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)
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();
227 * Ordered to improve efficiency of cpu_search() and switch().
228 * tdq_lock is padded to avoid false sharing with tdq_load and
231 struct mtx tdq_lock; /* run queue lock. */
232 char pad[64 - sizeof(struct mtx)];
233 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
234 volatile int tdq_load; /* Aggregate load. */
235 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
236 int tdq_sysload; /* For loadavg, !ITHD load. */
237 int tdq_transferable; /* Transferable thread count. */
238 short tdq_switchcnt; /* Switches this tick. */
239 short tdq_oldswitchcnt; /* Switches last tick. */
240 u_char tdq_lowpri; /* Lowest priority thread. */
241 u_char tdq_ipipending; /* IPI pending. */
242 u_char tdq_idx; /* Current insert index. */
243 u_char tdq_ridx; /* Current removal index. */
244 struct runq tdq_realtime; /* real-time run queue. */
245 struct runq tdq_timeshare; /* timeshare run queue. */
246 struct runq tdq_idle; /* Queue of IDLE threads. */
247 char tdq_name[TDQ_NAME_LEN];
249 char tdq_loadname[TDQ_LOADNAME_LEN];
253 /* Idle thread states and config. */
254 #define TDQ_RUNNING 1
258 struct cpu_group *cpu_top; /* CPU topology */
260 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
261 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
266 static int rebalance = 1;
267 static int balance_interval = 128; /* Default set in sched_initticks(). */
269 static int steal_idle = 1;
270 static int steal_thresh = 2;
273 * One thread queue per processor.
275 static struct tdq tdq_cpu[MAXCPU];
276 static struct tdq *balance_tdq;
277 static int balance_ticks;
278 static DPCPU_DEFINE(uint32_t, randomval);
280 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
281 #define TDQ_CPU(x) (&tdq_cpu[(x)])
282 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
284 static struct tdq tdq_cpu;
286 #define TDQ_ID(x) (0)
287 #define TDQ_SELF() (&tdq_cpu)
288 #define TDQ_CPU(x) (&tdq_cpu)
291 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
292 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
293 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
294 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
295 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
297 static void sched_priority(struct thread *);
298 static void sched_thread_priority(struct thread *, u_char);
299 static int sched_interact_score(struct thread *);
300 static void sched_interact_update(struct thread *);
301 static void sched_interact_fork(struct thread *);
302 static void sched_pctcpu_update(struct td_sched *, int);
304 /* Operations on per processor queues */
305 static struct thread *tdq_choose(struct tdq *);
306 static void tdq_setup(struct tdq *);
307 static void tdq_load_add(struct tdq *, struct thread *);
308 static void tdq_load_rem(struct tdq *, struct thread *);
309 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
310 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
311 static inline int sched_shouldpreempt(int, int, int);
312 void tdq_print(int cpu);
313 static void runq_print(struct runq *rq);
314 static void tdq_add(struct tdq *, struct thread *, int);
316 static int tdq_move(struct tdq *, struct tdq *);
317 static int tdq_idled(struct tdq *);
318 static void tdq_notify(struct tdq *, struct thread *);
319 static struct thread *tdq_steal(struct tdq *, int);
320 static struct thread *runq_steal(struct runq *, int);
321 static int sched_pickcpu(struct thread *, int);
322 static void sched_balance(void);
323 static int sched_balance_pair(struct tdq *, struct tdq *);
324 static inline struct tdq *sched_setcpu(struct thread *, int, int);
325 static inline void thread_unblock_switch(struct thread *, struct mtx *);
326 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
327 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
328 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
329 struct cpu_group *cg, int indent);
332 static void sched_setup(void *dummy);
333 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
335 static void sched_initticks(void *dummy);
336 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
339 SDT_PROVIDER_DEFINE(sched);
341 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
342 "struct proc *", "uint8_t");
343 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
344 "struct proc *", "void *");
345 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
346 "struct proc *", "void *", "int");
347 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
348 "struct proc *", "uint8_t", "struct thread *");
349 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
350 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
352 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
353 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
354 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
358 * Print the threads waiting on a run-queue.
361 runq_print(struct runq *rq)
369 for (i = 0; i < RQB_LEN; i++) {
370 printf("\t\trunq bits %d 0x%zx\n",
371 i, rq->rq_status.rqb_bits[i]);
372 for (j = 0; j < RQB_BPW; j++)
373 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
374 pri = j + (i << RQB_L2BPW);
375 rqh = &rq->rq_queues[pri];
376 TAILQ_FOREACH(td, rqh, td_runq) {
377 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
378 td, td->td_name, td->td_priority,
379 td->td_rqindex, pri);
386 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
395 printf("tdq %d:\n", TDQ_ID(tdq));
396 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
397 printf("\tLock name: %s\n", tdq->tdq_name);
398 printf("\tload: %d\n", tdq->tdq_load);
399 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
400 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
401 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
402 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
403 printf("\tload transferable: %d\n", tdq->tdq_transferable);
404 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
405 printf("\trealtime runq:\n");
406 runq_print(&tdq->tdq_realtime);
407 printf("\ttimeshare runq:\n");
408 runq_print(&tdq->tdq_timeshare);
409 printf("\tidle runq:\n");
410 runq_print(&tdq->tdq_idle);
414 sched_shouldpreempt(int pri, int cpri, int remote)
417 * If the new priority is not better than the current priority there is
423 * Always preempt idle.
425 if (cpri >= PRI_MIN_IDLE)
428 * If preemption is disabled don't preempt others.
430 if (preempt_thresh == 0)
433 * Preempt if we exceed the threshold.
435 if (pri <= preempt_thresh)
438 * If we're interactive or better and there is non-interactive
439 * or worse running preempt only remote processors.
441 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
447 * Add a thread to the actual run-queue. Keeps transferable counts up to
448 * date with what is actually on the run-queue. Selects the correct
449 * queue position for timeshare threads.
452 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
457 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
458 THREAD_LOCK_ASSERT(td, MA_OWNED);
460 pri = td->td_priority;
463 if (THREAD_CAN_MIGRATE(td)) {
464 tdq->tdq_transferable++;
465 ts->ts_flags |= TSF_XFERABLE;
467 if (pri < PRI_MIN_BATCH) {
468 ts->ts_runq = &tdq->tdq_realtime;
469 } else if (pri <= PRI_MAX_BATCH) {
470 ts->ts_runq = &tdq->tdq_timeshare;
471 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
472 ("Invalid priority %d on timeshare runq", pri));
474 * This queue contains only priorities between MIN and MAX
475 * realtime. Use the whole queue to represent these values.
477 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
478 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
479 pri = (pri + tdq->tdq_idx) % RQ_NQS;
481 * This effectively shortens the queue by one so we
482 * can have a one slot difference between idx and
483 * ridx while we wait for threads to drain.
485 if (tdq->tdq_ridx != tdq->tdq_idx &&
486 pri == tdq->tdq_ridx)
487 pri = (unsigned char)(pri - 1) % RQ_NQS;
490 runq_add_pri(ts->ts_runq, td, pri, flags);
493 ts->ts_runq = &tdq->tdq_idle;
494 runq_add(ts->ts_runq, td, flags);
498 * Remove a thread from a run-queue. This typically happens when a thread
499 * is selected to run. Running threads are not on the queue and the
500 * transferable count does not reflect them.
503 tdq_runq_rem(struct tdq *tdq, struct thread *td)
508 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
509 KASSERT(ts->ts_runq != NULL,
510 ("tdq_runq_remove: thread %p null ts_runq", td));
511 if (ts->ts_flags & TSF_XFERABLE) {
512 tdq->tdq_transferable--;
513 ts->ts_flags &= ~TSF_XFERABLE;
515 if (ts->ts_runq == &tdq->tdq_timeshare) {
516 if (tdq->tdq_idx != tdq->tdq_ridx)
517 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
519 runq_remove_idx(ts->ts_runq, td, NULL);
521 runq_remove(ts->ts_runq, td);
525 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
526 * for this thread to the referenced thread queue.
529 tdq_load_add(struct tdq *tdq, struct thread *td)
532 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
533 THREAD_LOCK_ASSERT(td, MA_OWNED);
536 if ((td->td_flags & TDF_NOLOAD) == 0)
538 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
539 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
543 * Remove the load from a thread that is transitioning to a sleep state or
547 tdq_load_rem(struct tdq *tdq, struct thread *td)
550 THREAD_LOCK_ASSERT(td, MA_OWNED);
551 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
552 KASSERT(tdq->tdq_load != 0,
553 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
556 if ((td->td_flags & TDF_NOLOAD) == 0)
558 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
559 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
563 * Set lowpri to its exact value by searching the run-queue and
564 * evaluating curthread. curthread may be passed as an optimization.
567 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
571 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
573 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
574 td = tdq_choose(tdq);
575 if (td == NULL || td->td_priority > ctd->td_priority)
576 tdq->tdq_lowpri = ctd->td_priority;
578 tdq->tdq_lowpri = td->td_priority;
585 int cs_pri; /* Min priority for low. */
586 int cs_limit; /* Max load for low, min load for high. */
591 #define CPU_SEARCH_LOWEST 0x1
592 #define CPU_SEARCH_HIGHEST 0x2
593 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
595 #define CPUSET_FOREACH(cpu, mask) \
596 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
597 if (CPU_ISSET(cpu, &mask))
599 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
600 struct cpu_search *high, const int match);
601 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
602 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
603 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
604 struct cpu_search *high);
607 * Search the tree of cpu_groups for the lowest or highest loaded cpu
608 * according to the match argument. This routine actually compares the
609 * load on all paths through the tree and finds the least loaded cpu on
610 * the least loaded path, which may differ from the least loaded cpu in
611 * the system. This balances work among caches and busses.
613 * This inline is instantiated in three forms below using constants for the
614 * match argument. It is reduced to the minimum set for each case. It is
615 * also recursive to the depth of the tree.
618 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
619 struct cpu_search *high, const int match)
621 struct cpu_search lgroup;
622 struct cpu_search hgroup;
624 struct cpu_group *child;
626 int cpu, i, hload, lload, load, total, rnd, *rndptr;
629 cpumask = cg->cg_mask;
630 if (match & CPU_SEARCH_LOWEST) {
634 if (match & CPU_SEARCH_HIGHEST) {
639 /* Iterate through the child CPU groups and then remaining CPUs. */
640 for (i = cg->cg_children, cpu = mp_maxid; i >= 0; ) {
642 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
648 child = &cg->cg_child[i - 1];
650 if (match & CPU_SEARCH_LOWEST)
652 if (match & CPU_SEARCH_HIGHEST)
654 if (child) { /* Handle child CPU group. */
655 CPU_NAND(&cpumask, &child->cg_mask);
657 case CPU_SEARCH_LOWEST:
658 load = cpu_search_lowest(child, &lgroup);
660 case CPU_SEARCH_HIGHEST:
661 load = cpu_search_highest(child, &hgroup);
663 case CPU_SEARCH_BOTH:
664 load = cpu_search_both(child, &lgroup, &hgroup);
667 } else { /* Handle child CPU. */
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))
722 * cpu_search instantiations must pass constants to maintain the inline
726 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
728 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
732 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
734 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
738 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
739 struct cpu_search *high)
741 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
745 * Find the cpu with the least load via the least loaded path that has a
746 * lowpri greater than pri pri. A pri of -1 indicates any priority is
750 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
753 struct cpu_search low;
756 low.cs_prefer = prefer;
759 low.cs_limit = maxload;
760 cpu_search_lowest(cg, &low);
765 * Find the cpu with the highest load via the highest loaded path.
768 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
770 struct cpu_search high;
774 high.cs_limit = minload;
775 cpu_search_highest(cg, &high);
780 * Simultaneously find the highest and lowest loaded cpu reachable via
784 sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
786 struct cpu_search high;
787 struct cpu_search low;
792 low.cs_limit = INT_MAX;
797 cpu_search_both(cg, &low, &high);
798 *lowcpu = low.cs_cpu;
799 *highcpu = high.cs_cpu;
804 sched_balance_group(struct cpu_group *cg)
806 cpuset_t hmask, lmask;
807 int high, low, anylow;
811 high = sched_highest(cg, hmask, 1);
812 /* Stop if there is no more CPU with transferrable threads. */
815 CPU_CLR(high, &hmask);
816 CPU_COPY(&hmask, &lmask);
817 /* Stop if there is no more CPU left for low. */
818 if (CPU_EMPTY(&lmask))
822 low = sched_lowest(cg, lmask, -1,
823 TDQ_CPU(high)->tdq_load - 1, high);
824 /* Stop if we looked well and found no less loaded CPU. */
825 if (anylow && low == -1)
827 /* Go to next high if we found no less loaded CPU. */
830 /* Transfer thread from high to low. */
831 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
832 /* CPU that got thread can no longer be a donor. */
833 CPU_CLR(low, &hmask);
836 * If failed, then there is no threads on high
837 * that can run on this low. Drop low from low
838 * mask and look for different one.
840 CPU_CLR(low, &lmask);
853 * Select a random time between .5 * balance_interval and
854 * 1.5 * balance_interval.
856 balance_ticks = max(balance_interval / 2, 1);
857 balance_ticks += random() % balance_interval;
858 if (smp_started == 0 || rebalance == 0)
862 sched_balance_group(cpu_top);
867 * Lock two thread queues using their address to maintain lock order.
870 tdq_lock_pair(struct tdq *one, struct tdq *two)
874 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
877 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
882 * Unlock two thread queues. Order is not important here.
885 tdq_unlock_pair(struct tdq *one, struct tdq *two)
892 * Transfer load between two imbalanced thread queues.
895 sched_balance_pair(struct tdq *high, struct tdq *low)
900 tdq_lock_pair(high, low);
903 * Determine what the imbalance is and then adjust that to how many
904 * threads we actually have to give up (transferable).
906 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
907 (moved = tdq_move(high, low)) > 0) {
909 * In case the target isn't the current cpu IPI it to force a
910 * reschedule with the new workload.
914 if (cpu != PCPU_GET(cpuid))
915 ipi_cpu(cpu, IPI_PREEMPT);
918 tdq_unlock_pair(high, low);
923 * Move a thread from one thread queue to another.
926 tdq_move(struct tdq *from, struct tdq *to)
933 TDQ_LOCK_ASSERT(from, MA_OWNED);
934 TDQ_LOCK_ASSERT(to, MA_OWNED);
938 td = tdq_steal(tdq, cpu);
943 * Although the run queue is locked the thread may be blocked. Lock
944 * it to clear this and acquire the run-queue lock.
947 /* Drop recursive lock on from acquired via thread_lock(). */
951 td->td_lock = TDQ_LOCKPTR(to);
952 tdq_add(to, td, SRQ_YIELDING);
957 * This tdq has idled. Try to steal a thread from another cpu and switch
961 tdq_idled(struct tdq *tdq)
963 struct cpu_group *cg;
969 if (smp_started == 0 || steal_idle == 0)
972 CPU_CLR(PCPU_GET(cpuid), &mask);
973 /* We don't want to be preempted while we're iterating. */
975 for (cg = tdq->tdq_cg; cg != NULL; ) {
976 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
977 thresh = steal_thresh;
980 cpu = sched_highest(cg, mask, thresh);
985 steal = TDQ_CPU(cpu);
987 tdq_lock_pair(tdq, steal);
988 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
989 tdq_unlock_pair(tdq, steal);
993 * If a thread was added while interrupts were disabled don't
994 * steal one here. If we fail to acquire one due to affinity
995 * restrictions loop again with this cpu removed from the
998 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
999 tdq_unlock_pair(tdq, steal);
1004 mi_switch(SW_VOL | SWT_IDLE, NULL);
1005 thread_unlock(curthread);
1014 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1017 tdq_notify(struct tdq *tdq, struct thread *td)
1023 if (tdq->tdq_ipipending)
1025 cpu = td->td_sched->ts_cpu;
1026 pri = td->td_priority;
1027 ctd = pcpu_find(cpu)->pc_curthread;
1028 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1030 if (TD_IS_IDLETHREAD(ctd)) {
1032 * If the MD code has an idle wakeup routine try that before
1033 * falling back to IPI.
1035 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1038 tdq->tdq_ipipending = 1;
1039 ipi_cpu(cpu, IPI_PREEMPT);
1043 * Steals load from a timeshare queue. Honors the rotating queue head
1046 static struct thread *
1047 runq_steal_from(struct runq *rq, int cpu, u_char start)
1051 struct thread *td, *first;
1056 rqb = &rq->rq_status;
1057 bit = start & (RQB_BPW -1);
1061 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1062 if (rqb->rqb_bits[i] == 0)
1065 for (pri = bit; pri < RQB_BPW; pri++)
1066 if (rqb->rqb_bits[i] & (1ul << pri))
1071 pri = RQB_FFS(rqb->rqb_bits[i]);
1072 pri += (i << RQB_L2BPW);
1073 rqh = &rq->rq_queues[pri];
1074 TAILQ_FOREACH(td, rqh, td_runq) {
1075 if (first && THREAD_CAN_MIGRATE(td) &&
1076 THREAD_CAN_SCHED(td, cpu))
1086 if (first && THREAD_CAN_MIGRATE(first) &&
1087 THREAD_CAN_SCHED(first, cpu))
1093 * Steals load from a standard linear queue.
1095 static struct thread *
1096 runq_steal(struct runq *rq, int cpu)
1104 rqb = &rq->rq_status;
1105 for (word = 0; word < RQB_LEN; word++) {
1106 if (rqb->rqb_bits[word] == 0)
1108 for (bit = 0; bit < RQB_BPW; bit++) {
1109 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1111 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1112 TAILQ_FOREACH(td, rqh, td_runq)
1113 if (THREAD_CAN_MIGRATE(td) &&
1114 THREAD_CAN_SCHED(td, cpu))
1122 * Attempt to steal a thread in priority order from a thread queue.
1124 static struct thread *
1125 tdq_steal(struct tdq *tdq, int cpu)
1129 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1130 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1132 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1133 cpu, tdq->tdq_ridx)) != NULL)
1135 return (runq_steal(&tdq->tdq_idle, cpu));
1139 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1140 * current lock and returns with the assigned queue locked.
1142 static inline struct tdq *
1143 sched_setcpu(struct thread *td, int cpu, int flags)
1148 THREAD_LOCK_ASSERT(td, MA_OWNED);
1150 td->td_sched->ts_cpu = cpu;
1152 * If the lock matches just return the queue.
1154 if (td->td_lock == TDQ_LOCKPTR(tdq))
1158 * If the thread isn't running its lockptr is a
1159 * turnstile or a sleepqueue. We can just lock_set without
1162 if (TD_CAN_RUN(td)) {
1164 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1169 * The hard case, migration, we need to block the thread first to
1170 * prevent order reversals with other cpus locks.
1173 thread_lock_block(td);
1175 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1180 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1181 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1182 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1183 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1184 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1185 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1188 sched_pickcpu(struct thread *td, int flags)
1190 struct cpu_group *cg, *ccg;
1191 struct td_sched *ts;
1196 self = PCPU_GET(cpuid);
1198 if (smp_started == 0)
1201 * Don't migrate a running thread from sched_switch().
1203 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1204 return (ts->ts_cpu);
1206 * Prefer to run interrupt threads on the processors that generate
1209 pri = td->td_priority;
1210 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1211 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1212 SCHED_STAT_INC(pickcpu_intrbind);
1214 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1215 SCHED_STAT_INC(pickcpu_affinity);
1216 return (ts->ts_cpu);
1220 * If the thread can run on the last cpu and the affinity has not
1221 * expired or it is idle run it there.
1223 tdq = TDQ_CPU(ts->ts_cpu);
1225 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1226 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1227 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1228 if (cg->cg_flags & CG_FLAG_THREAD) {
1229 CPUSET_FOREACH(cpu, cg->cg_mask) {
1230 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1235 if (cpu > mp_maxid) {
1236 SCHED_STAT_INC(pickcpu_idle_affinity);
1237 return (ts->ts_cpu);
1241 * Search for the last level cache CPU group in the tree.
1242 * Skip caches with expired affinity time and SMT groups.
1243 * Affinity to higher level caches will be handled less aggressively.
1245 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1246 if (cg->cg_flags & CG_FLAG_THREAD)
1248 if (!SCHED_AFFINITY(ts, cg->cg_level))
1255 /* Search the group for the less loaded idle CPU we can run now. */
1256 mask = td->td_cpuset->cs_mask;
1257 if (cg != NULL && cg != cpu_top &&
1258 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1259 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1260 INT_MAX, ts->ts_cpu);
1261 /* Search globally for the less loaded CPU we can run now. */
1263 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1264 /* Search globally for the less loaded CPU. */
1266 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1267 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1269 * Compare the lowest loaded cpu to current cpu.
1271 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1272 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1273 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1274 SCHED_STAT_INC(pickcpu_local);
1277 SCHED_STAT_INC(pickcpu_lowest);
1278 if (cpu != ts->ts_cpu)
1279 SCHED_STAT_INC(pickcpu_migration);
1285 * Pick the highest priority task we have and return it.
1287 static struct thread *
1288 tdq_choose(struct tdq *tdq)
1292 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1293 td = runq_choose(&tdq->tdq_realtime);
1296 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1298 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1299 ("tdq_choose: Invalid priority on timeshare queue %d",
1303 td = runq_choose(&tdq->tdq_idle);
1305 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1306 ("tdq_choose: Invalid priority on idle queue %d",
1315 * Initialize a thread queue.
1318 tdq_setup(struct tdq *tdq)
1322 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1323 runq_init(&tdq->tdq_realtime);
1324 runq_init(&tdq->tdq_timeshare);
1325 runq_init(&tdq->tdq_idle);
1326 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1327 "sched lock %d", (int)TDQ_ID(tdq));
1328 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1329 MTX_SPIN | MTX_RECURSE);
1331 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1332 "CPU %d load", (int)TDQ_ID(tdq));
1338 sched_setup_smp(void)
1343 cpu_top = smp_topo();
1347 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1348 if (tdq->tdq_cg == NULL)
1349 panic("Can't find cpu group for %d\n", i);
1351 balance_tdq = TDQ_SELF();
1357 * Setup the thread queues and initialize the topology based on MD
1361 sched_setup(void *dummy)
1372 /* Add thread0's load since it's running. */
1374 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1375 tdq_load_add(tdq, &thread0);
1376 tdq->tdq_lowpri = thread0.td_priority;
1381 * This routine determines time constants after stathz and hz are setup.
1385 sched_initticks(void *dummy)
1389 realstathz = stathz ? stathz : hz;
1390 sched_slice = realstathz / 10; /* ~100ms */
1391 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1395 * tickincr is shifted out by 10 to avoid rounding errors due to
1396 * hz not being evenly divisible by stathz on all platforms.
1398 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1400 * This does not work for values of stathz that are more than
1401 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1408 * Set the default balance interval now that we know
1409 * what realstathz is.
1411 balance_interval = realstathz;
1412 affinity = SCHED_AFFINITY_DEFAULT;
1414 if (sched_idlespinthresh < 0)
1415 sched_idlespinthresh = imax(16, 2 * hz / realstathz);
1420 * This is the core of the interactivity algorithm. Determines a score based
1421 * on past behavior. It is the ratio of sleep time to run time scaled to
1422 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1423 * differs from the cpu usage because it does not account for time spent
1424 * waiting on a run-queue. Would be prettier if we had floating point.
1427 sched_interact_score(struct thread *td)
1429 struct td_sched *ts;
1434 * The score is only needed if this is likely to be an interactive
1435 * task. Don't go through the expense of computing it if there's
1438 if (sched_interact <= SCHED_INTERACT_HALF &&
1439 ts->ts_runtime >= ts->ts_slptime)
1440 return (SCHED_INTERACT_HALF);
1442 if (ts->ts_runtime > ts->ts_slptime) {
1443 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1444 return (SCHED_INTERACT_HALF +
1445 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1447 if (ts->ts_slptime > ts->ts_runtime) {
1448 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1449 return (ts->ts_runtime / div);
1451 /* runtime == slptime */
1453 return (SCHED_INTERACT_HALF);
1456 * This can happen if slptime and runtime are 0.
1463 * Scale the scheduling priority according to the "interactivity" of this
1467 sched_priority(struct thread *td)
1472 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1475 * If the score is interactive we place the thread in the realtime
1476 * queue with a priority that is less than kernel and interrupt
1477 * priorities. These threads are not subject to nice restrictions.
1479 * Scores greater than this are placed on the normal timeshare queue
1480 * where the priority is partially decided by the most recent cpu
1481 * utilization and the rest is decided by nice value.
1483 * The nice value of the process has a linear effect on the calculated
1484 * score. Negative nice values make it easier for a thread to be
1485 * considered interactive.
1487 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1488 if (score < sched_interact) {
1489 pri = PRI_MIN_INTERACT;
1490 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1491 sched_interact) * score;
1492 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1493 ("sched_priority: invalid interactive priority %d score %d",
1496 pri = SCHED_PRI_MIN;
1497 if (td->td_sched->ts_ticks)
1498 pri += min(SCHED_PRI_TICKS(td->td_sched),
1500 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1501 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1502 ("sched_priority: invalid priority %d: nice %d, "
1503 "ticks %d ftick %d ltick %d tick pri %d",
1504 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1505 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1506 SCHED_PRI_TICKS(td->td_sched)));
1508 sched_user_prio(td, pri);
1514 * This routine enforces a maximum limit on the amount of scheduling history
1515 * kept. It is called after either the slptime or runtime is adjusted. This
1516 * function is ugly due to integer math.
1519 sched_interact_update(struct thread *td)
1521 struct td_sched *ts;
1525 sum = ts->ts_runtime + ts->ts_slptime;
1526 if (sum < SCHED_SLP_RUN_MAX)
1529 * This only happens from two places:
1530 * 1) We have added an unusual amount of run time from fork_exit.
1531 * 2) We have added an unusual amount of sleep time from sched_sleep().
1533 if (sum > SCHED_SLP_RUN_MAX * 2) {
1534 if (ts->ts_runtime > ts->ts_slptime) {
1535 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1538 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1544 * If we have exceeded by more than 1/5th then the algorithm below
1545 * will not bring us back into range. Dividing by two here forces
1546 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1548 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1549 ts->ts_runtime /= 2;
1550 ts->ts_slptime /= 2;
1553 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1554 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1558 * Scale back the interactivity history when a child thread is created. The
1559 * history is inherited from the parent but the thread may behave totally
1560 * differently. For example, a shell spawning a compiler process. We want
1561 * to learn that the compiler is behaving badly very quickly.
1564 sched_interact_fork(struct thread *td)
1569 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1570 if (sum > SCHED_SLP_RUN_FORK) {
1571 ratio = sum / SCHED_SLP_RUN_FORK;
1572 td->td_sched->ts_runtime /= ratio;
1573 td->td_sched->ts_slptime /= ratio;
1578 * Called from proc0_init() to setup the scheduler fields.
1585 * Set up the scheduler specific parts of proc0.
1587 proc0.p_sched = NULL; /* XXX */
1588 thread0.td_sched = &td_sched0;
1589 td_sched0.ts_ltick = ticks;
1590 td_sched0.ts_ftick = ticks;
1591 td_sched0.ts_slice = sched_slice;
1595 * This is only somewhat accurate since given many processes of the same
1596 * priority they will switch when their slices run out, which will be
1597 * at most sched_slice stathz ticks.
1600 sched_rr_interval(void)
1603 /* Convert sched_slice from stathz to hz. */
1604 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1608 * Update the percent cpu tracking information when it is requested or
1609 * the total history exceeds the maximum. We keep a sliding history of
1610 * tick counts that slowly decays. This is less precise than the 4BSD
1611 * mechanism since it happens with less regular and frequent events.
1614 sched_pctcpu_update(struct td_sched *ts, int run)
1618 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1620 ts->ts_ftick = t - SCHED_TICK_TARG;
1621 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1622 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1623 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1624 ts->ts_ftick = t - SCHED_TICK_TARG;
1627 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1632 * Adjust the priority of a thread. Move it to the appropriate run-queue
1633 * if necessary. This is the back-end for several priority related
1637 sched_thread_priority(struct thread *td, u_char prio)
1639 struct td_sched *ts;
1643 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1644 "prio:%d", td->td_priority, "new prio:%d", prio,
1645 KTR_ATTR_LINKED, sched_tdname(curthread));
1646 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
1647 if (td != curthread && prio < td->td_priority) {
1648 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1649 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1650 prio, KTR_ATTR_LINKED, sched_tdname(td));
1651 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
1655 THREAD_LOCK_ASSERT(td, MA_OWNED);
1656 if (td->td_priority == prio)
1659 * If the priority has been elevated due to priority
1660 * propagation, we may have to move ourselves to a new
1661 * queue. This could be optimized to not re-add in some
1664 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1666 td->td_priority = prio;
1667 sched_add(td, SRQ_BORROWING);
1671 * If the thread is currently running we may have to adjust the lowpri
1672 * information so other cpus are aware of our current priority.
1674 if (TD_IS_RUNNING(td)) {
1675 tdq = TDQ_CPU(ts->ts_cpu);
1676 oldpri = td->td_priority;
1677 td->td_priority = prio;
1678 if (prio < tdq->tdq_lowpri)
1679 tdq->tdq_lowpri = prio;
1680 else if (tdq->tdq_lowpri == oldpri)
1681 tdq_setlowpri(tdq, td);
1684 td->td_priority = prio;
1688 * Update a thread's priority when it is lent another thread's
1692 sched_lend_prio(struct thread *td, u_char prio)
1695 td->td_flags |= TDF_BORROWING;
1696 sched_thread_priority(td, prio);
1700 * Restore a thread's priority when priority propagation is
1701 * over. The prio argument is the minimum priority the thread
1702 * needs to have to satisfy other possible priority lending
1703 * requests. If the thread's regular priority is less
1704 * important than prio, the thread will keep a priority boost
1708 sched_unlend_prio(struct thread *td, u_char prio)
1712 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1713 td->td_base_pri <= PRI_MAX_TIMESHARE)
1714 base_pri = td->td_user_pri;
1716 base_pri = td->td_base_pri;
1717 if (prio >= base_pri) {
1718 td->td_flags &= ~TDF_BORROWING;
1719 sched_thread_priority(td, base_pri);
1721 sched_lend_prio(td, prio);
1725 * Standard entry for setting the priority to an absolute value.
1728 sched_prio(struct thread *td, u_char prio)
1732 /* First, update the base priority. */
1733 td->td_base_pri = prio;
1736 * If the thread is borrowing another thread's priority, don't
1737 * ever lower the priority.
1739 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1742 /* Change the real priority. */
1743 oldprio = td->td_priority;
1744 sched_thread_priority(td, prio);
1747 * If the thread is on a turnstile, then let the turnstile update
1750 if (TD_ON_LOCK(td) && oldprio != prio)
1751 turnstile_adjust(td, oldprio);
1755 * Set the base user priority, does not effect current running priority.
1758 sched_user_prio(struct thread *td, u_char prio)
1761 td->td_base_user_pri = prio;
1762 if (td->td_lend_user_pri <= prio)
1764 td->td_user_pri = prio;
1768 sched_lend_user_prio(struct thread *td, u_char prio)
1771 THREAD_LOCK_ASSERT(td, MA_OWNED);
1772 td->td_lend_user_pri = prio;
1773 td->td_user_pri = min(prio, td->td_base_user_pri);
1774 if (td->td_priority > td->td_user_pri)
1775 sched_prio(td, td->td_user_pri);
1776 else if (td->td_priority != td->td_user_pri)
1777 td->td_flags |= TDF_NEEDRESCHED;
1781 * Handle migration from sched_switch(). This happens only for
1785 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1789 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1791 tdq_load_rem(tdq, td);
1793 * Do the lock dance required to avoid LOR. We grab an extra
1794 * spinlock nesting to prevent preemption while we're
1795 * not holding either run-queue lock.
1798 thread_lock_block(td); /* This releases the lock on tdq. */
1801 * Acquire both run-queue locks before placing the thread on the new
1802 * run-queue to avoid deadlocks created by placing a thread with a
1803 * blocked lock on the run-queue of a remote processor. The deadlock
1804 * occurs when a third processor attempts to lock the two queues in
1805 * question while the target processor is spinning with its own
1806 * run-queue lock held while waiting for the blocked lock to clear.
1808 tdq_lock_pair(tdn, tdq);
1809 tdq_add(tdn, td, flags);
1810 tdq_notify(tdn, td);
1814 return (TDQ_LOCKPTR(tdn));
1818 * Variadic version of thread_lock_unblock() that does not assume td_lock
1822 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1824 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1829 * Switch threads. This function has to handle threads coming in while
1830 * blocked for some reason, running, or idle. It also must deal with
1831 * migrating a thread from one queue to another as running threads may
1832 * be assigned elsewhere via binding.
1835 sched_switch(struct thread *td, struct thread *newtd, int flags)
1838 struct td_sched *ts;
1841 int cpuid, preempted;
1843 THREAD_LOCK_ASSERT(td, MA_OWNED);
1844 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1846 cpuid = PCPU_GET(cpuid);
1847 tdq = TDQ_CPU(cpuid);
1850 sched_pctcpu_update(ts, 1);
1851 ts->ts_rltick = ticks;
1852 td->td_lastcpu = td->td_oncpu;
1853 td->td_oncpu = NOCPU;
1854 preempted = !(td->td_flags & TDF_SLICEEND);
1855 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1856 td->td_owepreempt = 0;
1857 tdq->tdq_switchcnt++;
1859 * The lock pointer in an idle thread should never change. Reset it
1860 * to CAN_RUN as well.
1862 if (TD_IS_IDLETHREAD(td)) {
1863 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1865 } else if (TD_IS_RUNNING(td)) {
1866 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1867 srqflag = preempted ?
1868 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1869 SRQ_OURSELF|SRQ_YIELDING;
1871 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1872 ts->ts_cpu = sched_pickcpu(td, 0);
1874 if (ts->ts_cpu == cpuid)
1875 tdq_runq_add(tdq, td, srqflag);
1877 KASSERT(THREAD_CAN_MIGRATE(td) ||
1878 (ts->ts_flags & TSF_BOUND) != 0,
1879 ("Thread %p shouldn't migrate", td));
1880 mtx = sched_switch_migrate(tdq, td, srqflag);
1883 /* This thread must be going to sleep. */
1885 mtx = thread_lock_block(td);
1886 tdq_load_rem(tdq, td);
1889 * We enter here with the thread blocked and assigned to the
1890 * appropriate cpu run-queue or sleep-queue and with the current
1891 * thread-queue locked.
1893 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1894 newtd = choosethread();
1896 * Call the MD code to switch contexts if necessary.
1900 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1901 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1903 SDT_PROBE2(sched, , , off_cpu, td, td->td_proc);
1904 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1905 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1906 sched_pctcpu_update(newtd->td_sched, 0);
1908 #ifdef KDTRACE_HOOKS
1910 * If DTrace has set the active vtime enum to anything
1911 * other than INACTIVE (0), then it should have set the
1914 if (dtrace_vtime_active)
1915 (*dtrace_vtime_switch_func)(newtd);
1918 cpu_switch(td, newtd, mtx);
1920 * We may return from cpu_switch on a different cpu. However,
1921 * we always return with td_lock pointing to the current cpu's
1924 cpuid = PCPU_GET(cpuid);
1925 tdq = TDQ_CPU(cpuid);
1926 lock_profile_obtain_lock_success(
1927 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1929 SDT_PROBE0(sched, , , on_cpu);
1931 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1932 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1935 thread_unblock_switch(td, mtx);
1936 SDT_PROBE0(sched, , , remain_cpu);
1939 * Assert that all went well and return.
1941 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1942 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1943 td->td_oncpu = cpuid;
1947 * Adjust thread priorities as a result of a nice request.
1950 sched_nice(struct proc *p, int nice)
1954 PROC_LOCK_ASSERT(p, MA_OWNED);
1957 FOREACH_THREAD_IN_PROC(p, td) {
1960 sched_prio(td, td->td_base_user_pri);
1966 * Record the sleep time for the interactivity scorer.
1969 sched_sleep(struct thread *td, int prio)
1972 THREAD_LOCK_ASSERT(td, MA_OWNED);
1974 td->td_slptick = ticks;
1975 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1976 td->td_flags |= TDF_CANSWAP;
1977 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1979 if (static_boost == 1 && prio)
1980 sched_prio(td, prio);
1981 else if (static_boost && td->td_priority > static_boost)
1982 sched_prio(td, static_boost);
1986 * Schedule a thread to resume execution and record how long it voluntarily
1987 * slept. We also update the pctcpu, interactivity, and priority.
1990 sched_wakeup(struct thread *td)
1992 struct td_sched *ts;
1995 THREAD_LOCK_ASSERT(td, MA_OWNED);
1997 td->td_flags &= ~TDF_CANSWAP;
1999 * If we slept for more than a tick update our interactivity and
2002 slptick = td->td_slptick;
2004 if (slptick && slptick != ticks) {
2005 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2006 sched_interact_update(td);
2007 sched_pctcpu_update(ts, 0);
2009 /* Reset the slice value after we sleep. */
2010 ts->ts_slice = sched_slice;
2011 sched_add(td, SRQ_BORING);
2015 * Penalize the parent for creating a new child and initialize the child's
2019 sched_fork(struct thread *td, struct thread *child)
2021 THREAD_LOCK_ASSERT(td, MA_OWNED);
2022 sched_pctcpu_update(td->td_sched, 1);
2023 sched_fork_thread(td, child);
2025 * Penalize the parent and child for forking.
2027 sched_interact_fork(child);
2028 sched_priority(child);
2029 td->td_sched->ts_runtime += tickincr;
2030 sched_interact_update(td);
2035 * Fork a new thread, may be within the same process.
2038 sched_fork_thread(struct thread *td, struct thread *child)
2040 struct td_sched *ts;
2041 struct td_sched *ts2;
2043 THREAD_LOCK_ASSERT(td, MA_OWNED);
2048 ts2 = child->td_sched;
2049 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
2050 child->td_cpuset = cpuset_ref(td->td_cpuset);
2051 ts2->ts_cpu = ts->ts_cpu;
2054 * Grab our parents cpu estimation information.
2056 ts2->ts_ticks = ts->ts_ticks;
2057 ts2->ts_ltick = ts->ts_ltick;
2058 ts2->ts_ftick = ts->ts_ftick;
2060 * Do not inherit any borrowed priority from the parent.
2062 child->td_priority = child->td_base_pri;
2064 * And update interactivity score.
2066 ts2->ts_slptime = ts->ts_slptime;
2067 ts2->ts_runtime = ts->ts_runtime;
2068 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
2070 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2075 * Adjust the priority class of a thread.
2078 sched_class(struct thread *td, int class)
2081 THREAD_LOCK_ASSERT(td, MA_OWNED);
2082 if (td->td_pri_class == class)
2084 td->td_pri_class = class;
2088 * Return some of the child's priority and interactivity to the parent.
2091 sched_exit(struct proc *p, struct thread *child)
2095 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2096 "prio:%d", child->td_priority);
2097 PROC_LOCK_ASSERT(p, MA_OWNED);
2098 td = FIRST_THREAD_IN_PROC(p);
2099 sched_exit_thread(td, child);
2103 * Penalize another thread for the time spent on this one. This helps to
2104 * worsen the priority and interactivity of processes which schedule batch
2105 * jobs such as make. This has little effect on the make process itself but
2106 * causes new processes spawned by it to receive worse scores immediately.
2109 sched_exit_thread(struct thread *td, struct thread *child)
2112 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2113 "prio:%d", child->td_priority);
2115 * Give the child's runtime to the parent without returning the
2116 * sleep time as a penalty to the parent. This causes shells that
2117 * launch expensive things to mark their children as expensive.
2120 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2121 sched_interact_update(td);
2127 sched_preempt(struct thread *td)
2131 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2135 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2136 tdq->tdq_ipipending = 0;
2137 if (td->td_priority > tdq->tdq_lowpri) {
2140 flags = SW_INVOL | SW_PREEMPT;
2141 if (td->td_critnest > 1)
2142 td->td_owepreempt = 1;
2143 else if (TD_IS_IDLETHREAD(td))
2144 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2146 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2152 * Fix priorities on return to user-space. Priorities may be elevated due
2153 * to static priorities in msleep() or similar.
2156 sched_userret(struct thread *td)
2159 * XXX we cheat slightly on the locking here to avoid locking in
2160 * the usual case. Setting td_priority here is essentially an
2161 * incomplete workaround for not setting it properly elsewhere.
2162 * Now that some interrupt handlers are threads, not setting it
2163 * properly elsewhere can clobber it in the window between setting
2164 * it here and returning to user mode, so don't waste time setting
2165 * it perfectly here.
2167 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2168 ("thread with borrowed priority returning to userland"));
2169 if (td->td_priority != td->td_user_pri) {
2171 td->td_priority = td->td_user_pri;
2172 td->td_base_pri = td->td_user_pri;
2173 tdq_setlowpri(TDQ_SELF(), td);
2179 * Handle a stathz tick. This is really only relevant for timeshare
2183 sched_clock(struct thread *td)
2186 struct td_sched *ts;
2188 THREAD_LOCK_ASSERT(td, MA_OWNED);
2192 * We run the long term load balancer infrequently on the first cpu.
2194 if (balance_tdq == tdq) {
2195 if (balance_ticks && --balance_ticks == 0)
2200 * Save the old switch count so we have a record of the last ticks
2201 * activity. Initialize the new switch count based on our load.
2202 * If there is some activity seed it to reflect that.
2204 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2205 tdq->tdq_switchcnt = tdq->tdq_load;
2207 * Advance the insert index once for each tick to ensure that all
2208 * threads get a chance to run.
2210 if (tdq->tdq_idx == tdq->tdq_ridx) {
2211 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2212 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2213 tdq->tdq_ridx = tdq->tdq_idx;
2216 sched_pctcpu_update(ts, 1);
2217 if (td->td_pri_class & PRI_FIFO_BIT)
2219 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2221 * We used a tick; charge it to the thread so
2222 * that we can compute our interactivity.
2224 td->td_sched->ts_runtime += tickincr;
2225 sched_interact_update(td);
2230 * Force a context switch if the current thread has used up a full
2231 * time slice (default is 100ms).
2233 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
2234 ts->ts_slice = sched_slice;
2235 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2240 * Called once per hz tick.
2249 * Return whether the current CPU has runnable tasks. Used for in-kernel
2250 * cooperative idle threads.
2253 sched_runnable(void)
2261 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2262 if (tdq->tdq_load > 0)
2265 if (tdq->tdq_load - 1 > 0)
2273 * Choose the highest priority thread to run. The thread is removed from
2274 * the run-queue while running however the load remains. For SMP we set
2275 * the tdq in the global idle bitmask if it idles here.
2284 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2285 td = tdq_choose(tdq);
2287 tdq_runq_rem(tdq, td);
2288 tdq->tdq_lowpri = td->td_priority;
2291 tdq->tdq_lowpri = PRI_MAX_IDLE;
2292 return (PCPU_GET(idlethread));
2296 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2297 * we always request it once we exit a critical section.
2300 sched_setpreempt(struct thread *td)
2306 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2309 pri = td->td_priority;
2310 cpri = ctd->td_priority;
2312 ctd->td_flags |= TDF_NEEDRESCHED;
2313 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2315 if (!sched_shouldpreempt(pri, cpri, 0))
2317 ctd->td_owepreempt = 1;
2321 * Add a thread to a thread queue. Select the appropriate runq and add the
2322 * thread to it. This is the internal function called when the tdq is
2326 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2329 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2330 KASSERT((td->td_inhibitors == 0),
2331 ("sched_add: trying to run inhibited thread"));
2332 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2333 ("sched_add: bad thread state"));
2334 KASSERT(td->td_flags & TDF_INMEM,
2335 ("sched_add: thread swapped out"));
2337 if (td->td_priority < tdq->tdq_lowpri)
2338 tdq->tdq_lowpri = td->td_priority;
2339 tdq_runq_add(tdq, td, flags);
2340 tdq_load_add(tdq, td);
2344 * Select the target thread queue and add a thread to it. Request
2345 * preemption or IPI a remote processor if required.
2348 sched_add(struct thread *td, int flags)
2355 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2356 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2357 sched_tdname(curthread));
2358 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2359 KTR_ATTR_LINKED, sched_tdname(td));
2360 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2361 flags & SRQ_PREEMPTED);
2362 THREAD_LOCK_ASSERT(td, MA_OWNED);
2364 * Recalculate the priority before we select the target cpu or
2367 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2371 * Pick the destination cpu and if it isn't ours transfer to the
2374 cpu = sched_pickcpu(td, flags);
2375 tdq = sched_setcpu(td, cpu, flags);
2376 tdq_add(tdq, td, flags);
2377 if (cpu != PCPU_GET(cpuid)) {
2378 tdq_notify(tdq, td);
2385 * Now that the thread is moving to the run-queue, set the lock
2386 * to the scheduler's lock.
2388 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2389 tdq_add(tdq, td, flags);
2391 if (!(flags & SRQ_YIELDING))
2392 sched_setpreempt(td);
2396 * Remove a thread from a run-queue without running it. This is used
2397 * when we're stealing a thread from a remote queue. Otherwise all threads
2398 * exit by calling sched_exit_thread() and sched_throw() themselves.
2401 sched_rem(struct thread *td)
2405 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2406 "prio:%d", td->td_priority);
2407 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2408 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2409 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2410 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2411 KASSERT(TD_ON_RUNQ(td),
2412 ("sched_rem: thread not on run queue"));
2413 tdq_runq_rem(tdq, td);
2414 tdq_load_rem(tdq, td);
2416 if (td->td_priority == tdq->tdq_lowpri)
2417 tdq_setlowpri(tdq, NULL);
2421 * Fetch cpu utilization information. Updates on demand.
2424 sched_pctcpu(struct thread *td)
2427 struct td_sched *ts;
2434 THREAD_LOCK_ASSERT(td, MA_OWNED);
2435 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2439 /* How many rtick per second ? */
2440 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2441 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2448 * Enforce affinity settings for a thread. Called after adjustments to
2452 sched_affinity(struct thread *td)
2455 struct td_sched *ts;
2457 THREAD_LOCK_ASSERT(td, MA_OWNED);
2459 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2461 if (TD_ON_RUNQ(td)) {
2463 sched_add(td, SRQ_BORING);
2466 if (!TD_IS_RUNNING(td))
2469 * Force a switch before returning to userspace. If the
2470 * target thread is not running locally send an ipi to force
2473 td->td_flags |= TDF_NEEDRESCHED;
2474 if (td != curthread)
2475 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2480 * Bind a thread to a target cpu.
2483 sched_bind(struct thread *td, int cpu)
2485 struct td_sched *ts;
2487 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2488 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2490 if (ts->ts_flags & TSF_BOUND)
2492 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2493 ts->ts_flags |= TSF_BOUND;
2495 if (PCPU_GET(cpuid) == cpu)
2498 /* When we return from mi_switch we'll be on the correct cpu. */
2499 mi_switch(SW_VOL, NULL);
2503 * Release a bound thread.
2506 sched_unbind(struct thread *td)
2508 struct td_sched *ts;
2510 THREAD_LOCK_ASSERT(td, MA_OWNED);
2511 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2513 if ((ts->ts_flags & TSF_BOUND) == 0)
2515 ts->ts_flags &= ~TSF_BOUND;
2520 sched_is_bound(struct thread *td)
2522 THREAD_LOCK_ASSERT(td, MA_OWNED);
2523 return (td->td_sched->ts_flags & TSF_BOUND);
2530 sched_relinquish(struct thread *td)
2533 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2538 * Return the total system load.
2549 total += TDQ_CPU(i)->tdq_sysload;
2552 return (TDQ_SELF()->tdq_sysload);
2557 sched_sizeof_proc(void)
2559 return (sizeof(struct proc));
2563 sched_sizeof_thread(void)
2565 return (sizeof(struct thread) + sizeof(struct td_sched));
2569 #define TDQ_IDLESPIN(tdq) \
2570 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2572 #define TDQ_IDLESPIN(tdq) 1
2576 * The actual idle process.
2579 sched_idletd(void *dummy)
2586 mtx_assert(&Giant, MA_NOTOWNED);
2589 THREAD_NO_SLEEPING();
2592 if (tdq_idled(tdq) == 0)
2595 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2597 * If we're switching very frequently, spin while checking
2598 * for load rather than entering a low power state that
2599 * may require an IPI. However, don't do any busy
2600 * loops while on SMT machines as this simply steals
2601 * cycles from cores doing useful work.
2603 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2604 for (i = 0; i < sched_idlespins; i++) {
2610 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2611 if (tdq->tdq_load == 0) {
2612 tdq->tdq_cpu_idle = 1;
2613 if (tdq->tdq_load == 0) {
2614 cpu_idle(switchcnt > sched_idlespinthresh * 4);
2615 tdq->tdq_switchcnt++;
2617 tdq->tdq_cpu_idle = 0;
2619 if (tdq->tdq_load) {
2621 mi_switch(SW_VOL | SWT_IDLE, NULL);
2628 * A CPU is entering for the first time or a thread is exiting.
2631 sched_throw(struct thread *td)
2633 struct thread *newtd;
2638 /* Correct spinlock nesting and acquire the correct lock. */
2641 PCPU_SET(switchtime, cpu_ticks());
2642 PCPU_SET(switchticks, ticks);
2644 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2645 tdq_load_rem(tdq, td);
2646 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2648 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2649 newtd = choosethread();
2650 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2651 cpu_throw(td, newtd); /* doesn't return */
2655 * This is called from fork_exit(). Just acquire the correct locks and
2656 * let fork do the rest of the work.
2659 sched_fork_exit(struct thread *td)
2661 struct td_sched *ts;
2666 * Finish setting up thread glue so that it begins execution in a
2667 * non-nested critical section with the scheduler lock held.
2669 cpuid = PCPU_GET(cpuid);
2670 tdq = TDQ_CPU(cpuid);
2672 if (TD_IS_IDLETHREAD(td))
2673 td->td_lock = TDQ_LOCKPTR(tdq);
2674 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2675 td->td_oncpu = cpuid;
2676 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2677 lock_profile_obtain_lock_success(
2678 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2682 * Create on first use to catch odd startup conditons.
2685 sched_tdname(struct thread *td)
2688 struct td_sched *ts;
2691 if (ts->ts_name[0] == '\0')
2692 snprintf(ts->ts_name, sizeof(ts->ts_name),
2693 "%s tid %d", td->td_name, td->td_tid);
2694 return (ts->ts_name);
2696 return (td->td_name);
2702 sched_clear_tdname(struct thread *td)
2704 struct td_sched *ts;
2707 ts->ts_name[0] = '\0';
2714 * Build the CPU topology dump string. Is recursively called to collect
2715 * the topology tree.
2718 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2721 char cpusetbuf[CPUSETBUFSIZ];
2724 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2725 "", 1 + indent / 2, cg->cg_level);
2726 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2727 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2729 for (i = 0; i < MAXCPU; i++) {
2730 if (CPU_ISSET(i, &cg->cg_mask)) {
2732 sbuf_printf(sb, ", ");
2735 sbuf_printf(sb, "%d", i);
2738 sbuf_printf(sb, "</cpu>\n");
2740 if (cg->cg_flags != 0) {
2741 sbuf_printf(sb, "%*s <flags>", indent, "");
2742 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2743 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2744 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2745 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2746 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2747 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2748 sbuf_printf(sb, "</flags>\n");
2751 if (cg->cg_children > 0) {
2752 sbuf_printf(sb, "%*s <children>\n", indent, "");
2753 for (i = 0; i < cg->cg_children; i++)
2754 sysctl_kern_sched_topology_spec_internal(sb,
2755 &cg->cg_child[i], indent+2);
2756 sbuf_printf(sb, "%*s </children>\n", indent, "");
2758 sbuf_printf(sb, "%*s</group>\n", indent, "");
2763 * Sysctl handler for retrieving topology dump. It's a wrapper for
2764 * the recursive sysctl_kern_smp_topology_spec_internal().
2767 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2772 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2774 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2778 sbuf_printf(topo, "<groups>\n");
2779 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2780 sbuf_printf(topo, "</groups>\n");
2784 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2793 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2795 int error, new_val, period;
2797 period = 1000000 / realstathz;
2798 new_val = period * sched_slice;
2799 error = sysctl_handle_int(oidp, &new_val, 0, req);
2800 if (error != 0 || req->newptr == NULL)
2804 sched_slice = imax(1, (new_val + period / 2) / period);
2805 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2810 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2811 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2813 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2814 NULL, 0, sysctl_kern_quantum, "I",
2815 "Quantum for timeshare threads in microseconds");
2816 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2817 "Quantum for timeshare threads in stathz ticks");
2818 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2819 "Interactivity score threshold");
2820 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2822 "Maximal (lowest) priority for preemption");
2823 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2824 "Assign static kernel priorities to sleeping threads");
2825 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2826 "Number of times idle thread will spin waiting for new work");
2827 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2828 &sched_idlespinthresh, 0,
2829 "Threshold before we will permit idle thread spinning");
2831 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2832 "Number of hz ticks to keep thread affinity for");
2833 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2834 "Enables the long-term load balancer");
2835 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2836 &balance_interval, 0,
2837 "Average period in stathz ticks to run the long-term balancer");
2838 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2839 "Attempts to steal work from other cores before idling");
2840 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2841 "Minimum load on remote CPU before we'll steal");
2842 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2843 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2844 "XML dump of detected CPU topology");
2847 /* ps compat. All cpu percentages from ULE are weighted. */
2848 static int ccpu = 0;
2849 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");