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>
58 #include <sys/sysctl.h>
59 #include <sys/sysproto.h>
60 #include <sys/turnstile.h>
62 #include <sys/vmmeter.h>
63 #include <sys/cpuset.h>
67 #include <sys/pmckern.h>
71 #include <sys/dtrace_bsd.h>
72 int dtrace_vtime_active;
73 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
76 #include <machine/cpu.h>
77 #include <machine/smp.h>
79 #if defined(__sparc64__)
80 #error "This architecture is not currently compatible with ULE"
85 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
86 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
87 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
90 * Thread scheduler specific section. All fields are protected
94 struct runq *ts_runq; /* Run-queue we're queued on. */
95 short ts_flags; /* TSF_* flags. */
96 u_char ts_cpu; /* CPU that we have affinity for. */
97 int ts_rltick; /* Real last tick, for affinity. */
98 int ts_slice; /* Ticks of slice remaining. */
99 u_int ts_slptime; /* Number of ticks we vol. slept */
100 u_int ts_runtime; /* Number of ticks we were running */
101 int ts_ltick; /* Last tick that we were running on */
102 int ts_incrtick; /* Last tick that we incremented 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)
129 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
130 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
131 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
132 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
135 * Cpu percentage computation macros and defines.
137 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
138 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
139 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
140 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
141 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
142 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
144 #define SCHED_TICK_SECS 10
145 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
146 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
147 #define SCHED_TICK_SHIFT 10
148 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
149 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
152 * These macros determine priorities for non-interactive threads. They are
153 * assigned a priority based on their recent cpu utilization as expressed
154 * by the ratio of ticks to the tick total. NHALF priorities at the start
155 * and end of the MIN to MAX timeshare range are only reachable with negative
156 * or positive nice respectively.
158 * PRI_RANGE: Priority range for utilization dependent priorities.
159 * PRI_NRESV: Number of nice values.
160 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
161 * PRI_NICE: Determines the part of the priority inherited from nice.
163 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
164 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
165 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
166 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
167 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
168 #define SCHED_PRI_TICKS(ts) \
169 (SCHED_TICK_HZ((ts)) / \
170 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
171 #define SCHED_PRI_NICE(nice) (nice)
174 * These determine the interactivity of a process. Interactivity differs from
175 * cpu utilization in that it expresses the voluntary time slept vs time ran
176 * while cpu utilization includes all time not running. This more accurately
177 * models the intent of the thread.
179 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
180 * before throttling back.
181 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
182 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
183 * INTERACT_THRESH: Threshold for placement on the current runq.
185 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
186 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
187 #define SCHED_INTERACT_MAX (100)
188 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
189 #define SCHED_INTERACT_THRESH (30)
192 * tickincr: Converts a stathz tick into a hz domain scaled by
193 * the shift factor. Without the shift the error rate
194 * due to rounding would be unacceptably high.
195 * realstathz: stathz is sometimes 0 and run off of hz.
196 * sched_slice: Runtime of each thread before rescheduling.
197 * preempt_thresh: Priority threshold for preemption and remote IPIs.
199 static int sched_interact = SCHED_INTERACT_THRESH;
200 static int realstathz;
202 static int sched_slice = 1;
204 #ifdef FULL_PREEMPTION
205 static int preempt_thresh = PRI_MAX_IDLE;
207 static int preempt_thresh = PRI_MIN_KERN;
210 static int preempt_thresh = 0;
212 static int static_boost = PRI_MIN_BATCH;
213 static int sched_idlespins = 10000;
214 static int sched_idlespinthresh = 16;
217 * tdq - per processor runqs and statistics. All fields are protected by the
218 * tdq_lock. The load and lowpri may be accessed without to avoid excess
219 * locking in sched_pickcpu();
222 /* Ordered to improve efficiency of cpu_search() and switch(). */
223 struct mtx tdq_lock; /* run queue lock. */
224 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
225 volatile int tdq_load; /* Aggregate load. */
226 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
227 int tdq_sysload; /* For loadavg, !ITHD load. */
228 int tdq_transferable; /* Transferable thread count. */
229 short tdq_switchcnt; /* Switches this tick. */
230 short tdq_oldswitchcnt; /* Switches last tick. */
231 u_char tdq_lowpri; /* Lowest priority thread. */
232 u_char tdq_ipipending; /* IPI pending. */
233 u_char tdq_idx; /* Current insert index. */
234 u_char tdq_ridx; /* Current removal index. */
235 struct runq tdq_realtime; /* real-time run queue. */
236 struct runq tdq_timeshare; /* timeshare run queue. */
237 struct runq tdq_idle; /* Queue of IDLE threads. */
238 char tdq_name[TDQ_NAME_LEN];
240 char tdq_loadname[TDQ_LOADNAME_LEN];
244 /* Idle thread states and config. */
245 #define TDQ_RUNNING 1
249 struct cpu_group *cpu_top; /* CPU topology */
251 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
252 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
257 static int rebalance = 1;
258 static int balance_interval = 128; /* Default set in sched_initticks(). */
260 static int steal_htt = 1;
261 static int steal_idle = 1;
262 static int steal_thresh = 2;
265 * One thread queue per processor.
267 static struct tdq tdq_cpu[MAXCPU];
268 static struct tdq *balance_tdq;
269 static int balance_ticks;
271 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
272 #define TDQ_CPU(x) (&tdq_cpu[(x)])
273 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
275 static struct tdq tdq_cpu;
277 #define TDQ_ID(x) (0)
278 #define TDQ_SELF() (&tdq_cpu)
279 #define TDQ_CPU(x) (&tdq_cpu)
282 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
283 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
284 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
285 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
286 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
288 static void sched_priority(struct thread *);
289 static void sched_thread_priority(struct thread *, u_char);
290 static int sched_interact_score(struct thread *);
291 static void sched_interact_update(struct thread *);
292 static void sched_interact_fork(struct thread *);
293 static void sched_pctcpu_update(struct td_sched *);
295 /* Operations on per processor queues */
296 static struct thread *tdq_choose(struct tdq *);
297 static void tdq_setup(struct tdq *);
298 static void tdq_load_add(struct tdq *, struct thread *);
299 static void tdq_load_rem(struct tdq *, struct thread *);
300 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
301 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
302 static inline int sched_shouldpreempt(int, int, int);
303 void tdq_print(int cpu);
304 static void runq_print(struct runq *rq);
305 static void tdq_add(struct tdq *, struct thread *, int);
307 static int tdq_move(struct tdq *, struct tdq *);
308 static int tdq_idled(struct tdq *);
309 static void tdq_notify(struct tdq *, struct thread *);
310 static struct thread *tdq_steal(struct tdq *, int);
311 static struct thread *runq_steal(struct runq *, int);
312 static int sched_pickcpu(struct thread *, int);
313 static void sched_balance(void);
314 static int sched_balance_pair(struct tdq *, struct tdq *);
315 static inline struct tdq *sched_setcpu(struct thread *, int, int);
316 static inline void thread_unblock_switch(struct thread *, struct mtx *);
317 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
318 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
319 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
320 struct cpu_group *cg, int indent);
323 static void sched_setup(void *dummy);
324 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
326 static void sched_initticks(void *dummy);
327 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
331 * Print the threads waiting on a run-queue.
334 runq_print(struct runq *rq)
342 for (i = 0; i < RQB_LEN; i++) {
343 printf("\t\trunq bits %d 0x%zx\n",
344 i, rq->rq_status.rqb_bits[i]);
345 for (j = 0; j < RQB_BPW; j++)
346 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
347 pri = j + (i << RQB_L2BPW);
348 rqh = &rq->rq_queues[pri];
349 TAILQ_FOREACH(td, rqh, td_runq) {
350 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
351 td, td->td_name, td->td_priority,
352 td->td_rqindex, pri);
359 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
368 printf("tdq %d:\n", TDQ_ID(tdq));
369 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
370 printf("\tLock name: %s\n", tdq->tdq_name);
371 printf("\tload: %d\n", tdq->tdq_load);
372 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
373 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
374 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
375 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
376 printf("\tload transferable: %d\n", tdq->tdq_transferable);
377 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
378 printf("\trealtime runq:\n");
379 runq_print(&tdq->tdq_realtime);
380 printf("\ttimeshare runq:\n");
381 runq_print(&tdq->tdq_timeshare);
382 printf("\tidle runq:\n");
383 runq_print(&tdq->tdq_idle);
387 sched_shouldpreempt(int pri, int cpri, int remote)
390 * If the new priority is not better than the current priority there is
396 * Always preempt idle.
398 if (cpri >= PRI_MIN_IDLE)
401 * If preemption is disabled don't preempt others.
403 if (preempt_thresh == 0)
406 * Preempt if we exceed the threshold.
408 if (pri <= preempt_thresh)
411 * If we're interactive or better and there is non-interactive
412 * or worse running preempt only remote processors.
414 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
419 #define TS_RQ_PPQ (((PRI_MAX_BATCH - PRI_MIN_BATCH) + 1) / RQ_NQS)
421 * Add a thread to the actual run-queue. Keeps transferable counts up to
422 * date with what is actually on the run-queue. Selects the correct
423 * queue position for timeshare threads.
426 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
431 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
432 THREAD_LOCK_ASSERT(td, MA_OWNED);
434 pri = td->td_priority;
437 if (THREAD_CAN_MIGRATE(td)) {
438 tdq->tdq_transferable++;
439 ts->ts_flags |= TSF_XFERABLE;
441 if (pri < PRI_MIN_BATCH) {
442 ts->ts_runq = &tdq->tdq_realtime;
443 } else if (pri <= PRI_MAX_BATCH) {
444 ts->ts_runq = &tdq->tdq_timeshare;
445 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
446 ("Invalid priority %d on timeshare runq", pri));
448 * This queue contains only priorities between MIN and MAX
449 * realtime. Use the whole queue to represent these values.
451 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
452 pri = (pri - PRI_MIN_BATCH) / TS_RQ_PPQ;
453 pri = (pri + tdq->tdq_idx) % RQ_NQS;
455 * This effectively shortens the queue by one so we
456 * can have a one slot difference between idx and
457 * ridx while we wait for threads to drain.
459 if (tdq->tdq_ridx != tdq->tdq_idx &&
460 pri == tdq->tdq_ridx)
461 pri = (unsigned char)(pri - 1) % RQ_NQS;
464 runq_add_pri(ts->ts_runq, td, pri, flags);
467 ts->ts_runq = &tdq->tdq_idle;
468 runq_add(ts->ts_runq, td, flags);
472 * Remove a thread from a run-queue. This typically happens when a thread
473 * is selected to run. Running threads are not on the queue and the
474 * transferable count does not reflect them.
477 tdq_runq_rem(struct tdq *tdq, struct thread *td)
482 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
483 KASSERT(ts->ts_runq != NULL,
484 ("tdq_runq_remove: thread %p null ts_runq", td));
485 if (ts->ts_flags & TSF_XFERABLE) {
486 tdq->tdq_transferable--;
487 ts->ts_flags &= ~TSF_XFERABLE;
489 if (ts->ts_runq == &tdq->tdq_timeshare) {
490 if (tdq->tdq_idx != tdq->tdq_ridx)
491 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
493 runq_remove_idx(ts->ts_runq, td, NULL);
495 runq_remove(ts->ts_runq, td);
499 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
500 * for this thread to the referenced thread queue.
503 tdq_load_add(struct tdq *tdq, struct thread *td)
506 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
507 THREAD_LOCK_ASSERT(td, MA_OWNED);
510 if ((td->td_flags & TDF_NOLOAD) == 0)
512 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
516 * Remove the load from a thread that is transitioning to a sleep state or
520 tdq_load_rem(struct tdq *tdq, struct thread *td)
523 THREAD_LOCK_ASSERT(td, MA_OWNED);
524 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
525 KASSERT(tdq->tdq_load != 0,
526 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
529 if ((td->td_flags & TDF_NOLOAD) == 0)
531 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
535 * Set lowpri to its exact value by searching the run-queue and
536 * evaluating curthread. curthread may be passed as an optimization.
539 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
543 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
545 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
546 td = tdq_choose(tdq);
547 if (td == NULL || td->td_priority > ctd->td_priority)
548 tdq->tdq_lowpri = ctd->td_priority;
550 tdq->tdq_lowpri = td->td_priority;
558 int cs_limit; /* Min priority for low min load for high. */
561 #define CPU_SEARCH_LOWEST 0x1
562 #define CPU_SEARCH_HIGHEST 0x2
563 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
565 #define CPUSET_FOREACH(cpu, mask) \
566 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
567 if (CPU_ISSET(cpu, &mask))
569 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
570 struct cpu_search *high, const int match);
571 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
572 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
573 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
574 struct cpu_search *high);
577 * This routine compares according to the match argument and should be
578 * reduced in actual instantiations via constant propagation and dead code
582 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
588 if (match & CPU_SEARCH_LOWEST)
589 if (CPU_ISSET(cpu, &low->cs_mask) &&
590 tdq->tdq_load < low->cs_load &&
591 tdq->tdq_lowpri > low->cs_limit) {
593 low->cs_load = tdq->tdq_load;
595 if (match & CPU_SEARCH_HIGHEST)
596 if (CPU_ISSET(cpu, &high->cs_mask) &&
597 tdq->tdq_load >= high->cs_limit &&
598 tdq->tdq_load > high->cs_load &&
599 tdq->tdq_transferable) {
601 high->cs_load = tdq->tdq_load;
603 return (tdq->tdq_load);
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(struct cpu_group *cg, struct cpu_search *low,
619 struct cpu_search *high, const int match)
624 if (cg->cg_children) {
625 struct cpu_search lgroup;
626 struct cpu_search hgroup;
627 struct cpu_group *child;
635 for (i = 0; i < cg->cg_children; i++) {
636 child = &cg->cg_child[i];
637 if (match & CPU_SEARCH_LOWEST) {
641 if (match & CPU_SEARCH_HIGHEST) {
646 case CPU_SEARCH_LOWEST:
647 load = cpu_search_lowest(child, &lgroup);
649 case CPU_SEARCH_HIGHEST:
650 load = cpu_search_highest(child, &hgroup);
652 case CPU_SEARCH_BOTH:
653 load = cpu_search_both(child, &lgroup, &hgroup);
657 if (match & CPU_SEARCH_LOWEST)
658 if (load < lload || low->cs_cpu == -1) {
662 if (match & CPU_SEARCH_HIGHEST)
663 if (load > hload || high->cs_cpu == -1) {
671 CPUSET_FOREACH(cpu, cg->cg_mask)
672 total += cpu_compare(cpu, low, high, match);
678 * cpu_search instantiations must pass constants to maintain the inline
682 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
684 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
688 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
690 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
694 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
695 struct cpu_search *high)
697 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
701 * Find the cpu with the least load via the least loaded path that has a
702 * lowpri greater than pri pri. A pri of -1 indicates any priority is
706 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
708 struct cpu_search low;
714 cpu_search_lowest(cg, &low);
719 * Find the cpu with the highest load via the highest loaded path.
722 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
724 struct cpu_search high;
729 high.cs_limit = minload;
730 cpu_search_highest(cg, &high);
735 * Simultaneously find the highest and lowest loaded cpu reachable via
739 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
741 struct cpu_search high;
742 struct cpu_search low;
752 cpu_search_both(cg, &low, &high);
753 *lowcpu = low.cs_cpu;
754 *highcpu = high.cs_cpu;
759 sched_balance_group(struct cpu_group *cg)
768 sched_both(cg, mask, &low, &high);
769 if (low == high || low == -1 || high == -1)
771 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
774 * If we failed to move any threads determine which cpu
775 * to kick out of the set and try again.
777 if (TDQ_CPU(high)->tdq_transferable == 0)
778 CPU_CLR(high, &mask);
783 for (i = 0; i < cg->cg_children; i++)
784 sched_balance_group(&cg->cg_child[i]);
793 * Select a random time between .5 * balance_interval and
794 * 1.5 * balance_interval.
796 balance_ticks = max(balance_interval / 2, 1);
797 balance_ticks += random() % balance_interval;
798 if (smp_started == 0 || rebalance == 0)
802 sched_balance_group(cpu_top);
807 * Lock two thread queues using their address to maintain lock order.
810 tdq_lock_pair(struct tdq *one, struct tdq *two)
814 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
817 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
822 * Unlock two thread queues. Order is not important here.
825 tdq_unlock_pair(struct tdq *one, struct tdq *two)
832 * Transfer load between two imbalanced thread queues.
835 sched_balance_pair(struct tdq *high, struct tdq *low)
845 tdq_lock_pair(high, low);
846 transferable = high->tdq_transferable;
847 high_load = high->tdq_load;
848 low_load = low->tdq_load;
851 * Determine what the imbalance is and then adjust that to how many
852 * threads we actually have to give up (transferable).
854 if (transferable != 0) {
855 diff = high_load - low_load;
859 move = min(move, transferable);
860 for (i = 0; i < move; i++)
861 moved += tdq_move(high, low);
863 * IPI the target cpu to force it to reschedule with the new
866 ipi_cpu(TDQ_ID(low), IPI_PREEMPT);
868 tdq_unlock_pair(high, low);
873 * Move a thread from one thread queue to another.
876 tdq_move(struct tdq *from, struct tdq *to)
883 TDQ_LOCK_ASSERT(from, MA_OWNED);
884 TDQ_LOCK_ASSERT(to, MA_OWNED);
888 td = tdq_steal(tdq, cpu);
893 * Although the run queue is locked the thread may be blocked. Lock
894 * it to clear this and acquire the run-queue lock.
897 /* Drop recursive lock on from acquired via thread_lock(). */
901 td->td_lock = TDQ_LOCKPTR(to);
902 tdq_add(to, td, SRQ_YIELDING);
907 * This tdq has idled. Try to steal a thread from another cpu and switch
911 tdq_idled(struct tdq *tdq)
913 struct cpu_group *cg;
919 if (smp_started == 0 || steal_idle == 0)
922 CPU_CLR(PCPU_GET(cpuid), &mask);
923 /* We don't want to be preempted while we're iterating. */
925 for (cg = tdq->tdq_cg; cg != NULL; ) {
926 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
927 thresh = steal_thresh;
930 cpu = sched_highest(cg, mask, thresh);
935 steal = TDQ_CPU(cpu);
937 tdq_lock_pair(tdq, steal);
938 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
939 tdq_unlock_pair(tdq, steal);
943 * If a thread was added while interrupts were disabled don't
944 * steal one here. If we fail to acquire one due to affinity
945 * restrictions loop again with this cpu removed from the
948 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
949 tdq_unlock_pair(tdq, steal);
954 mi_switch(SW_VOL | SWT_IDLE, NULL);
955 thread_unlock(curthread);
964 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
967 tdq_notify(struct tdq *tdq, struct thread *td)
973 if (tdq->tdq_ipipending)
975 cpu = td->td_sched->ts_cpu;
976 pri = td->td_priority;
977 ctd = pcpu_find(cpu)->pc_curthread;
978 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
980 if (TD_IS_IDLETHREAD(ctd)) {
982 * If the MD code has an idle wakeup routine try that before
983 * falling back to IPI.
985 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
988 tdq->tdq_ipipending = 1;
989 ipi_cpu(cpu, IPI_PREEMPT);
993 * Steals load from a timeshare queue. Honors the rotating queue head
996 static struct thread *
997 runq_steal_from(struct runq *rq, int cpu, u_char start)
1007 rqb = &rq->rq_status;
1008 bit = start & (RQB_BPW -1);
1012 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1013 if (rqb->rqb_bits[i] == 0)
1016 for (pri = bit; pri < RQB_BPW; pri++)
1017 if (rqb->rqb_bits[i] & (1ul << pri))
1022 pri = RQB_FFS(rqb->rqb_bits[i]);
1023 pri += (i << RQB_L2BPW);
1024 rqh = &rq->rq_queues[pri];
1025 TAILQ_FOREACH(td, rqh, td_runq) {
1026 if (first && THREAD_CAN_MIGRATE(td) &&
1027 THREAD_CAN_SCHED(td, cpu))
1041 * Steals load from a standard linear queue.
1043 static struct thread *
1044 runq_steal(struct runq *rq, int cpu)
1052 rqb = &rq->rq_status;
1053 for (word = 0; word < RQB_LEN; word++) {
1054 if (rqb->rqb_bits[word] == 0)
1056 for (bit = 0; bit < RQB_BPW; bit++) {
1057 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1059 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1060 TAILQ_FOREACH(td, rqh, td_runq)
1061 if (THREAD_CAN_MIGRATE(td) &&
1062 THREAD_CAN_SCHED(td, cpu))
1070 * Attempt to steal a thread in priority order from a thread queue.
1072 static struct thread *
1073 tdq_steal(struct tdq *tdq, int cpu)
1077 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1078 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1080 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1081 cpu, tdq->tdq_ridx)) != NULL)
1083 return (runq_steal(&tdq->tdq_idle, cpu));
1087 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1088 * current lock and returns with the assigned queue locked.
1090 static inline struct tdq *
1091 sched_setcpu(struct thread *td, int cpu, int flags)
1096 THREAD_LOCK_ASSERT(td, MA_OWNED);
1098 td->td_sched->ts_cpu = cpu;
1100 * If the lock matches just return the queue.
1102 if (td->td_lock == TDQ_LOCKPTR(tdq))
1106 * If the thread isn't running its lockptr is a
1107 * turnstile or a sleepqueue. We can just lock_set without
1110 if (TD_CAN_RUN(td)) {
1112 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1117 * The hard case, migration, we need to block the thread first to
1118 * prevent order reversals with other cpus locks.
1121 thread_lock_block(td);
1123 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1128 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1129 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1130 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1131 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1132 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1133 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1136 sched_pickcpu(struct thread *td, int flags)
1138 struct cpu_group *cg;
1139 struct td_sched *ts;
1146 self = PCPU_GET(cpuid);
1148 if (smp_started == 0)
1151 * Don't migrate a running thread from sched_switch().
1153 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1154 return (ts->ts_cpu);
1156 * Prefer to run interrupt threads on the processors that generate
1159 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1160 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1161 SCHED_STAT_INC(pickcpu_intrbind);
1165 * If the thread can run on the last cpu and the affinity has not
1166 * expired or it is idle run it there.
1168 pri = td->td_priority;
1169 tdq = TDQ_CPU(ts->ts_cpu);
1170 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1171 if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1172 SCHED_STAT_INC(pickcpu_idle_affinity);
1173 return (ts->ts_cpu);
1175 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1176 SCHED_STAT_INC(pickcpu_affinity);
1177 return (ts->ts_cpu);
1181 * Search for the highest level in the tree that still has affinity.
1184 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1185 if (SCHED_AFFINITY(ts, cg->cg_level))
1188 mask = td->td_cpuset->cs_mask;
1190 cpu = sched_lowest(cg, mask, pri);
1192 cpu = sched_lowest(cpu_top, mask, -1);
1194 * Compare the lowest loaded cpu to current cpu.
1196 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1197 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1198 SCHED_STAT_INC(pickcpu_local);
1201 SCHED_STAT_INC(pickcpu_lowest);
1202 if (cpu != ts->ts_cpu)
1203 SCHED_STAT_INC(pickcpu_migration);
1204 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1210 * Pick the highest priority task we have and return it.
1212 static struct thread *
1213 tdq_choose(struct tdq *tdq)
1217 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1218 td = runq_choose(&tdq->tdq_realtime);
1221 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1223 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1224 ("tdq_choose: Invalid priority on timeshare queue %d",
1228 td = runq_choose(&tdq->tdq_idle);
1230 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1231 ("tdq_choose: Invalid priority on idle queue %d",
1240 * Initialize a thread queue.
1243 tdq_setup(struct tdq *tdq)
1247 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1248 runq_init(&tdq->tdq_realtime);
1249 runq_init(&tdq->tdq_timeshare);
1250 runq_init(&tdq->tdq_idle);
1251 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1252 "sched lock %d", (int)TDQ_ID(tdq));
1253 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1254 MTX_SPIN | MTX_RECURSE);
1256 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1257 "CPU %d load", (int)TDQ_ID(tdq));
1263 sched_setup_smp(void)
1268 cpu_top = smp_topo();
1272 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1273 if (tdq->tdq_cg == NULL)
1274 panic("Can't find cpu group for %d\n", i);
1276 balance_tdq = TDQ_SELF();
1282 * Setup the thread queues and initialize the topology based on MD
1286 sched_setup(void *dummy)
1297 * To avoid divide-by-zero, we set realstathz a dummy value
1298 * in case which sched_clock() called before sched_initticks().
1301 sched_slice = (realstathz/10); /* ~100ms */
1302 tickincr = 1 << SCHED_TICK_SHIFT;
1304 /* Add thread0's load since it's running. */
1306 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1307 tdq_load_add(tdq, &thread0);
1308 tdq->tdq_lowpri = thread0.td_priority;
1313 * This routine determines the tickincr after stathz and hz are setup.
1317 sched_initticks(void *dummy)
1321 realstathz = stathz ? stathz : hz;
1322 sched_slice = (realstathz/10); /* ~100ms */
1325 * tickincr is shifted out by 10 to avoid rounding errors due to
1326 * hz not being evenly divisible by stathz on all platforms.
1328 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1330 * This does not work for values of stathz that are more than
1331 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1338 * Set the default balance interval now that we know
1339 * what realstathz is.
1341 balance_interval = realstathz;
1343 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1344 * This prevents excess thrashing on large machines and excess idle
1345 * on smaller machines.
1347 steal_thresh = min(fls(mp_ncpus) - 1, 3);
1348 affinity = SCHED_AFFINITY_DEFAULT;
1354 * This is the core of the interactivity algorithm. Determines a score based
1355 * on past behavior. It is the ratio of sleep time to run time scaled to
1356 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1357 * differs from the cpu usage because it does not account for time spent
1358 * waiting on a run-queue. Would be prettier if we had floating point.
1361 sched_interact_score(struct thread *td)
1363 struct td_sched *ts;
1368 * The score is only needed if this is likely to be an interactive
1369 * task. Don't go through the expense of computing it if there's
1372 if (sched_interact <= SCHED_INTERACT_HALF &&
1373 ts->ts_runtime >= ts->ts_slptime)
1374 return (SCHED_INTERACT_HALF);
1376 if (ts->ts_runtime > ts->ts_slptime) {
1377 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1378 return (SCHED_INTERACT_HALF +
1379 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1381 if (ts->ts_slptime > ts->ts_runtime) {
1382 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1383 return (ts->ts_runtime / div);
1385 /* runtime == slptime */
1387 return (SCHED_INTERACT_HALF);
1390 * This can happen if slptime and runtime are 0.
1397 * Scale the scheduling priority according to the "interactivity" of this
1401 sched_priority(struct thread *td)
1406 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1409 * If the score is interactive we place the thread in the realtime
1410 * queue with a priority that is less than kernel and interrupt
1411 * priorities. These threads are not subject to nice restrictions.
1413 * Scores greater than this are placed on the normal timeshare queue
1414 * where the priority is partially decided by the most recent cpu
1415 * utilization and the rest is decided by nice value.
1417 * The nice value of the process has a linear effect on the calculated
1418 * score. Negative nice values make it easier for a thread to be
1419 * considered interactive.
1421 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1422 if (score < sched_interact) {
1423 pri = PRI_MIN_INTERACT;
1424 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1425 sched_interact) * score;
1426 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1427 ("sched_priority: invalid interactive priority %d score %d",
1430 pri = SCHED_PRI_MIN;
1431 if (td->td_sched->ts_ticks)
1432 pri += SCHED_PRI_TICKS(td->td_sched);
1433 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1434 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1435 ("sched_priority: invalid priority %d: nice %d, "
1436 "ticks %d ftick %d ltick %d tick pri %d",
1437 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1438 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1439 SCHED_PRI_TICKS(td->td_sched)));
1441 sched_user_prio(td, pri);
1447 * This routine enforces a maximum limit on the amount of scheduling history
1448 * kept. It is called after either the slptime or runtime is adjusted. This
1449 * function is ugly due to integer math.
1452 sched_interact_update(struct thread *td)
1454 struct td_sched *ts;
1458 sum = ts->ts_runtime + ts->ts_slptime;
1459 if (sum < SCHED_SLP_RUN_MAX)
1462 * This only happens from two places:
1463 * 1) We have added an unusual amount of run time from fork_exit.
1464 * 2) We have added an unusual amount of sleep time from sched_sleep().
1466 if (sum > SCHED_SLP_RUN_MAX * 2) {
1467 if (ts->ts_runtime > ts->ts_slptime) {
1468 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1471 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1477 * If we have exceeded by more than 1/5th then the algorithm below
1478 * will not bring us back into range. Dividing by two here forces
1479 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1481 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1482 ts->ts_runtime /= 2;
1483 ts->ts_slptime /= 2;
1486 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1487 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1491 * Scale back the interactivity history when a child thread is created. The
1492 * history is inherited from the parent but the thread may behave totally
1493 * differently. For example, a shell spawning a compiler process. We want
1494 * to learn that the compiler is behaving badly very quickly.
1497 sched_interact_fork(struct thread *td)
1502 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1503 if (sum > SCHED_SLP_RUN_FORK) {
1504 ratio = sum / SCHED_SLP_RUN_FORK;
1505 td->td_sched->ts_runtime /= ratio;
1506 td->td_sched->ts_slptime /= ratio;
1511 * Called from proc0_init() to setup the scheduler fields.
1518 * Set up the scheduler specific parts of proc0.
1520 proc0.p_sched = NULL; /* XXX */
1521 thread0.td_sched = &td_sched0;
1522 td_sched0.ts_ltick = ticks;
1523 td_sched0.ts_ftick = ticks;
1524 td_sched0.ts_slice = sched_slice;
1528 * This is only somewhat accurate since given many processes of the same
1529 * priority they will switch when their slices run out, which will be
1530 * at most sched_slice stathz ticks.
1533 sched_rr_interval(void)
1536 /* Convert sched_slice to hz */
1537 return (hz/(realstathz/sched_slice));
1541 * Update the percent cpu tracking information when it is requested or
1542 * the total history exceeds the maximum. We keep a sliding history of
1543 * tick counts that slowly decays. This is less precise than the 4BSD
1544 * mechanism since it happens with less regular and frequent events.
1547 sched_pctcpu_update(struct td_sched *ts)
1550 if (ts->ts_ticks == 0)
1552 if (ticks - (hz / 10) < ts->ts_ltick &&
1553 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1556 * Adjust counters and watermark for pctcpu calc.
1558 if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1559 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1563 ts->ts_ltick = ticks;
1564 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1568 * Adjust the priority of a thread. Move it to the appropriate run-queue
1569 * if necessary. This is the back-end for several priority related
1573 sched_thread_priority(struct thread *td, u_char prio)
1575 struct td_sched *ts;
1579 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1580 "prio:%d", td->td_priority, "new prio:%d", prio,
1581 KTR_ATTR_LINKED, sched_tdname(curthread));
1582 if (td != curthread && prio > td->td_priority) {
1583 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1584 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1585 prio, KTR_ATTR_LINKED, sched_tdname(td));
1588 THREAD_LOCK_ASSERT(td, MA_OWNED);
1589 if (td->td_priority == prio)
1592 * If the priority has been elevated due to priority
1593 * propagation, we may have to move ourselves to a new
1594 * queue. This could be optimized to not re-add in some
1597 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1599 td->td_priority = prio;
1600 sched_add(td, SRQ_BORROWING);
1604 * If the thread is currently running we may have to adjust the lowpri
1605 * information so other cpus are aware of our current priority.
1607 if (TD_IS_RUNNING(td)) {
1608 tdq = TDQ_CPU(ts->ts_cpu);
1609 oldpri = td->td_priority;
1610 td->td_priority = prio;
1611 if (prio < tdq->tdq_lowpri)
1612 tdq->tdq_lowpri = prio;
1613 else if (tdq->tdq_lowpri == oldpri)
1614 tdq_setlowpri(tdq, td);
1617 td->td_priority = prio;
1621 * Update a thread's priority when it is lent another thread's
1625 sched_lend_prio(struct thread *td, u_char prio)
1628 td->td_flags |= TDF_BORROWING;
1629 sched_thread_priority(td, prio);
1633 * Restore a thread's priority when priority propagation is
1634 * over. The prio argument is the minimum priority the thread
1635 * needs to have to satisfy other possible priority lending
1636 * requests. If the thread's regular priority is less
1637 * important than prio, the thread will keep a priority boost
1641 sched_unlend_prio(struct thread *td, u_char prio)
1645 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1646 td->td_base_pri <= PRI_MAX_TIMESHARE)
1647 base_pri = td->td_user_pri;
1649 base_pri = td->td_base_pri;
1650 if (prio >= base_pri) {
1651 td->td_flags &= ~TDF_BORROWING;
1652 sched_thread_priority(td, base_pri);
1654 sched_lend_prio(td, prio);
1658 * Standard entry for setting the priority to an absolute value.
1661 sched_prio(struct thread *td, u_char prio)
1665 /* First, update the base priority. */
1666 td->td_base_pri = prio;
1669 * If the thread is borrowing another thread's priority, don't
1670 * ever lower the priority.
1672 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1675 /* Change the real priority. */
1676 oldprio = td->td_priority;
1677 sched_thread_priority(td, prio);
1680 * If the thread is on a turnstile, then let the turnstile update
1683 if (TD_ON_LOCK(td) && oldprio != prio)
1684 turnstile_adjust(td, oldprio);
1688 * Set the base user priority, does not effect current running priority.
1691 sched_user_prio(struct thread *td, u_char prio)
1694 td->td_base_user_pri = prio;
1695 if (td->td_lend_user_pri <= prio)
1697 td->td_user_pri = prio;
1701 sched_lend_user_prio(struct thread *td, u_char prio)
1704 THREAD_LOCK_ASSERT(td, MA_OWNED);
1705 td->td_lend_user_pri = prio;
1706 td->td_user_pri = min(prio, td->td_base_user_pri);
1707 if (td->td_priority > td->td_user_pri)
1708 sched_prio(td, td->td_user_pri);
1709 else if (td->td_priority != td->td_user_pri)
1710 td->td_flags |= TDF_NEEDRESCHED;
1714 * Handle migration from sched_switch(). This happens only for
1718 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1722 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1724 tdq_load_rem(tdq, td);
1726 * Do the lock dance required to avoid LOR. We grab an extra
1727 * spinlock nesting to prevent preemption while we're
1728 * not holding either run-queue lock.
1731 thread_lock_block(td); /* This releases the lock on tdq. */
1734 * Acquire both run-queue locks before placing the thread on the new
1735 * run-queue to avoid deadlocks created by placing a thread with a
1736 * blocked lock on the run-queue of a remote processor. The deadlock
1737 * occurs when a third processor attempts to lock the two queues in
1738 * question while the target processor is spinning with its own
1739 * run-queue lock held while waiting for the blocked lock to clear.
1741 tdq_lock_pair(tdn, tdq);
1742 tdq_add(tdn, td, flags);
1743 tdq_notify(tdn, td);
1747 return (TDQ_LOCKPTR(tdn));
1751 * Variadic version of thread_lock_unblock() that does not assume td_lock
1755 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1757 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1762 * Switch threads. This function has to handle threads coming in while
1763 * blocked for some reason, running, or idle. It also must deal with
1764 * migrating a thread from one queue to another as running threads may
1765 * be assigned elsewhere via binding.
1768 sched_switch(struct thread *td, struct thread *newtd, int flags)
1771 struct td_sched *ts;
1776 THREAD_LOCK_ASSERT(td, MA_OWNED);
1777 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1779 cpuid = PCPU_GET(cpuid);
1780 tdq = TDQ_CPU(cpuid);
1783 ts->ts_rltick = ticks;
1784 td->td_lastcpu = td->td_oncpu;
1785 td->td_oncpu = NOCPU;
1786 if (!(flags & SW_PREEMPT))
1787 td->td_flags &= ~TDF_NEEDRESCHED;
1788 td->td_owepreempt = 0;
1789 tdq->tdq_switchcnt++;
1791 * The lock pointer in an idle thread should never change. Reset it
1792 * to CAN_RUN as well.
1794 if (TD_IS_IDLETHREAD(td)) {
1795 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1797 } else if (TD_IS_RUNNING(td)) {
1798 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1799 srqflag = (flags & SW_PREEMPT) ?
1800 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1801 SRQ_OURSELF|SRQ_YIELDING;
1803 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1804 ts->ts_cpu = sched_pickcpu(td, 0);
1806 if (ts->ts_cpu == cpuid)
1807 tdq_runq_add(tdq, td, srqflag);
1809 KASSERT(THREAD_CAN_MIGRATE(td) ||
1810 (ts->ts_flags & TSF_BOUND) != 0,
1811 ("Thread %p shouldn't migrate", td));
1812 mtx = sched_switch_migrate(tdq, td, srqflag);
1815 /* This thread must be going to sleep. */
1817 mtx = thread_lock_block(td);
1818 tdq_load_rem(tdq, td);
1821 * We enter here with the thread blocked and assigned to the
1822 * appropriate cpu run-queue or sleep-queue and with the current
1823 * thread-queue locked.
1825 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1826 newtd = choosethread();
1828 * Call the MD code to switch contexts if necessary.
1832 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1833 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1835 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1836 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1838 #ifdef KDTRACE_HOOKS
1840 * If DTrace has set the active vtime enum to anything
1841 * other than INACTIVE (0), then it should have set the
1844 if (dtrace_vtime_active)
1845 (*dtrace_vtime_switch_func)(newtd);
1848 cpu_switch(td, newtd, mtx);
1850 * We may return from cpu_switch on a different cpu. However,
1851 * we always return with td_lock pointing to the current cpu's
1854 cpuid = PCPU_GET(cpuid);
1855 tdq = TDQ_CPU(cpuid);
1856 lock_profile_obtain_lock_success(
1857 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1859 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1860 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1863 thread_unblock_switch(td, mtx);
1865 * Assert that all went well and return.
1867 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1868 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1869 td->td_oncpu = cpuid;
1873 * Adjust thread priorities as a result of a nice request.
1876 sched_nice(struct proc *p, int nice)
1880 PROC_LOCK_ASSERT(p, MA_OWNED);
1883 FOREACH_THREAD_IN_PROC(p, td) {
1886 sched_prio(td, td->td_base_user_pri);
1892 * Record the sleep time for the interactivity scorer.
1895 sched_sleep(struct thread *td, int prio)
1898 THREAD_LOCK_ASSERT(td, MA_OWNED);
1900 td->td_slptick = ticks;
1901 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1902 td->td_flags |= TDF_CANSWAP;
1903 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1905 if (static_boost == 1 && prio)
1906 sched_prio(td, prio);
1907 else if (static_boost && td->td_priority > static_boost)
1908 sched_prio(td, static_boost);
1912 * Schedule a thread to resume execution and record how long it voluntarily
1913 * slept. We also update the pctcpu, interactivity, and priority.
1916 sched_wakeup(struct thread *td)
1918 struct td_sched *ts;
1921 THREAD_LOCK_ASSERT(td, MA_OWNED);
1923 td->td_flags &= ~TDF_CANSWAP;
1925 * If we slept for more than a tick update our interactivity and
1928 slptick = td->td_slptick;
1930 if (slptick && slptick != ticks) {
1933 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1934 ts->ts_slptime += hzticks;
1935 sched_interact_update(td);
1936 sched_pctcpu_update(ts);
1938 /* Reset the slice value after we sleep. */
1939 ts->ts_slice = sched_slice;
1940 sched_add(td, SRQ_BORING);
1944 * Penalize the parent for creating a new child and initialize the child's
1948 sched_fork(struct thread *td, struct thread *child)
1950 THREAD_LOCK_ASSERT(td, MA_OWNED);
1951 sched_fork_thread(td, child);
1953 * Penalize the parent and child for forking.
1955 sched_interact_fork(child);
1956 sched_priority(child);
1957 td->td_sched->ts_runtime += tickincr;
1958 sched_interact_update(td);
1963 * Fork a new thread, may be within the same process.
1966 sched_fork_thread(struct thread *td, struct thread *child)
1968 struct td_sched *ts;
1969 struct td_sched *ts2;
1971 THREAD_LOCK_ASSERT(td, MA_OWNED);
1976 ts2 = child->td_sched;
1977 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1978 child->td_cpuset = cpuset_ref(td->td_cpuset);
1979 ts2->ts_cpu = ts->ts_cpu;
1982 * Grab our parents cpu estimation information.
1984 ts2->ts_ticks = ts->ts_ticks;
1985 ts2->ts_ltick = ts->ts_ltick;
1986 ts2->ts_incrtick = ts->ts_incrtick;
1987 ts2->ts_ftick = ts->ts_ftick;
1989 * Do not inherit any borrowed priority from the parent.
1991 child->td_priority = child->td_base_pri;
1993 * And update interactivity score.
1995 ts2->ts_slptime = ts->ts_slptime;
1996 ts2->ts_runtime = ts->ts_runtime;
1997 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
1999 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2004 * Adjust the priority class of a thread.
2007 sched_class(struct thread *td, int class)
2010 THREAD_LOCK_ASSERT(td, MA_OWNED);
2011 if (td->td_pri_class == class)
2013 td->td_pri_class = class;
2017 * Return some of the child's priority and interactivity to the parent.
2020 sched_exit(struct proc *p, struct thread *child)
2024 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2025 "prio:%d", child->td_priority);
2026 PROC_LOCK_ASSERT(p, MA_OWNED);
2027 td = FIRST_THREAD_IN_PROC(p);
2028 sched_exit_thread(td, child);
2032 * Penalize another thread for the time spent on this one. This helps to
2033 * worsen the priority and interactivity of processes which schedule batch
2034 * jobs such as make. This has little effect on the make process itself but
2035 * causes new processes spawned by it to receive worse scores immediately.
2038 sched_exit_thread(struct thread *td, struct thread *child)
2041 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2042 "prio:%d", child->td_priority);
2044 * Give the child's runtime to the parent without returning the
2045 * sleep time as a penalty to the parent. This causes shells that
2046 * launch expensive things to mark their children as expensive.
2049 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2050 sched_interact_update(td);
2056 sched_preempt(struct thread *td)
2062 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2063 tdq->tdq_ipipending = 0;
2064 if (td->td_priority > tdq->tdq_lowpri) {
2067 flags = SW_INVOL | SW_PREEMPT;
2068 if (td->td_critnest > 1)
2069 td->td_owepreempt = 1;
2070 else if (TD_IS_IDLETHREAD(td))
2071 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2073 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2079 * Fix priorities on return to user-space. Priorities may be elevated due
2080 * to static priorities in msleep() or similar.
2083 sched_userret(struct thread *td)
2086 * XXX we cheat slightly on the locking here to avoid locking in
2087 * the usual case. Setting td_priority here is essentially an
2088 * incomplete workaround for not setting it properly elsewhere.
2089 * Now that some interrupt handlers are threads, not setting it
2090 * properly elsewhere can clobber it in the window between setting
2091 * it here and returning to user mode, so don't waste time setting
2092 * it perfectly here.
2094 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2095 ("thread with borrowed priority returning to userland"));
2096 if (td->td_priority != td->td_user_pri) {
2098 td->td_priority = td->td_user_pri;
2099 td->td_base_pri = td->td_user_pri;
2100 tdq_setlowpri(TDQ_SELF(), td);
2106 * Handle a stathz tick. This is really only relevant for timeshare
2110 sched_clock(struct thread *td)
2113 struct td_sched *ts;
2115 THREAD_LOCK_ASSERT(td, MA_OWNED);
2119 * We run the long term load balancer infrequently on the first cpu.
2121 if (balance_tdq == tdq) {
2122 if (balance_ticks && --balance_ticks == 0)
2127 * Save the old switch count so we have a record of the last ticks
2128 * activity. Initialize the new switch count based on our load.
2129 * If there is some activity seed it to reflect that.
2131 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2132 tdq->tdq_switchcnt = tdq->tdq_load;
2134 * Advance the insert index once for each tick to ensure that all
2135 * threads get a chance to run.
2137 if (tdq->tdq_idx == tdq->tdq_ridx) {
2138 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2139 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2140 tdq->tdq_ridx = tdq->tdq_idx;
2143 if (td->td_pri_class & PRI_FIFO_BIT)
2145 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2147 * We used a tick; charge it to the thread so
2148 * that we can compute our interactivity.
2150 td->td_sched->ts_runtime += tickincr;
2151 sched_interact_update(td);
2155 * We used up one time slice.
2157 if (--ts->ts_slice > 0)
2160 * We're out of time, force a requeue at userret().
2162 ts->ts_slice = sched_slice;
2163 td->td_flags |= TDF_NEEDRESCHED;
2167 * Called once per hz tick. Used for cpu utilization information. This
2168 * is easier than trying to scale based on stathz.
2173 struct td_sched *ts;
2175 ts = curthread->td_sched;
2177 * Ticks is updated asynchronously on a single cpu. Check here to
2178 * avoid incrementing ts_ticks multiple times in a single tick.
2180 if (ts->ts_incrtick == ticks)
2182 /* Adjust ticks for pctcpu */
2183 ts->ts_ticks += cnt << SCHED_TICK_SHIFT;
2184 ts->ts_ltick = ticks;
2185 ts->ts_incrtick = ticks;
2187 * Update if we've exceeded our desired tick threshold by over one
2190 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2191 sched_pctcpu_update(ts);
2195 * Return whether the current CPU has runnable tasks. Used for in-kernel
2196 * cooperative idle threads.
2199 sched_runnable(void)
2207 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2208 if (tdq->tdq_load > 0)
2211 if (tdq->tdq_load - 1 > 0)
2219 * Choose the highest priority thread to run. The thread is removed from
2220 * the run-queue while running however the load remains. For SMP we set
2221 * the tdq in the global idle bitmask if it idles here.
2230 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2231 td = tdq_choose(tdq);
2233 td->td_sched->ts_ltick = ticks;
2234 tdq_runq_rem(tdq, td);
2235 tdq->tdq_lowpri = td->td_priority;
2238 tdq->tdq_lowpri = PRI_MAX_IDLE;
2239 return (PCPU_GET(idlethread));
2243 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2244 * we always request it once we exit a critical section.
2247 sched_setpreempt(struct thread *td)
2253 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2256 pri = td->td_priority;
2257 cpri = ctd->td_priority;
2259 ctd->td_flags |= TDF_NEEDRESCHED;
2260 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2262 if (!sched_shouldpreempt(pri, cpri, 0))
2264 ctd->td_owepreempt = 1;
2268 * Add a thread to a thread queue. Select the appropriate runq and add the
2269 * thread to it. This is the internal function called when the tdq is
2273 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2276 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2277 KASSERT((td->td_inhibitors == 0),
2278 ("sched_add: trying to run inhibited thread"));
2279 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2280 ("sched_add: bad thread state"));
2281 KASSERT(td->td_flags & TDF_INMEM,
2282 ("sched_add: thread swapped out"));
2284 if (td->td_priority < tdq->tdq_lowpri)
2285 tdq->tdq_lowpri = td->td_priority;
2286 tdq_runq_add(tdq, td, flags);
2287 tdq_load_add(tdq, td);
2291 * Select the target thread queue and add a thread to it. Request
2292 * preemption or IPI a remote processor if required.
2295 sched_add(struct thread *td, int flags)
2302 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2303 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2304 sched_tdname(curthread));
2305 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2306 KTR_ATTR_LINKED, sched_tdname(td));
2307 THREAD_LOCK_ASSERT(td, MA_OWNED);
2309 * Recalculate the priority before we select the target cpu or
2312 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2316 * Pick the destination cpu and if it isn't ours transfer to the
2319 cpu = sched_pickcpu(td, flags);
2320 tdq = sched_setcpu(td, cpu, flags);
2321 tdq_add(tdq, td, flags);
2322 if (cpu != PCPU_GET(cpuid)) {
2323 tdq_notify(tdq, td);
2330 * Now that the thread is moving to the run-queue, set the lock
2331 * to the scheduler's lock.
2333 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2334 tdq_add(tdq, td, flags);
2336 if (!(flags & SRQ_YIELDING))
2337 sched_setpreempt(td);
2341 * Remove a thread from a run-queue without running it. This is used
2342 * when we're stealing a thread from a remote queue. Otherwise all threads
2343 * exit by calling sched_exit_thread() and sched_throw() themselves.
2346 sched_rem(struct thread *td)
2350 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2351 "prio:%d", td->td_priority);
2352 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2353 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2354 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2355 KASSERT(TD_ON_RUNQ(td),
2356 ("sched_rem: thread not on run queue"));
2357 tdq_runq_rem(tdq, td);
2358 tdq_load_rem(tdq, td);
2360 if (td->td_priority == tdq->tdq_lowpri)
2361 tdq_setlowpri(tdq, NULL);
2365 * Fetch cpu utilization information. Updates on demand.
2368 sched_pctcpu(struct thread *td)
2371 struct td_sched *ts;
2378 THREAD_LOCK_ASSERT(td, MA_OWNED);
2382 sched_pctcpu_update(ts);
2383 /* How many rtick per second ? */
2384 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2385 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2392 * Enforce affinity settings for a thread. Called after adjustments to
2396 sched_affinity(struct thread *td)
2399 struct td_sched *ts;
2401 THREAD_LOCK_ASSERT(td, MA_OWNED);
2403 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2405 if (TD_ON_RUNQ(td)) {
2407 sched_add(td, SRQ_BORING);
2410 if (!TD_IS_RUNNING(td))
2413 * Force a switch before returning to userspace. If the
2414 * target thread is not running locally send an ipi to force
2417 td->td_flags |= TDF_NEEDRESCHED;
2418 if (td != curthread)
2419 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2424 * Bind a thread to a target cpu.
2427 sched_bind(struct thread *td, int cpu)
2429 struct td_sched *ts;
2431 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2432 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2434 if (ts->ts_flags & TSF_BOUND)
2436 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2437 ts->ts_flags |= TSF_BOUND;
2439 if (PCPU_GET(cpuid) == cpu)
2442 /* When we return from mi_switch we'll be on the correct cpu. */
2443 mi_switch(SW_VOL, NULL);
2447 * Release a bound thread.
2450 sched_unbind(struct thread *td)
2452 struct td_sched *ts;
2454 THREAD_LOCK_ASSERT(td, MA_OWNED);
2455 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2457 if ((ts->ts_flags & TSF_BOUND) == 0)
2459 ts->ts_flags &= ~TSF_BOUND;
2464 sched_is_bound(struct thread *td)
2466 THREAD_LOCK_ASSERT(td, MA_OWNED);
2467 return (td->td_sched->ts_flags & TSF_BOUND);
2474 sched_relinquish(struct thread *td)
2477 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2482 * Return the total system load.
2493 total += TDQ_CPU(i)->tdq_sysload;
2496 return (TDQ_SELF()->tdq_sysload);
2501 sched_sizeof_proc(void)
2503 return (sizeof(struct proc));
2507 sched_sizeof_thread(void)
2509 return (sizeof(struct thread) + sizeof(struct td_sched));
2513 #define TDQ_IDLESPIN(tdq) \
2514 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2516 #define TDQ_IDLESPIN(tdq) 1
2520 * The actual idle process.
2523 sched_idletd(void *dummy)
2530 mtx_assert(&Giant, MA_NOTOWNED);
2535 if (tdq_idled(tdq) == 0)
2538 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2540 * If we're switching very frequently, spin while checking
2541 * for load rather than entering a low power state that
2542 * may require an IPI. However, don't do any busy
2543 * loops while on SMT machines as this simply steals
2544 * cycles from cores doing useful work.
2546 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2547 for (i = 0; i < sched_idlespins; i++) {
2553 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2554 if (tdq->tdq_load == 0) {
2555 tdq->tdq_cpu_idle = 1;
2556 if (tdq->tdq_load == 0) {
2557 cpu_idle(switchcnt > sched_idlespinthresh * 4);
2558 tdq->tdq_switchcnt++;
2560 tdq->tdq_cpu_idle = 0;
2562 if (tdq->tdq_load) {
2564 mi_switch(SW_VOL | SWT_IDLE, NULL);
2571 * A CPU is entering for the first time or a thread is exiting.
2574 sched_throw(struct thread *td)
2576 struct thread *newtd;
2581 /* Correct spinlock nesting and acquire the correct lock. */
2585 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2586 tdq_load_rem(tdq, td);
2587 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2589 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2590 newtd = choosethread();
2591 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2592 PCPU_SET(switchtime, cpu_ticks());
2593 PCPU_SET(switchticks, ticks);
2594 cpu_throw(td, newtd); /* doesn't return */
2598 * This is called from fork_exit(). Just acquire the correct locks and
2599 * let fork do the rest of the work.
2602 sched_fork_exit(struct thread *td)
2604 struct td_sched *ts;
2609 * Finish setting up thread glue so that it begins execution in a
2610 * non-nested critical section with the scheduler lock held.
2612 cpuid = PCPU_GET(cpuid);
2613 tdq = TDQ_CPU(cpuid);
2615 if (TD_IS_IDLETHREAD(td))
2616 td->td_lock = TDQ_LOCKPTR(tdq);
2617 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2618 td->td_oncpu = cpuid;
2619 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2620 lock_profile_obtain_lock_success(
2621 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2625 * Create on first use to catch odd startup conditons.
2628 sched_tdname(struct thread *td)
2631 struct td_sched *ts;
2634 if (ts->ts_name[0] == '\0')
2635 snprintf(ts->ts_name, sizeof(ts->ts_name),
2636 "%s tid %d", td->td_name, td->td_tid);
2637 return (ts->ts_name);
2639 return (td->td_name);
2646 * Build the CPU topology dump string. Is recursively called to collect
2647 * the topology tree.
2650 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2653 char cpusetbuf[CPUSETBUFSIZ];
2656 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2657 "", 1 + indent / 2, cg->cg_level);
2658 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2659 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2661 for (i = 0; i < MAXCPU; i++) {
2662 if (CPU_ISSET(i, &cg->cg_mask)) {
2664 sbuf_printf(sb, ", ");
2667 sbuf_printf(sb, "%d", i);
2670 sbuf_printf(sb, "</cpu>\n");
2672 if (cg->cg_flags != 0) {
2673 sbuf_printf(sb, "%*s <flags>", indent, "");
2674 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2675 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2676 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2677 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2678 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2679 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2680 sbuf_printf(sb, "</flags>\n");
2683 if (cg->cg_children > 0) {
2684 sbuf_printf(sb, "%*s <children>\n", indent, "");
2685 for (i = 0; i < cg->cg_children; i++)
2686 sysctl_kern_sched_topology_spec_internal(sb,
2687 &cg->cg_child[i], indent+2);
2688 sbuf_printf(sb, "%*s </children>\n", indent, "");
2690 sbuf_printf(sb, "%*s</group>\n", indent, "");
2695 * Sysctl handler for retrieving topology dump. It's a wrapper for
2696 * the recursive sysctl_kern_smp_topology_spec_internal().
2699 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2704 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2706 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2710 sbuf_printf(topo, "<groups>\n");
2711 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2712 sbuf_printf(topo, "</groups>\n");
2716 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2724 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2725 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2727 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2728 "Slice size for timeshare threads");
2729 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2730 "Interactivity score threshold");
2731 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2732 0,"Min priority for preemption, lower priorities have greater precedence");
2733 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2734 0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2735 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2736 0,"Number of times idle will spin waiting for new work.");
2737 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2738 0,"Threshold before we will permit idle spinning.");
2740 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2741 "Number of hz ticks to keep thread affinity for");
2742 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2743 "Enables the long-term load balancer");
2744 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2745 &balance_interval, 0,
2746 "Average frequency in stathz ticks to run the long-term balancer");
2747 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2748 "Steals work from another hyper-threaded core on idle");
2749 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2750 "Attempts to steal work from other cores before idling");
2751 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2752 "Minimum load on remote cpu before we'll steal");
2754 /* Retrieve SMP topology */
2755 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2756 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2757 "XML dump of detected CPU topology");
2761 /* ps compat. All cpu percentages from ULE are weighted. */
2762 static int ccpu = 0;
2763 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");