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(__powerpc__) && defined(E500)
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)
128 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
130 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
131 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
136 * Cpu percentage computation macros and defines.
138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
145 #define SCHED_TICK_SECS 10
146 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
147 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
148 #define SCHED_TICK_SHIFT 10
149 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
153 * These macros determine priorities for non-interactive threads. They are
154 * assigned a priority based on their recent cpu utilization as expressed
155 * by the ratio of ticks to the tick total. NHALF priorities at the start
156 * and end of the MIN to MAX timeshare range are only reachable with negative
157 * or positive nice respectively.
159 * PRI_RANGE: Priority range for utilization dependent priorities.
160 * PRI_NRESV: Number of nice values.
161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
162 * PRI_NICE: Determines the part of the priority inherited from nice.
164 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
165 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
166 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define SCHED_PRI_TICKS(ts) \
170 (SCHED_TICK_HZ((ts)) / \
171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define SCHED_PRI_NICE(nice) (nice)
175 * These determine the interactivity of a process. Interactivity differs from
176 * cpu utilization in that it expresses the voluntary time slept vs time ran
177 * while cpu utilization includes all time not running. This more accurately
178 * models the intent of the thread.
180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
181 * before throttling back.
182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
183 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
184 * INTERACT_THRESH: Threshold for placement on the current runq.
186 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
187 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
188 #define SCHED_INTERACT_MAX (100)
189 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
190 #define SCHED_INTERACT_THRESH (30)
193 * tickincr: Converts a stathz tick into a hz domain scaled by
194 * the shift factor. Without the shift the error rate
195 * due to rounding would be unacceptably high.
196 * realstathz: stathz is sometimes 0 and run off of hz.
197 * sched_slice: Runtime of each thread before rescheduling.
198 * preempt_thresh: Priority threshold for preemption and remote IPIs.
200 static int sched_interact = SCHED_INTERACT_THRESH;
201 static int realstathz;
203 static int sched_slice = 1;
205 #ifdef FULL_PREEMPTION
206 static int preempt_thresh = PRI_MAX_IDLE;
208 static int preempt_thresh = PRI_MIN_KERN;
211 static int preempt_thresh = 0;
213 static int static_boost = PRI_MIN_BATCH;
214 static int sched_idlespins = 10000;
215 static int sched_idlespinthresh = 16;
218 * tdq - per processor runqs and statistics. All fields are protected by the
219 * tdq_lock. The load and lowpri may be accessed without to avoid excess
220 * locking in sched_pickcpu();
223 /* Ordered to improve efficiency of cpu_search() and switch(). */
224 struct mtx tdq_lock; /* run queue lock. */
225 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
226 volatile int tdq_load; /* Aggregate load. */
227 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
228 int tdq_sysload; /* For loadavg, !ITHD load. */
229 int tdq_transferable; /* Transferable thread count. */
230 short tdq_switchcnt; /* Switches this tick. */
231 short tdq_oldswitchcnt; /* Switches last tick. */
232 u_char tdq_lowpri; /* Lowest priority thread. */
233 u_char tdq_ipipending; /* IPI pending. */
234 u_char tdq_idx; /* Current insert index. */
235 u_char tdq_ridx; /* Current removal index. */
236 struct runq tdq_realtime; /* real-time run queue. */
237 struct runq tdq_timeshare; /* timeshare run queue. */
238 struct runq tdq_idle; /* Queue of IDLE threads. */
239 char tdq_name[TDQ_NAME_LEN];
241 char tdq_loadname[TDQ_LOADNAME_LEN];
245 /* Idle thread states and config. */
246 #define TDQ_RUNNING 1
250 struct cpu_group *cpu_top; /* CPU topology */
252 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
253 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
258 static int rebalance = 1;
259 static int balance_interval = 128; /* Default set in sched_initticks(). */
261 static int steal_htt = 1;
262 static int steal_idle = 1;
263 static int steal_thresh = 2;
266 * One thread queue per processor.
268 static struct tdq tdq_cpu[MAXCPU];
269 static struct tdq *balance_tdq;
270 static int balance_ticks;
272 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
273 #define TDQ_CPU(x) (&tdq_cpu[(x)])
274 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
276 static struct tdq tdq_cpu;
278 #define TDQ_ID(x) (0)
279 #define TDQ_SELF() (&tdq_cpu)
280 #define TDQ_CPU(x) (&tdq_cpu)
283 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
284 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
285 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
286 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
287 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
289 static void sched_priority(struct thread *);
290 static void sched_thread_priority(struct thread *, u_char);
291 static int sched_interact_score(struct thread *);
292 static void sched_interact_update(struct thread *);
293 static void sched_interact_fork(struct thread *);
294 static void sched_pctcpu_update(struct td_sched *);
296 /* Operations on per processor queues */
297 static struct thread *tdq_choose(struct tdq *);
298 static void tdq_setup(struct tdq *);
299 static void tdq_load_add(struct tdq *, struct thread *);
300 static void tdq_load_rem(struct tdq *, struct thread *);
301 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
302 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
303 static inline int sched_shouldpreempt(int, int, int);
304 void tdq_print(int cpu);
305 static void runq_print(struct runq *rq);
306 static void tdq_add(struct tdq *, struct thread *, int);
308 static int tdq_move(struct tdq *, struct tdq *);
309 static int tdq_idled(struct tdq *);
310 static void tdq_notify(struct tdq *, struct thread *);
311 static struct thread *tdq_steal(struct tdq *, int);
312 static struct thread *runq_steal(struct runq *, int);
313 static int sched_pickcpu(struct thread *, int);
314 static void sched_balance(void);
315 static int sched_balance_pair(struct tdq *, struct tdq *);
316 static inline struct tdq *sched_setcpu(struct thread *, int, int);
317 static inline void thread_unblock_switch(struct thread *, struct mtx *);
318 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
319 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
320 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
321 struct cpu_group *cg, int indent);
324 static void sched_setup(void *dummy);
325 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
327 static void sched_initticks(void *dummy);
328 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
332 * Print the threads waiting on a run-queue.
335 runq_print(struct runq *rq)
343 for (i = 0; i < RQB_LEN; i++) {
344 printf("\t\trunq bits %d 0x%zx\n",
345 i, rq->rq_status.rqb_bits[i]);
346 for (j = 0; j < RQB_BPW; j++)
347 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
348 pri = j + (i << RQB_L2BPW);
349 rqh = &rq->rq_queues[pri];
350 TAILQ_FOREACH(td, rqh, td_runq) {
351 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
352 td, td->td_name, td->td_priority,
353 td->td_rqindex, pri);
360 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
369 printf("tdq %d:\n", TDQ_ID(tdq));
370 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
371 printf("\tLock name: %s\n", tdq->tdq_name);
372 printf("\tload: %d\n", tdq->tdq_load);
373 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
374 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
375 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
376 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
377 printf("\tload transferable: %d\n", tdq->tdq_transferable);
378 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
379 printf("\trealtime runq:\n");
380 runq_print(&tdq->tdq_realtime);
381 printf("\ttimeshare runq:\n");
382 runq_print(&tdq->tdq_timeshare);
383 printf("\tidle runq:\n");
384 runq_print(&tdq->tdq_idle);
388 sched_shouldpreempt(int pri, int cpri, int remote)
391 * If the new priority is not better than the current priority there is
397 * Always preempt idle.
399 if (cpri >= PRI_MIN_IDLE)
402 * If preemption is disabled don't preempt others.
404 if (preempt_thresh == 0)
407 * Preempt if we exceed the threshold.
409 if (pri <= preempt_thresh)
412 * If we're interactive or better and there is non-interactive
413 * or worse running preempt only remote processors.
415 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
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 = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
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)
846 tdq_lock_pair(high, low);
847 transferable = high->tdq_transferable;
848 high_load = high->tdq_load;
849 low_load = low->tdq_load;
852 * Determine what the imbalance is and then adjust that to how many
853 * threads we actually have to give up (transferable).
855 if (transferable != 0) {
856 diff = high_load - low_load;
860 move = min(move, transferable);
861 for (i = 0; i < move; i++)
862 moved += tdq_move(high, low);
864 * In case the target isn't the current cpu IPI it to force a
865 * reschedule with the new workload.
869 if (cpu != PCPU_GET(cpuid))
870 ipi_cpu(cpu, IPI_PREEMPT);
873 tdq_unlock_pair(high, low);
878 * Move a thread from one thread queue to another.
881 tdq_move(struct tdq *from, struct tdq *to)
888 TDQ_LOCK_ASSERT(from, MA_OWNED);
889 TDQ_LOCK_ASSERT(to, MA_OWNED);
893 td = tdq_steal(tdq, cpu);
898 * Although the run queue is locked the thread may be blocked. Lock
899 * it to clear this and acquire the run-queue lock.
902 /* Drop recursive lock on from acquired via thread_lock(). */
906 td->td_lock = TDQ_LOCKPTR(to);
907 tdq_add(to, td, SRQ_YIELDING);
912 * This tdq has idled. Try to steal a thread from another cpu and switch
916 tdq_idled(struct tdq *tdq)
918 struct cpu_group *cg;
924 if (smp_started == 0 || steal_idle == 0)
927 CPU_CLR(PCPU_GET(cpuid), &mask);
928 /* We don't want to be preempted while we're iterating. */
930 for (cg = tdq->tdq_cg; cg != NULL; ) {
931 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
932 thresh = steal_thresh;
935 cpu = sched_highest(cg, mask, thresh);
940 steal = TDQ_CPU(cpu);
942 tdq_lock_pair(tdq, steal);
943 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
944 tdq_unlock_pair(tdq, steal);
948 * If a thread was added while interrupts were disabled don't
949 * steal one here. If we fail to acquire one due to affinity
950 * restrictions loop again with this cpu removed from the
953 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
954 tdq_unlock_pair(tdq, steal);
959 mi_switch(SW_VOL | SWT_IDLE, NULL);
960 thread_unlock(curthread);
969 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
972 tdq_notify(struct tdq *tdq, struct thread *td)
978 if (tdq->tdq_ipipending)
980 cpu = td->td_sched->ts_cpu;
981 pri = td->td_priority;
982 ctd = pcpu_find(cpu)->pc_curthread;
983 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
985 if (TD_IS_IDLETHREAD(ctd)) {
987 * If the MD code has an idle wakeup routine try that before
988 * falling back to IPI.
990 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
993 tdq->tdq_ipipending = 1;
994 ipi_cpu(cpu, IPI_PREEMPT);
998 * Steals load from a timeshare queue. Honors the rotating queue head
1001 static struct thread *
1002 runq_steal_from(struct runq *rq, int cpu, u_char start)
1012 rqb = &rq->rq_status;
1013 bit = start & (RQB_BPW -1);
1017 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1018 if (rqb->rqb_bits[i] == 0)
1021 for (pri = bit; pri < RQB_BPW; pri++)
1022 if (rqb->rqb_bits[i] & (1ul << pri))
1027 pri = RQB_FFS(rqb->rqb_bits[i]);
1028 pri += (i << RQB_L2BPW);
1029 rqh = &rq->rq_queues[pri];
1030 TAILQ_FOREACH(td, rqh, td_runq) {
1031 if (first && THREAD_CAN_MIGRATE(td) &&
1032 THREAD_CAN_SCHED(td, cpu))
1046 * Steals load from a standard linear queue.
1048 static struct thread *
1049 runq_steal(struct runq *rq, int cpu)
1057 rqb = &rq->rq_status;
1058 for (word = 0; word < RQB_LEN; word++) {
1059 if (rqb->rqb_bits[word] == 0)
1061 for (bit = 0; bit < RQB_BPW; bit++) {
1062 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1064 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1065 TAILQ_FOREACH(td, rqh, td_runq)
1066 if (THREAD_CAN_MIGRATE(td) &&
1067 THREAD_CAN_SCHED(td, cpu))
1075 * Attempt to steal a thread in priority order from a thread queue.
1077 static struct thread *
1078 tdq_steal(struct tdq *tdq, int cpu)
1082 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1083 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1085 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1086 cpu, tdq->tdq_ridx)) != NULL)
1088 return (runq_steal(&tdq->tdq_idle, cpu));
1092 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1093 * current lock and returns with the assigned queue locked.
1095 static inline struct tdq *
1096 sched_setcpu(struct thread *td, int cpu, int flags)
1101 THREAD_LOCK_ASSERT(td, MA_OWNED);
1103 td->td_sched->ts_cpu = cpu;
1105 * If the lock matches just return the queue.
1107 if (td->td_lock == TDQ_LOCKPTR(tdq))
1111 * If the thread isn't running its lockptr is a
1112 * turnstile or a sleepqueue. We can just lock_set without
1115 if (TD_CAN_RUN(td)) {
1117 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1122 * The hard case, migration, we need to block the thread first to
1123 * prevent order reversals with other cpus locks.
1126 thread_lock_block(td);
1128 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1133 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1134 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1135 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1136 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1137 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1138 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1141 sched_pickcpu(struct thread *td, int flags)
1143 struct cpu_group *cg;
1144 struct td_sched *ts;
1151 self = PCPU_GET(cpuid);
1153 if (smp_started == 0)
1156 * Don't migrate a running thread from sched_switch().
1158 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1159 return (ts->ts_cpu);
1161 * Prefer to run interrupt threads on the processors that generate
1164 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1165 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1166 SCHED_STAT_INC(pickcpu_intrbind);
1170 * If the thread can run on the last cpu and the affinity has not
1171 * expired or it is idle run it there.
1173 pri = td->td_priority;
1174 tdq = TDQ_CPU(ts->ts_cpu);
1175 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1176 if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1177 SCHED_STAT_INC(pickcpu_idle_affinity);
1178 return (ts->ts_cpu);
1180 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1181 SCHED_STAT_INC(pickcpu_affinity);
1182 return (ts->ts_cpu);
1186 * Search for the highest level in the tree that still has affinity.
1189 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1190 if (SCHED_AFFINITY(ts, cg->cg_level))
1193 mask = td->td_cpuset->cs_mask;
1195 cpu = sched_lowest(cg, mask, pri);
1197 cpu = sched_lowest(cpu_top, mask, -1);
1199 * Compare the lowest loaded cpu to current cpu.
1201 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1202 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1203 SCHED_STAT_INC(pickcpu_local);
1206 SCHED_STAT_INC(pickcpu_lowest);
1207 if (cpu != ts->ts_cpu)
1208 SCHED_STAT_INC(pickcpu_migration);
1209 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1215 * Pick the highest priority task we have and return it.
1217 static struct thread *
1218 tdq_choose(struct tdq *tdq)
1222 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1223 td = runq_choose(&tdq->tdq_realtime);
1226 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1228 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1229 ("tdq_choose: Invalid priority on timeshare queue %d",
1233 td = runq_choose(&tdq->tdq_idle);
1235 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1236 ("tdq_choose: Invalid priority on idle queue %d",
1245 * Initialize a thread queue.
1248 tdq_setup(struct tdq *tdq)
1252 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1253 runq_init(&tdq->tdq_realtime);
1254 runq_init(&tdq->tdq_timeshare);
1255 runq_init(&tdq->tdq_idle);
1256 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1257 "sched lock %d", (int)TDQ_ID(tdq));
1258 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1259 MTX_SPIN | MTX_RECURSE);
1261 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1262 "CPU %d load", (int)TDQ_ID(tdq));
1268 sched_setup_smp(void)
1273 cpu_top = smp_topo();
1277 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1278 if (tdq->tdq_cg == NULL)
1279 panic("Can't find cpu group for %d\n", i);
1281 balance_tdq = TDQ_SELF();
1287 * Setup the thread queues and initialize the topology based on MD
1291 sched_setup(void *dummy)
1302 * To avoid divide-by-zero, we set realstathz a dummy value
1303 * in case which sched_clock() called before sched_initticks().
1306 sched_slice = (realstathz/10); /* ~100ms */
1307 tickincr = 1 << SCHED_TICK_SHIFT;
1309 /* Add thread0's load since it's running. */
1311 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1312 tdq_load_add(tdq, &thread0);
1313 tdq->tdq_lowpri = thread0.td_priority;
1318 * This routine determines the tickincr after stathz and hz are setup.
1322 sched_initticks(void *dummy)
1326 realstathz = stathz ? stathz : hz;
1327 sched_slice = (realstathz/10); /* ~100ms */
1330 * tickincr is shifted out by 10 to avoid rounding errors due to
1331 * hz not being evenly divisible by stathz on all platforms.
1333 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1335 * This does not work for values of stathz that are more than
1336 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1343 * Set the default balance interval now that we know
1344 * what realstathz is.
1346 balance_interval = realstathz;
1348 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1349 * This prevents excess thrashing on large machines and excess idle
1350 * on smaller machines.
1352 steal_thresh = min(fls(mp_ncpus) - 1, 3);
1353 affinity = SCHED_AFFINITY_DEFAULT;
1359 * This is the core of the interactivity algorithm. Determines a score based
1360 * on past behavior. It is the ratio of sleep time to run time scaled to
1361 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1362 * differs from the cpu usage because it does not account for time spent
1363 * waiting on a run-queue. Would be prettier if we had floating point.
1366 sched_interact_score(struct thread *td)
1368 struct td_sched *ts;
1373 * The score is only needed if this is likely to be an interactive
1374 * task. Don't go through the expense of computing it if there's
1377 if (sched_interact <= SCHED_INTERACT_HALF &&
1378 ts->ts_runtime >= ts->ts_slptime)
1379 return (SCHED_INTERACT_HALF);
1381 if (ts->ts_runtime > ts->ts_slptime) {
1382 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1383 return (SCHED_INTERACT_HALF +
1384 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1386 if (ts->ts_slptime > ts->ts_runtime) {
1387 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1388 return (ts->ts_runtime / div);
1390 /* runtime == slptime */
1392 return (SCHED_INTERACT_HALF);
1395 * This can happen if slptime and runtime are 0.
1402 * Scale the scheduling priority according to the "interactivity" of this
1406 sched_priority(struct thread *td)
1411 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1414 * If the score is interactive we place the thread in the realtime
1415 * queue with a priority that is less than kernel and interrupt
1416 * priorities. These threads are not subject to nice restrictions.
1418 * Scores greater than this are placed on the normal timeshare queue
1419 * where the priority is partially decided by the most recent cpu
1420 * utilization and the rest is decided by nice value.
1422 * The nice value of the process has a linear effect on the calculated
1423 * score. Negative nice values make it easier for a thread to be
1424 * considered interactive.
1426 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1427 if (score < sched_interact) {
1428 pri = PRI_MIN_INTERACT;
1429 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1430 sched_interact) * score;
1431 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1432 ("sched_priority: invalid interactive priority %d score %d",
1435 pri = SCHED_PRI_MIN;
1436 if (td->td_sched->ts_ticks)
1437 pri += min(SCHED_PRI_TICKS(td->td_sched),
1439 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1440 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1441 ("sched_priority: invalid priority %d: nice %d, "
1442 "ticks %d ftick %d ltick %d tick pri %d",
1443 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1444 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1445 SCHED_PRI_TICKS(td->td_sched)));
1447 sched_user_prio(td, pri);
1453 * This routine enforces a maximum limit on the amount of scheduling history
1454 * kept. It is called after either the slptime or runtime is adjusted. This
1455 * function is ugly due to integer math.
1458 sched_interact_update(struct thread *td)
1460 struct td_sched *ts;
1464 sum = ts->ts_runtime + ts->ts_slptime;
1465 if (sum < SCHED_SLP_RUN_MAX)
1468 * This only happens from two places:
1469 * 1) We have added an unusual amount of run time from fork_exit.
1470 * 2) We have added an unusual amount of sleep time from sched_sleep().
1472 if (sum > SCHED_SLP_RUN_MAX * 2) {
1473 if (ts->ts_runtime > ts->ts_slptime) {
1474 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1477 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1483 * If we have exceeded by more than 1/5th then the algorithm below
1484 * will not bring us back into range. Dividing by two here forces
1485 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1487 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1488 ts->ts_runtime /= 2;
1489 ts->ts_slptime /= 2;
1492 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1493 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1497 * Scale back the interactivity history when a child thread is created. The
1498 * history is inherited from the parent but the thread may behave totally
1499 * differently. For example, a shell spawning a compiler process. We want
1500 * to learn that the compiler is behaving badly very quickly.
1503 sched_interact_fork(struct thread *td)
1508 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1509 if (sum > SCHED_SLP_RUN_FORK) {
1510 ratio = sum / SCHED_SLP_RUN_FORK;
1511 td->td_sched->ts_runtime /= ratio;
1512 td->td_sched->ts_slptime /= ratio;
1517 * Called from proc0_init() to setup the scheduler fields.
1524 * Set up the scheduler specific parts of proc0.
1526 proc0.p_sched = NULL; /* XXX */
1527 thread0.td_sched = &td_sched0;
1528 td_sched0.ts_ltick = ticks;
1529 td_sched0.ts_ftick = ticks;
1530 td_sched0.ts_slice = sched_slice;
1534 * This is only somewhat accurate since given many processes of the same
1535 * priority they will switch when their slices run out, which will be
1536 * at most sched_slice stathz ticks.
1539 sched_rr_interval(void)
1542 /* Convert sched_slice to hz */
1543 return (hz/(realstathz/sched_slice));
1547 * Update the percent cpu tracking information when it is requested or
1548 * the total history exceeds the maximum. We keep a sliding history of
1549 * tick counts that slowly decays. This is less precise than the 4BSD
1550 * mechanism since it happens with less regular and frequent events.
1553 sched_pctcpu_update(struct td_sched *ts)
1556 if (ts->ts_ticks == 0)
1558 if (ticks - (hz / 10) < ts->ts_ltick &&
1559 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1562 * Adjust counters and watermark for pctcpu calc.
1564 if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1565 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1569 ts->ts_ltick = ticks;
1570 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1574 * Adjust the priority of a thread. Move it to the appropriate run-queue
1575 * if necessary. This is the back-end for several priority related
1579 sched_thread_priority(struct thread *td, u_char prio)
1581 struct td_sched *ts;
1585 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1586 "prio:%d", td->td_priority, "new prio:%d", prio,
1587 KTR_ATTR_LINKED, sched_tdname(curthread));
1588 if (td != curthread && prio > td->td_priority) {
1589 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1590 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1591 prio, KTR_ATTR_LINKED, sched_tdname(td));
1594 THREAD_LOCK_ASSERT(td, MA_OWNED);
1595 if (td->td_priority == prio)
1598 * If the priority has been elevated due to priority
1599 * propagation, we may have to move ourselves to a new
1600 * queue. This could be optimized to not re-add in some
1603 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1605 td->td_priority = prio;
1606 sched_add(td, SRQ_BORROWING);
1610 * If the thread is currently running we may have to adjust the lowpri
1611 * information so other cpus are aware of our current priority.
1613 if (TD_IS_RUNNING(td)) {
1614 tdq = TDQ_CPU(ts->ts_cpu);
1615 oldpri = td->td_priority;
1616 td->td_priority = prio;
1617 if (prio < tdq->tdq_lowpri)
1618 tdq->tdq_lowpri = prio;
1619 else if (tdq->tdq_lowpri == oldpri)
1620 tdq_setlowpri(tdq, td);
1623 td->td_priority = prio;
1627 * Update a thread's priority when it is lent another thread's
1631 sched_lend_prio(struct thread *td, u_char prio)
1634 td->td_flags |= TDF_BORROWING;
1635 sched_thread_priority(td, prio);
1639 * Restore a thread's priority when priority propagation is
1640 * over. The prio argument is the minimum priority the thread
1641 * needs to have to satisfy other possible priority lending
1642 * requests. If the thread's regular priority is less
1643 * important than prio, the thread will keep a priority boost
1647 sched_unlend_prio(struct thread *td, u_char prio)
1651 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1652 td->td_base_pri <= PRI_MAX_TIMESHARE)
1653 base_pri = td->td_user_pri;
1655 base_pri = td->td_base_pri;
1656 if (prio >= base_pri) {
1657 td->td_flags &= ~TDF_BORROWING;
1658 sched_thread_priority(td, base_pri);
1660 sched_lend_prio(td, prio);
1664 * Standard entry for setting the priority to an absolute value.
1667 sched_prio(struct thread *td, u_char prio)
1671 /* First, update the base priority. */
1672 td->td_base_pri = prio;
1675 * If the thread is borrowing another thread's priority, don't
1676 * ever lower the priority.
1678 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1681 /* Change the real priority. */
1682 oldprio = td->td_priority;
1683 sched_thread_priority(td, prio);
1686 * If the thread is on a turnstile, then let the turnstile update
1689 if (TD_ON_LOCK(td) && oldprio != prio)
1690 turnstile_adjust(td, oldprio);
1694 * Set the base user priority, does not effect current running priority.
1697 sched_user_prio(struct thread *td, u_char prio)
1700 td->td_base_user_pri = prio;
1701 if (td->td_lend_user_pri <= prio)
1703 td->td_user_pri = prio;
1707 sched_lend_user_prio(struct thread *td, u_char prio)
1710 THREAD_LOCK_ASSERT(td, MA_OWNED);
1711 td->td_lend_user_pri = prio;
1712 td->td_user_pri = min(prio, td->td_base_user_pri);
1713 if (td->td_priority > td->td_user_pri)
1714 sched_prio(td, td->td_user_pri);
1715 else if (td->td_priority != td->td_user_pri)
1716 td->td_flags |= TDF_NEEDRESCHED;
1720 * Handle migration from sched_switch(). This happens only for
1724 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1728 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1730 tdq_load_rem(tdq, td);
1732 * Do the lock dance required to avoid LOR. We grab an extra
1733 * spinlock nesting to prevent preemption while we're
1734 * not holding either run-queue lock.
1737 thread_lock_block(td); /* This releases the lock on tdq. */
1740 * Acquire both run-queue locks before placing the thread on the new
1741 * run-queue to avoid deadlocks created by placing a thread with a
1742 * blocked lock on the run-queue of a remote processor. The deadlock
1743 * occurs when a third processor attempts to lock the two queues in
1744 * question while the target processor is spinning with its own
1745 * run-queue lock held while waiting for the blocked lock to clear.
1747 tdq_lock_pair(tdn, tdq);
1748 tdq_add(tdn, td, flags);
1749 tdq_notify(tdn, td);
1753 return (TDQ_LOCKPTR(tdn));
1757 * Variadic version of thread_lock_unblock() that does not assume td_lock
1761 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1763 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1768 * Switch threads. This function has to handle threads coming in while
1769 * blocked for some reason, running, or idle. It also must deal with
1770 * migrating a thread from one queue to another as running threads may
1771 * be assigned elsewhere via binding.
1774 sched_switch(struct thread *td, struct thread *newtd, int flags)
1777 struct td_sched *ts;
1782 THREAD_LOCK_ASSERT(td, MA_OWNED);
1783 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1785 cpuid = PCPU_GET(cpuid);
1786 tdq = TDQ_CPU(cpuid);
1789 ts->ts_rltick = ticks;
1790 td->td_lastcpu = td->td_oncpu;
1791 td->td_oncpu = NOCPU;
1792 if (!(flags & SW_PREEMPT))
1793 td->td_flags &= ~TDF_NEEDRESCHED;
1794 td->td_owepreempt = 0;
1795 tdq->tdq_switchcnt++;
1797 * The lock pointer in an idle thread should never change. Reset it
1798 * to CAN_RUN as well.
1800 if (TD_IS_IDLETHREAD(td)) {
1801 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1803 } else if (TD_IS_RUNNING(td)) {
1804 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1805 srqflag = (flags & SW_PREEMPT) ?
1806 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1807 SRQ_OURSELF|SRQ_YIELDING;
1809 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1810 ts->ts_cpu = sched_pickcpu(td, 0);
1812 if (ts->ts_cpu == cpuid)
1813 tdq_runq_add(tdq, td, srqflag);
1815 KASSERT(THREAD_CAN_MIGRATE(td) ||
1816 (ts->ts_flags & TSF_BOUND) != 0,
1817 ("Thread %p shouldn't migrate", td));
1818 mtx = sched_switch_migrate(tdq, td, srqflag);
1821 /* This thread must be going to sleep. */
1823 mtx = thread_lock_block(td);
1824 tdq_load_rem(tdq, td);
1827 * We enter here with the thread blocked and assigned to the
1828 * appropriate cpu run-queue or sleep-queue and with the current
1829 * thread-queue locked.
1831 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1832 newtd = choosethread();
1834 * Call the MD code to switch contexts if necessary.
1838 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1839 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1841 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1842 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1844 #ifdef KDTRACE_HOOKS
1846 * If DTrace has set the active vtime enum to anything
1847 * other than INACTIVE (0), then it should have set the
1850 if (dtrace_vtime_active)
1851 (*dtrace_vtime_switch_func)(newtd);
1854 cpu_switch(td, newtd, mtx);
1856 * We may return from cpu_switch on a different cpu. However,
1857 * we always return with td_lock pointing to the current cpu's
1860 cpuid = PCPU_GET(cpuid);
1861 tdq = TDQ_CPU(cpuid);
1862 lock_profile_obtain_lock_success(
1863 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1865 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1866 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1869 thread_unblock_switch(td, mtx);
1871 * Assert that all went well and return.
1873 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1874 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1875 td->td_oncpu = cpuid;
1879 * Adjust thread priorities as a result of a nice request.
1882 sched_nice(struct proc *p, int nice)
1886 PROC_LOCK_ASSERT(p, MA_OWNED);
1889 FOREACH_THREAD_IN_PROC(p, td) {
1892 sched_prio(td, td->td_base_user_pri);
1898 * Record the sleep time for the interactivity scorer.
1901 sched_sleep(struct thread *td, int prio)
1904 THREAD_LOCK_ASSERT(td, MA_OWNED);
1906 td->td_slptick = ticks;
1907 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1908 td->td_flags |= TDF_CANSWAP;
1909 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1911 if (static_boost == 1 && prio)
1912 sched_prio(td, prio);
1913 else if (static_boost && td->td_priority > static_boost)
1914 sched_prio(td, static_boost);
1918 * Schedule a thread to resume execution and record how long it voluntarily
1919 * slept. We also update the pctcpu, interactivity, and priority.
1922 sched_wakeup(struct thread *td)
1924 struct td_sched *ts;
1927 THREAD_LOCK_ASSERT(td, MA_OWNED);
1929 td->td_flags &= ~TDF_CANSWAP;
1931 * If we slept for more than a tick update our interactivity and
1934 slptick = td->td_slptick;
1936 if (slptick && slptick != ticks) {
1939 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1940 ts->ts_slptime += hzticks;
1941 sched_interact_update(td);
1942 sched_pctcpu_update(ts);
1944 /* Reset the slice value after we sleep. */
1945 ts->ts_slice = sched_slice;
1946 sched_add(td, SRQ_BORING);
1950 * Penalize the parent for creating a new child and initialize the child's
1954 sched_fork(struct thread *td, struct thread *child)
1956 THREAD_LOCK_ASSERT(td, MA_OWNED);
1957 sched_fork_thread(td, child);
1959 * Penalize the parent and child for forking.
1961 sched_interact_fork(child);
1962 sched_priority(child);
1963 td->td_sched->ts_runtime += tickincr;
1964 sched_interact_update(td);
1969 * Fork a new thread, may be within the same process.
1972 sched_fork_thread(struct thread *td, struct thread *child)
1974 struct td_sched *ts;
1975 struct td_sched *ts2;
1977 THREAD_LOCK_ASSERT(td, MA_OWNED);
1982 ts2 = child->td_sched;
1983 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1984 child->td_cpuset = cpuset_ref(td->td_cpuset);
1985 ts2->ts_cpu = ts->ts_cpu;
1988 * Grab our parents cpu estimation information.
1990 ts2->ts_ticks = ts->ts_ticks;
1991 ts2->ts_ltick = ts->ts_ltick;
1992 ts2->ts_incrtick = ts->ts_incrtick;
1993 ts2->ts_ftick = ts->ts_ftick;
1995 * Do not inherit any borrowed priority from the parent.
1997 child->td_priority = child->td_base_pri;
1999 * And update interactivity score.
2001 ts2->ts_slptime = ts->ts_slptime;
2002 ts2->ts_runtime = ts->ts_runtime;
2003 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
2005 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2010 * Adjust the priority class of a thread.
2013 sched_class(struct thread *td, int class)
2016 THREAD_LOCK_ASSERT(td, MA_OWNED);
2017 if (td->td_pri_class == class)
2019 td->td_pri_class = class;
2023 * Return some of the child's priority and interactivity to the parent.
2026 sched_exit(struct proc *p, struct thread *child)
2030 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2031 "prio:%d", child->td_priority);
2032 PROC_LOCK_ASSERT(p, MA_OWNED);
2033 td = FIRST_THREAD_IN_PROC(p);
2034 sched_exit_thread(td, child);
2038 * Penalize another thread for the time spent on this one. This helps to
2039 * worsen the priority and interactivity of processes which schedule batch
2040 * jobs such as make. This has little effect on the make process itself but
2041 * causes new processes spawned by it to receive worse scores immediately.
2044 sched_exit_thread(struct thread *td, struct thread *child)
2047 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2048 "prio:%d", child->td_priority);
2050 * Give the child's runtime to the parent without returning the
2051 * sleep time as a penalty to the parent. This causes shells that
2052 * launch expensive things to mark their children as expensive.
2055 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2056 sched_interact_update(td);
2062 sched_preempt(struct thread *td)
2068 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2069 tdq->tdq_ipipending = 0;
2070 if (td->td_priority > tdq->tdq_lowpri) {
2073 flags = SW_INVOL | SW_PREEMPT;
2074 if (td->td_critnest > 1)
2075 td->td_owepreempt = 1;
2076 else if (TD_IS_IDLETHREAD(td))
2077 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2079 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2085 * Fix priorities on return to user-space. Priorities may be elevated due
2086 * to static priorities in msleep() or similar.
2089 sched_userret(struct thread *td)
2092 * XXX we cheat slightly on the locking here to avoid locking in
2093 * the usual case. Setting td_priority here is essentially an
2094 * incomplete workaround for not setting it properly elsewhere.
2095 * Now that some interrupt handlers are threads, not setting it
2096 * properly elsewhere can clobber it in the window between setting
2097 * it here and returning to user mode, so don't waste time setting
2098 * it perfectly here.
2100 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2101 ("thread with borrowed priority returning to userland"));
2102 if (td->td_priority != td->td_user_pri) {
2104 td->td_priority = td->td_user_pri;
2105 td->td_base_pri = td->td_user_pri;
2106 tdq_setlowpri(TDQ_SELF(), td);
2112 * Handle a stathz tick. This is really only relevant for timeshare
2116 sched_clock(struct thread *td)
2119 struct td_sched *ts;
2121 THREAD_LOCK_ASSERT(td, MA_OWNED);
2125 * We run the long term load balancer infrequently on the first cpu.
2127 if (balance_tdq == tdq) {
2128 if (balance_ticks && --balance_ticks == 0)
2133 * Save the old switch count so we have a record of the last ticks
2134 * activity. Initialize the new switch count based on our load.
2135 * If there is some activity seed it to reflect that.
2137 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2138 tdq->tdq_switchcnt = tdq->tdq_load;
2140 * Advance the insert index once for each tick to ensure that all
2141 * threads get a chance to run.
2143 if (tdq->tdq_idx == tdq->tdq_ridx) {
2144 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2145 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2146 tdq->tdq_ridx = tdq->tdq_idx;
2149 if (td->td_pri_class & PRI_FIFO_BIT)
2151 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2153 * We used a tick; charge it to the thread so
2154 * that we can compute our interactivity.
2156 td->td_sched->ts_runtime += tickincr;
2157 sched_interact_update(td);
2161 * We used up one time slice.
2163 if (--ts->ts_slice > 0)
2166 * We're out of time, force a requeue at userret().
2168 ts->ts_slice = sched_slice;
2169 td->td_flags |= TDF_NEEDRESCHED;
2173 * Called once per hz tick. Used for cpu utilization information. This
2174 * is easier than trying to scale based on stathz.
2179 struct td_sched *ts;
2181 ts = curthread->td_sched;
2183 * Ticks is updated asynchronously on a single cpu. Check here to
2184 * avoid incrementing ts_ticks multiple times in a single tick.
2186 if (ts->ts_incrtick == ticks)
2188 /* Adjust ticks for pctcpu */
2189 ts->ts_ticks += cnt << SCHED_TICK_SHIFT;
2190 ts->ts_ltick = ticks;
2191 ts->ts_incrtick = ticks;
2193 * Update if we've exceeded our desired tick threshold by over one
2196 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2197 sched_pctcpu_update(ts);
2201 * Return whether the current CPU has runnable tasks. Used for in-kernel
2202 * cooperative idle threads.
2205 sched_runnable(void)
2213 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2214 if (tdq->tdq_load > 0)
2217 if (tdq->tdq_load - 1 > 0)
2225 * Choose the highest priority thread to run. The thread is removed from
2226 * the run-queue while running however the load remains. For SMP we set
2227 * the tdq in the global idle bitmask if it idles here.
2236 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2237 td = tdq_choose(tdq);
2239 td->td_sched->ts_ltick = ticks;
2240 tdq_runq_rem(tdq, td);
2241 tdq->tdq_lowpri = td->td_priority;
2244 tdq->tdq_lowpri = PRI_MAX_IDLE;
2245 return (PCPU_GET(idlethread));
2249 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2250 * we always request it once we exit a critical section.
2253 sched_setpreempt(struct thread *td)
2259 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2262 pri = td->td_priority;
2263 cpri = ctd->td_priority;
2265 ctd->td_flags |= TDF_NEEDRESCHED;
2266 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2268 if (!sched_shouldpreempt(pri, cpri, 0))
2270 ctd->td_owepreempt = 1;
2274 * Add a thread to a thread queue. Select the appropriate runq and add the
2275 * thread to it. This is the internal function called when the tdq is
2279 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2282 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2283 KASSERT((td->td_inhibitors == 0),
2284 ("sched_add: trying to run inhibited thread"));
2285 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2286 ("sched_add: bad thread state"));
2287 KASSERT(td->td_flags & TDF_INMEM,
2288 ("sched_add: thread swapped out"));
2290 if (td->td_priority < tdq->tdq_lowpri)
2291 tdq->tdq_lowpri = td->td_priority;
2292 tdq_runq_add(tdq, td, flags);
2293 tdq_load_add(tdq, td);
2297 * Select the target thread queue and add a thread to it. Request
2298 * preemption or IPI a remote processor if required.
2301 sched_add(struct thread *td, int flags)
2308 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2309 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2310 sched_tdname(curthread));
2311 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2312 KTR_ATTR_LINKED, sched_tdname(td));
2313 THREAD_LOCK_ASSERT(td, MA_OWNED);
2315 * Recalculate the priority before we select the target cpu or
2318 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2322 * Pick the destination cpu and if it isn't ours transfer to the
2325 cpu = sched_pickcpu(td, flags);
2326 tdq = sched_setcpu(td, cpu, flags);
2327 tdq_add(tdq, td, flags);
2328 if (cpu != PCPU_GET(cpuid)) {
2329 tdq_notify(tdq, td);
2336 * Now that the thread is moving to the run-queue, set the lock
2337 * to the scheduler's lock.
2339 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2340 tdq_add(tdq, td, flags);
2342 if (!(flags & SRQ_YIELDING))
2343 sched_setpreempt(td);
2347 * Remove a thread from a run-queue without running it. This is used
2348 * when we're stealing a thread from a remote queue. Otherwise all threads
2349 * exit by calling sched_exit_thread() and sched_throw() themselves.
2352 sched_rem(struct thread *td)
2356 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2357 "prio:%d", td->td_priority);
2358 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2359 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2360 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2361 KASSERT(TD_ON_RUNQ(td),
2362 ("sched_rem: thread not on run queue"));
2363 tdq_runq_rem(tdq, td);
2364 tdq_load_rem(tdq, td);
2366 if (td->td_priority == tdq->tdq_lowpri)
2367 tdq_setlowpri(tdq, NULL);
2371 * Fetch cpu utilization information. Updates on demand.
2374 sched_pctcpu(struct thread *td)
2377 struct td_sched *ts;
2384 THREAD_LOCK_ASSERT(td, MA_OWNED);
2388 sched_pctcpu_update(ts);
2389 /* How many rtick per second ? */
2390 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2391 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2398 * Enforce affinity settings for a thread. Called after adjustments to
2402 sched_affinity(struct thread *td)
2405 struct td_sched *ts;
2407 THREAD_LOCK_ASSERT(td, MA_OWNED);
2409 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2411 if (TD_ON_RUNQ(td)) {
2413 sched_add(td, SRQ_BORING);
2416 if (!TD_IS_RUNNING(td))
2419 * Force a switch before returning to userspace. If the
2420 * target thread is not running locally send an ipi to force
2423 td->td_flags |= TDF_NEEDRESCHED;
2424 if (td != curthread)
2425 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2430 * Bind a thread to a target cpu.
2433 sched_bind(struct thread *td, int cpu)
2435 struct td_sched *ts;
2437 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2438 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2440 if (ts->ts_flags & TSF_BOUND)
2442 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2443 ts->ts_flags |= TSF_BOUND;
2445 if (PCPU_GET(cpuid) == cpu)
2448 /* When we return from mi_switch we'll be on the correct cpu. */
2449 mi_switch(SW_VOL, NULL);
2453 * Release a bound thread.
2456 sched_unbind(struct thread *td)
2458 struct td_sched *ts;
2460 THREAD_LOCK_ASSERT(td, MA_OWNED);
2461 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2463 if ((ts->ts_flags & TSF_BOUND) == 0)
2465 ts->ts_flags &= ~TSF_BOUND;
2470 sched_is_bound(struct thread *td)
2472 THREAD_LOCK_ASSERT(td, MA_OWNED);
2473 return (td->td_sched->ts_flags & TSF_BOUND);
2480 sched_relinquish(struct thread *td)
2483 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2488 * Return the total system load.
2499 total += TDQ_CPU(i)->tdq_sysload;
2502 return (TDQ_SELF()->tdq_sysload);
2507 sched_sizeof_proc(void)
2509 return (sizeof(struct proc));
2513 sched_sizeof_thread(void)
2515 return (sizeof(struct thread) + sizeof(struct td_sched));
2519 #define TDQ_IDLESPIN(tdq) \
2520 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2522 #define TDQ_IDLESPIN(tdq) 1
2526 * The actual idle process.
2529 sched_idletd(void *dummy)
2536 mtx_assert(&Giant, MA_NOTOWNED);
2541 if (tdq_idled(tdq) == 0)
2544 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2546 * If we're switching very frequently, spin while checking
2547 * for load rather than entering a low power state that
2548 * may require an IPI. However, don't do any busy
2549 * loops while on SMT machines as this simply steals
2550 * cycles from cores doing useful work.
2552 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2553 for (i = 0; i < sched_idlespins; i++) {
2559 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2560 if (tdq->tdq_load == 0) {
2561 tdq->tdq_cpu_idle = 1;
2562 if (tdq->tdq_load == 0) {
2563 cpu_idle(switchcnt > sched_idlespinthresh * 4);
2564 tdq->tdq_switchcnt++;
2566 tdq->tdq_cpu_idle = 0;
2568 if (tdq->tdq_load) {
2570 mi_switch(SW_VOL | SWT_IDLE, NULL);
2577 * A CPU is entering for the first time or a thread is exiting.
2580 sched_throw(struct thread *td)
2582 struct thread *newtd;
2587 /* Correct spinlock nesting and acquire the correct lock. */
2590 PCPU_SET(switchtime, cpu_ticks());
2591 PCPU_SET(switchticks, ticks);
2593 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2594 tdq_load_rem(tdq, td);
2595 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2597 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2598 newtd = choosethread();
2599 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2600 cpu_throw(td, newtd); /* doesn't return */
2604 * This is called from fork_exit(). Just acquire the correct locks and
2605 * let fork do the rest of the work.
2608 sched_fork_exit(struct thread *td)
2610 struct td_sched *ts;
2615 * Finish setting up thread glue so that it begins execution in a
2616 * non-nested critical section with the scheduler lock held.
2618 cpuid = PCPU_GET(cpuid);
2619 tdq = TDQ_CPU(cpuid);
2621 if (TD_IS_IDLETHREAD(td))
2622 td->td_lock = TDQ_LOCKPTR(tdq);
2623 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2624 td->td_oncpu = cpuid;
2625 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2626 lock_profile_obtain_lock_success(
2627 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2631 * Create on first use to catch odd startup conditons.
2634 sched_tdname(struct thread *td)
2637 struct td_sched *ts;
2640 if (ts->ts_name[0] == '\0')
2641 snprintf(ts->ts_name, sizeof(ts->ts_name),
2642 "%s tid %d", td->td_name, td->td_tid);
2643 return (ts->ts_name);
2645 return (td->td_name);
2652 * Build the CPU topology dump string. Is recursively called to collect
2653 * the topology tree.
2656 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2659 char cpusetbuf[CPUSETBUFSIZ];
2662 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2663 "", 1 + indent / 2, cg->cg_level);
2664 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2665 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2667 for (i = 0; i < MAXCPU; i++) {
2668 if (CPU_ISSET(i, &cg->cg_mask)) {
2670 sbuf_printf(sb, ", ");
2673 sbuf_printf(sb, "%d", i);
2676 sbuf_printf(sb, "</cpu>\n");
2678 if (cg->cg_flags != 0) {
2679 sbuf_printf(sb, "%*s <flags>", indent, "");
2680 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2681 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2682 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2683 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2684 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2685 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2686 sbuf_printf(sb, "</flags>\n");
2689 if (cg->cg_children > 0) {
2690 sbuf_printf(sb, "%*s <children>\n", indent, "");
2691 for (i = 0; i < cg->cg_children; i++)
2692 sysctl_kern_sched_topology_spec_internal(sb,
2693 &cg->cg_child[i], indent+2);
2694 sbuf_printf(sb, "%*s </children>\n", indent, "");
2696 sbuf_printf(sb, "%*s</group>\n", indent, "");
2701 * Sysctl handler for retrieving topology dump. It's a wrapper for
2702 * the recursive sysctl_kern_smp_topology_spec_internal().
2705 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2710 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2712 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2716 sbuf_printf(topo, "<groups>\n");
2717 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2718 sbuf_printf(topo, "</groups>\n");
2722 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2730 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2731 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2733 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2734 "Slice size for timeshare threads");
2735 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2736 "Interactivity score threshold");
2737 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2738 0,"Min priority for preemption, lower priorities have greater precedence");
2739 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2740 0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2741 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2742 0,"Number of times idle will spin waiting for new work.");
2743 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2744 0,"Threshold before we will permit idle spinning.");
2746 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2747 "Number of hz ticks to keep thread affinity for");
2748 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2749 "Enables the long-term load balancer");
2750 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2751 &balance_interval, 0,
2752 "Average frequency in stathz ticks to run the long-term balancer");
2753 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2754 "Steals work from another hyper-threaded core on idle");
2755 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2756 "Attempts to steal work from other cores before idling");
2757 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2758 "Minimum load on remote cpu before we'll steal");
2760 /* Retrieve SMP topology */
2761 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2762 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2763 "XML dump of detected CPU topology");
2767 /* ps compat. All cpu percentages from ULE are weighted. */
2768 static int ccpu = 0;
2769 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");