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/ktrace.h>
71 #include <sys/pmckern.h>
75 #include <sys/dtrace_bsd.h>
76 int dtrace_vtime_active;
77 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
80 #include <machine/cpu.h>
81 #include <machine/smp.h>
83 #if defined(__sparc64__) || defined(__mips__)
84 #error "This architecture is not currently compatible with ULE"
89 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
90 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
91 #define TDQ_LOADNAME_LEN (PCPU_NAME_LEN + sizeof(" load"))
94 * Thread scheduler specific section. All fields are protected
98 struct runq *ts_runq; /* Run-queue we're queued on. */
99 short ts_flags; /* TSF_* flags. */
100 u_char ts_cpu; /* CPU that we have affinity for. */
101 int ts_rltick; /* Real last tick, for affinity. */
102 int ts_slice; /* Ticks of slice remaining. */
103 u_int ts_slptime; /* Number of ticks we vol. slept */
104 u_int ts_runtime; /* Number of ticks we were running */
105 int ts_ltick; /* Last tick that we were running on */
106 int ts_incrtick; /* Last tick that we incremented on */
107 int ts_ftick; /* First tick that we were running on */
108 int ts_ticks; /* Tick count */
110 char ts_name[TS_NAME_LEN];
113 /* flags kept in ts_flags */
114 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
115 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
117 static struct td_sched td_sched0;
119 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
120 #define THREAD_CAN_SCHED(td, cpu) \
121 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
124 * Cpu percentage computation macros and defines.
126 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
127 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
128 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
129 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
130 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
131 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
133 #define SCHED_TICK_SECS 10
134 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
135 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
136 #define SCHED_TICK_SHIFT 10
137 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
138 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
141 * These macros determine priorities for non-interactive threads. They are
142 * assigned a priority based on their recent cpu utilization as expressed
143 * by the ratio of ticks to the tick total. NHALF priorities at the start
144 * and end of the MIN to MAX timeshare range are only reachable with negative
145 * or positive nice respectively.
147 * PRI_RANGE: Priority range for utilization dependent priorities.
148 * PRI_NRESV: Number of nice values.
149 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
150 * PRI_NICE: Determines the part of the priority inherited from nice.
152 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
153 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
154 #define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
155 #define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
156 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN)
157 #define SCHED_PRI_TICKS(ts) \
158 (SCHED_TICK_HZ((ts)) / \
159 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
160 #define SCHED_PRI_NICE(nice) (nice)
163 * These determine the interactivity of a process. Interactivity differs from
164 * cpu utilization in that it expresses the voluntary time slept vs time ran
165 * while cpu utilization includes all time not running. This more accurately
166 * models the intent of the thread.
168 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
169 * before throttling back.
170 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
171 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
172 * INTERACT_THRESH: Threshhold for placement on the current runq.
174 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
175 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
176 #define SCHED_INTERACT_MAX (100)
177 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
178 #define SCHED_INTERACT_THRESH (30)
181 * tickincr: Converts a stathz tick into a hz domain scaled by
182 * the shift factor. Without the shift the error rate
183 * due to rounding would be unacceptably high.
184 * realstathz: stathz is sometimes 0 and run off of hz.
185 * sched_slice: Runtime of each thread before rescheduling.
186 * preempt_thresh: Priority threshold for preemption and remote IPIs.
188 static int sched_interact = SCHED_INTERACT_THRESH;
189 static int realstathz;
191 static int sched_slice = 1;
193 #ifdef FULL_PREEMPTION
194 static int preempt_thresh = PRI_MAX_IDLE;
196 static int preempt_thresh = PRI_MIN_KERN;
199 static int preempt_thresh = 0;
201 static int static_boost = PRI_MIN_TIMESHARE;
202 static int sched_idlespins = 10000;
203 static int sched_idlespinthresh = 4;
206 * tdq - per processor runqs and statistics. All fields are protected by the
207 * tdq_lock. The load and lowpri may be accessed without to avoid excess
208 * locking in sched_pickcpu();
211 /* Ordered to improve efficiency of cpu_search() and switch(). */
212 struct mtx tdq_lock; /* run queue lock. */
213 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
214 volatile int tdq_load; /* Aggregate load. */
215 int tdq_sysload; /* For loadavg, !ITHD load. */
216 int tdq_transferable; /* Transferable thread count. */
217 short tdq_switchcnt; /* Switches this tick. */
218 short tdq_oldswitchcnt; /* Switches last tick. */
219 u_char tdq_lowpri; /* Lowest priority thread. */
220 u_char tdq_ipipending; /* IPI pending. */
221 u_char tdq_idx; /* Current insert index. */
222 u_char tdq_ridx; /* Current removal index. */
223 struct runq tdq_realtime; /* real-time run queue. */
224 struct runq tdq_timeshare; /* timeshare run queue. */
225 struct runq tdq_idle; /* Queue of IDLE threads. */
226 char tdq_name[TDQ_NAME_LEN];
228 char tdq_loadname[TDQ_LOADNAME_LEN];
232 /* Idle thread states and config. */
233 #define TDQ_RUNNING 1
237 struct cpu_group *cpu_top; /* CPU topology */
239 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
240 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
245 static int rebalance = 1;
246 static int balance_interval = 128; /* Default set in sched_initticks(). */
248 static int steal_htt = 1;
249 static int steal_idle = 1;
250 static int steal_thresh = 2;
253 * One thread queue per processor.
255 static struct tdq tdq_cpu[MAXCPU];
256 static struct tdq *balance_tdq;
257 static int balance_ticks;
259 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
260 #define TDQ_CPU(x) (&tdq_cpu[(x)])
261 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
263 static struct tdq tdq_cpu;
265 #define TDQ_ID(x) (0)
266 #define TDQ_SELF() (&tdq_cpu)
267 #define TDQ_CPU(x) (&tdq_cpu)
270 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
271 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
272 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
273 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
274 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
276 static void sched_priority(struct thread *);
277 static void sched_thread_priority(struct thread *, u_char);
278 static int sched_interact_score(struct thread *);
279 static void sched_interact_update(struct thread *);
280 static void sched_interact_fork(struct thread *);
281 static void sched_pctcpu_update(struct td_sched *);
283 /* Operations on per processor queues */
284 static struct thread *tdq_choose(struct tdq *);
285 static void tdq_setup(struct tdq *);
286 static void tdq_load_add(struct tdq *, struct thread *);
287 static void tdq_load_rem(struct tdq *, struct thread *);
288 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
289 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
290 static inline int sched_shouldpreempt(int, int, int);
291 void tdq_print(int cpu);
292 static void runq_print(struct runq *rq);
293 static void tdq_add(struct tdq *, struct thread *, int);
295 static int tdq_move(struct tdq *, struct tdq *);
296 static int tdq_idled(struct tdq *);
297 static void tdq_notify(struct tdq *, struct thread *);
298 static struct thread *tdq_steal(struct tdq *, int);
299 static struct thread *runq_steal(struct runq *, int);
300 static int sched_pickcpu(struct thread *, int);
301 static void sched_balance(void);
302 static int sched_balance_pair(struct tdq *, struct tdq *);
303 static inline struct tdq *sched_setcpu(struct thread *, int, int);
304 static inline struct mtx *thread_block_switch(struct thread *);
305 static inline void thread_unblock_switch(struct thread *, struct mtx *);
306 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
307 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
308 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
309 struct cpu_group *cg, int indent);
312 static void sched_setup(void *dummy);
313 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
315 static void sched_initticks(void *dummy);
316 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
320 * Print the threads waiting on a run-queue.
323 runq_print(struct runq *rq)
331 for (i = 0; i < RQB_LEN; i++) {
332 printf("\t\trunq bits %d 0x%zx\n",
333 i, rq->rq_status.rqb_bits[i]);
334 for (j = 0; j < RQB_BPW; j++)
335 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
336 pri = j + (i << RQB_L2BPW);
337 rqh = &rq->rq_queues[pri];
338 TAILQ_FOREACH(td, rqh, td_runq) {
339 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
340 td, td->td_name, td->td_priority,
341 td->td_rqindex, pri);
348 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
357 printf("tdq %d:\n", TDQ_ID(tdq));
358 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
359 printf("\tLock name: %s\n", tdq->tdq_name);
360 printf("\tload: %d\n", tdq->tdq_load);
361 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
362 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
363 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
364 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
365 printf("\tload transferable: %d\n", tdq->tdq_transferable);
366 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
367 printf("\trealtime runq:\n");
368 runq_print(&tdq->tdq_realtime);
369 printf("\ttimeshare runq:\n");
370 runq_print(&tdq->tdq_timeshare);
371 printf("\tidle runq:\n");
372 runq_print(&tdq->tdq_idle);
376 sched_shouldpreempt(int pri, int cpri, int remote)
379 * If the new priority is not better than the current priority there is
385 * Always preempt idle.
387 if (cpri >= PRI_MIN_IDLE)
390 * If preemption is disabled don't preempt others.
392 if (preempt_thresh == 0)
395 * Preempt if we exceed the threshold.
397 if (pri <= preempt_thresh)
400 * If we're realtime or better and there is timeshare or worse running
401 * preempt only remote processors.
403 if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
408 #define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
410 * Add a thread to the actual run-queue. Keeps transferable counts up to
411 * date with what is actually on the run-queue. Selects the correct
412 * queue position for timeshare threads.
415 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
420 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
421 THREAD_LOCK_ASSERT(td, MA_OWNED);
423 pri = td->td_priority;
426 if (THREAD_CAN_MIGRATE(td)) {
427 tdq->tdq_transferable++;
428 ts->ts_flags |= TSF_XFERABLE;
430 if (pri <= PRI_MAX_REALTIME) {
431 ts->ts_runq = &tdq->tdq_realtime;
432 } else if (pri <= PRI_MAX_TIMESHARE) {
433 ts->ts_runq = &tdq->tdq_timeshare;
434 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
435 ("Invalid priority %d on timeshare runq", pri));
437 * This queue contains only priorities between MIN and MAX
438 * realtime. Use the whole queue to represent these values.
440 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
441 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
442 pri = (pri + tdq->tdq_idx) % RQ_NQS;
444 * This effectively shortens the queue by one so we
445 * can have a one slot difference between idx and
446 * ridx while we wait for threads to drain.
448 if (tdq->tdq_ridx != tdq->tdq_idx &&
449 pri == tdq->tdq_ridx)
450 pri = (unsigned char)(pri - 1) % RQ_NQS;
453 runq_add_pri(ts->ts_runq, td, pri, flags);
456 ts->ts_runq = &tdq->tdq_idle;
457 runq_add(ts->ts_runq, td, flags);
461 * Remove a thread from a run-queue. This typically happens when a thread
462 * is selected to run. Running threads are not on the queue and the
463 * transferable count does not reflect them.
466 tdq_runq_rem(struct tdq *tdq, struct thread *td)
471 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
472 KASSERT(ts->ts_runq != NULL,
473 ("tdq_runq_remove: thread %p null ts_runq", td));
474 if (ts->ts_flags & TSF_XFERABLE) {
475 tdq->tdq_transferable--;
476 ts->ts_flags &= ~TSF_XFERABLE;
478 if (ts->ts_runq == &tdq->tdq_timeshare) {
479 if (tdq->tdq_idx != tdq->tdq_ridx)
480 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
482 runq_remove_idx(ts->ts_runq, td, NULL);
484 runq_remove(ts->ts_runq, td);
488 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
489 * for this thread to the referenced thread queue.
492 tdq_load_add(struct tdq *tdq, struct thread *td)
495 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
496 THREAD_LOCK_ASSERT(td, MA_OWNED);
499 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
501 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
505 * Remove the load from a thread that is transitioning to a sleep state or
509 tdq_load_rem(struct tdq *tdq, struct thread *td)
512 THREAD_LOCK_ASSERT(td, MA_OWNED);
513 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
514 KASSERT(tdq->tdq_load != 0,
515 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
518 if ((td->td_proc->p_flag & P_NOLOAD) == 0)
520 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
524 * Set lowpri to its exact value by searching the run-queue and
525 * evaluating curthread. curthread may be passed as an optimization.
528 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
532 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
534 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
535 td = tdq_choose(tdq);
536 if (td == NULL || td->td_priority > ctd->td_priority)
537 tdq->tdq_lowpri = ctd->td_priority;
539 tdq->tdq_lowpri = td->td_priority;
547 int cs_limit; /* Min priority for low min load for high. */
550 #define CPU_SEARCH_LOWEST 0x1
551 #define CPU_SEARCH_HIGHEST 0x2
552 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
554 #define CPUSET_FOREACH(cpu, mask) \
555 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
556 if ((mask) & 1 << (cpu))
558 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
559 struct cpu_search *high, const int match);
560 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
561 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
562 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
563 struct cpu_search *high);
566 * This routine compares according to the match argument and should be
567 * reduced in actual instantiations via constant propagation and dead code
571 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
577 if (match & CPU_SEARCH_LOWEST)
578 if (CPU_ISSET(cpu, &low->cs_mask) &&
579 tdq->tdq_load < low->cs_load &&
580 tdq->tdq_lowpri > low->cs_limit) {
582 low->cs_load = tdq->tdq_load;
584 if (match & CPU_SEARCH_HIGHEST)
585 if (CPU_ISSET(cpu, &high->cs_mask) &&
586 tdq->tdq_load >= high->cs_limit &&
587 tdq->tdq_load > high->cs_load &&
588 tdq->tdq_transferable) {
590 high->cs_load = tdq->tdq_load;
592 return (tdq->tdq_load);
596 * Search the tree of cpu_groups for the lowest or highest loaded cpu
597 * according to the match argument. This routine actually compares the
598 * load on all paths through the tree and finds the least loaded cpu on
599 * the least loaded path, which may differ from the least loaded cpu in
600 * the system. This balances work among caches and busses.
602 * This inline is instantiated in three forms below using constants for the
603 * match argument. It is reduced to the minimum set for each case. It is
604 * also recursive to the depth of the tree.
607 cpu_search(struct cpu_group *cg, struct cpu_search *low,
608 struct cpu_search *high, const int match)
613 if (cg->cg_children) {
614 struct cpu_search lgroup;
615 struct cpu_search hgroup;
616 struct cpu_group *child;
624 for (i = 0; i < cg->cg_children; i++) {
625 child = &cg->cg_child[i];
626 if (match & CPU_SEARCH_LOWEST) {
630 if (match & CPU_SEARCH_HIGHEST) {
635 case CPU_SEARCH_LOWEST:
636 load = cpu_search_lowest(child, &lgroup);
638 case CPU_SEARCH_HIGHEST:
639 load = cpu_search_highest(child, &hgroup);
641 case CPU_SEARCH_BOTH:
642 load = cpu_search_both(child, &lgroup, &hgroup);
646 if (match & CPU_SEARCH_LOWEST)
647 if (load < lload || low->cs_cpu == -1) {
651 if (match & CPU_SEARCH_HIGHEST)
652 if (load > hload || high->cs_cpu == -1) {
660 CPUSET_FOREACH(cpu, cg->cg_mask)
661 total += cpu_compare(cpu, low, high, match);
667 * cpu_search instantiations must pass constants to maintain the inline
671 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
673 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
677 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
679 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
683 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
684 struct cpu_search *high)
686 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
690 * Find the cpu with the least load via the least loaded path that has a
691 * lowpri greater than pri pri. A pri of -1 indicates any priority is
695 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
697 struct cpu_search low;
703 cpu_search_lowest(cg, &low);
708 * Find the cpu with the highest load via the highest loaded path.
711 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
713 struct cpu_search high;
718 high.cs_limit = minload;
719 cpu_search_highest(cg, &high);
724 * Simultaneously find the highest and lowest loaded cpu reachable via
728 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
730 struct cpu_search high;
731 struct cpu_search low;
741 cpu_search_both(cg, &low, &high);
742 *lowcpu = low.cs_cpu;
743 *highcpu = high.cs_cpu;
748 sched_balance_group(struct cpu_group *cg)
757 sched_both(cg, mask, &low, &high);
758 if (low == high || low == -1 || high == -1)
760 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
763 * If we failed to move any threads determine which cpu
764 * to kick out of the set and try again.
766 if (TDQ_CPU(high)->tdq_transferable == 0)
767 CPU_CLR(high, &mask);
772 for (i = 0; i < cg->cg_children; i++)
773 sched_balance_group(&cg->cg_child[i]);
782 * Select a random time between .5 * balance_interval and
783 * 1.5 * balance_interval.
785 balance_ticks = max(balance_interval / 2, 1);
786 balance_ticks += random() % balance_interval;
787 if (smp_started == 0 || rebalance == 0)
791 sched_balance_group(cpu_top);
796 * Lock two thread queues using their address to maintain lock order.
799 tdq_lock_pair(struct tdq *one, struct tdq *two)
803 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
806 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
811 * Unlock two thread queues. Order is not important here.
814 tdq_unlock_pair(struct tdq *one, struct tdq *two)
821 * Transfer load between two imbalanced thread queues.
824 sched_balance_pair(struct tdq *high, struct tdq *low)
834 tdq_lock_pair(high, low);
835 transferable = high->tdq_transferable;
836 high_load = high->tdq_load;
837 low_load = low->tdq_load;
840 * Determine what the imbalance is and then adjust that to how many
841 * threads we actually have to give up (transferable).
843 if (transferable != 0) {
844 diff = high_load - low_load;
848 move = min(move, transferable);
849 for (i = 0; i < move; i++)
850 moved += tdq_move(high, low);
852 * IPI the target cpu to force it to reschedule with the new
855 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT);
857 tdq_unlock_pair(high, low);
862 * Move a thread from one thread queue to another.
865 tdq_move(struct tdq *from, struct tdq *to)
872 TDQ_LOCK_ASSERT(from, MA_OWNED);
873 TDQ_LOCK_ASSERT(to, MA_OWNED);
877 td = tdq_steal(tdq, cpu);
882 * Although the run queue is locked the thread may be blocked. Lock
883 * it to clear this and acquire the run-queue lock.
886 /* Drop recursive lock on from acquired via thread_lock(). */
890 td->td_lock = TDQ_LOCKPTR(to);
891 tdq_add(to, td, SRQ_YIELDING);
896 * This tdq has idled. Try to steal a thread from another cpu and switch
900 tdq_idled(struct tdq *tdq)
902 struct cpu_group *cg;
908 if (smp_started == 0 || steal_idle == 0)
911 CPU_CLR(PCPU_GET(cpuid), &mask);
912 /* We don't want to be preempted while we're iterating. */
914 for (cg = tdq->tdq_cg; cg != NULL; ) {
915 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
916 thresh = steal_thresh;
919 cpu = sched_highest(cg, mask, thresh);
924 steal = TDQ_CPU(cpu);
926 tdq_lock_pair(tdq, steal);
927 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
928 tdq_unlock_pair(tdq, steal);
932 * If a thread was added while interrupts were disabled don't
933 * steal one here. If we fail to acquire one due to affinity
934 * restrictions loop again with this cpu removed from the
937 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
938 tdq_unlock_pair(tdq, steal);
943 mi_switch(SW_VOL | SWT_IDLE, NULL);
944 thread_unlock(curthread);
953 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
956 tdq_notify(struct tdq *tdq, struct thread *td)
962 if (tdq->tdq_ipipending)
964 cpu = td->td_sched->ts_cpu;
965 pri = td->td_priority;
966 ctd = pcpu_find(cpu)->pc_curthread;
967 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
969 if (TD_IS_IDLETHREAD(ctd)) {
971 * If the MD code has an idle wakeup routine try that before
972 * falling back to IPI.
974 if (cpu_idle_wakeup(cpu))
977 tdq->tdq_ipipending = 1;
978 ipi_selected(1 << cpu, IPI_PREEMPT);
982 * Steals load from a timeshare queue. Honors the rotating queue head
985 static struct thread *
986 runq_steal_from(struct runq *rq, int cpu, u_char start)
996 rqb = &rq->rq_status;
997 bit = start & (RQB_BPW -1);
1001 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1002 if (rqb->rqb_bits[i] == 0)
1005 for (pri = bit; pri < RQB_BPW; pri++)
1006 if (rqb->rqb_bits[i] & (1ul << pri))
1011 pri = RQB_FFS(rqb->rqb_bits[i]);
1012 pri += (i << RQB_L2BPW);
1013 rqh = &rq->rq_queues[pri];
1014 TAILQ_FOREACH(td, rqh, td_runq) {
1015 if (first && THREAD_CAN_MIGRATE(td) &&
1016 THREAD_CAN_SCHED(td, cpu))
1030 * Steals load from a standard linear queue.
1032 static struct thread *
1033 runq_steal(struct runq *rq, int cpu)
1041 rqb = &rq->rq_status;
1042 for (word = 0; word < RQB_LEN; word++) {
1043 if (rqb->rqb_bits[word] == 0)
1045 for (bit = 0; bit < RQB_BPW; bit++) {
1046 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1048 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1049 TAILQ_FOREACH(td, rqh, td_runq)
1050 if (THREAD_CAN_MIGRATE(td) &&
1051 THREAD_CAN_SCHED(td, cpu))
1059 * Attempt to steal a thread in priority order from a thread queue.
1061 static struct thread *
1062 tdq_steal(struct tdq *tdq, int cpu)
1066 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1067 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1069 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1070 cpu, tdq->tdq_ridx)) != NULL)
1072 return (runq_steal(&tdq->tdq_idle, cpu));
1076 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1077 * current lock and returns with the assigned queue locked.
1079 static inline struct tdq *
1080 sched_setcpu(struct thread *td, int cpu, int flags)
1085 THREAD_LOCK_ASSERT(td, MA_OWNED);
1087 td->td_sched->ts_cpu = cpu;
1089 * If the lock matches just return the queue.
1091 if (td->td_lock == TDQ_LOCKPTR(tdq))
1095 * If the thread isn't running its lockptr is a
1096 * turnstile or a sleepqueue. We can just lock_set without
1099 if (TD_CAN_RUN(td)) {
1101 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1106 * The hard case, migration, we need to block the thread first to
1107 * prevent order reversals with other cpus locks.
1109 thread_lock_block(td);
1111 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1115 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1116 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1117 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1118 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1119 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1120 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1123 sched_pickcpu(struct thread *td, int flags)
1125 struct cpu_group *cg;
1126 struct td_sched *ts;
1133 self = PCPU_GET(cpuid);
1135 if (smp_started == 0)
1138 * Don't migrate a running thread from sched_switch().
1140 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1141 return (ts->ts_cpu);
1143 * Prefer to run interrupt threads on the processors that generate
1146 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1147 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1148 SCHED_STAT_INC(pickcpu_intrbind);
1152 * If the thread can run on the last cpu and the affinity has not
1153 * expired or it is idle run it there.
1155 pri = td->td_priority;
1156 tdq = TDQ_CPU(ts->ts_cpu);
1157 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1158 if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1159 SCHED_STAT_INC(pickcpu_idle_affinity);
1160 return (ts->ts_cpu);
1162 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1163 SCHED_STAT_INC(pickcpu_affinity);
1164 return (ts->ts_cpu);
1168 * Search for the highest level in the tree that still has affinity.
1171 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1172 if (SCHED_AFFINITY(ts, cg->cg_level))
1175 mask = td->td_cpuset->cs_mask;
1177 cpu = sched_lowest(cg, mask, pri);
1179 cpu = sched_lowest(cpu_top, mask, -1);
1181 * Compare the lowest loaded cpu to current cpu.
1183 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1184 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1185 SCHED_STAT_INC(pickcpu_local);
1188 SCHED_STAT_INC(pickcpu_lowest);
1189 if (cpu != ts->ts_cpu)
1190 SCHED_STAT_INC(pickcpu_migration);
1191 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1197 * Pick the highest priority task we have and return it.
1199 static struct thread *
1200 tdq_choose(struct tdq *tdq)
1204 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1205 td = runq_choose(&tdq->tdq_realtime);
1208 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1210 KASSERT(td->td_priority >= PRI_MIN_TIMESHARE,
1211 ("tdq_choose: Invalid priority on timeshare queue %d",
1215 td = runq_choose(&tdq->tdq_idle);
1217 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1218 ("tdq_choose: Invalid priority on idle queue %d",
1227 * Initialize a thread queue.
1230 tdq_setup(struct tdq *tdq)
1234 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1235 runq_init(&tdq->tdq_realtime);
1236 runq_init(&tdq->tdq_timeshare);
1237 runq_init(&tdq->tdq_idle);
1238 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1239 "sched lock %d", (int)TDQ_ID(tdq));
1240 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1241 MTX_SPIN | MTX_RECURSE);
1243 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1244 "CPU %d load", (int)TDQ_ID(tdq));
1250 sched_setup_smp(void)
1255 cpu_top = smp_topo();
1256 for (i = 0; i < MAXCPU; i++) {
1261 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1262 if (tdq->tdq_cg == NULL)
1263 panic("Can't find cpu group for %d\n", i);
1265 balance_tdq = TDQ_SELF();
1271 * Setup the thread queues and initialize the topology based on MD
1275 sched_setup(void *dummy)
1286 * To avoid divide-by-zero, we set realstathz a dummy value
1287 * in case which sched_clock() called before sched_initticks().
1290 sched_slice = (realstathz/10); /* ~100ms */
1291 tickincr = 1 << SCHED_TICK_SHIFT;
1293 /* Add thread0's load since it's running. */
1295 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1296 tdq_load_add(tdq, &thread0);
1297 tdq->tdq_lowpri = thread0.td_priority;
1302 * This routine determines the tickincr after stathz and hz are setup.
1306 sched_initticks(void *dummy)
1310 realstathz = stathz ? stathz : hz;
1311 sched_slice = (realstathz/10); /* ~100ms */
1314 * tickincr is shifted out by 10 to avoid rounding errors due to
1315 * hz not being evenly divisible by stathz on all platforms.
1317 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1319 * This does not work for values of stathz that are more than
1320 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1327 * Set the default balance interval now that we know
1328 * what realstathz is.
1330 balance_interval = realstathz;
1332 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1333 * This prevents excess thrashing on large machines and excess idle
1334 * on smaller machines.
1336 steal_thresh = min(fls(mp_ncpus) - 1, 3);
1337 affinity = SCHED_AFFINITY_DEFAULT;
1343 * This is the core of the interactivity algorithm. Determines a score based
1344 * on past behavior. It is the ratio of sleep time to run time scaled to
1345 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1346 * differs from the cpu usage because it does not account for time spent
1347 * waiting on a run-queue. Would be prettier if we had floating point.
1350 sched_interact_score(struct thread *td)
1352 struct td_sched *ts;
1357 * The score is only needed if this is likely to be an interactive
1358 * task. Don't go through the expense of computing it if there's
1361 if (sched_interact <= SCHED_INTERACT_HALF &&
1362 ts->ts_runtime >= ts->ts_slptime)
1363 return (SCHED_INTERACT_HALF);
1365 if (ts->ts_runtime > ts->ts_slptime) {
1366 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1367 return (SCHED_INTERACT_HALF +
1368 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1370 if (ts->ts_slptime > ts->ts_runtime) {
1371 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1372 return (ts->ts_runtime / div);
1374 /* runtime == slptime */
1376 return (SCHED_INTERACT_HALF);
1379 * This can happen if slptime and runtime are 0.
1386 * Scale the scheduling priority according to the "interactivity" of this
1390 sched_priority(struct thread *td)
1395 if (td->td_pri_class != PRI_TIMESHARE)
1398 * If the score is interactive we place the thread in the realtime
1399 * queue with a priority that is less than kernel and interrupt
1400 * priorities. These threads are not subject to nice restrictions.
1402 * Scores greater than this are placed on the normal timeshare queue
1403 * where the priority is partially decided by the most recent cpu
1404 * utilization and the rest is decided by nice value.
1406 * The nice value of the process has a linear effect on the calculated
1407 * score. Negative nice values make it easier for a thread to be
1408 * considered interactive.
1410 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1411 if (score < sched_interact) {
1412 pri = PRI_MIN_REALTIME;
1413 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1415 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1416 ("sched_priority: invalid interactive priority %d score %d",
1419 pri = SCHED_PRI_MIN;
1420 if (td->td_sched->ts_ticks)
1421 pri += SCHED_PRI_TICKS(td->td_sched);
1422 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1423 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1424 ("sched_priority: invalid priority %d: nice %d, "
1425 "ticks %d ftick %d ltick %d tick pri %d",
1426 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1427 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1428 SCHED_PRI_TICKS(td->td_sched)));
1430 sched_user_prio(td, pri);
1436 * This routine enforces a maximum limit on the amount of scheduling history
1437 * kept. It is called after either the slptime or runtime is adjusted. This
1438 * function is ugly due to integer math.
1441 sched_interact_update(struct thread *td)
1443 struct td_sched *ts;
1447 sum = ts->ts_runtime + ts->ts_slptime;
1448 if (sum < SCHED_SLP_RUN_MAX)
1451 * This only happens from two places:
1452 * 1) We have added an unusual amount of run time from fork_exit.
1453 * 2) We have added an unusual amount of sleep time from sched_sleep().
1455 if (sum > SCHED_SLP_RUN_MAX * 2) {
1456 if (ts->ts_runtime > ts->ts_slptime) {
1457 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1460 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1466 * If we have exceeded by more than 1/5th then the algorithm below
1467 * will not bring us back into range. Dividing by two here forces
1468 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1470 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1471 ts->ts_runtime /= 2;
1472 ts->ts_slptime /= 2;
1475 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1476 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1480 * Scale back the interactivity history when a child thread is created. The
1481 * history is inherited from the parent but the thread may behave totally
1482 * differently. For example, a shell spawning a compiler process. We want
1483 * to learn that the compiler is behaving badly very quickly.
1486 sched_interact_fork(struct thread *td)
1491 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1492 if (sum > SCHED_SLP_RUN_FORK) {
1493 ratio = sum / SCHED_SLP_RUN_FORK;
1494 td->td_sched->ts_runtime /= ratio;
1495 td->td_sched->ts_slptime /= ratio;
1500 * Called from proc0_init() to setup the scheduler fields.
1507 * Set up the scheduler specific parts of proc0.
1509 proc0.p_sched = NULL; /* XXX */
1510 thread0.td_sched = &td_sched0;
1511 td_sched0.ts_ltick = ticks;
1512 td_sched0.ts_ftick = ticks;
1513 td_sched0.ts_slice = sched_slice;
1517 * This is only somewhat accurate since given many processes of the same
1518 * priority they will switch when their slices run out, which will be
1519 * at most sched_slice stathz ticks.
1522 sched_rr_interval(void)
1525 /* Convert sched_slice to hz */
1526 return (hz/(realstathz/sched_slice));
1530 * Update the percent cpu tracking information when it is requested or
1531 * the total history exceeds the maximum. We keep a sliding history of
1532 * tick counts that slowly decays. This is less precise than the 4BSD
1533 * mechanism since it happens with less regular and frequent events.
1536 sched_pctcpu_update(struct td_sched *ts)
1539 if (ts->ts_ticks == 0)
1541 if (ticks - (hz / 10) < ts->ts_ltick &&
1542 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1545 * Adjust counters and watermark for pctcpu calc.
1547 if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1548 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1552 ts->ts_ltick = ticks;
1553 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1557 * Adjust the priority of a thread. Move it to the appropriate run-queue
1558 * if necessary. This is the back-end for several priority related
1562 sched_thread_priority(struct thread *td, u_char prio)
1564 struct td_sched *ts;
1568 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1569 "prio:%d", td->td_priority, "new prio:%d", prio,
1570 KTR_ATTR_LINKED, sched_tdname(curthread));
1571 if (td != curthread && prio > td->td_priority) {
1572 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1573 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1574 prio, KTR_ATTR_LINKED, sched_tdname(td));
1577 THREAD_LOCK_ASSERT(td, MA_OWNED);
1578 if (td->td_priority == prio)
1581 * If the priority has been elevated due to priority
1582 * propagation, we may have to move ourselves to a new
1583 * queue. This could be optimized to not re-add in some
1586 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1588 td->td_priority = prio;
1589 sched_add(td, SRQ_BORROWING);
1593 * If the thread is currently running we may have to adjust the lowpri
1594 * information so other cpus are aware of our current priority.
1596 if (TD_IS_RUNNING(td)) {
1597 tdq = TDQ_CPU(ts->ts_cpu);
1598 oldpri = td->td_priority;
1599 td->td_priority = prio;
1600 if (prio < tdq->tdq_lowpri)
1601 tdq->tdq_lowpri = prio;
1602 else if (tdq->tdq_lowpri == oldpri)
1603 tdq_setlowpri(tdq, td);
1606 td->td_priority = prio;
1610 * Update a thread's priority when it is lent another thread's
1614 sched_lend_prio(struct thread *td, u_char prio)
1617 td->td_flags |= TDF_BORROWING;
1618 sched_thread_priority(td, prio);
1622 * Restore a thread's priority when priority propagation is
1623 * over. The prio argument is the minimum priority the thread
1624 * needs to have to satisfy other possible priority lending
1625 * requests. If the thread's regular priority is less
1626 * important than prio, the thread will keep a priority boost
1630 sched_unlend_prio(struct thread *td, u_char prio)
1634 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1635 td->td_base_pri <= PRI_MAX_TIMESHARE)
1636 base_pri = td->td_user_pri;
1638 base_pri = td->td_base_pri;
1639 if (prio >= base_pri) {
1640 td->td_flags &= ~TDF_BORROWING;
1641 sched_thread_priority(td, base_pri);
1643 sched_lend_prio(td, prio);
1647 * Standard entry for setting the priority to an absolute value.
1650 sched_prio(struct thread *td, u_char prio)
1654 /* First, update the base priority. */
1655 td->td_base_pri = prio;
1658 * If the thread is borrowing another thread's priority, don't
1659 * ever lower the priority.
1661 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1664 /* Change the real priority. */
1665 oldprio = td->td_priority;
1666 sched_thread_priority(td, prio);
1669 * If the thread is on a turnstile, then let the turnstile update
1672 if (TD_ON_LOCK(td) && oldprio != prio)
1673 turnstile_adjust(td, oldprio);
1677 * Set the base user priority, does not effect current running priority.
1680 sched_user_prio(struct thread *td, u_char prio)
1684 td->td_base_user_pri = prio;
1685 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1687 oldprio = td->td_user_pri;
1688 td->td_user_pri = prio;
1692 sched_lend_user_prio(struct thread *td, u_char prio)
1696 THREAD_LOCK_ASSERT(td, MA_OWNED);
1697 td->td_flags |= TDF_UBORROWING;
1698 oldprio = td->td_user_pri;
1699 td->td_user_pri = prio;
1703 sched_unlend_user_prio(struct thread *td, u_char prio)
1707 THREAD_LOCK_ASSERT(td, MA_OWNED);
1708 base_pri = td->td_base_user_pri;
1709 if (prio >= base_pri) {
1710 td->td_flags &= ~TDF_UBORROWING;
1711 sched_user_prio(td, base_pri);
1713 sched_lend_user_prio(td, prio);
1718 * Block a thread for switching. Similar to thread_block() but does not
1719 * bump the spin count.
1721 static inline struct mtx *
1722 thread_block_switch(struct thread *td)
1726 THREAD_LOCK_ASSERT(td, MA_OWNED);
1728 td->td_lock = &blocked_lock;
1729 mtx_unlock_spin(lock);
1735 * Handle migration from sched_switch(). This happens only for
1739 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1743 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1745 tdq_load_rem(tdq, td);
1747 * Do the lock dance required to avoid LOR. We grab an extra
1748 * spinlock nesting to prevent preemption while we're
1749 * not holding either run-queue lock.
1752 thread_block_switch(td); /* This releases the lock on tdq. */
1755 * Acquire both run-queue locks before placing the thread on the new
1756 * run-queue to avoid deadlocks created by placing a thread with a
1757 * blocked lock on the run-queue of a remote processor. The deadlock
1758 * occurs when a third processor attempts to lock the two queues in
1759 * question while the target processor is spinning with its own
1760 * run-queue lock held while waiting for the blocked lock to clear.
1762 tdq_lock_pair(tdn, tdq);
1763 tdq_add(tdn, td, flags);
1764 tdq_notify(tdn, td);
1768 return (TDQ_LOCKPTR(tdn));
1772 * Release a thread that was blocked with thread_block_switch().
1775 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1777 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1782 * Switch threads. This function has to handle threads coming in while
1783 * blocked for some reason, running, or idle. It also must deal with
1784 * migrating a thread from one queue to another as running threads may
1785 * be assigned elsewhere via binding.
1788 sched_switch(struct thread *td, struct thread *newtd, int flags)
1791 struct td_sched *ts;
1796 THREAD_LOCK_ASSERT(td, MA_OWNED);
1797 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1799 cpuid = PCPU_GET(cpuid);
1800 tdq = TDQ_CPU(cpuid);
1803 ts->ts_rltick = ticks;
1804 td->td_lastcpu = td->td_oncpu;
1805 td->td_oncpu = NOCPU;
1806 td->td_flags &= ~TDF_NEEDRESCHED;
1807 td->td_owepreempt = 0;
1808 tdq->tdq_switchcnt++;
1810 * The lock pointer in an idle thread should never change. Reset it
1811 * to CAN_RUN as well.
1813 if (TD_IS_IDLETHREAD(td)) {
1814 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1816 } else if (TD_IS_RUNNING(td)) {
1817 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1818 srqflag = (flags & SW_PREEMPT) ?
1819 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1820 SRQ_OURSELF|SRQ_YIELDING;
1821 if (ts->ts_cpu == cpuid)
1822 tdq_runq_add(tdq, td, srqflag);
1824 mtx = sched_switch_migrate(tdq, td, srqflag);
1826 /* This thread must be going to sleep. */
1828 mtx = thread_block_switch(td);
1829 tdq_load_rem(tdq, td);
1832 * We enter here with the thread blocked and assigned to the
1833 * appropriate cpu run-queue or sleep-queue and with the current
1834 * thread-queue locked.
1836 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1837 newtd = choosethread();
1839 * Call the MD code to switch contexts if necessary.
1843 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1844 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1846 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1847 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1849 #ifdef KDTRACE_HOOKS
1851 * If DTrace has set the active vtime enum to anything
1852 * other than INACTIVE (0), then it should have set the
1855 if (dtrace_vtime_active)
1856 (*dtrace_vtime_switch_func)(newtd);
1859 cpu_switch(td, newtd, mtx);
1861 * We may return from cpu_switch on a different cpu. However,
1862 * we always return with td_lock pointing to the current cpu's
1865 cpuid = PCPU_GET(cpuid);
1866 tdq = TDQ_CPU(cpuid);
1867 lock_profile_obtain_lock_success(
1868 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1870 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1871 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1874 thread_unblock_switch(td, mtx);
1876 * Assert that all went well and return.
1878 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1879 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1880 td->td_oncpu = cpuid;
1884 * Adjust thread priorities as a result of a nice request.
1887 sched_nice(struct proc *p, int nice)
1891 PROC_LOCK_ASSERT(p, MA_OWNED);
1894 FOREACH_THREAD_IN_PROC(p, td) {
1897 sched_prio(td, td->td_base_user_pri);
1903 * Record the sleep time for the interactivity scorer.
1906 sched_sleep(struct thread *td, int prio)
1909 THREAD_LOCK_ASSERT(td, MA_OWNED);
1911 td->td_slptick = ticks;
1912 if (TD_IS_SUSPENDED(td) || prio <= PSOCK)
1913 td->td_flags |= TDF_CANSWAP;
1914 if (static_boost == 1 && prio)
1915 sched_prio(td, prio);
1916 else if (static_boost && td->td_priority > static_boost)
1917 sched_prio(td, static_boost);
1921 * Schedule a thread to resume execution and record how long it voluntarily
1922 * slept. We also update the pctcpu, interactivity, and priority.
1925 sched_wakeup(struct thread *td)
1927 struct td_sched *ts;
1930 THREAD_LOCK_ASSERT(td, MA_OWNED);
1932 td->td_flags &= ~TDF_CANSWAP;
1934 * If we slept for more than a tick update our interactivity and
1937 slptick = td->td_slptick;
1939 if (slptick && slptick != ticks) {
1942 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1943 ts->ts_slptime += hzticks;
1944 sched_interact_update(td);
1945 sched_pctcpu_update(ts);
1947 /* Reset the slice value after we sleep. */
1948 ts->ts_slice = sched_slice;
1949 sched_add(td, SRQ_BORING);
1953 * Penalize the parent for creating a new child and initialize the child's
1957 sched_fork(struct thread *td, struct thread *child)
1959 THREAD_LOCK_ASSERT(td, MA_OWNED);
1960 sched_fork_thread(td, child);
1962 * Penalize the parent and child for forking.
1964 sched_interact_fork(child);
1965 sched_priority(child);
1966 td->td_sched->ts_runtime += tickincr;
1967 sched_interact_update(td);
1972 * Fork a new thread, may be within the same process.
1975 sched_fork_thread(struct thread *td, struct thread *child)
1977 struct td_sched *ts;
1978 struct td_sched *ts2;
1980 THREAD_LOCK_ASSERT(td, MA_OWNED);
1985 ts2 = child->td_sched;
1986 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1987 child->td_cpuset = cpuset_ref(td->td_cpuset);
1988 ts2->ts_cpu = ts->ts_cpu;
1991 * Grab our parents cpu estimation information and priority.
1993 ts2->ts_ticks = ts->ts_ticks;
1994 ts2->ts_ltick = ts->ts_ltick;
1995 ts2->ts_incrtick = ts->ts_incrtick;
1996 ts2->ts_ftick = ts->ts_ftick;
1997 child->td_user_pri = td->td_user_pri;
1998 child->td_base_user_pri = td->td_base_user_pri;
2000 * And update interactivity score.
2002 ts2->ts_slptime = ts->ts_slptime;
2003 ts2->ts_runtime = ts->ts_runtime;
2004 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
2006 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2011 * Adjust the priority class of a thread.
2014 sched_class(struct thread *td, int class)
2017 THREAD_LOCK_ASSERT(td, MA_OWNED);
2018 if (td->td_pri_class == class)
2020 td->td_pri_class = class;
2024 * Return some of the child's priority and interactivity to the parent.
2027 sched_exit(struct proc *p, struct thread *child)
2031 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2032 "prio:td", child->td_priority);
2033 PROC_LOCK_ASSERT(p, MA_OWNED);
2034 td = FIRST_THREAD_IN_PROC(p);
2035 sched_exit_thread(td, child);
2039 * Penalize another thread for the time spent on this one. This helps to
2040 * worsen the priority and interactivity of processes which schedule batch
2041 * jobs such as make. This has little effect on the make process itself but
2042 * causes new processes spawned by it to receive worse scores immediately.
2045 sched_exit_thread(struct thread *td, struct thread *child)
2048 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2049 "prio:td", child->td_priority);
2051 * Give the child's runtime to the parent without returning the
2052 * sleep time as a penalty to the parent. This causes shells that
2053 * launch expensive things to mark their children as expensive.
2056 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2057 sched_interact_update(td);
2063 sched_preempt(struct thread *td)
2069 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2070 tdq->tdq_ipipending = 0;
2071 if (td->td_priority > tdq->tdq_lowpri) {
2074 flags = SW_INVOL | SW_PREEMPT;
2075 if (td->td_critnest > 1)
2076 td->td_owepreempt = 1;
2077 else if (TD_IS_IDLETHREAD(td))
2078 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2080 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2086 * Fix priorities on return to user-space. Priorities may be elevated due
2087 * to static priorities in msleep() or similar.
2090 sched_userret(struct thread *td)
2093 * XXX we cheat slightly on the locking here to avoid locking in
2094 * the usual case. Setting td_priority here is essentially an
2095 * incomplete workaround for not setting it properly elsewhere.
2096 * Now that some interrupt handlers are threads, not setting it
2097 * properly elsewhere can clobber it in the window between setting
2098 * it here and returning to user mode, so don't waste time setting
2099 * it perfectly here.
2101 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2102 ("thread with borrowed priority returning to userland"));
2103 if (td->td_priority != td->td_user_pri) {
2105 td->td_priority = td->td_user_pri;
2106 td->td_base_pri = td->td_user_pri;
2107 tdq_setlowpri(TDQ_SELF(), td);
2113 * Handle a stathz tick. This is really only relevant for timeshare
2117 sched_clock(struct thread *td)
2120 struct td_sched *ts;
2122 THREAD_LOCK_ASSERT(td, MA_OWNED);
2126 * We run the long term load balancer infrequently on the first cpu.
2128 if (balance_tdq == tdq) {
2129 if (balance_ticks && --balance_ticks == 0)
2134 * Save the old switch count so we have a record of the last ticks
2135 * activity. Initialize the new switch count based on our load.
2136 * If there is some activity seed it to reflect that.
2138 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2139 tdq->tdq_switchcnt = tdq->tdq_load;
2141 * Advance the insert index once for each tick to ensure that all
2142 * threads get a chance to run.
2144 if (tdq->tdq_idx == tdq->tdq_ridx) {
2145 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2146 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2147 tdq->tdq_ridx = tdq->tdq_idx;
2150 if (td->td_pri_class & PRI_FIFO_BIT)
2152 if (td->td_pri_class == PRI_TIMESHARE) {
2154 * We used a tick; charge it to the thread so
2155 * that we can compute our interactivity.
2157 td->td_sched->ts_runtime += tickincr;
2158 sched_interact_update(td);
2162 * We used up one time slice.
2164 if (--ts->ts_slice > 0)
2167 * We're out of time, force a requeue at userret().
2169 ts->ts_slice = sched_slice;
2170 td->td_flags |= TDF_NEEDRESCHED;
2174 * Called once per hz tick. Used for cpu utilization information. This
2175 * is easier than trying to scale based on stathz.
2180 struct td_sched *ts;
2182 ts = curthread->td_sched;
2184 * Ticks is updated asynchronously on a single cpu. Check here to
2185 * avoid incrementing ts_ticks multiple times in a single tick.
2187 if (ts->ts_incrtick == ticks)
2189 /* Adjust ticks for pctcpu */
2190 ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2191 ts->ts_ltick = ticks;
2192 ts->ts_incrtick = ticks;
2194 * Update if we've exceeded our desired tick threshhold by over one
2197 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2198 sched_pctcpu_update(ts);
2202 * Return whether the current CPU has runnable tasks. Used for in-kernel
2203 * cooperative idle threads.
2206 sched_runnable(void)
2214 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2215 if (tdq->tdq_load > 0)
2218 if (tdq->tdq_load - 1 > 0)
2226 * Choose the highest priority thread to run. The thread is removed from
2227 * the run-queue while running however the load remains. For SMP we set
2228 * the tdq in the global idle bitmask if it idles here.
2237 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2238 td = tdq_choose(tdq);
2240 td->td_sched->ts_ltick = ticks;
2241 tdq_runq_rem(tdq, td);
2242 tdq->tdq_lowpri = td->td_priority;
2245 tdq->tdq_lowpri = PRI_MAX_IDLE;
2246 return (PCPU_GET(idlethread));
2250 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2251 * we always request it once we exit a critical section.
2254 sched_setpreempt(struct thread *td)
2260 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2263 pri = td->td_priority;
2264 cpri = ctd->td_priority;
2266 ctd->td_flags |= TDF_NEEDRESCHED;
2267 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2269 if (!sched_shouldpreempt(pri, cpri, 0))
2271 ctd->td_owepreempt = 1;
2275 * Add a thread to a thread queue. Select the appropriate runq and add the
2276 * thread to it. This is the internal function called when the tdq is
2280 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2283 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2284 KASSERT((td->td_inhibitors == 0),
2285 ("sched_add: trying to run inhibited thread"));
2286 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2287 ("sched_add: bad thread state"));
2288 KASSERT(td->td_flags & TDF_INMEM,
2289 ("sched_add: thread swapped out"));
2291 if (td->td_priority < tdq->tdq_lowpri)
2292 tdq->tdq_lowpri = td->td_priority;
2293 tdq_runq_add(tdq, td, flags);
2294 tdq_load_add(tdq, td);
2298 * Select the target thread queue and add a thread to it. Request
2299 * preemption or IPI a remote processor if required.
2302 sched_add(struct thread *td, int flags)
2309 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2310 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2311 sched_tdname(curthread));
2312 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2313 KTR_ATTR_LINKED, sched_tdname(td));
2314 THREAD_LOCK_ASSERT(td, MA_OWNED);
2316 * Recalculate the priority before we select the target cpu or
2319 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2323 * Pick the destination cpu and if it isn't ours transfer to the
2326 cpu = sched_pickcpu(td, flags);
2327 tdq = sched_setcpu(td, cpu, flags);
2328 tdq_add(tdq, td, flags);
2329 if (cpu != PCPU_GET(cpuid)) {
2330 tdq_notify(tdq, td);
2337 * Now that the thread is moving to the run-queue, set the lock
2338 * to the scheduler's lock.
2340 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2341 tdq_add(tdq, td, flags);
2343 if (!(flags & SRQ_YIELDING))
2344 sched_setpreempt(td);
2348 * Remove a thread from a run-queue without running it. This is used
2349 * when we're stealing a thread from a remote queue. Otherwise all threads
2350 * exit by calling sched_exit_thread() and sched_throw() themselves.
2353 sched_rem(struct thread *td)
2357 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2358 "prio:%d", td->td_priority);
2359 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2360 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2361 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2362 KASSERT(TD_ON_RUNQ(td),
2363 ("sched_rem: thread not on run queue"));
2364 tdq_runq_rem(tdq, td);
2365 tdq_load_rem(tdq, td);
2367 if (td->td_priority == tdq->tdq_lowpri)
2368 tdq_setlowpri(tdq, NULL);
2372 * Fetch cpu utilization information. Updates on demand.
2375 sched_pctcpu(struct thread *td)
2378 struct td_sched *ts;
2389 sched_pctcpu_update(ts);
2390 /* How many rtick per second ? */
2391 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2392 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2400 * Enforce affinity settings for a thread. Called after adjustments to
2404 sched_affinity(struct thread *td)
2407 struct td_sched *ts;
2410 THREAD_LOCK_ASSERT(td, MA_OWNED);
2412 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2414 if (TD_ON_RUNQ(td)) {
2416 sched_add(td, SRQ_BORING);
2419 if (!TD_IS_RUNNING(td))
2421 td->td_flags |= TDF_NEEDRESCHED;
2422 if (!THREAD_CAN_MIGRATE(td))
2425 * Assign the new cpu and force a switch before returning to
2426 * userspace. If the target thread is not running locally send
2427 * an ipi to force the issue.
2430 ts->ts_cpu = sched_pickcpu(td, 0);
2431 if (cpu != PCPU_GET(cpuid))
2432 ipi_selected(1 << cpu, IPI_PREEMPT);
2437 * Bind a thread to a target cpu.
2440 sched_bind(struct thread *td, int cpu)
2442 struct td_sched *ts;
2444 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2446 if (ts->ts_flags & TSF_BOUND)
2448 ts->ts_flags |= TSF_BOUND;
2450 if (PCPU_GET(cpuid) == cpu)
2453 /* When we return from mi_switch we'll be on the correct cpu. */
2454 mi_switch(SW_VOL, NULL);
2458 * Release a bound thread.
2461 sched_unbind(struct thread *td)
2463 struct td_sched *ts;
2465 THREAD_LOCK_ASSERT(td, MA_OWNED);
2467 if ((ts->ts_flags & TSF_BOUND) == 0)
2469 ts->ts_flags &= ~TSF_BOUND;
2474 sched_is_bound(struct thread *td)
2476 THREAD_LOCK_ASSERT(td, MA_OWNED);
2477 return (td->td_sched->ts_flags & TSF_BOUND);
2484 sched_relinquish(struct thread *td)
2487 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2492 * Return the total system load.
2502 for (i = 0; i <= mp_maxid; i++)
2503 total += TDQ_CPU(i)->tdq_sysload;
2506 return (TDQ_SELF()->tdq_sysload);
2511 sched_sizeof_proc(void)
2513 return (sizeof(struct proc));
2517 sched_sizeof_thread(void)
2519 return (sizeof(struct thread) + sizeof(struct td_sched));
2523 #define TDQ_IDLESPIN(tdq) \
2524 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2526 #define TDQ_IDLESPIN(tdq) 1
2530 * The actual idle process.
2533 sched_idletd(void *dummy)
2540 mtx_assert(&Giant, MA_NOTOWNED);
2545 if (tdq_idled(tdq) == 0)
2548 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2550 * If we're switching very frequently, spin while checking
2551 * for load rather than entering a low power state that
2552 * may require an IPI. However, don't do any busy
2553 * loops while on SMT machines as this simply steals
2554 * cycles from cores doing useful work.
2556 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2557 for (i = 0; i < sched_idlespins; i++) {
2563 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2564 if (tdq->tdq_load == 0)
2565 cpu_idle(switchcnt > 1);
2566 if (tdq->tdq_load) {
2568 mi_switch(SW_VOL | SWT_IDLE, NULL);
2575 * A CPU is entering for the first time or a thread is exiting.
2578 sched_throw(struct thread *td)
2580 struct thread *newtd;
2585 /* Correct spinlock nesting and acquire the correct lock. */
2589 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2590 tdq_load_rem(tdq, td);
2591 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2593 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2594 newtd = choosethread();
2595 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2596 PCPU_SET(switchtime, cpu_ticks());
2597 PCPU_SET(switchticks, ticks);
2598 cpu_throw(td, newtd); /* doesn't return */
2602 * This is called from fork_exit(). Just acquire the correct locks and
2603 * let fork do the rest of the work.
2606 sched_fork_exit(struct thread *td)
2608 struct td_sched *ts;
2613 * Finish setting up thread glue so that it begins execution in a
2614 * non-nested critical section with the scheduler lock held.
2616 cpuid = PCPU_GET(cpuid);
2617 tdq = TDQ_CPU(cpuid);
2619 if (TD_IS_IDLETHREAD(td))
2620 td->td_lock = TDQ_LOCKPTR(tdq);
2621 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2622 td->td_oncpu = cpuid;
2623 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2624 lock_profile_obtain_lock_success(
2625 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2629 * Create on first use to catch odd startup conditons.
2632 sched_tdname(struct thread *td)
2635 struct td_sched *ts;
2638 if (ts->ts_name[0] == '\0')
2639 snprintf(ts->ts_name, sizeof(ts->ts_name),
2640 "%s tid %d", td->td_name, td->td_tid);
2641 return (ts->ts_name);
2643 return (td->td_name);
2650 * Build the CPU topology dump string. Is recursively called to collect
2651 * the topology tree.
2654 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2659 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2660 "", indent, cg->cg_level);
2661 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2662 cg->cg_count, cg->cg_mask);
2664 for (i = 0; i < MAXCPU; i++) {
2665 if ((cg->cg_mask & (1 << i)) != 0) {
2667 sbuf_printf(sb, ", ");
2670 sbuf_printf(sb, "%d", i);
2673 sbuf_printf(sb, "</cpu>\n");
2675 sbuf_printf(sb, "%*s <flags>", indent, "");
2676 if (cg->cg_flags != 0) {
2677 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2678 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>\n");
2679 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2680 sbuf_printf(sb, "<flag name=\"THREAD\">SMT group</flag>\n");
2682 sbuf_printf(sb, "</flags>\n");
2684 if (cg->cg_children > 0) {
2685 sbuf_printf(sb, "%*s <children>\n", indent, "");
2686 for (i = 0; i < cg->cg_children; i++)
2687 sysctl_kern_sched_topology_spec_internal(sb,
2688 &cg->cg_child[i], indent+2);
2689 sbuf_printf(sb, "%*s </children>\n", indent, "");
2691 sbuf_printf(sb, "%*s</group>\n", indent, "");
2696 * Sysctl handler for retrieving topology dump. It's a wrapper for
2697 * the recursive sysctl_kern_smp_topology_spec_internal().
2700 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2705 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2707 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2711 sbuf_printf(topo, "<groups>\n");
2712 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2713 sbuf_printf(topo, "</groups>\n");
2717 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");
2760 /* ps compat. All cpu percentages from ULE are weighted. */
2761 static int ccpu = 0;
2762 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");