct td_sched *ts, int run)
1621 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1623 ts->ts_ftick = t - SCHED_TICK_TARG;
1624 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1625 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1626 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1627 ts->ts_ftick = t - SCHED_TICK_TARG;
1630 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1635 * Adjust the priority of a thread. Move it to the appropriate run-queue
1636 * if necessary. This is the back-end for several priority related
1640 sched_thread_priority(struct thread *td, u_char prio)
1642 struct td_sched *ts;
1646 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1647 "prio:%d", td->td_priority, "new prio:%d", prio,
1648 KTR_ATTR_LINKED, sched_tdname(curthread));
1649 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
1650 if (td != curthread && prio > td->td_priority) {
1651 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1652 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1653 prio, KTR_ATTR_LINKED, sched_tdname(td));
1654 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
1658 THREAD_LOCK_ASSERT(td, MA_OWNED);
1659 if (td->td_priority == prio)
1662 * If the priority has been elevated due to priority
1663 * propagation, we may have to move ourselves to a new
1664 * queue. This could be optimized to not re-add in some
1667 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1669 td->td_priority = prio;
1670 sched_add(td, SRQ_BORROWING);
1674 * If the thread is currently running we may have to adjust the lowpri
1675 * information so other cpus are aware of our current priority.
1677 if (TD_IS_RUNNING(td)) {
1678 tdq = TDQ_CPU(ts->ts_cpu);
1679 oldpri = td->td_priority;
1680 td->td_priority = prio;
1681 if (prio < tdq->tdq_lowpri)
1682 tdq->tdq_lowpri = prio;
1683 else if (tdq->tdq_lowpri == oldpri)
1684 tdq_setlowpri(tdq, td);
1687 td->td_priority = prio;
1691 * Update a thread's priority when it is lent another thread's
1695 sched_lend_prio(struct thread *td, u_char prio)
1698 td->td_flags |= TDF_BORROWING;
1699 sched_thread_priority(td, prio);
1703 * Restore a thread's priority when priority propagation is
1704 * over. The prio argument is the minimum priority the thread
1705 * needs to have to satisfy other possible priority lending
1706 * requests. If the thread's regular priority is less
1707 * important than prio, the thread will keep a priority boost
1711 sched_unlend_prio(struct thread *td, u_char prio)
1715 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1716 td->td_base_pri <= PRI_MAX_TIMESHARE)
1717 base_pri = td->td_user_pri;
1719 base_pri = td->td_base_pri;
1720 if (prio >= base_pri) {
1721 td->td_flags &= ~TDF_BORROWING;
1722 sched_thread_priority(td, base_pri);
1724 sched_lend_prio(td, prio);
1728 * Standard entry for setting the priority to an absolute value.
1731 sched_prio(struct thread *td, u_char prio)
1735 /* First, update the base priority. */
1736 td->td_base_pri = prio;
1739 * If the thread is borrowing another thread's priority, don't
1740 * ever lower the priority.
1742 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1745 /* Change the real priority. */
1746 oldprio = td->td_priority;
1747 sched_thread_priority(td, prio);
1750 * If the thread is on a turnstile, then let the turnstile update
1753 if (TD_ON_LOCK(td) && oldprio != prio)
1754 turnstile_adjust(td, oldprio);
1758 * Set the base user priority, does not effect current running priority.
1761 sched_user_prio(struct thread *td, u_char prio)
1764 td->td_base_user_pri = prio;
1765 if (td->td_lend_user_pri <= prio)
1767 td->td_user_pri = prio;
1771 sched_lend_user_prio(struct thread *td, u_char prio)
1774 THREAD_LOCK_ASSERT(td, MA_OWNED);
1775 td->td_lend_user_pri = prio;
1776 td->td_user_pri = min(prio, td->td_base_user_pri);
1777 if (td->td_priority > td->td_user_pri)
1778 sched_prio(td, td->td_user_pri);
1779 else if (td->td_priority != td->td_user_pri)
1780 td->td_flags |= TDF_NEEDRESCHED;
1784 * Handle migration from sched_switch(). This happens only for
1788 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1792 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1794 tdq_load_rem(tdq, td);
1796 * Do the lock dance required to avoid LOR. We grab an extra
1797 * spinlock nesting to prevent preemption while we're
1798 * not holding either run-queue lock.
1801 thread_lock_block(td); /* This releases the lock on tdq. */
1804 * Acquire both run-queue locks before placing the thread on the new
1805 * run-queue to avoid deadlocks created by placing a thread with a
1806 * blocked lock on the run-queue of a remote processor. The deadlock
1807 * occurs when a third processor attempts to lock the two queues in
1808 * question while the target processor is spinning with its own
1809 * run-queue lock held while waiting for the blocked lock to clear.
1811 tdq_lock_pair(tdn, tdq);
1812 tdq_add(tdn, td, flags);
1813 tdq_notify(tdn, td);
1817 return (TDQ_LOCKPTR(tdn));
1821 * Variadic version of thread_lock_unblock() that does not assume td_lock
1825 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1827 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1832 * Switch threads. This function has to handle threads coming in while
1833 * blocked for some reason, running, or idle. It also must deal with
1834 * migrating a thread from one queue to another as running threads may
1835 * be assigned elsewhere via binding.
1838 sched_switch(struct thread *td, struct thread *newtd, int flags)
1841 struct td_sched *ts;
1846 THREAD_LOCK_ASSERT(td, MA_OWNED);
1847 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1849 cpuid = PCPU_GET(cpuid);
1850 tdq = TDQ_CPU(cpuid);
1853 sched_pctcpu_update(ts, 1);
1854 ts->ts_rltick = ticks;
1855 td->td_lastcpu = td->td_oncpu;
1856 td->td_oncpu = NOCPU;
1857 if (!(flags & SW_PREEMPT))
1858 td->td_flags &= ~TDF_NEEDRESCHED;
1859 td->td_owepreempt = 0;
1860 tdq->tdq_switchcnt++;
1862 * The lock pointer in an idle thread should never change. Reset it
1863 * to CAN_RUN as well.
1865 if (TD_IS_IDLETHREAD(td)) {
1866 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1868 } else if (TD_IS_RUNNING(td)) {
1869 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1870 srqflag = (flags & SW_PREEMPT) ?
1871 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1872 SRQ_OURSELF|SRQ_YIELDING;
1874 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1875 ts->ts_cpu = sched_pickcpu(td, 0);
1877 if (ts->ts_cpu == cpuid)
1878 tdq_runq_add(tdq, td, srqflag);
1880 KASSERT(THREAD_CAN_MIGRATE(td) ||
1881 (ts->ts_flags & TSF_BOUND) != 0,
1882 ("Thread %p shouldn't migrate", td));
1883 mtx = sched_switch_migrate(tdq, td, srqflag);
1886 /* This thread must be going to sleep. */
1888 mtx = thread_lock_block(td);
1889 tdq_load_rem(tdq, td);
1892 * We enter here with the thread blocked and assigned to the
1893 * appropriate cpu run-queue or sleep-queue and with the current
1894 * thread-queue locked.
1896 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1897 newtd = choosethread();
1899 * Call the MD code to switch contexts if necessary.
1903 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1904 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1906 SDT_PROBE2(sched, , , off_cpu, td, td->td_proc);
1907 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1908 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1909 sched_pctcpu_update(newtd->td_sched, 0);
1911 #ifdef KDTRACE_HOOKS
1913 * If DTrace has set the active vtime enum to anything
1914 * other than INACTIVE (0), then it should have set the
1917 if (dtrace_vtime_active)
1918 (*dtrace_vtime_switch_func)(newtd);
1921 cpu_switch(td, newtd, mtx);
1923 * We may return from cpu_switch on a different cpu. However,
1924 * we always return with td_lock pointing to the current cpu's
1927 cpuid = PCPU_GET(cpuid);
1928 tdq = TDQ_CPU(cpuid);
1929 lock_profile_obtain_lock_success(
1930 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1932 SDT_PROBE0(sched, , , on_cpu);
1934 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1935 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1938 thread_unblock_switch(td, mtx);
1939 SDT_PROBE0(sched, , , remain_cpu);
1942 * Assert that all went well and return.
1944 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1945 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1946 td->td_oncpu = cpuid;
1950 * Adjust thread priorities as a result of a nice request.
1953 sched_nice(struct proc *p, int nice)
1957 PROC_LOCK_ASSERT(p, MA_OWNED);
1960 FOREACH_THREAD_IN_PROC(p, td) {
1963 sched_prio(td, td->td_base_user_pri);
1969 * Record the sleep time for the interactivity scorer.
1972 sched_sleep(struct thread *td, int prio)
1975 THREAD_LOCK_ASSERT(td, MA_OWNED);
1977 td->td_slptick = ticks;
1978 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1979 td->td_flags |= TDF_CANSWAP;
1980 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1982 if (static_boost == 1 && prio)
1983 sched_prio(td, prio);
1984 else if (static_boost && td->td_priority > static_boost)
1985 sched_prio(td, static_boost);
1989 * Schedule a thread to resume execution and record how long it voluntarily
1990 * slept. We also update the pctcpu, interactivity, and priority.
1993 sched_wakeup(struct thread *td)
1995 struct td_sched *ts;
1998 THREAD_LOCK_ASSERT(td, MA_OWNED);
2000 td->td_flags &= ~TDF_CANSWAP;
2002 * If we slept for more than a tick update our interactivity and
2005 slptick = td->td_slptick;
2007 if (slptick && slptick != ticks) {
2008 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2009 sched_interact_update(td);
2010 sched_pctcpu_update(ts, 0);
2012 /* Reset the slice value after we sleep. */
2013 ts->ts_slice = sched_slice;
2014 sched_add(td, SRQ_BORING);
2018 * Penalize the parent for creating a new child and initialize the child's
2022 sched_fork(struct thread *td, struct thread *child)
2024 THREAD_LOCK_ASSERT(td, MA_OWNED);
2025 sched_pctcpu_update(td->td_sched, 1);
2026 sched_fork_thread(td, child);
2028 * Penalize the parent and child for forking.
2030 sched_interact_fork(child);
2031 sched_priority(child);
2032 td->td_sched->ts_runtime += tickincr;
2033 sched_interact_update(td);
2038 * Fork a new thread, may be within the same process.
2041 sched_fork_thread(struct thread *td, struct thread *child)
2043 struct td_sched *ts;
2044 struct td_sched *ts2;
2046 THREAD_LOCK_ASSERT(td, MA_OWNED);
2051 ts2 = child->td_sched;
2052 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
2053 child->td_cpuset = cpuset_ref(td->td_cpuset);
2054 ts2->ts_cpu = ts->ts_cpu;
2057 * Grab our parents cpu estimation information.
2059 ts2->ts_ticks = ts->ts_ticks;
2060 ts2->ts_ltick = ts->ts_ltick;
2061 ts2->ts_ftick = ts->ts_ftick;
2063 * Do not inherit any borrowed priority from the parent.
2065 child->td_priority = child->td_base_pri;
2067 * And update interactivity score.
2069 ts2->ts_slptime = ts->ts_slptime;
2070 ts2->ts_runtime = ts->ts_runtime;
2071 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
2073 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2078 * Adjust the priority class of a thread.
2081 sched_class(struct thread *td, int class)
2084 THREAD_LOCK_ASSERT(td, MA_OWNED);
2085 if (td->td_pri_class == class)
2087 td->td_pri_class = class;
2091 * Return some of the child's priority and interactivity to the parent.
2094 sched_exit(struct proc *p, struct thread *child)
2098 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2099 "prio:%d", child->td_priority);
2100 PROC_LOCK_ASSERT(p, MA_OWNED);
2101 td = FIRST_THREAD_IN_PROC(p);
2102 sched_exit_thread(td, child);
2106 * Penalize another thread for the time spent on this one. This helps to
2107 * worsen the priority and interactivity of processes which schedule batch
2108 * jobs such as make. This has little effect on the make process itself but
2109 * causes new processes spawned by it to receive worse scores immediately.
2112 sched_exit_thread(struct thread *td, struct thread *child)
2115 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2116 "prio:%d", child->td_priority);
2118 * Give the child's runtime to the parent without returning the
2119 * sleep time as a penalty to the parent. This causes shells that
2120 * launch expensive things to mark their children as expensive.
2123 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2124 sched_interact_update(td);
2130 sched_preempt(struct thread *td)
2134 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2138 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2139 tdq->tdq_ipipending = 0;
2140 if (td->td_priority > tdq->tdq_lowpri) {
2143 flags = SW_INVOL | SW_PREEMPT;
2144 if (td->td_critnest > 1)
2145 td->td_owepreempt = 1;
2146 else if (TD_IS_IDLETHREAD(td))
2147 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2149 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2155 * Fix priorities on return to user-space. Priorities may be elevated due
2156 * to static priorities in msleep() or similar.
2159 sched_userret(struct thread *td)
2162 * XXX we cheat slightly on the locking here to avoid locking in
2163 * the usual case. Setting td_priority here is essentially an
2164 * incomplete workaround for not setting it properly elsewhere.
2165 * Now that some interrupt handlers are threads, not setting it
2166 * properly elsewhere can clobber it in the window between setting
2167 * it here and returning to user mode, so don't waste time setting
2168 * it perfectly here.
2170 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2171 ("thread with borrowed priority returning to userland"));
2172 if (td->td_priority != td->td_user_pri) {
2174 td->td_priority = td->td_user_pri;
2175 td->td_base_pri = td->td_user_pri;
2176 tdq_setlowpri(TDQ_SELF(), td);
2182 * Handle a stathz tick. This is really only relevant for timeshare
2186 sched_clock(struct thread *td)
2189 struct td_sched *ts;
2191 THREAD_LOCK_ASSERT(td, MA_OWNED);
2195 * We run the long term load balancer infrequently on the first cpu.
2197 if (balance_tdq == tdq) {
2198 if (balance_ticks && --balance_ticks == 0)
2203 * Save the old switch count so we have a record of the last ticks
2204 * activity. Initialize the new switch count based on our load.
2205 * If there is some activity seed it to reflect that.
2207 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2208 tdq->tdq_switchcnt = tdq->tdq_load;
2210 * Advance the insert index once for each tick to ensure that all
2211 * threads get a chance to run.
2213 if (tdq->tdq_idx == tdq->tdq_ridx) {
2214 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2215 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2216 tdq->tdq_ridx = tdq->tdq_idx;
2219 sched_pctcpu_update(ts, 1);
2220 if (td->td_pri_class & PRI_FIFO_BIT)
2222 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2224 * We used a tick; charge it to the thread so
2225 * that we can compute our interactivity.
2227 td->td_sched->ts_runtime += tickincr;
2228 sched_interact_update(td);
2232 * We used up one time slice.
2234 if (--ts->ts_slice > 0)
2237 * We're out of time, force a requeue at userret().
2239 ts->ts_slice = sched_slice;
2240 td->td_flags |= TDF_NEEDRESCHED;
2244 * Called once per hz tick.
2253 * Return whether the current CPU has runnable tasks. Used for in-kernel
2254 * cooperative idle threads.
2257 sched_runnable(void)
2265 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2266 if (tdq->tdq_load > 0)
2269 if (tdq->tdq_load - 1 > 0)
2277 * Choose the highest priority thread to run. The thread is removed from
2278 * the run-queue while running however the load remains. For SMP we set
2279 * the tdq in the global idle bitmask if it idles here.
2288 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2289 td = tdq_choose(tdq);
2291 tdq_runq_rem(tdq, td);
2292 tdq->tdq_lowpri = td->td_priority;
2295 tdq->tdq_lowpri = PRI_MAX_IDLE;
2296 return (PCPU_GET(idlethread));
2300 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2301 * we always request it once we exit a critical section.
2304 sched_setpreempt(struct thread *td)
2310 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2313 pri = td->td_priority;
2314 cpri = ctd->td_priority;
2316 ctd->td_flags |= TDF_NEEDRESCHED;
2317 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2319 if (!sched_shouldpreempt(pri, cpri, 0))
2321 ctd->td_owepreempt = 1;
2325 * Add a thread to a thread queue. Select the appropriate runq and add the
2326 * thread to it. This is the internal function called when the tdq is
2330 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2333 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2334 KASSERT((td->td_inhibitors == 0),
2335 ("sched_add: trying to run inhibited thread"));
2336 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2337 ("sched_add: bad thread state"));
2338 KASSERT(td->td_flags & TDF_INMEM,
2339 ("sched_add: thread swapped out"));
2341 if (td->td_priority < tdq->tdq_lowpri)
2342 tdq->tdq_lowpri = td->td_priority;
2343 tdq_runq_add(tdq, td, flags);
2344 tdq_load_add(tdq, td);
2348 * Select the target thread queue and add a thread to it. Request
2349 * preemption or IPI a remote processor if required.
2352 sched_add(struct thread *td, int flags)
2359 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2360 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2361 sched_tdname(curthread));
2362 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2363 KTR_ATTR_LINKED, sched_tdname(td));
2364 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2365 flags & SRQ_PREEMPTED);
2366 THREAD_LOCK_ASSERT(td, MA_OWNED);
2368 * Recalculate the priority before we select the target cpu or
2371 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2375 * Pick the destination cpu and if it isn't ours transfer to the
2378 cpu = sched_pickcpu(td, flags);
2379 tdq = sched_setcpu(td, cpu, flags);
2380 tdq_add(tdq, td, flags);
2381 if (cpu != PCPU_GET(cpuid)) {
2382 tdq_notify(tdq, td);
2389 * Now that the thread is moving to the run-queue, set the lock
2390 * to the scheduler's lock.
2392 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2393 tdq_add(tdq, td, flags);
2395 if (!(flags & SRQ_YIELDING))
2396 sched_setpreempt(td);
2400 * Remove a thread from a run-queue without running it. This is used
2401 * when we're stealing a thread from a remote queue. Otherwise all threads
2402 * exit by calling sched_exit_thread() and sched_throw() themselves.
2405 sched_rem(struct thread *td)
2409 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2410 "prio:%d", td->td_priority);
2411 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2412 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2413 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2414 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2415 KASSERT(TD_ON_RUNQ(td),
2416 ("sched_rem: thread not on run queue"));
2417 tdq_runq_rem(tdq, td);
2418 tdq_load_rem(tdq, td);
2420 if (td->td_priority == tdq->tdq_lowpri)
2421 tdq_setlowpri(tdq, NULL);
2425 * Fetch cpu utilization information. Updates on demand.
2428 sched_pctcpu(struct thread *td)
2431 struct td_sched *ts;
2438 THREAD_LOCK_ASSERT(td, MA_OWNED);
2439 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2443 /* How many rtick per second ? */
2444 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2445 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2452 * Enforce affinity settings for a thread. Called after adjustments to
2456 sched_affinity(struct thread *td)
2459 struct td_sched *ts;
2461 THREAD_LOCK_ASSERT(td, MA_OWNED);
2463 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2465 if (TD_ON_RUNQ(td)) {
2467 sched_add(td, SRQ_BORING);
2470 if (!TD_IS_RUNNING(td))
2473 * Force a switch before returning to userspace. If the
2474 * target thread is not running locally send an ipi to force
2477 td->td_flags |= TDF_NEEDRESCHED;
2478 if (td != curthread)
2479 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2484 * Bind a thread to a target cpu.
2487 sched_bind(struct thread *td, int cpu)
2489 struct td_sched *ts;
2491 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2492 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2494 if (ts->ts_flags & TSF_BOUND)
2496 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2497 ts->ts_flags |= TSF_BOUND;
2499 if (PCPU_GET(cpuid) == cpu)
2502 /* When we return from mi_switch we'll be on the correct cpu. */
2503 mi_switch(SW_VOL, NULL);
2507 * Release a bound thread.
2510 sched_unbind(struct thread *td)
2512 struct td_sched *ts;
2514 THREAD_LOCK_ASSERT(td, MA_OWNED);
2515 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2517 if ((ts->ts_flags & TSF_BOUND) == 0)
2519 ts->ts_flags &= ~TSF_BOUND;
2524 sched_is_bound(struct thread *td)
2526 THREAD_LOCK_ASSERT(td, MA_OWNED);
2527 return (td->td_sched->ts_flags & TSF_BOUND);
2534 sched_relinquish(struct thread *td)
2537 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2542 * Return the total system load.
2553 total += TDQ_CPU(i)->tdq_sysload;
2556 return (TDQ_SELF()->tdq_sysload);
2561 sched_sizeof_proc(void)
2563 return (sizeof(struct proc));
2567 sched_sizeof_thread(void)
2569 return (sizeof(struct thread) + sizeof(struct td_sched));
2573 #define TDQ_IDLESPIN(tdq) \
2574 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2576 #define TDQ_IDLESPIN(tdq) 1
2580 * The actual idle process.
2583 sched_idletd(void *dummy)
2590 mtx_assert(&Giant, MA_NOTOWNED);
2595 if (tdq_idled(tdq) == 0)
2598 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2600 * If we're switching very frequently, spin while checking
2601 * for load rather than entering a low power state that
2602 * may require an IPI. However, don't do any busy
2603 * loops while on SMT machines as this simply steals
2604 * cycles from cores doing useful work.
2606 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2607 for (i = 0; i < sched_idlespins; i++) {
2613 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2614 if (tdq->tdq_load == 0) {
2615 tdq->tdq_cpu_idle = 1;
2616 if (tdq->tdq_load == 0) {
2617 cpu_idle(switchcnt > sched_idlespinthresh * 4);
2618 tdq->tdq_switchcnt++;
2620 tdq->tdq_cpu_idle = 0;
2622 if (tdq->tdq_load) {
2624 mi_switch(SW_VOL | SWT_IDLE, NULL);
2631 * A CPU is entering for the first time or a thread is exiting.
2634 sched_throw(struct thread *td)
2636 struct thread *newtd;
2641 /* Correct spinlock nesting and acquire the correct lock. */
2644 PCPU_SET(switchtime, cpu_ticks());
2645 PCPU_SET(switchticks, ticks);
2647 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2648 tdq_load_rem(tdq, td);
2649 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2651 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2652 newtd = choosethread();
2653 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2654 cpu_throw(td, newtd); /* doesn't return */
2658 * This is called from fork_exit(). Just acquire the correct locks and
2659 * let fork do the rest of the work.
2662 sched_fork_exit(struct thread *td)
2664 struct td_sched *ts;
2669 * Finish setting up thread glue so that it begins execution in a
2670 * non-nested critical section with the scheduler lock held.
2672 cpuid = PCPU_GET(cpuid);
2673 tdq = TDQ_CPU(cpuid);
2675 if (TD_IS_IDLETHREAD(td))
2676 td->td_lock = TDQ_LOCKPTR(tdq);
2677 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2678 td->td_oncpu = cpuid;
2679 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2680 lock_profile_obtain_lock_success(
2681 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2685 * Create on first use to catch odd startup conditons.
2688 sched_tdname(struct thread *td)
2691 struct td_sched *ts;
2694 if (ts->ts_name[0] == '\0')
2695 snprintf(ts->ts_name, sizeof(ts->ts_name),
2696 "%s tid %d", td->td_name, td->td_tid);
2697 return (ts->ts_name);
2699 return (td->td_name);
2705 sched_clear_tdname(struct thread *td)
2707 struct td_sched *ts;
2710 ts->ts_name[0] = '\0';
2717 * Build the CPU topology dump string. Is recursively called to collect
2718 * the topology tree.
2721 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2724 char cpusetbuf[CPUSETBUFSIZ];
2727 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2728 "", 1 + indent / 2, cg->cg_level);
2729 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2730 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2732 for (i = 0; i < MAXCPU; i++) {
2733 if (CPU_ISSET(i, &cg->cg_mask)) {
2735 sbuf_printf(sb, ", ");
2738 sbuf_printf(sb, "%d", i);
2741 sbuf_printf(sb, "</cpu>\n");
2743 if (cg->cg_flags != 0) {
2744 sbuf_printf(sb, "%*s <flags>", indent, "");
2745 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2746 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2747 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2748 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2749 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2750 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2751 sbuf_printf(sb, "</flags>\n");
2754 if (cg->cg_children > 0) {
2755 sbuf_printf(sb, "%*s <children>\n", indent, "");
2756 for (i = 0; i < cg->cg_children; i++)
2757 sysctl_kern_sched_topology_spec_internal(sb,
2758 &cg->cg_child[i], indent+2);
2759 sbuf_printf(sb, "%*s </children>\n", indent, "");
2761 sbuf_printf(sb, "%*s</group>\n", indent, "");
2766 * Sysctl handler for retrieving topology dump. It's a wrapper for
2767 * the recursive sysctl_kern_smp_topology_spec_internal().
2770 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2775 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2777 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2781 sbuf_printf(topo, "<groups>\n");
2782 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2783 sbuf_printf(topo, "</groups>\n");
2787 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2795 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2796 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2798 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2799 "Slice size for timeshare threads");
2800 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2801 "Interactivity score threshold");
2802 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2803 0,"Min priority for preemption, lower priorities have greater precedence");
2804 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2805 0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2806 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2807 0,"Number of times idle will spin waiting for new work.");
2808 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2809 0,"Threshold before we will permit idle spinning.");
2811 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2812 "Number of hz ticks to keep thread affinity for");
2813 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2814 "Enables the long-term load balancer");
2815 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2816 &balance_interval, 0,
2817 "Average frequency in stathz ticks to run the long-term balancer");
2818 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2819 "Attempts to steal work from other cores before idling");
2820 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2821 "Minimum load on remote cpu before we'll steal");
2823 /* Retrieve SMP topology */
2824 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2825 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2826 "XML dump of detected CPU topology");
2830 /* ps compat. All cpu percentages from ULE are weighted. */
2831 static int ccpu = 0;
2832 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");