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(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)
192 /* Flags kept in td_flags. */
193 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
196 * tickincr: Converts a stathz tick into a hz domain scaled by
197 * the shift factor. Without the shift the error rate
198 * due to rounding would be unacceptably high.
199 * realstathz: stathz is sometimes 0 and run off of hz.
200 * sched_slice: Runtime of each thread before rescheduling.
201 * preempt_thresh: Priority threshold for preemption and remote IPIs.
203 static int sched_interact = SCHED_INTERACT_THRESH;
204 static int realstathz = 127;
205 static int tickincr = 8 << SCHED_TICK_SHIFT;
206 static int sched_slice = 12;
208 #ifdef FULL_PREEMPTION
209 static int preempt_thresh = PRI_MAX_IDLE;
211 static int preempt_thresh = PRI_MIN_KERN;
214 static int preempt_thresh = 0;
216 static int static_boost = PRI_MIN_BATCH;
217 static int sched_idlespins = 10000;
218 static int sched_idlespinthresh = -1;
221 * tdq - per processor runqs and statistics. All fields are protected by the
222 * tdq_lock. The load and lowpri may be accessed without to avoid excess
223 * locking in sched_pickcpu();
226 /* Ordered to improve efficiency of cpu_search() and switch(). */
227 struct mtx tdq_lock; /* run queue lock. */
228 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
229 volatile int tdq_load; /* Aggregate load. */
230 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
231 int tdq_sysload; /* For loadavg, !ITHD load. */
232 int tdq_transferable; /* Transferable thread count. */
233 short tdq_switchcnt; /* Switches this tick. */
234 short tdq_oldswitchcnt; /* Switches last tick. */
235 u_char tdq_lowpri; /* Lowest priority thread. */
236 u_char tdq_ipipending; /* IPI pending. */
237 u_char tdq_idx; /* Current insert index. */
238 u_char tdq_ridx; /* Current removal index. */
239 struct runq tdq_realtime; /* real-time run queue. */
240 struct runq tdq_timeshare; /* timeshare run queue. */
241 struct runq tdq_idle; /* Queue of IDLE threads. */
242 char tdq_name[TDQ_NAME_LEN];
244 char tdq_loadname[TDQ_LOADNAME_LEN];
248 /* Idle thread states and config. */
249 #define TDQ_RUNNING 1
253 struct cpu_group *cpu_top; /* CPU topology */
255 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
256 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
261 static int rebalance = 1;
262 static int balance_interval = 128; /* Default set in sched_initticks(). */
264 static int steal_idle = 1;
265 static int steal_thresh = 2;
268 * One thread queue per processor.
270 static struct tdq tdq_cpu[MAXCPU];
271 static struct tdq *balance_tdq;
272 static int balance_ticks;
273 static DPCPU_DEFINE(uint32_t, randomval);
275 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
276 #define TDQ_CPU(x) (&tdq_cpu[(x)])
277 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
279 static struct tdq tdq_cpu;
281 #define TDQ_ID(x) (0)
282 #define TDQ_SELF() (&tdq_cpu)
283 #define TDQ_CPU(x) (&tdq_cpu)
286 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
287 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
288 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
289 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
290 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
292 static void sched_priority(struct thread *);
293 static void sched_thread_priority(struct thread *, u_char);
294 static int sched_interact_score(struct thread *);
295 static void sched_interact_update(struct thread *);
296 static void sched_interact_fork(struct thread *);
297 static void sched_pctcpu_update(struct td_sched *, int);
299 /* Operations on per processor queues */
300 static struct thread *tdq_choose(struct tdq *);
301 static void tdq_setup(struct tdq *);
302 static void tdq_load_add(struct tdq *, struct thread *);
303 static void tdq_load_rem(struct tdq *, struct thread *);
304 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
305 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
306 static inline int sched_shouldpreempt(int, int, int);
307 void tdq_print(int cpu);
308 static void runq_print(struct runq *rq);
309 static void tdq_add(struct tdq *, struct thread *, int);
311 static int tdq_move(struct tdq *, struct tdq *);
312 static int tdq_idled(struct tdq *);
313 static void tdq_notify(struct tdq *, struct thread *);
314 static struct thread *tdq_steal(struct tdq *, int);
315 static struct thread *runq_steal(struct runq *, int);
316 static int sched_pickcpu(struct thread *, int);
317 static void sched_balance(void);
318 static int sched_balance_pair(struct tdq *, struct tdq *);
319 static inline struct tdq *sched_setcpu(struct thread *, int, int);
320 static inline void thread_unblock_switch(struct thread *, struct mtx *);
321 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
322 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
323 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
324 struct cpu_group *cg, int indent);
327 static void sched_setup(void *dummy);
328 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
330 static void sched_initticks(void *dummy);
331 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
334 SDT_PROVIDER_DEFINE(sched);
336 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
337 "struct proc *", "uint8_t");
338 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
339 "struct proc *", "void *");
340 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
341 "struct proc *", "void *", "int");
342 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
343 "struct proc *", "uint8_t", "struct thread *");
344 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
345 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
347 SDT_PROBE_DEFINE(sched, , , on__cpu);
348 SDT_PROBE_DEFINE(sched, , , remain__cpu);
349 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
353 * Print the threads waiting on a run-queue.
356 runq_print(struct runq *rq)
364 for (i = 0; i < RQB_LEN; i++) {
365 printf("\t\trunq bits %d 0x%zx\n",
366 i, rq->rq_status.rqb_bits[i]);
367 for (j = 0; j < RQB_BPW; j++)
368 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
369 pri = j + (i << RQB_L2BPW);
370 rqh = &rq->rq_queues[pri];
371 TAILQ_FOREACH(td, rqh, td_runq) {
372 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
373 td, td->td_name, td->td_priority,
374 td->td_rqindex, pri);
381 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
390 printf("tdq %d:\n", TDQ_ID(tdq));
391 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
392 printf("\tLock name: %s\n", tdq->tdq_name);
393 printf("\tload: %d\n", tdq->tdq_load);
394 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
395 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
396 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
397 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
398 printf("\tload transferable: %d\n", tdq->tdq_transferable);
399 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
400 printf("\trealtime runq:\n");
401 runq_print(&tdq->tdq_realtime);
402 printf("\ttimeshare runq:\n");
403 runq_print(&tdq->tdq_timeshare);
404 printf("\tidle runq:\n");
405 runq_print(&tdq->tdq_idle);
409 sched_shouldpreempt(int pri, int cpri, int remote)
412 * If the new priority is not better than the current priority there is
418 * Always preempt idle.
420 if (cpri >= PRI_MIN_IDLE)
423 * If preemption is disabled don't preempt others.
425 if (preempt_thresh == 0)
428 * Preempt if we exceed the threshold.
430 if (pri <= preempt_thresh)
433 * If we're interactive or better and there is non-interactive
434 * or worse running preempt only remote processors.
436 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
442 * Add a thread to the actual run-queue. Keeps transferable counts up to
443 * date with what is actually on the run-queue. Selects the correct
444 * queue position for timeshare threads.
447 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
452 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
453 THREAD_LOCK_ASSERT(td, MA_OWNED);
455 pri = td->td_priority;
458 if (THREAD_CAN_MIGRATE(td)) {
459 tdq->tdq_transferable++;
460 ts->ts_flags |= TSF_XFERABLE;
462 if (pri < PRI_MIN_BATCH) {
463 ts->ts_runq = &tdq->tdq_realtime;
464 } else if (pri <= PRI_MAX_BATCH) {
465 ts->ts_runq = &tdq->tdq_timeshare;
466 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
467 ("Invalid priority %d on timeshare runq", pri));
469 * This queue contains only priorities between MIN and MAX
470 * realtime. Use the whole queue to represent these values.
472 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
473 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
474 pri = (pri + tdq->tdq_idx) % RQ_NQS;
476 * This effectively shortens the queue by one so we
477 * can have a one slot difference between idx and
478 * ridx while we wait for threads to drain.
480 if (tdq->tdq_ridx != tdq->tdq_idx &&
481 pri == tdq->tdq_ridx)
482 pri = (unsigned char)(pri - 1) % RQ_NQS;
485 runq_add_pri(ts->ts_runq, td, pri, flags);
488 ts->ts_runq = &tdq->tdq_idle;
489 runq_add(ts->ts_runq, td, flags);
493 * Remove a thread from a run-queue. This typically happens when a thread
494 * is selected to run. Running threads are not on the queue and the
495 * transferable count does not reflect them.
498 tdq_runq_rem(struct tdq *tdq, struct thread *td)
503 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
504 KASSERT(ts->ts_runq != NULL,
505 ("tdq_runq_remove: thread %p null ts_runq", td));
506 if (ts->ts_flags & TSF_XFERABLE) {
507 tdq->tdq_transferable--;
508 ts->ts_flags &= ~TSF_XFERABLE;
510 if (ts->ts_runq == &tdq->tdq_timeshare) {
511 if (tdq->tdq_idx != tdq->tdq_ridx)
512 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
514 runq_remove_idx(ts->ts_runq, td, NULL);
516 runq_remove(ts->ts_runq, td);
520 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
521 * for this thread to the referenced thread queue.
524 tdq_load_add(struct tdq *tdq, struct thread *td)
527 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
528 THREAD_LOCK_ASSERT(td, MA_OWNED);
531 if ((td->td_flags & TDF_NOLOAD) == 0)
533 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
534 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
538 * Remove the load from a thread that is transitioning to a sleep state or
542 tdq_load_rem(struct tdq *tdq, struct thread *td)
545 THREAD_LOCK_ASSERT(td, MA_OWNED);
546 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
547 KASSERT(tdq->tdq_load != 0,
548 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
551 if ((td->td_flags & TDF_NOLOAD) == 0)
553 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
554 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
558 * Set lowpri to its exact value by searching the run-queue and
559 * evaluating curthread. curthread may be passed as an optimization.
562 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
566 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
568 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
569 td = tdq_choose(tdq);
570 if (td == NULL || td->td_priority > ctd->td_priority)
571 tdq->tdq_lowpri = ctd->td_priority;
573 tdq->tdq_lowpri = td->td_priority;
580 int cs_pri; /* Min priority for low. */
581 int cs_limit; /* Max load for low, min load for high. */
586 #define CPU_SEARCH_LOWEST 0x1
587 #define CPU_SEARCH_HIGHEST 0x2
588 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
590 #define CPUSET_FOREACH(cpu, mask) \
591 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
592 if (CPU_ISSET(cpu, &mask))
594 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
595 struct cpu_search *high, const int match);
596 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
597 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
598 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
599 struct cpu_search *high);
602 * Search the tree of cpu_groups for the lowest or highest loaded cpu
603 * according to the match argument. This routine actually compares the
604 * load on all paths through the tree and finds the least loaded cpu on
605 * the least loaded path, which may differ from the least loaded cpu in
606 * the system. This balances work among caches and busses.
608 * This inline is instantiated in three forms below using constants for the
609 * match argument. It is reduced to the minimum set for each case. It is
610 * also recursive to the depth of the tree.
613 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
614 struct cpu_search *high, const int match)
616 struct cpu_search lgroup;
617 struct cpu_search hgroup;
619 struct cpu_group *child;
621 int cpu, i, hload, lload, load, total, rnd, *rndptr;
624 cpumask = cg->cg_mask;
625 if (match & CPU_SEARCH_LOWEST) {
629 if (match & CPU_SEARCH_HIGHEST) {
634 /* Iterate through the child CPU groups and then remaining CPUs. */
635 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
637 #ifdef HAVE_INLINE_FFSL
638 cpu = CPU_FFS(&cpumask) - 1;
640 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
647 child = &cg->cg_child[i - 1];
649 if (match & CPU_SEARCH_LOWEST)
651 if (match & CPU_SEARCH_HIGHEST)
653 if (child) { /* Handle child CPU group. */
654 CPU_NAND(&cpumask, &child->cg_mask);
656 case CPU_SEARCH_LOWEST:
657 load = cpu_search_lowest(child, &lgroup);
659 case CPU_SEARCH_HIGHEST:
660 load = cpu_search_highest(child, &hgroup);
662 case CPU_SEARCH_BOTH:
663 load = cpu_search_both(child, &lgroup, &hgroup);
666 } else { /* Handle child CPU. */
667 CPU_CLR(cpu, &cpumask);
669 load = tdq->tdq_load * 256;
670 rndptr = DPCPU_PTR(randomval);
671 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
672 if (match & CPU_SEARCH_LOWEST) {
673 if (cpu == low->cs_prefer)
675 /* If that CPU is allowed and get data. */
676 if (tdq->tdq_lowpri > lgroup.cs_pri &&
677 tdq->tdq_load <= lgroup.cs_limit &&
678 CPU_ISSET(cpu, &lgroup.cs_mask)) {
680 lgroup.cs_load = load - rnd;
683 if (match & CPU_SEARCH_HIGHEST)
684 if (tdq->tdq_load >= hgroup.cs_limit &&
685 tdq->tdq_transferable &&
686 CPU_ISSET(cpu, &hgroup.cs_mask)) {
688 hgroup.cs_load = load - rnd;
693 /* We have info about child item. Compare it. */
694 if (match & CPU_SEARCH_LOWEST) {
695 if (lgroup.cs_cpu >= 0 &&
697 (load == lload && lgroup.cs_load < low->cs_load))) {
699 low->cs_cpu = lgroup.cs_cpu;
700 low->cs_load = lgroup.cs_load;
703 if (match & CPU_SEARCH_HIGHEST)
704 if (hgroup.cs_cpu >= 0 &&
706 (load == hload && hgroup.cs_load > high->cs_load))) {
708 high->cs_cpu = hgroup.cs_cpu;
709 high->cs_load = hgroup.cs_load;
713 if (i == 0 && CPU_EMPTY(&cpumask))
716 #ifndef HAVE_INLINE_FFSL
725 * cpu_search instantiations must pass constants to maintain the inline
729 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
731 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
735 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
737 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
741 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
742 struct cpu_search *high)
744 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
748 * Find the cpu with the least load via the least loaded path that has a
749 * lowpri greater than pri pri. A pri of -1 indicates any priority is
753 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
756 struct cpu_search low;
759 low.cs_prefer = prefer;
762 low.cs_limit = maxload;
763 cpu_search_lowest(cg, &low);
768 * Find the cpu with the highest load via the highest loaded path.
771 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
773 struct cpu_search high;
777 high.cs_limit = minload;
778 cpu_search_highest(cg, &high);
783 sched_balance_group(struct cpu_group *cg)
785 cpuset_t hmask, lmask;
786 int high, low, anylow;
790 high = sched_highest(cg, hmask, 1);
791 /* Stop if there is no more CPU with transferrable threads. */
794 CPU_CLR(high, &hmask);
795 CPU_COPY(&hmask, &lmask);
796 /* Stop if there is no more CPU left for low. */
797 if (CPU_EMPTY(&lmask))
801 low = sched_lowest(cg, lmask, -1,
802 TDQ_CPU(high)->tdq_load - 1, high);
803 /* Stop if we looked well and found no less loaded CPU. */
804 if (anylow && low == -1)
806 /* Go to next high if we found no less loaded CPU. */
809 /* Transfer thread from high to low. */
810 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
811 /* CPU that got thread can no longer be a donor. */
812 CPU_CLR(low, &hmask);
815 * If failed, then there is no threads on high
816 * that can run on this low. Drop low from low
817 * mask and look for different one.
819 CPU_CLR(low, &lmask);
832 * Select a random time between .5 * balance_interval and
833 * 1.5 * balance_interval.
835 balance_ticks = max(balance_interval / 2, 1);
836 balance_ticks += random() % balance_interval;
837 if (smp_started == 0 || rebalance == 0)
841 sched_balance_group(cpu_top);
846 * Lock two thread queues using their address to maintain lock order.
849 tdq_lock_pair(struct tdq *one, struct tdq *two)
853 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
856 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
861 * Unlock two thread queues. Order is not important here.
864 tdq_unlock_pair(struct tdq *one, struct tdq *two)
871 * Transfer load between two imbalanced thread queues.
874 sched_balance_pair(struct tdq *high, struct tdq *low)
879 tdq_lock_pair(high, low);
882 * Determine what the imbalance is and then adjust that to how many
883 * threads we actually have to give up (transferable).
885 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
886 (moved = tdq_move(high, low)) > 0) {
888 * In case the target isn't the current cpu IPI it to force a
889 * reschedule with the new workload.
893 if (cpu != PCPU_GET(cpuid))
894 ipi_cpu(cpu, IPI_PREEMPT);
897 tdq_unlock_pair(high, low);
902 * Move a thread from one thread queue to another.
905 tdq_move(struct tdq *from, struct tdq *to)
912 TDQ_LOCK_ASSERT(from, MA_OWNED);
913 TDQ_LOCK_ASSERT(to, MA_OWNED);
917 td = tdq_steal(tdq, cpu);
922 * Although the run queue is locked the thread may be blocked. Lock
923 * it to clear this and acquire the run-queue lock.
926 /* Drop recursive lock on from acquired via thread_lock(). */
930 td->td_lock = TDQ_LOCKPTR(to);
931 tdq_add(to, td, SRQ_YIELDING);
936 * This tdq has idled. Try to steal a thread from another cpu and switch
940 tdq_idled(struct tdq *tdq)
942 struct cpu_group *cg;
948 if (smp_started == 0 || steal_idle == 0)
951 CPU_CLR(PCPU_GET(cpuid), &mask);
952 /* We don't want to be preempted while we're iterating. */
954 for (cg = tdq->tdq_cg; cg != NULL; ) {
955 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
956 thresh = steal_thresh;
959 cpu = sched_highest(cg, mask, thresh);
964 steal = TDQ_CPU(cpu);
966 tdq_lock_pair(tdq, steal);
967 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
968 tdq_unlock_pair(tdq, steal);
972 * If a thread was added while interrupts were disabled don't
973 * steal one here. If we fail to acquire one due to affinity
974 * restrictions loop again with this cpu removed from the
977 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
978 tdq_unlock_pair(tdq, steal);
983 mi_switch(SW_VOL | SWT_IDLE, NULL);
984 thread_unlock(curthread);
993 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
996 tdq_notify(struct tdq *tdq, struct thread *td)
1002 if (tdq->tdq_ipipending)
1004 cpu = td->td_sched->ts_cpu;
1005 pri = td->td_priority;
1006 ctd = pcpu_find(cpu)->pc_curthread;
1007 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1011 * Make sure that tdq_load updated before calling this function
1012 * is globally visible before we read tdq_cpu_idle. Idle thread
1013 * accesses both of them without locks, and the order is important.
1017 if (TD_IS_IDLETHREAD(ctd)) {
1019 * If the MD code has an idle wakeup routine try that before
1020 * falling back to IPI.
1022 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1025 tdq->tdq_ipipending = 1;
1026 ipi_cpu(cpu, IPI_PREEMPT);
1030 * Steals load from a timeshare queue. Honors the rotating queue head
1033 static struct thread *
1034 runq_steal_from(struct runq *rq, int cpu, u_char start)
1038 struct thread *td, *first;
1043 rqb = &rq->rq_status;
1044 bit = start & (RQB_BPW -1);
1048 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1049 if (rqb->rqb_bits[i] == 0)
1052 for (pri = bit; pri < RQB_BPW; pri++)
1053 if (rqb->rqb_bits[i] & (1ul << pri))
1058 pri = RQB_FFS(rqb->rqb_bits[i]);
1059 pri += (i << RQB_L2BPW);
1060 rqh = &rq->rq_queues[pri];
1061 TAILQ_FOREACH(td, rqh, td_runq) {
1062 if (first && THREAD_CAN_MIGRATE(td) &&
1063 THREAD_CAN_SCHED(td, cpu))
1073 if (first && THREAD_CAN_MIGRATE(first) &&
1074 THREAD_CAN_SCHED(first, cpu))
1080 * Steals load from a standard linear queue.
1082 static struct thread *
1083 runq_steal(struct runq *rq, int cpu)
1091 rqb = &rq->rq_status;
1092 for (word = 0; word < RQB_LEN; word++) {
1093 if (rqb->rqb_bits[word] == 0)
1095 for (bit = 0; bit < RQB_BPW; bit++) {
1096 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1098 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1099 TAILQ_FOREACH(td, rqh, td_runq)
1100 if (THREAD_CAN_MIGRATE(td) &&
1101 THREAD_CAN_SCHED(td, cpu))
1109 * Attempt to steal a thread in priority order from a thread queue.
1111 static struct thread *
1112 tdq_steal(struct tdq *tdq, int cpu)
1116 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1117 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1119 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1120 cpu, tdq->tdq_ridx)) != NULL)
1122 return (runq_steal(&tdq->tdq_idle, cpu));
1126 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1127 * current lock and returns with the assigned queue locked.
1129 static inline struct tdq *
1130 sched_setcpu(struct thread *td, int cpu, int flags)
1135 THREAD_LOCK_ASSERT(td, MA_OWNED);
1137 td->td_sched->ts_cpu = cpu;
1139 * If the lock matches just return the queue.
1141 if (td->td_lock == TDQ_LOCKPTR(tdq))
1145 * If the thread isn't running its lockptr is a
1146 * turnstile or a sleepqueue. We can just lock_set without
1149 if (TD_CAN_RUN(td)) {
1151 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1156 * The hard case, migration, we need to block the thread first to
1157 * prevent order reversals with other cpus locks.
1160 thread_lock_block(td);
1162 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1167 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1168 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1169 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1170 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1171 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1172 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1175 sched_pickcpu(struct thread *td, int flags)
1177 struct cpu_group *cg, *ccg;
1178 struct td_sched *ts;
1183 self = PCPU_GET(cpuid);
1185 if (smp_started == 0)
1188 * Don't migrate a running thread from sched_switch().
1190 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1191 return (ts->ts_cpu);
1193 * Prefer to run interrupt threads on the processors that generate
1196 pri = td->td_priority;
1197 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1198 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1199 SCHED_STAT_INC(pickcpu_intrbind);
1201 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1202 SCHED_STAT_INC(pickcpu_affinity);
1203 return (ts->ts_cpu);
1207 * If the thread can run on the last cpu and the affinity has not
1208 * expired or it is idle run it there.
1210 tdq = TDQ_CPU(ts->ts_cpu);
1212 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1213 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1214 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1215 if (cg->cg_flags & CG_FLAG_THREAD) {
1216 CPUSET_FOREACH(cpu, cg->cg_mask) {
1217 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1222 if (cpu > mp_maxid) {
1223 SCHED_STAT_INC(pickcpu_idle_affinity);
1224 return (ts->ts_cpu);
1228 * Search for the last level cache CPU group in the tree.
1229 * Skip caches with expired affinity time and SMT groups.
1230 * Affinity to higher level caches will be handled less aggressively.
1232 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1233 if (cg->cg_flags & CG_FLAG_THREAD)
1235 if (!SCHED_AFFINITY(ts, cg->cg_level))
1242 /* Search the group for the less loaded idle CPU we can run now. */
1243 mask = td->td_cpuset->cs_mask;
1244 if (cg != NULL && cg != cpu_top &&
1245 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1246 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1247 INT_MAX, ts->ts_cpu);
1248 /* Search globally for the less loaded CPU we can run now. */
1250 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1251 /* Search globally for the less loaded CPU. */
1253 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1254 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1256 * Compare the lowest loaded cpu to current cpu.
1258 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1259 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1260 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1261 SCHED_STAT_INC(pickcpu_local);
1264 SCHED_STAT_INC(pickcpu_lowest);
1265 if (cpu != ts->ts_cpu)
1266 SCHED_STAT_INC(pickcpu_migration);
1272 * Pick the highest priority task we have and return it.
1274 static struct thread *
1275 tdq_choose(struct tdq *tdq)
1279 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1280 td = runq_choose(&tdq->tdq_realtime);
1283 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1285 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1286 ("tdq_choose: Invalid priority on timeshare queue %d",
1290 td = runq_choose(&tdq->tdq_idle);
1292 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1293 ("tdq_choose: Invalid priority on idle queue %d",
1302 * Initialize a thread queue.
1305 tdq_setup(struct tdq *tdq)
1309 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1310 runq_init(&tdq->tdq_realtime);
1311 runq_init(&tdq->tdq_timeshare);
1312 runq_init(&tdq->tdq_idle);
1313 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1314 "sched lock %d", (int)TDQ_ID(tdq));
1315 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1316 MTX_SPIN | MTX_RECURSE);
1318 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1319 "CPU %d load", (int)TDQ_ID(tdq));
1325 sched_setup_smp(void)
1330 cpu_top = smp_topo();
1334 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1335 if (tdq->tdq_cg == NULL)
1336 panic("Can't find cpu group for %d\n", i);
1338 balance_tdq = TDQ_SELF();
1344 * Setup the thread queues and initialize the topology based on MD
1348 sched_setup(void *dummy)
1359 /* Add thread0's load since it's running. */
1361 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1362 tdq_load_add(tdq, &thread0);
1363 tdq->tdq_lowpri = thread0.td_priority;
1368 * This routine determines time constants after stathz and hz are setup.
1372 sched_initticks(void *dummy)
1376 realstathz = stathz ? stathz : hz;
1377 sched_slice = realstathz / 10; /* ~100ms */
1378 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1382 * tickincr is shifted out by 10 to avoid rounding errors due to
1383 * hz not being evenly divisible by stathz on all platforms.
1385 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1387 * This does not work for values of stathz that are more than
1388 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1395 * Set the default balance interval now that we know
1396 * what realstathz is.
1398 balance_interval = realstathz;
1399 affinity = SCHED_AFFINITY_DEFAULT;
1401 if (sched_idlespinthresh < 0)
1402 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1407 * This is the core of the interactivity algorithm. Determines a score based
1408 * on past behavior. It is the ratio of sleep time to run time scaled to
1409 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1410 * differs from the cpu usage because it does not account for time spent
1411 * waiting on a run-queue. Would be prettier if we had floating point.
1414 sched_interact_score(struct thread *td)
1416 struct td_sched *ts;
1421 * The score is only needed if this is likely to be an interactive
1422 * task. Don't go through the expense of computing it if there's
1425 if (sched_interact <= SCHED_INTERACT_HALF &&
1426 ts->ts_runtime >= ts->ts_slptime)
1427 return (SCHED_INTERACT_HALF);
1429 if (ts->ts_runtime > ts->ts_slptime) {
1430 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1431 return (SCHED_INTERACT_HALF +
1432 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1434 if (ts->ts_slptime > ts->ts_runtime) {
1435 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1436 return (ts->ts_runtime / div);
1438 /* runtime == slptime */
1440 return (SCHED_INTERACT_HALF);
1443 * This can happen if slptime and runtime are 0.
1450 * Scale the scheduling priority according to the "interactivity" of this
1454 sched_priority(struct thread *td)
1459 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1462 * If the score is interactive we place the thread in the realtime
1463 * queue with a priority that is less than kernel and interrupt
1464 * priorities. These threads are not subject to nice restrictions.
1466 * Scores greater than this are placed on the normal timeshare queue
1467 * where the priority is partially decided by the most recent cpu
1468 * utilization and the rest is decided by nice value.
1470 * The nice value of the process has a linear effect on the calculated
1471 * score. Negative nice values make it easier for a thread to be
1472 * considered interactive.
1474 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1475 if (score < sched_interact) {
1476 pri = PRI_MIN_INTERACT;
1477 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1478 sched_interact) * score;
1479 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1480 ("sched_priority: invalid interactive priority %d score %d",
1483 pri = SCHED_PRI_MIN;
1484 if (td->td_sched->ts_ticks)
1485 pri += min(SCHED_PRI_TICKS(td->td_sched),
1486 SCHED_PRI_RANGE - 1);
1487 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1488 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1489 ("sched_priority: invalid priority %d: nice %d, "
1490 "ticks %d ftick %d ltick %d tick pri %d",
1491 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1492 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1493 SCHED_PRI_TICKS(td->td_sched)));
1495 sched_user_prio(td, pri);
1501 * This routine enforces a maximum limit on the amount of scheduling history
1502 * kept. It is called after either the slptime or runtime is adjusted. This
1503 * function is ugly due to integer math.
1506 sched_interact_update(struct thread *td)
1508 struct td_sched *ts;
1512 sum = ts->ts_runtime + ts->ts_slptime;
1513 if (sum < SCHED_SLP_RUN_MAX)
1516 * This only happens from two places:
1517 * 1) We have added an unusual amount of run time from fork_exit.
1518 * 2) We have added an unusual amount of sleep time from sched_sleep().
1520 if (sum > SCHED_SLP_RUN_MAX * 2) {
1521 if (ts->ts_runtime > ts->ts_slptime) {
1522 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1525 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1531 * If we have exceeded by more than 1/5th then the algorithm below
1532 * will not bring us back into range. Dividing by two here forces
1533 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1535 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1536 ts->ts_runtime /= 2;
1537 ts->ts_slptime /= 2;
1540 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1541 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1545 * Scale back the interactivity history when a child thread is created. The
1546 * history is inherited from the parent but the thread may behave totally
1547 * differently. For example, a shell spawning a compiler process. We want
1548 * to learn that the compiler is behaving badly very quickly.
1551 sched_interact_fork(struct thread *td)
1556 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1557 if (sum > SCHED_SLP_RUN_FORK) {
1558 ratio = sum / SCHED_SLP_RUN_FORK;
1559 td->td_sched->ts_runtime /= ratio;
1560 td->td_sched->ts_slptime /= ratio;
1565 * Called from proc0_init() to setup the scheduler fields.
1572 * Set up the scheduler specific parts of proc0.
1574 proc0.p_sched = NULL; /* XXX */
1575 thread0.td_sched = &td_sched0;
1576 td_sched0.ts_ltick = ticks;
1577 td_sched0.ts_ftick = ticks;
1578 td_sched0.ts_slice = sched_slice;
1582 * This is only somewhat accurate since given many processes of the same
1583 * priority they will switch when their slices run out, which will be
1584 * at most sched_slice stathz ticks.
1587 sched_rr_interval(void)
1590 /* Convert sched_slice from stathz to hz. */
1591 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1595 * Update the percent cpu tracking information when it is requested or
1596 * the total history exceeds the maximum. We keep a sliding history of
1597 * tick counts that slowly decays. This is less precise than the 4BSD
1598 * mechanism since it happens with less regular and frequent events.
1601 sched_pctcpu_update(struct td_sched *ts, int run)
1605 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1607 ts->ts_ftick = t - SCHED_TICK_TARG;
1608 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1609 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1610 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1611 ts->ts_ftick = t - SCHED_TICK_TARG;
1614 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1619 * Adjust the priority of a thread. Move it to the appropriate run-queue
1620 * if necessary. This is the back-end for several priority related
1624 sched_thread_priority(struct thread *td, u_char prio)
1626 struct td_sched *ts;
1630 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1631 "prio:%d", td->td_priority, "new prio:%d", prio,
1632 KTR_ATTR_LINKED, sched_tdname(curthread));
1633 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1634 if (td != curthread && prio < td->td_priority) {
1635 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1636 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1637 prio, KTR_ATTR_LINKED, sched_tdname(td));
1638 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1642 THREAD_LOCK_ASSERT(td, MA_OWNED);
1643 if (td->td_priority == prio)
1646 * If the priority has been elevated due to priority
1647 * propagation, we may have to move ourselves to a new
1648 * queue. This could be optimized to not re-add in some
1651 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1653 td->td_priority = prio;
1654 sched_add(td, SRQ_BORROWING);
1658 * If the thread is currently running we may have to adjust the lowpri
1659 * information so other cpus are aware of our current priority.
1661 if (TD_IS_RUNNING(td)) {
1662 tdq = TDQ_CPU(ts->ts_cpu);
1663 oldpri = td->td_priority;
1664 td->td_priority = prio;
1665 if (prio < tdq->tdq_lowpri)
1666 tdq->tdq_lowpri = prio;
1667 else if (tdq->tdq_lowpri == oldpri)
1668 tdq_setlowpri(tdq, td);
1671 td->td_priority = prio;
1675 * Update a thread's priority when it is lent another thread's
1679 sched_lend_prio(struct thread *td, u_char prio)
1682 td->td_flags |= TDF_BORROWING;
1683 sched_thread_priority(td, prio);
1687 * Restore a thread's priority when priority propagation is
1688 * over. The prio argument is the minimum priority the thread
1689 * needs to have to satisfy other possible priority lending
1690 * requests. If the thread's regular priority is less
1691 * important than prio, the thread will keep a priority boost
1695 sched_unlend_prio(struct thread *td, u_char prio)
1699 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1700 td->td_base_pri <= PRI_MAX_TIMESHARE)
1701 base_pri = td->td_user_pri;
1703 base_pri = td->td_base_pri;
1704 if (prio >= base_pri) {
1705 td->td_flags &= ~TDF_BORROWING;
1706 sched_thread_priority(td, base_pri);
1708 sched_lend_prio(td, prio);
1712 * Standard entry for setting the priority to an absolute value.
1715 sched_prio(struct thread *td, u_char prio)
1719 /* First, update the base priority. */
1720 td->td_base_pri = prio;
1723 * If the thread is borrowing another thread's priority, don't
1724 * ever lower the priority.
1726 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1729 /* Change the real priority. */
1730 oldprio = td->td_priority;
1731 sched_thread_priority(td, prio);
1734 * If the thread is on a turnstile, then let the turnstile update
1737 if (TD_ON_LOCK(td) && oldprio != prio)
1738 turnstile_adjust(td, oldprio);
1742 * Set the base user priority, does not effect current running priority.
1745 sched_user_prio(struct thread *td, u_char prio)
1748 td->td_base_user_pri = prio;
1749 if (td->td_lend_user_pri <= prio)
1751 td->td_user_pri = prio;
1755 sched_lend_user_prio(struct thread *td, u_char prio)
1758 THREAD_LOCK_ASSERT(td, MA_OWNED);
1759 td->td_lend_user_pri = prio;
1760 td->td_user_pri = min(prio, td->td_base_user_pri);
1761 if (td->td_priority > td->td_user_pri)
1762 sched_prio(td, td->td_user_pri);
1763 else if (td->td_priority != td->td_user_pri)
1764 td->td_flags |= TDF_NEEDRESCHED;
1768 * Handle migration from sched_switch(). This happens only for
1772 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1776 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1778 tdq_load_rem(tdq, td);
1780 * Do the lock dance required to avoid LOR. We grab an extra
1781 * spinlock nesting to prevent preemption while we're
1782 * not holding either run-queue lock.
1785 thread_lock_block(td); /* This releases the lock on tdq. */
1788 * Acquire both run-queue locks before placing the thread on the new
1789 * run-queue to avoid deadlocks created by placing a thread with a
1790 * blocked lock on the run-queue of a remote processor. The deadlock
1791 * occurs when a third processor attempts to lock the two queues in
1792 * question while the target processor is spinning with its own
1793 * run-queue lock held while waiting for the blocked lock to clear.
1795 tdq_lock_pair(tdn, tdq);
1796 tdq_add(tdn, td, flags);
1797 tdq_notify(tdn, td);
1801 return (TDQ_LOCKPTR(tdn));
1805 * Variadic version of thread_lock_unblock() that does not assume td_lock
1809 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1811 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1816 * Switch threads. This function has to handle threads coming in while
1817 * blocked for some reason, running, or idle. It also must deal with
1818 * migrating a thread from one queue to another as running threads may
1819 * be assigned elsewhere via binding.
1822 sched_switch(struct thread *td, struct thread *newtd, int flags)
1825 struct td_sched *ts;
1828 int cpuid, preempted;
1830 THREAD_LOCK_ASSERT(td, MA_OWNED);
1831 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1833 cpuid = PCPU_GET(cpuid);
1834 tdq = TDQ_CPU(cpuid);
1837 sched_pctcpu_update(ts, 1);
1838 ts->ts_rltick = ticks;
1839 td->td_lastcpu = td->td_oncpu;
1840 td->td_oncpu = NOCPU;
1841 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
1842 (flags & SW_PREEMPT) != 0;
1843 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1844 td->td_owepreempt = 0;
1845 if (!TD_IS_IDLETHREAD(td))
1846 tdq->tdq_switchcnt++;
1848 * The lock pointer in an idle thread should never change. Reset it
1849 * to CAN_RUN as well.
1851 if (TD_IS_IDLETHREAD(td)) {
1852 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1854 } else if (TD_IS_RUNNING(td)) {
1855 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1856 srqflag = preempted ?
1857 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1858 SRQ_OURSELF|SRQ_YIELDING;
1860 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1861 ts->ts_cpu = sched_pickcpu(td, 0);
1863 if (ts->ts_cpu == cpuid)
1864 tdq_runq_add(tdq, td, srqflag);
1866 KASSERT(THREAD_CAN_MIGRATE(td) ||
1867 (ts->ts_flags & TSF_BOUND) != 0,
1868 ("Thread %p shouldn't migrate", td));
1869 mtx = sched_switch_migrate(tdq, td, srqflag);
1872 /* This thread must be going to sleep. */
1874 mtx = thread_lock_block(td);
1875 tdq_load_rem(tdq, td);
1878 * We enter here with the thread blocked and assigned to the
1879 * appropriate cpu run-queue or sleep-queue and with the current
1880 * thread-queue locked.
1882 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1883 newtd = choosethread();
1885 * Call the MD code to switch contexts if necessary.
1889 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1890 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1892 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1893 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1894 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1895 sched_pctcpu_update(newtd->td_sched, 0);
1897 #ifdef KDTRACE_HOOKS
1899 * If DTrace has set the active vtime enum to anything
1900 * other than INACTIVE (0), then it should have set the
1903 if (dtrace_vtime_active)
1904 (*dtrace_vtime_switch_func)(newtd);
1907 cpu_switch(td, newtd, mtx);
1909 * We may return from cpu_switch on a different cpu. However,
1910 * we always return with td_lock pointing to the current cpu's
1913 cpuid = PCPU_GET(cpuid);
1914 tdq = TDQ_CPU(cpuid);
1915 lock_profile_obtain_lock_success(
1916 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1918 SDT_PROBE0(sched, , , on__cpu);
1920 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1921 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1924 thread_unblock_switch(td, mtx);
1925 SDT_PROBE0(sched, , , remain__cpu);
1928 * Assert that all went well and return.
1930 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1931 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1932 td->td_oncpu = cpuid;
1936 * Adjust thread priorities as a result of a nice request.
1939 sched_nice(struct proc *p, int nice)
1943 PROC_LOCK_ASSERT(p, MA_OWNED);
1946 FOREACH_THREAD_IN_PROC(p, td) {
1949 sched_prio(td, td->td_base_user_pri);
1955 * Record the sleep time for the interactivity scorer.
1958 sched_sleep(struct thread *td, int prio)
1961 THREAD_LOCK_ASSERT(td, MA_OWNED);
1963 td->td_slptick = ticks;
1964 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1965 td->td_flags |= TDF_CANSWAP;
1966 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1968 if (static_boost == 1 && prio)
1969 sched_prio(td, prio);
1970 else if (static_boost && td->td_priority > static_boost)
1971 sched_prio(td, static_boost);
1975 * Schedule a thread to resume execution and record how long it voluntarily
1976 * slept. We also update the pctcpu, interactivity, and priority.
1979 sched_wakeup(struct thread *td)
1981 struct td_sched *ts;
1984 THREAD_LOCK_ASSERT(td, MA_OWNED);
1986 td->td_flags &= ~TDF_CANSWAP;
1988 * If we slept for more than a tick update our interactivity and
1991 slptick = td->td_slptick;
1993 if (slptick && slptick != ticks) {
1994 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
1995 sched_interact_update(td);
1996 sched_pctcpu_update(ts, 0);
1998 /* Reset the slice value after we sleep. */
1999 ts->ts_slice = sched_slice;
2000 sched_add(td, SRQ_BORING);
2004 * Penalize the parent for creating a new child and initialize the child's
2008 sched_fork(struct thread *td, struct thread *child)
2010 THREAD_LOCK_ASSERT(td, MA_OWNED);
2011 sched_pctcpu_update(td->td_sched, 1);
2012 sched_fork_thread(td, child);
2014 * Penalize the parent and child for forking.
2016 sched_interact_fork(child);
2017 sched_priority(child);
2018 td->td_sched->ts_runtime += tickincr;
2019 sched_interact_update(td);
2024 * Fork a new thread, may be within the same process.
2027 sched_fork_thread(struct thread *td, struct thread *child)
2029 struct td_sched *ts;
2030 struct td_sched *ts2;
2032 THREAD_LOCK_ASSERT(td, MA_OWNED);
2037 ts2 = child->td_sched;
2038 child->td_oncpu = NOCPU;
2039 child->td_lastcpu = NOCPU;
2040 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
2041 child->td_cpuset = cpuset_ref(td->td_cpuset);
2042 ts2->ts_cpu = ts->ts_cpu;
2045 * Grab our parents cpu estimation information.
2047 ts2->ts_ticks = ts->ts_ticks;
2048 ts2->ts_ltick = ts->ts_ltick;
2049 ts2->ts_ftick = ts->ts_ftick;
2051 * Do not inherit any borrowed priority from the parent.
2053 child->td_priority = child->td_base_pri;
2055 * And update interactivity score.
2057 ts2->ts_slptime = ts->ts_slptime;
2058 ts2->ts_runtime = ts->ts_runtime;
2059 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
2061 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2066 * Adjust the priority class of a thread.
2069 sched_class(struct thread *td, int class)
2072 THREAD_LOCK_ASSERT(td, MA_OWNED);
2073 if (td->td_pri_class == class)
2075 td->td_pri_class = class;
2079 * Return some of the child's priority and interactivity to the parent.
2082 sched_exit(struct proc *p, struct thread *child)
2086 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2087 "prio:%d", child->td_priority);
2088 PROC_LOCK_ASSERT(p, MA_OWNED);
2089 td = FIRST_THREAD_IN_PROC(p);
2090 sched_exit_thread(td, child);
2094 * Penalize another thread for the time spent on this one. This helps to
2095 * worsen the priority and interactivity of processes which schedule batch
2096 * jobs such as make. This has little effect on the make process itself but
2097 * causes new processes spawned by it to receive worse scores immediately.
2100 sched_exit_thread(struct thread *td, struct thread *child)
2103 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2104 "prio:%d", child->td_priority);
2106 * Give the child's runtime to the parent without returning the
2107 * sleep time as a penalty to the parent. This causes shells that
2108 * launch expensive things to mark their children as expensive.
2111 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2112 sched_interact_update(td);
2118 sched_preempt(struct thread *td)
2122 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2126 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2127 tdq->tdq_ipipending = 0;
2128 if (td->td_priority > tdq->tdq_lowpri) {
2131 flags = SW_INVOL | SW_PREEMPT;
2132 if (td->td_critnest > 1)
2133 td->td_owepreempt = 1;
2134 else if (TD_IS_IDLETHREAD(td))
2135 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2137 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2143 * Fix priorities on return to user-space. Priorities may be elevated due
2144 * to static priorities in msleep() or similar.
2147 sched_userret(struct thread *td)
2150 * XXX we cheat slightly on the locking here to avoid locking in
2151 * the usual case. Setting td_priority here is essentially an
2152 * incomplete workaround for not setting it properly elsewhere.
2153 * Now that some interrupt handlers are threads, not setting it
2154 * properly elsewhere can clobber it in the window between setting
2155 * it here and returning to user mode, so don't waste time setting
2156 * it perfectly here.
2158 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2159 ("thread with borrowed priority returning to userland"));
2160 if (td->td_priority != td->td_user_pri) {
2162 td->td_priority = td->td_user_pri;
2163 td->td_base_pri = td->td_user_pri;
2164 tdq_setlowpri(TDQ_SELF(), td);
2170 * Handle a stathz tick. This is really only relevant for timeshare
2174 sched_clock(struct thread *td)
2177 struct td_sched *ts;
2179 THREAD_LOCK_ASSERT(td, MA_OWNED);
2183 * We run the long term load balancer infrequently on the first cpu.
2185 if (balance_tdq == tdq) {
2186 if (balance_ticks && --balance_ticks == 0)
2191 * Save the old switch count so we have a record of the last ticks
2192 * activity. Initialize the new switch count based on our load.
2193 * If there is some activity seed it to reflect that.
2195 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2196 tdq->tdq_switchcnt = tdq->tdq_load;
2198 * Advance the insert index once for each tick to ensure that all
2199 * threads get a chance to run.
2201 if (tdq->tdq_idx == tdq->tdq_ridx) {
2202 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2203 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2204 tdq->tdq_ridx = tdq->tdq_idx;
2207 sched_pctcpu_update(ts, 1);
2208 if (td->td_pri_class & PRI_FIFO_BIT)
2210 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2212 * We used a tick; charge it to the thread so
2213 * that we can compute our interactivity.
2215 td->td_sched->ts_runtime += tickincr;
2216 sched_interact_update(td);
2221 * Force a context switch if the current thread has used up a full
2222 * time slice (default is 100ms).
2224 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
2225 ts->ts_slice = sched_slice;
2226 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2231 * Called once per hz tick.
2240 * Return whether the current CPU has runnable tasks. Used for in-kernel
2241 * cooperative idle threads.
2244 sched_runnable(void)
2252 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2253 if (tdq->tdq_load > 0)
2256 if (tdq->tdq_load - 1 > 0)
2264 * Choose the highest priority thread to run. The thread is removed from
2265 * the run-queue while running however the load remains. For SMP we set
2266 * the tdq in the global idle bitmask if it idles here.
2275 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2276 td = tdq_choose(tdq);
2278 tdq_runq_rem(tdq, td);
2279 tdq->tdq_lowpri = td->td_priority;
2282 tdq->tdq_lowpri = PRI_MAX_IDLE;
2283 return (PCPU_GET(idlethread));
2287 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2288 * we always request it once we exit a critical section.
2291 sched_setpreempt(struct thread *td)
2297 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2300 pri = td->td_priority;
2301 cpri = ctd->td_priority;
2303 ctd->td_flags |= TDF_NEEDRESCHED;
2304 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2306 if (!sched_shouldpreempt(pri, cpri, 0))
2308 ctd->td_owepreempt = 1;
2312 * Add a thread to a thread queue. Select the appropriate runq and add the
2313 * thread to it. This is the internal function called when the tdq is
2317 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2320 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2321 KASSERT((td->td_inhibitors == 0),
2322 ("sched_add: trying to run inhibited thread"));
2323 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2324 ("sched_add: bad thread state"));
2325 KASSERT(td->td_flags & TDF_INMEM,
2326 ("sched_add: thread swapped out"));
2328 if (td->td_priority < tdq->tdq_lowpri)
2329 tdq->tdq_lowpri = td->td_priority;
2330 tdq_runq_add(tdq, td, flags);
2331 tdq_load_add(tdq, td);
2335 * Select the target thread queue and add a thread to it. Request
2336 * preemption or IPI a remote processor if required.
2339 sched_add(struct thread *td, int flags)
2346 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2347 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2348 sched_tdname(curthread));
2349 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2350 KTR_ATTR_LINKED, sched_tdname(td));
2351 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2352 flags & SRQ_PREEMPTED);
2353 THREAD_LOCK_ASSERT(td, MA_OWNED);
2355 * Recalculate the priority before we select the target cpu or
2358 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2362 * Pick the destination cpu and if it isn't ours transfer to the
2365 cpu = sched_pickcpu(td, flags);
2366 tdq = sched_setcpu(td, cpu, flags);
2367 tdq_add(tdq, td, flags);
2368 if (cpu != PCPU_GET(cpuid)) {
2369 tdq_notify(tdq, td);
2376 * Now that the thread is moving to the run-queue, set the lock
2377 * to the scheduler's lock.
2379 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2380 tdq_add(tdq, td, flags);
2382 if (!(flags & SRQ_YIELDING))
2383 sched_setpreempt(td);
2387 * Remove a thread from a run-queue without running it. This is used
2388 * when we're stealing a thread from a remote queue. Otherwise all threads
2389 * exit by calling sched_exit_thread() and sched_throw() themselves.
2392 sched_rem(struct thread *td)
2396 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2397 "prio:%d", td->td_priority);
2398 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2399 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2400 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2401 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2402 KASSERT(TD_ON_RUNQ(td),
2403 ("sched_rem: thread not on run queue"));
2404 tdq_runq_rem(tdq, td);
2405 tdq_load_rem(tdq, td);
2407 if (td->td_priority == tdq->tdq_lowpri)
2408 tdq_setlowpri(tdq, NULL);
2412 * Fetch cpu utilization information. Updates on demand.
2415 sched_pctcpu(struct thread *td)
2418 struct td_sched *ts;
2425 THREAD_LOCK_ASSERT(td, MA_OWNED);
2426 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2430 /* How many rtick per second ? */
2431 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2432 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2439 * Enforce affinity settings for a thread. Called after adjustments to
2443 sched_affinity(struct thread *td)
2446 struct td_sched *ts;
2448 THREAD_LOCK_ASSERT(td, MA_OWNED);
2450 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2452 if (TD_ON_RUNQ(td)) {
2454 sched_add(td, SRQ_BORING);
2457 if (!TD_IS_RUNNING(td))
2460 * Force a switch before returning to userspace. If the
2461 * target thread is not running locally send an ipi to force
2464 td->td_flags |= TDF_NEEDRESCHED;
2465 if (td != curthread)
2466 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2471 * Bind a thread to a target cpu.
2474 sched_bind(struct thread *td, int cpu)
2476 struct td_sched *ts;
2478 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2479 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2481 if (ts->ts_flags & TSF_BOUND)
2483 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2484 ts->ts_flags |= TSF_BOUND;
2486 if (PCPU_GET(cpuid) == cpu)
2489 /* When we return from mi_switch we'll be on the correct cpu. */
2490 mi_switch(SW_VOL, NULL);
2494 * Release a bound thread.
2497 sched_unbind(struct thread *td)
2499 struct td_sched *ts;
2501 THREAD_LOCK_ASSERT(td, MA_OWNED);
2502 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2504 if ((ts->ts_flags & TSF_BOUND) == 0)
2506 ts->ts_flags &= ~TSF_BOUND;
2511 sched_is_bound(struct thread *td)
2513 THREAD_LOCK_ASSERT(td, MA_OWNED);
2514 return (td->td_sched->ts_flags & TSF_BOUND);
2521 sched_relinquish(struct thread *td)
2524 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2529 * Return the total system load.
2540 total += TDQ_CPU(i)->tdq_sysload;
2543 return (TDQ_SELF()->tdq_sysload);
2548 sched_sizeof_proc(void)
2550 return (sizeof(struct proc));
2554 sched_sizeof_thread(void)
2556 return (sizeof(struct thread) + sizeof(struct td_sched));
2560 #define TDQ_IDLESPIN(tdq) \
2561 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2563 #define TDQ_IDLESPIN(tdq) 1
2567 * The actual idle process.
2570 sched_idletd(void *dummy)
2574 int oldswitchcnt, switchcnt;
2577 mtx_assert(&Giant, MA_NOTOWNED);
2582 if (tdq->tdq_load) {
2584 mi_switch(SW_VOL | SWT_IDLE, NULL);
2587 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2589 if (switchcnt != oldswitchcnt) {
2590 oldswitchcnt = switchcnt;
2591 if (tdq_idled(tdq) == 0)
2594 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2596 oldswitchcnt = switchcnt;
2599 * If we're switching very frequently, spin while checking
2600 * for load rather than entering a low power state that
2601 * may require an IPI. However, don't do any busy
2602 * loops while on SMT machines as this simply steals
2603 * cycles from cores doing useful work.
2605 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2606 for (i = 0; i < sched_idlespins; i++) {
2613 /* If there was context switch during spin, restart it. */
2614 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2615 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2618 /* Run main MD idle handler. */
2619 tdq->tdq_cpu_idle = 1;
2621 * Make sure that tdq_cpu_idle update is globally visible
2622 * before cpu_idle() read tdq_load. The order is important
2623 * to avoid race with tdq_notify.
2626 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2627 tdq->tdq_cpu_idle = 0;
2630 * Account thread-less hardware interrupts and
2631 * other wakeup reasons equal to context switches.
2633 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2634 if (switchcnt != oldswitchcnt)
2636 tdq->tdq_switchcnt++;
2642 * A CPU is entering for the first time or a thread is exiting.
2645 sched_throw(struct thread *td)
2647 struct thread *newtd;
2652 /* Correct spinlock nesting and acquire the correct lock. */
2655 PCPU_SET(switchtime, cpu_ticks());
2656 PCPU_SET(switchticks, ticks);
2658 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2659 tdq_load_rem(tdq, td);
2660 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2661 td->td_lastcpu = td->td_oncpu;
2662 td->td_oncpu = NOCPU;
2664 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2665 newtd = choosethread();
2666 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2667 cpu_throw(td, newtd); /* doesn't return */
2671 * This is called from fork_exit(). Just acquire the correct locks and
2672 * let fork do the rest of the work.
2675 sched_fork_exit(struct thread *td)
2677 struct td_sched *ts;
2682 * Finish setting up thread glue so that it begins execution in a
2683 * non-nested critical section with the scheduler lock held.
2685 cpuid = PCPU_GET(cpuid);
2686 tdq = TDQ_CPU(cpuid);
2688 if (TD_IS_IDLETHREAD(td))
2689 td->td_lock = TDQ_LOCKPTR(tdq);
2690 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2691 td->td_oncpu = cpuid;
2692 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2693 lock_profile_obtain_lock_success(
2694 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2698 * Create on first use to catch odd startup conditons.
2701 sched_tdname(struct thread *td)
2704 struct td_sched *ts;
2707 if (ts->ts_name[0] == '\0')
2708 snprintf(ts->ts_name, sizeof(ts->ts_name),
2709 "%s tid %d", td->td_name, td->td_tid);
2710 return (ts->ts_name);
2712 return (td->td_name);
2718 sched_clear_tdname(struct thread *td)
2720 struct td_sched *ts;
2723 ts->ts_name[0] = '\0';
2730 * Build the CPU topology dump string. Is recursively called to collect
2731 * the topology tree.
2734 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2737 char cpusetbuf[CPUSETBUFSIZ];
2740 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2741 "", 1 + indent / 2, cg->cg_level);
2742 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2743 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2745 for (i = 0; i < MAXCPU; i++) {
2746 if (CPU_ISSET(i, &cg->cg_mask)) {
2748 sbuf_printf(sb, ", ");
2751 sbuf_printf(sb, "%d", i);
2754 sbuf_printf(sb, "</cpu>\n");
2756 if (cg->cg_flags != 0) {
2757 sbuf_printf(sb, "%*s <flags>", indent, "");
2758 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2759 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2760 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2761 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2762 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2763 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2764 sbuf_printf(sb, "</flags>\n");
2767 if (cg->cg_children > 0) {
2768 sbuf_printf(sb, "%*s <children>\n", indent, "");
2769 for (i = 0; i < cg->cg_children; i++)
2770 sysctl_kern_sched_topology_spec_internal(sb,
2771 &cg->cg_child[i], indent+2);
2772 sbuf_printf(sb, "%*s </children>\n", indent, "");
2774 sbuf_printf(sb, "%*s</group>\n", indent, "");
2779 * Sysctl handler for retrieving topology dump. It's a wrapper for
2780 * the recursive sysctl_kern_smp_topology_spec_internal().
2783 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2788 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2790 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2794 sbuf_printf(topo, "<groups>\n");
2795 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2796 sbuf_printf(topo, "</groups>\n");
2800 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2809 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2811 int error, new_val, period;
2813 period = 1000000 / realstathz;
2814 new_val = period * sched_slice;
2815 error = sysctl_handle_int(oidp, &new_val, 0, req);
2816 if (error != 0 || req->newptr == NULL)
2820 sched_slice = imax(1, (new_val + period / 2) / period);
2821 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2826 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2827 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2829 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2830 NULL, 0, sysctl_kern_quantum, "I",
2831 "Quantum for timeshare threads in microseconds");
2832 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2833 "Quantum for timeshare threads in stathz ticks");
2834 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2835 "Interactivity score threshold");
2836 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2838 "Maximal (lowest) priority for preemption");
2839 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2840 "Assign static kernel priorities to sleeping threads");
2841 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2842 "Number of times idle thread will spin waiting for new work");
2843 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2844 &sched_idlespinthresh, 0,
2845 "Threshold before we will permit idle thread spinning");
2847 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2848 "Number of hz ticks to keep thread affinity for");
2849 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2850 "Enables the long-term load balancer");
2851 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2852 &balance_interval, 0,
2853 "Average period in stathz ticks to run the long-term balancer");
2854 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2855 "Attempts to steal work from other cores before idling");
2856 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2857 "Minimum load on remote CPU before we'll steal");
2858 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2859 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2860 "XML dump of detected CPU topology");
2863 /* ps compat. All cpu percentages from ULE are weighted. */
2864 static int ccpu = 0;
2865 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");