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__)
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 * Priority ranges used for interactive and non-interactive timeshare
125 * threads. Interactive threads use realtime priorities.
127 #define PRI_MIN_INTERACT PRI_MIN_REALTIME
128 #define PRI_MAX_INTERACT PRI_MAX_REALTIME
129 #define PRI_MIN_BATCH PRI_MIN_TIMESHARE
130 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
133 * Cpu percentage computation macros and defines.
135 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
136 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
137 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
138 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
139 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
140 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
142 #define SCHED_TICK_SECS 10
143 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
144 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
145 #define SCHED_TICK_SHIFT 10
146 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
147 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
150 * These macros determine priorities for non-interactive threads. They are
151 * assigned a priority based on their recent cpu utilization as expressed
152 * by the ratio of ticks to the tick total. NHALF priorities at the start
153 * and end of the MIN to MAX timeshare range are only reachable with negative
154 * or positive nice respectively.
156 * PRI_RANGE: Priority range for utilization dependent priorities.
157 * PRI_NRESV: Number of nice values.
158 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
159 * PRI_NICE: Determines the part of the priority inherited from nice.
161 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
162 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
163 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
164 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
165 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
166 #define SCHED_PRI_TICKS(ts) \
167 (SCHED_TICK_HZ((ts)) / \
168 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
169 #define SCHED_PRI_NICE(nice) (nice)
172 * These determine the interactivity of a process. Interactivity differs from
173 * cpu utilization in that it expresses the voluntary time slept vs time ran
174 * while cpu utilization includes all time not running. This more accurately
175 * models the intent of the thread.
177 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
178 * before throttling back.
179 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
180 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
181 * INTERACT_THRESH: Threshold for placement on the current runq.
183 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
184 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
185 #define SCHED_INTERACT_MAX (100)
186 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
187 #define SCHED_INTERACT_THRESH (30)
190 * tickincr: Converts a stathz tick into a hz domain scaled by
191 * the shift factor. Without the shift the error rate
192 * due to rounding would be unacceptably high.
193 * realstathz: stathz is sometimes 0 and run off of hz.
194 * sched_slice: Runtime of each thread before rescheduling.
195 * preempt_thresh: Priority threshold for preemption and remote IPIs.
197 static int sched_interact = SCHED_INTERACT_THRESH;
198 static int realstathz;
200 static int sched_slice = 1;
202 #ifdef FULL_PREEMPTION
203 static int preempt_thresh = PRI_MAX_IDLE;
205 static int preempt_thresh = PRI_MIN_KERN;
208 static int preempt_thresh = 0;
210 static int static_boost = PRI_MIN_BATCH;
211 static int sched_idlespins = 10000;
212 static int sched_idlespinthresh = 4;
215 * tdq - per processor runqs and statistics. All fields are protected by the
216 * tdq_lock. The load and lowpri may be accessed without to avoid excess
217 * locking in sched_pickcpu();
220 /* Ordered to improve efficiency of cpu_search() and switch(). */
221 struct mtx tdq_lock; /* run queue lock. */
222 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
223 volatile int tdq_load; /* Aggregate load. */
224 int tdq_sysload; /* For loadavg, !ITHD load. */
225 int tdq_transferable; /* Transferable thread count. */
226 short tdq_switchcnt; /* Switches this tick. */
227 short tdq_oldswitchcnt; /* Switches last tick. */
228 u_char tdq_lowpri; /* Lowest priority thread. */
229 u_char tdq_ipipending; /* IPI pending. */
230 u_char tdq_idx; /* Current insert index. */
231 u_char tdq_ridx; /* Current removal index. */
232 struct runq tdq_realtime; /* real-time run queue. */
233 struct runq tdq_timeshare; /* timeshare run queue. */
234 struct runq tdq_idle; /* Queue of IDLE threads. */
235 char tdq_name[TDQ_NAME_LEN];
237 char tdq_loadname[TDQ_LOADNAME_LEN];
241 /* Idle thread states and config. */
242 #define TDQ_RUNNING 1
246 struct cpu_group *cpu_top; /* CPU topology */
248 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
249 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
254 static int rebalance = 1;
255 static int balance_interval = 128; /* Default set in sched_initticks(). */
257 static int steal_htt = 1;
258 static int steal_idle = 1;
259 static int steal_thresh = 2;
262 * One thread queue per processor.
264 static struct tdq tdq_cpu[MAXCPU];
265 static struct tdq *balance_tdq;
266 static int balance_ticks;
268 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
269 #define TDQ_CPU(x) (&tdq_cpu[(x)])
270 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
272 static struct tdq tdq_cpu;
274 #define TDQ_ID(x) (0)
275 #define TDQ_SELF() (&tdq_cpu)
276 #define TDQ_CPU(x) (&tdq_cpu)
279 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
280 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
281 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
282 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
283 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
285 static void sched_priority(struct thread *);
286 static void sched_thread_priority(struct thread *, u_char);
287 static int sched_interact_score(struct thread *);
288 static void sched_interact_update(struct thread *);
289 static void sched_interact_fork(struct thread *);
290 static void sched_pctcpu_update(struct td_sched *);
292 /* Operations on per processor queues */
293 static struct thread *tdq_choose(struct tdq *);
294 static void tdq_setup(struct tdq *);
295 static void tdq_load_add(struct tdq *, struct thread *);
296 static void tdq_load_rem(struct tdq *, struct thread *);
297 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
298 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
299 static inline int sched_shouldpreempt(int, int, int);
300 void tdq_print(int cpu);
301 static void runq_print(struct runq *rq);
302 static void tdq_add(struct tdq *, struct thread *, int);
304 static int tdq_move(struct tdq *, struct tdq *);
305 static int tdq_idled(struct tdq *);
306 static void tdq_notify(struct tdq *, struct thread *);
307 static struct thread *tdq_steal(struct tdq *, int);
308 static struct thread *runq_steal(struct runq *, int);
309 static int sched_pickcpu(struct thread *, int);
310 static void sched_balance(void);
311 static int sched_balance_pair(struct tdq *, struct tdq *);
312 static inline struct tdq *sched_setcpu(struct thread *, int, int);
313 static inline void thread_unblock_switch(struct thread *, struct mtx *);
314 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
315 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
316 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
317 struct cpu_group *cg, int indent);
320 static void sched_setup(void *dummy);
321 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
323 static void sched_initticks(void *dummy);
324 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
328 * Print the threads waiting on a run-queue.
331 runq_print(struct runq *rq)
339 for (i = 0; i < RQB_LEN; i++) {
340 printf("\t\trunq bits %d 0x%zx\n",
341 i, rq->rq_status.rqb_bits[i]);
342 for (j = 0; j < RQB_BPW; j++)
343 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
344 pri = j + (i << RQB_L2BPW);
345 rqh = &rq->rq_queues[pri];
346 TAILQ_FOREACH(td, rqh, td_runq) {
347 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
348 td, td->td_name, td->td_priority,
349 td->td_rqindex, pri);
356 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
365 printf("tdq %d:\n", TDQ_ID(tdq));
366 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
367 printf("\tLock name: %s\n", tdq->tdq_name);
368 printf("\tload: %d\n", tdq->tdq_load);
369 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
370 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
371 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
372 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
373 printf("\tload transferable: %d\n", tdq->tdq_transferable);
374 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
375 printf("\trealtime runq:\n");
376 runq_print(&tdq->tdq_realtime);
377 printf("\ttimeshare runq:\n");
378 runq_print(&tdq->tdq_timeshare);
379 printf("\tidle runq:\n");
380 runq_print(&tdq->tdq_idle);
384 sched_shouldpreempt(int pri, int cpri, int remote)
387 * If the new priority is not better than the current priority there is
393 * Always preempt idle.
395 if (cpri >= PRI_MIN_IDLE)
398 * If preemption is disabled don't preempt others.
400 if (preempt_thresh == 0)
403 * Preempt if we exceed the threshold.
405 if (pri <= preempt_thresh)
408 * If we're interactive or better and there is non-interactive
409 * or worse running preempt only remote processors.
411 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
417 * Add a thread to the actual run-queue. Keeps transferable counts up to
418 * date with what is actually on the run-queue. Selects the correct
419 * queue position for timeshare threads.
422 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
427 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
428 THREAD_LOCK_ASSERT(td, MA_OWNED);
430 pri = td->td_priority;
433 if (THREAD_CAN_MIGRATE(td)) {
434 tdq->tdq_transferable++;
435 ts->ts_flags |= TSF_XFERABLE;
437 if (pri < PRI_MIN_BATCH) {
438 ts->ts_runq = &tdq->tdq_realtime;
439 } else if (pri <= PRI_MAX_BATCH) {
440 ts->ts_runq = &tdq->tdq_timeshare;
441 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
442 ("Invalid priority %d on timeshare runq", pri));
444 * This queue contains only priorities between MIN and MAX
445 * realtime. Use the whole queue to represent these values.
447 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
448 pri = RQ_NQS * (pri - PRI_MIN_BATCH) /
449 (PRI_MAX_BATCH - PRI_MIN_BATCH + 1);
450 pri = (pri + tdq->tdq_idx) % RQ_NQS;
452 * This effectively shortens the queue by one so we
453 * can have a one slot difference between idx and
454 * ridx while we wait for threads to drain.
456 if (tdq->tdq_ridx != tdq->tdq_idx &&
457 pri == tdq->tdq_ridx)
458 pri = (unsigned char)(pri - 1) % RQ_NQS;
461 runq_add_pri(ts->ts_runq, td, pri, flags);
464 ts->ts_runq = &tdq->tdq_idle;
465 runq_add(ts->ts_runq, td, flags);
469 * Remove a thread from a run-queue. This typically happens when a thread
470 * is selected to run. Running threads are not on the queue and the
471 * transferable count does not reflect them.
474 tdq_runq_rem(struct tdq *tdq, struct thread *td)
479 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
480 KASSERT(ts->ts_runq != NULL,
481 ("tdq_runq_remove: thread %p null ts_runq", td));
482 if (ts->ts_flags & TSF_XFERABLE) {
483 tdq->tdq_transferable--;
484 ts->ts_flags &= ~TSF_XFERABLE;
486 if (ts->ts_runq == &tdq->tdq_timeshare) {
487 if (tdq->tdq_idx != tdq->tdq_ridx)
488 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
490 runq_remove_idx(ts->ts_runq, td, NULL);
492 runq_remove(ts->ts_runq, td);
496 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
497 * for this thread to the referenced thread queue.
500 tdq_load_add(struct tdq *tdq, struct thread *td)
503 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
504 THREAD_LOCK_ASSERT(td, MA_OWNED);
507 if ((td->td_flags & TDF_NOLOAD) == 0)
509 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
513 * Remove the load from a thread that is transitioning to a sleep state or
517 tdq_load_rem(struct tdq *tdq, struct thread *td)
520 THREAD_LOCK_ASSERT(td, MA_OWNED);
521 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
522 KASSERT(tdq->tdq_load != 0,
523 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
526 if ((td->td_flags & TDF_NOLOAD) == 0)
528 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
532 * Set lowpri to its exact value by searching the run-queue and
533 * evaluating curthread. curthread may be passed as an optimization.
536 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
540 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
542 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
543 td = tdq_choose(tdq);
544 if (td == NULL || td->td_priority > ctd->td_priority)
545 tdq->tdq_lowpri = ctd->td_priority;
547 tdq->tdq_lowpri = td->td_priority;
555 int cs_limit; /* Min priority for low min load for high. */
558 #define CPU_SEARCH_LOWEST 0x1
559 #define CPU_SEARCH_HIGHEST 0x2
560 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
562 #define CPUSET_FOREACH(cpu, mask) \
563 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
564 if ((mask) & 1 << (cpu))
566 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
567 struct cpu_search *high, const int match);
568 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
569 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
570 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
571 struct cpu_search *high);
574 * This routine compares according to the match argument and should be
575 * reduced in actual instantiations via constant propagation and dead code
579 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
585 if (match & CPU_SEARCH_LOWEST)
586 if (CPU_ISSET(cpu, &low->cs_mask) &&
587 tdq->tdq_load < low->cs_load &&
588 tdq->tdq_lowpri > low->cs_limit) {
590 low->cs_load = tdq->tdq_load;
592 if (match & CPU_SEARCH_HIGHEST)
593 if (CPU_ISSET(cpu, &high->cs_mask) &&
594 tdq->tdq_load >= high->cs_limit &&
595 tdq->tdq_load > high->cs_load &&
596 tdq->tdq_transferable) {
598 high->cs_load = tdq->tdq_load;
600 return (tdq->tdq_load);
604 * Search the tree of cpu_groups for the lowest or highest loaded cpu
605 * according to the match argument. This routine actually compares the
606 * load on all paths through the tree and finds the least loaded cpu on
607 * the least loaded path, which may differ from the least loaded cpu in
608 * the system. This balances work among caches and busses.
610 * This inline is instantiated in three forms below using constants for the
611 * match argument. It is reduced to the minimum set for each case. It is
612 * also recursive to the depth of the tree.
615 cpu_search(struct cpu_group *cg, struct cpu_search *low,
616 struct cpu_search *high, const int match)
621 if (cg->cg_children) {
622 struct cpu_search lgroup;
623 struct cpu_search hgroup;
624 struct cpu_group *child;
632 for (i = 0; i < cg->cg_children; i++) {
633 child = &cg->cg_child[i];
634 if (match & CPU_SEARCH_LOWEST) {
638 if (match & CPU_SEARCH_HIGHEST) {
643 case CPU_SEARCH_LOWEST:
644 load = cpu_search_lowest(child, &lgroup);
646 case CPU_SEARCH_HIGHEST:
647 load = cpu_search_highest(child, &hgroup);
649 case CPU_SEARCH_BOTH:
650 load = cpu_search_both(child, &lgroup, &hgroup);
654 if (match & CPU_SEARCH_LOWEST)
655 if (load < lload || low->cs_cpu == -1) {
659 if (match & CPU_SEARCH_HIGHEST)
660 if (load > hload || high->cs_cpu == -1) {
668 CPUSET_FOREACH(cpu, cg->cg_mask)
669 total += cpu_compare(cpu, low, high, match);
675 * cpu_search instantiations must pass constants to maintain the inline
679 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
681 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
685 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
687 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
691 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
692 struct cpu_search *high)
694 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
698 * Find the cpu with the least load via the least loaded path that has a
699 * lowpri greater than pri pri. A pri of -1 indicates any priority is
703 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
705 struct cpu_search low;
711 cpu_search_lowest(cg, &low);
716 * Find the cpu with the highest load via the highest loaded path.
719 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
721 struct cpu_search high;
726 high.cs_limit = minload;
727 cpu_search_highest(cg, &high);
732 * Simultaneously find the highest and lowest loaded cpu reachable via
736 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
738 struct cpu_search high;
739 struct cpu_search low;
749 cpu_search_both(cg, &low, &high);
750 *lowcpu = low.cs_cpu;
751 *highcpu = high.cs_cpu;
756 sched_balance_group(struct cpu_group *cg)
765 sched_both(cg, mask, &low, &high);
766 if (low == high || low == -1 || high == -1)
768 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
771 * If we failed to move any threads determine which cpu
772 * to kick out of the set and try again.
774 if (TDQ_CPU(high)->tdq_transferable == 0)
775 CPU_CLR(high, &mask);
780 for (i = 0; i < cg->cg_children; i++)
781 sched_balance_group(&cg->cg_child[i]);
790 * Select a random time between .5 * balance_interval and
791 * 1.5 * balance_interval.
793 balance_ticks = max(balance_interval / 2, 1);
794 balance_ticks += random() % balance_interval;
795 if (smp_started == 0 || rebalance == 0)
799 sched_balance_group(cpu_top);
804 * Lock two thread queues using their address to maintain lock order.
807 tdq_lock_pair(struct tdq *one, struct tdq *two)
811 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
814 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
819 * Unlock two thread queues. Order is not important here.
822 tdq_unlock_pair(struct tdq *one, struct tdq *two)
829 * Transfer load between two imbalanced thread queues.
832 sched_balance_pair(struct tdq *high, struct tdq *low)
842 tdq_lock_pair(high, low);
843 transferable = high->tdq_transferable;
844 high_load = high->tdq_load;
845 low_load = low->tdq_load;
848 * Determine what the imbalance is and then adjust that to how many
849 * threads we actually have to give up (transferable).
851 if (transferable != 0) {
852 diff = high_load - low_load;
856 move = min(move, transferable);
857 for (i = 0; i < move; i++)
858 moved += tdq_move(high, low);
860 * IPI the target cpu to force it to reschedule with the new
863 ipi_cpu(TDQ_ID(low), IPI_PREEMPT);
865 tdq_unlock_pair(high, low);
870 * Move a thread from one thread queue to another.
873 tdq_move(struct tdq *from, struct tdq *to)
880 TDQ_LOCK_ASSERT(from, MA_OWNED);
881 TDQ_LOCK_ASSERT(to, MA_OWNED);
885 td = tdq_steal(tdq, cpu);
890 * Although the run queue is locked the thread may be blocked. Lock
891 * it to clear this and acquire the run-queue lock.
894 /* Drop recursive lock on from acquired via thread_lock(). */
898 td->td_lock = TDQ_LOCKPTR(to);
899 tdq_add(to, td, SRQ_YIELDING);
904 * This tdq has idled. Try to steal a thread from another cpu and switch
908 tdq_idled(struct tdq *tdq)
910 struct cpu_group *cg;
916 if (smp_started == 0 || steal_idle == 0)
919 CPU_CLR(PCPU_GET(cpuid), &mask);
920 /* We don't want to be preempted while we're iterating. */
922 for (cg = tdq->tdq_cg; cg != NULL; ) {
923 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
924 thresh = steal_thresh;
927 cpu = sched_highest(cg, mask, thresh);
932 steal = TDQ_CPU(cpu);
934 tdq_lock_pair(tdq, steal);
935 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
936 tdq_unlock_pair(tdq, steal);
940 * If a thread was added while interrupts were disabled don't
941 * steal one here. If we fail to acquire one due to affinity
942 * restrictions loop again with this cpu removed from the
945 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
946 tdq_unlock_pair(tdq, steal);
951 mi_switch(SW_VOL | SWT_IDLE, NULL);
952 thread_unlock(curthread);
961 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
964 tdq_notify(struct tdq *tdq, struct thread *td)
970 if (tdq->tdq_ipipending)
972 cpu = td->td_sched->ts_cpu;
973 pri = td->td_priority;
974 ctd = pcpu_find(cpu)->pc_curthread;
975 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
977 if (TD_IS_IDLETHREAD(ctd)) {
979 * If the MD code has an idle wakeup routine try that before
980 * falling back to IPI.
982 if (cpu_idle_wakeup(cpu))
985 tdq->tdq_ipipending = 1;
986 ipi_cpu(cpu, IPI_PREEMPT);
990 * Steals load from a timeshare queue. Honors the rotating queue head
993 static struct thread *
994 runq_steal_from(struct runq *rq, int cpu, u_char start)
1004 rqb = &rq->rq_status;
1005 bit = start & (RQB_BPW -1);
1009 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1010 if (rqb->rqb_bits[i] == 0)
1013 for (pri = bit; pri < RQB_BPW; pri++)
1014 if (rqb->rqb_bits[i] & (1ul << pri))
1019 pri = RQB_FFS(rqb->rqb_bits[i]);
1020 pri += (i << RQB_L2BPW);
1021 rqh = &rq->rq_queues[pri];
1022 TAILQ_FOREACH(td, rqh, td_runq) {
1023 if (first && THREAD_CAN_MIGRATE(td) &&
1024 THREAD_CAN_SCHED(td, cpu))
1038 * Steals load from a standard linear queue.
1040 static struct thread *
1041 runq_steal(struct runq *rq, int cpu)
1049 rqb = &rq->rq_status;
1050 for (word = 0; word < RQB_LEN; word++) {
1051 if (rqb->rqb_bits[word] == 0)
1053 for (bit = 0; bit < RQB_BPW; bit++) {
1054 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1056 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1057 TAILQ_FOREACH(td, rqh, td_runq)
1058 if (THREAD_CAN_MIGRATE(td) &&
1059 THREAD_CAN_SCHED(td, cpu))
1067 * Attempt to steal a thread in priority order from a thread queue.
1069 static struct thread *
1070 tdq_steal(struct tdq *tdq, int cpu)
1074 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1075 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1077 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1078 cpu, tdq->tdq_ridx)) != NULL)
1080 return (runq_steal(&tdq->tdq_idle, cpu));
1084 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1085 * current lock and returns with the assigned queue locked.
1087 static inline struct tdq *
1088 sched_setcpu(struct thread *td, int cpu, int flags)
1093 THREAD_LOCK_ASSERT(td, MA_OWNED);
1095 td->td_sched->ts_cpu = cpu;
1097 * If the lock matches just return the queue.
1099 if (td->td_lock == TDQ_LOCKPTR(tdq))
1103 * If the thread isn't running its lockptr is a
1104 * turnstile or a sleepqueue. We can just lock_set without
1107 if (TD_CAN_RUN(td)) {
1109 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1114 * The hard case, migration, we need to block the thread first to
1115 * prevent order reversals with other cpus locks.
1118 thread_lock_block(td);
1120 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1125 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1126 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1127 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1128 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1129 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1130 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1133 sched_pickcpu(struct thread *td, int flags)
1135 struct cpu_group *cg;
1136 struct td_sched *ts;
1143 self = PCPU_GET(cpuid);
1145 if (smp_started == 0)
1148 * Don't migrate a running thread from sched_switch().
1150 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1151 return (ts->ts_cpu);
1153 * Prefer to run interrupt threads on the processors that generate
1156 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1157 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1158 SCHED_STAT_INC(pickcpu_intrbind);
1162 * If the thread can run on the last cpu and the affinity has not
1163 * expired or it is idle run it there.
1165 pri = td->td_priority;
1166 tdq = TDQ_CPU(ts->ts_cpu);
1167 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1168 if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1169 SCHED_STAT_INC(pickcpu_idle_affinity);
1170 return (ts->ts_cpu);
1172 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1173 SCHED_STAT_INC(pickcpu_affinity);
1174 return (ts->ts_cpu);
1178 * Search for the highest level in the tree that still has affinity.
1181 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1182 if (SCHED_AFFINITY(ts, cg->cg_level))
1185 mask = td->td_cpuset->cs_mask;
1187 cpu = sched_lowest(cg, mask, pri);
1189 cpu = sched_lowest(cpu_top, mask, -1);
1191 * Compare the lowest loaded cpu to current cpu.
1193 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1194 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1195 SCHED_STAT_INC(pickcpu_local);
1198 SCHED_STAT_INC(pickcpu_lowest);
1199 if (cpu != ts->ts_cpu)
1200 SCHED_STAT_INC(pickcpu_migration);
1201 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1207 * Pick the highest priority task we have and return it.
1209 static struct thread *
1210 tdq_choose(struct tdq *tdq)
1214 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1215 td = runq_choose(&tdq->tdq_realtime);
1218 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1220 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1221 ("tdq_choose: Invalid priority on timeshare queue %d",
1225 td = runq_choose(&tdq->tdq_idle);
1227 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1228 ("tdq_choose: Invalid priority on idle queue %d",
1237 * Initialize a thread queue.
1240 tdq_setup(struct tdq *tdq)
1244 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1245 runq_init(&tdq->tdq_realtime);
1246 runq_init(&tdq->tdq_timeshare);
1247 runq_init(&tdq->tdq_idle);
1248 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1249 "sched lock %d", (int)TDQ_ID(tdq));
1250 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1251 MTX_SPIN | MTX_RECURSE);
1253 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1254 "CPU %d load", (int)TDQ_ID(tdq));
1260 sched_setup_smp(void)
1265 cpu_top = smp_topo();
1269 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1270 if (tdq->tdq_cg == NULL)
1271 panic("Can't find cpu group for %d\n", i);
1273 balance_tdq = TDQ_SELF();
1279 * Setup the thread queues and initialize the topology based on MD
1283 sched_setup(void *dummy)
1294 * To avoid divide-by-zero, we set realstathz a dummy value
1295 * in case which sched_clock() called before sched_initticks().
1298 sched_slice = (realstathz/10); /* ~100ms */
1299 tickincr = 1 << SCHED_TICK_SHIFT;
1301 /* Add thread0's load since it's running. */
1303 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1304 tdq_load_add(tdq, &thread0);
1305 tdq->tdq_lowpri = thread0.td_priority;
1310 * This routine determines the tickincr after stathz and hz are setup.
1314 sched_initticks(void *dummy)
1318 realstathz = stathz ? stathz : hz;
1319 sched_slice = (realstathz/10); /* ~100ms */
1322 * tickincr is shifted out by 10 to avoid rounding errors due to
1323 * hz not being evenly divisible by stathz on all platforms.
1325 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1327 * This does not work for values of stathz that are more than
1328 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1335 * Set the default balance interval now that we know
1336 * what realstathz is.
1338 balance_interval = realstathz;
1340 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1341 * This prevents excess thrashing on large machines and excess idle
1342 * on smaller machines.
1344 steal_thresh = min(fls(mp_ncpus) - 1, 3);
1345 affinity = SCHED_AFFINITY_DEFAULT;
1351 * This is the core of the interactivity algorithm. Determines a score based
1352 * on past behavior. It is the ratio of sleep time to run time scaled to
1353 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1354 * differs from the cpu usage because it does not account for time spent
1355 * waiting on a run-queue. Would be prettier if we had floating point.
1358 sched_interact_score(struct thread *td)
1360 struct td_sched *ts;
1365 * The score is only needed if this is likely to be an interactive
1366 * task. Don't go through the expense of computing it if there's
1369 if (sched_interact <= SCHED_INTERACT_HALF &&
1370 ts->ts_runtime >= ts->ts_slptime)
1371 return (SCHED_INTERACT_HALF);
1373 if (ts->ts_runtime > ts->ts_slptime) {
1374 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1375 return (SCHED_INTERACT_HALF +
1376 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1378 if (ts->ts_slptime > ts->ts_runtime) {
1379 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1380 return (ts->ts_runtime / div);
1382 /* runtime == slptime */
1384 return (SCHED_INTERACT_HALF);
1387 * This can happen if slptime and runtime are 0.
1394 * Scale the scheduling priority according to the "interactivity" of this
1398 sched_priority(struct thread *td)
1403 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1406 * If the score is interactive we place the thread in the realtime
1407 * queue with a priority that is less than kernel and interrupt
1408 * priorities. These threads are not subject to nice restrictions.
1410 * Scores greater than this are placed on the normal timeshare queue
1411 * where the priority is partially decided by the most recent cpu
1412 * utilization and the rest is decided by nice value.
1414 * The nice value of the process has a linear effect on the calculated
1415 * score. Negative nice values make it easier for a thread to be
1416 * considered interactive.
1418 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1419 if (score < sched_interact) {
1420 pri = PRI_MIN_INTERACT;
1421 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1422 sched_interact) * score;
1423 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1424 ("sched_priority: invalid interactive priority %d score %d",
1427 pri = SCHED_PRI_MIN;
1428 if (td->td_sched->ts_ticks)
1429 pri += min(SCHED_PRI_TICKS(td->td_sched),
1431 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1432 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1433 ("sched_priority: invalid priority %d: nice %d, "
1434 "ticks %d ftick %d ltick %d tick pri %d",
1435 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1436 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1437 SCHED_PRI_TICKS(td->td_sched)));
1439 sched_user_prio(td, pri);
1445 * This routine enforces a maximum limit on the amount of scheduling history
1446 * kept. It is called after either the slptime or runtime is adjusted. This
1447 * function is ugly due to integer math.
1450 sched_interact_update(struct thread *td)
1452 struct td_sched *ts;
1456 sum = ts->ts_runtime + ts->ts_slptime;
1457 if (sum < SCHED_SLP_RUN_MAX)
1460 * This only happens from two places:
1461 * 1) We have added an unusual amount of run time from fork_exit.
1462 * 2) We have added an unusual amount of sleep time from sched_sleep().
1464 if (sum > SCHED_SLP_RUN_MAX * 2) {
1465 if (ts->ts_runtime > ts->ts_slptime) {
1466 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1469 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1475 * If we have exceeded by more than 1/5th then the algorithm below
1476 * will not bring us back into range. Dividing by two here forces
1477 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1479 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1480 ts->ts_runtime /= 2;
1481 ts->ts_slptime /= 2;
1484 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1485 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1489 * Scale back the interactivity history when a child thread is created. The
1490 * history is inherited from the parent but the thread may behave totally
1491 * differently. For example, a shell spawning a compiler process. We want
1492 * to learn that the compiler is behaving badly very quickly.
1495 sched_interact_fork(struct thread *td)
1500 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1501 if (sum > SCHED_SLP_RUN_FORK) {
1502 ratio = sum / SCHED_SLP_RUN_FORK;
1503 td->td_sched->ts_runtime /= ratio;
1504 td->td_sched->ts_slptime /= ratio;
1509 * Called from proc0_init() to setup the scheduler fields.
1516 * Set up the scheduler specific parts of proc0.
1518 proc0.p_sched = NULL; /* XXX */
1519 thread0.td_sched = &td_sched0;
1520 td_sched0.ts_ltick = ticks;
1521 td_sched0.ts_ftick = ticks;
1522 td_sched0.ts_slice = sched_slice;
1526 * This is only somewhat accurate since given many processes of the same
1527 * priority they will switch when their slices run out, which will be
1528 * at most sched_slice stathz ticks.
1531 sched_rr_interval(void)
1534 /* Convert sched_slice to hz */
1535 return (hz/(realstathz/sched_slice));
1539 * Update the percent cpu tracking information when it is requested or
1540 * the total history exceeds the maximum. We keep a sliding history of
1541 * tick counts that slowly decays. This is less precise than the 4BSD
1542 * mechanism since it happens with less regular and frequent events.
1545 sched_pctcpu_update(struct td_sched *ts)
1548 if (ts->ts_ticks == 0)
1550 if (ticks - (hz / 10) < ts->ts_ltick &&
1551 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1554 * Adjust counters and watermark for pctcpu calc.
1556 if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1557 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1561 ts->ts_ltick = ticks;
1562 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1566 * Adjust the priority of a thread. Move it to the appropriate run-queue
1567 * if necessary. This is the back-end for several priority related
1571 sched_thread_priority(struct thread *td, u_char prio)
1573 struct td_sched *ts;
1577 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1578 "prio:%d", td->td_priority, "new prio:%d", prio,
1579 KTR_ATTR_LINKED, sched_tdname(curthread));
1580 if (td != curthread && prio > td->td_priority) {
1581 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1582 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1583 prio, KTR_ATTR_LINKED, sched_tdname(td));
1586 THREAD_LOCK_ASSERT(td, MA_OWNED);
1587 if (td->td_priority == prio)
1590 * If the priority has been elevated due to priority
1591 * propagation, we may have to move ourselves to a new
1592 * queue. This could be optimized to not re-add in some
1595 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1597 td->td_priority = prio;
1598 sched_add(td, SRQ_BORROWING);
1602 * If the thread is currently running we may have to adjust the lowpri
1603 * information so other cpus are aware of our current priority.
1605 if (TD_IS_RUNNING(td)) {
1606 tdq = TDQ_CPU(ts->ts_cpu);
1607 oldpri = td->td_priority;
1608 td->td_priority = prio;
1609 if (prio < tdq->tdq_lowpri)
1610 tdq->tdq_lowpri = prio;
1611 else if (tdq->tdq_lowpri == oldpri)
1612 tdq_setlowpri(tdq, td);
1615 td->td_priority = prio;
1619 * Update a thread's priority when it is lent another thread's
1623 sched_lend_prio(struct thread *td, u_char prio)
1626 td->td_flags |= TDF_BORROWING;
1627 sched_thread_priority(td, prio);
1631 * Restore a thread's priority when priority propagation is
1632 * over. The prio argument is the minimum priority the thread
1633 * needs to have to satisfy other possible priority lending
1634 * requests. If the thread's regular priority is less
1635 * important than prio, the thread will keep a priority boost
1639 sched_unlend_prio(struct thread *td, u_char prio)
1643 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1644 td->td_base_pri <= PRI_MAX_TIMESHARE)
1645 base_pri = td->td_user_pri;
1647 base_pri = td->td_base_pri;
1648 if (prio >= base_pri) {
1649 td->td_flags &= ~TDF_BORROWING;
1650 sched_thread_priority(td, base_pri);
1652 sched_lend_prio(td, prio);
1656 * Standard entry for setting the priority to an absolute value.
1659 sched_prio(struct thread *td, u_char prio)
1663 /* First, update the base priority. */
1664 td->td_base_pri = prio;
1667 * If the thread is borrowing another thread's priority, don't
1668 * ever lower the priority.
1670 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1673 /* Change the real priority. */
1674 oldprio = td->td_priority;
1675 sched_thread_priority(td, prio);
1678 * If the thread is on a turnstile, then let the turnstile update
1681 if (TD_ON_LOCK(td) && oldprio != prio)
1682 turnstile_adjust(td, oldprio);
1686 * Set the base user priority, does not effect current running priority.
1689 sched_user_prio(struct thread *td, u_char prio)
1693 td->td_base_user_pri = prio;
1694 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1696 oldprio = td->td_user_pri;
1697 td->td_user_pri = prio;
1701 sched_lend_user_prio(struct thread *td, u_char prio)
1705 THREAD_LOCK_ASSERT(td, MA_OWNED);
1706 td->td_flags |= TDF_UBORROWING;
1707 oldprio = td->td_user_pri;
1708 td->td_user_pri = prio;
1712 sched_unlend_user_prio(struct thread *td, u_char prio)
1716 THREAD_LOCK_ASSERT(td, MA_OWNED);
1717 base_pri = td->td_base_user_pri;
1718 if (prio >= base_pri) {
1719 td->td_flags &= ~TDF_UBORROWING;
1720 sched_user_prio(td, base_pri);
1722 sched_lend_user_prio(td, prio);
1727 * Handle migration from sched_switch(). This happens only for
1731 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1735 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1737 tdq_load_rem(tdq, td);
1739 * Do the lock dance required to avoid LOR. We grab an extra
1740 * spinlock nesting to prevent preemption while we're
1741 * not holding either run-queue lock.
1744 thread_lock_block(td); /* This releases the lock on tdq. */
1747 * Acquire both run-queue locks before placing the thread on the new
1748 * run-queue to avoid deadlocks created by placing a thread with a
1749 * blocked lock on the run-queue of a remote processor. The deadlock
1750 * occurs when a third processor attempts to lock the two queues in
1751 * question while the target processor is spinning with its own
1752 * run-queue lock held while waiting for the blocked lock to clear.
1754 tdq_lock_pair(tdn, tdq);
1755 tdq_add(tdn, td, flags);
1756 tdq_notify(tdn, td);
1760 return (TDQ_LOCKPTR(tdn));
1764 * Variadic version of thread_lock_unblock() that does not assume td_lock
1768 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1770 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1775 * Switch threads. This function has to handle threads coming in while
1776 * blocked for some reason, running, or idle. It also must deal with
1777 * migrating a thread from one queue to another as running threads may
1778 * be assigned elsewhere via binding.
1781 sched_switch(struct thread *td, struct thread *newtd, int flags)
1784 struct td_sched *ts;
1789 THREAD_LOCK_ASSERT(td, MA_OWNED);
1790 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1792 cpuid = PCPU_GET(cpuid);
1793 tdq = TDQ_CPU(cpuid);
1796 ts->ts_rltick = ticks;
1797 td->td_lastcpu = td->td_oncpu;
1798 td->td_oncpu = NOCPU;
1799 if (!(flags & SW_PREEMPT))
1800 td->td_flags &= ~TDF_NEEDRESCHED;
1801 td->td_owepreempt = 0;
1802 tdq->tdq_switchcnt++;
1804 * The lock pointer in an idle thread should never change. Reset it
1805 * to CAN_RUN as well.
1807 if (TD_IS_IDLETHREAD(td)) {
1808 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1810 } else if (TD_IS_RUNNING(td)) {
1811 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1812 srqflag = (flags & SW_PREEMPT) ?
1813 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1814 SRQ_OURSELF|SRQ_YIELDING;
1816 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1817 ts->ts_cpu = sched_pickcpu(td, 0);
1819 if (ts->ts_cpu == cpuid)
1820 tdq_runq_add(tdq, td, srqflag);
1822 KASSERT(THREAD_CAN_MIGRATE(td) ||
1823 (ts->ts_flags & TSF_BOUND) != 0,
1824 ("Thread %p shouldn't migrate", td));
1825 mtx = sched_switch_migrate(tdq, td, srqflag);
1828 /* This thread must be going to sleep. */
1830 mtx = thread_lock_block(td);
1831 tdq_load_rem(tdq, td);
1834 * We enter here with the thread blocked and assigned to the
1835 * appropriate cpu run-queue or sleep-queue and with the current
1836 * thread-queue locked.
1838 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1839 newtd = choosethread();
1841 * Call the MD code to switch contexts if necessary.
1845 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1846 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1848 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1849 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1851 #ifdef KDTRACE_HOOKS
1853 * If DTrace has set the active vtime enum to anything
1854 * other than INACTIVE (0), then it should have set the
1857 if (dtrace_vtime_active)
1858 (*dtrace_vtime_switch_func)(newtd);
1861 cpu_switch(td, newtd, mtx);
1863 * We may return from cpu_switch on a different cpu. However,
1864 * we always return with td_lock pointing to the current cpu's
1867 cpuid = PCPU_GET(cpuid);
1868 tdq = TDQ_CPU(cpuid);
1869 lock_profile_obtain_lock_success(
1870 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1872 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1873 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1876 thread_unblock_switch(td, mtx);
1878 * Assert that all went well and return.
1880 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1881 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1882 td->td_oncpu = cpuid;
1886 * Adjust thread priorities as a result of a nice request.
1889 sched_nice(struct proc *p, int nice)
1893 PROC_LOCK_ASSERT(p, MA_OWNED);
1896 FOREACH_THREAD_IN_PROC(p, td) {
1899 sched_prio(td, td->td_base_user_pri);
1905 * Record the sleep time for the interactivity scorer.
1908 sched_sleep(struct thread *td, int prio)
1911 THREAD_LOCK_ASSERT(td, MA_OWNED);
1913 td->td_slptick = ticks;
1914 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1915 td->td_flags |= TDF_CANSWAP;
1916 if (static_boost == 1 && prio)
1917 sched_prio(td, prio);
1918 else if (static_boost && td->td_priority > static_boost)
1919 sched_prio(td, static_boost);
1923 * Schedule a thread to resume execution and record how long it voluntarily
1924 * slept. We also update the pctcpu, interactivity, and priority.
1927 sched_wakeup(struct thread *td)
1929 struct td_sched *ts;
1932 THREAD_LOCK_ASSERT(td, MA_OWNED);
1934 td->td_flags &= ~TDF_CANSWAP;
1936 * If we slept for more than a tick update our interactivity and
1939 slptick = td->td_slptick;
1941 if (slptick && slptick != ticks) {
1944 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1945 ts->ts_slptime += hzticks;
1946 sched_interact_update(td);
1947 sched_pctcpu_update(ts);
1949 /* Reset the slice value after we sleep. */
1950 ts->ts_slice = sched_slice;
1951 sched_add(td, SRQ_BORING);
1955 * Penalize the parent for creating a new child and initialize the child's
1959 sched_fork(struct thread *td, struct thread *child)
1961 THREAD_LOCK_ASSERT(td, MA_OWNED);
1962 sched_fork_thread(td, child);
1964 * Penalize the parent and child for forking.
1966 sched_interact_fork(child);
1967 sched_priority(child);
1968 td->td_sched->ts_runtime += tickincr;
1969 sched_interact_update(td);
1974 * Fork a new thread, may be within the same process.
1977 sched_fork_thread(struct thread *td, struct thread *child)
1979 struct td_sched *ts;
1980 struct td_sched *ts2;
1982 THREAD_LOCK_ASSERT(td, MA_OWNED);
1987 ts2 = child->td_sched;
1988 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1989 child->td_cpuset = cpuset_ref(td->td_cpuset);
1990 ts2->ts_cpu = ts->ts_cpu;
1993 * Grab our parents cpu estimation information.
1995 ts2->ts_ticks = ts->ts_ticks;
1996 ts2->ts_ltick = ts->ts_ltick;
1997 ts2->ts_incrtick = ts->ts_incrtick;
1998 ts2->ts_ftick = ts->ts_ftick;
2000 * Do not inherit any borrowed priority from the parent.
2002 child->td_priority = child->td_base_pri;
2004 * And update interactivity score.
2006 ts2->ts_slptime = ts->ts_slptime;
2007 ts2->ts_runtime = ts->ts_runtime;
2008 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
2010 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2015 * Adjust the priority class of a thread.
2018 sched_class(struct thread *td, int class)
2021 THREAD_LOCK_ASSERT(td, MA_OWNED);
2022 if (td->td_pri_class == class)
2024 td->td_pri_class = class;
2028 * Return some of the child's priority and interactivity to the parent.
2031 sched_exit(struct proc *p, struct thread *child)
2035 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2036 "prio:%d", child->td_priority);
2037 PROC_LOCK_ASSERT(p, MA_OWNED);
2038 td = FIRST_THREAD_IN_PROC(p);
2039 sched_exit_thread(td, child);
2043 * Penalize another thread for the time spent on this one. This helps to
2044 * worsen the priority and interactivity of processes which schedule batch
2045 * jobs such as make. This has little effect on the make process itself but
2046 * causes new processes spawned by it to receive worse scores immediately.
2049 sched_exit_thread(struct thread *td, struct thread *child)
2052 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2053 "prio:%d", child->td_priority);
2055 * Give the child's runtime to the parent without returning the
2056 * sleep time as a penalty to the parent. This causes shells that
2057 * launch expensive things to mark their children as expensive.
2060 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2061 sched_interact_update(td);
2067 sched_preempt(struct thread *td)
2073 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2074 tdq->tdq_ipipending = 0;
2075 if (td->td_priority > tdq->tdq_lowpri) {
2078 flags = SW_INVOL | SW_PREEMPT;
2079 if (td->td_critnest > 1)
2080 td->td_owepreempt = 1;
2081 else if (TD_IS_IDLETHREAD(td))
2082 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2084 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2090 * Fix priorities on return to user-space. Priorities may be elevated due
2091 * to static priorities in msleep() or similar.
2094 sched_userret(struct thread *td)
2097 * XXX we cheat slightly on the locking here to avoid locking in
2098 * the usual case. Setting td_priority here is essentially an
2099 * incomplete workaround for not setting it properly elsewhere.
2100 * Now that some interrupt handlers are threads, not setting it
2101 * properly elsewhere can clobber it in the window between setting
2102 * it here and returning to user mode, so don't waste time setting
2103 * it perfectly here.
2105 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2106 ("thread with borrowed priority returning to userland"));
2107 if (td->td_priority != td->td_user_pri) {
2109 td->td_priority = td->td_user_pri;
2110 td->td_base_pri = td->td_user_pri;
2111 tdq_setlowpri(TDQ_SELF(), td);
2117 * Handle a stathz tick. This is really only relevant for timeshare
2121 sched_clock(struct thread *td)
2124 struct td_sched *ts;
2126 THREAD_LOCK_ASSERT(td, MA_OWNED);
2130 * We run the long term load balancer infrequently on the first cpu.
2132 if (balance_tdq == tdq) {
2133 if (balance_ticks && --balance_ticks == 0)
2138 * Save the old switch count so we have a record of the last ticks
2139 * activity. Initialize the new switch count based on our load.
2140 * If there is some activity seed it to reflect that.
2142 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2143 tdq->tdq_switchcnt = tdq->tdq_load;
2145 * Advance the insert index once for each tick to ensure that all
2146 * threads get a chance to run.
2148 if (tdq->tdq_idx == tdq->tdq_ridx) {
2149 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2150 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2151 tdq->tdq_ridx = tdq->tdq_idx;
2154 if (td->td_pri_class & PRI_FIFO_BIT)
2156 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2158 * We used a tick; charge it to the thread so
2159 * that we can compute our interactivity.
2161 td->td_sched->ts_runtime += tickincr;
2162 sched_interact_update(td);
2166 * We used up one time slice.
2168 if (--ts->ts_slice > 0)
2171 * We're out of time, force a requeue at userret().
2173 ts->ts_slice = sched_slice;
2174 td->td_flags |= TDF_NEEDRESCHED;
2178 * Called once per hz tick. Used for cpu utilization information. This
2179 * is easier than trying to scale based on stathz.
2184 struct td_sched *ts;
2186 ts = curthread->td_sched;
2188 * Ticks is updated asynchronously on a single cpu. Check here to
2189 * avoid incrementing ts_ticks multiple times in a single tick.
2191 if (ts->ts_incrtick == ticks)
2193 /* Adjust ticks for pctcpu */
2194 ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2195 ts->ts_ltick = ticks;
2196 ts->ts_incrtick = ticks;
2198 * Update if we've exceeded our desired tick threshold by over one
2201 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2202 sched_pctcpu_update(ts);
2206 * Return whether the current CPU has runnable tasks. Used for in-kernel
2207 * cooperative idle threads.
2210 sched_runnable(void)
2218 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2219 if (tdq->tdq_load > 0)
2222 if (tdq->tdq_load - 1 > 0)
2230 * Choose the highest priority thread to run. The thread is removed from
2231 * the run-queue while running however the load remains. For SMP we set
2232 * the tdq in the global idle bitmask if it idles here.
2241 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2242 td = tdq_choose(tdq);
2244 td->td_sched->ts_ltick = ticks;
2245 tdq_runq_rem(tdq, td);
2246 tdq->tdq_lowpri = td->td_priority;
2249 tdq->tdq_lowpri = PRI_MAX_IDLE;
2250 return (PCPU_GET(idlethread));
2254 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2255 * we always request it once we exit a critical section.
2258 sched_setpreempt(struct thread *td)
2264 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2267 pri = td->td_priority;
2268 cpri = ctd->td_priority;
2270 ctd->td_flags |= TDF_NEEDRESCHED;
2271 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2273 if (!sched_shouldpreempt(pri, cpri, 0))
2275 ctd->td_owepreempt = 1;
2279 * Add a thread to a thread queue. Select the appropriate runq and add the
2280 * thread to it. This is the internal function called when the tdq is
2284 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2287 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2288 KASSERT((td->td_inhibitors == 0),
2289 ("sched_add: trying to run inhibited thread"));
2290 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2291 ("sched_add: bad thread state"));
2292 KASSERT(td->td_flags & TDF_INMEM,
2293 ("sched_add: thread swapped out"));
2295 if (td->td_priority < tdq->tdq_lowpri)
2296 tdq->tdq_lowpri = td->td_priority;
2297 tdq_runq_add(tdq, td, flags);
2298 tdq_load_add(tdq, td);
2302 * Select the target thread queue and add a thread to it. Request
2303 * preemption or IPI a remote processor if required.
2306 sched_add(struct thread *td, int flags)
2313 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2314 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2315 sched_tdname(curthread));
2316 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2317 KTR_ATTR_LINKED, sched_tdname(td));
2318 THREAD_LOCK_ASSERT(td, MA_OWNED);
2320 * Recalculate the priority before we select the target cpu or
2323 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2327 * Pick the destination cpu and if it isn't ours transfer to the
2330 cpu = sched_pickcpu(td, flags);
2331 tdq = sched_setcpu(td, cpu, flags);
2332 tdq_add(tdq, td, flags);
2333 if (cpu != PCPU_GET(cpuid)) {
2334 tdq_notify(tdq, td);
2341 * Now that the thread is moving to the run-queue, set the lock
2342 * to the scheduler's lock.
2344 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2345 tdq_add(tdq, td, flags);
2347 if (!(flags & SRQ_YIELDING))
2348 sched_setpreempt(td);
2352 * Remove a thread from a run-queue without running it. This is used
2353 * when we're stealing a thread from a remote queue. Otherwise all threads
2354 * exit by calling sched_exit_thread() and sched_throw() themselves.
2357 sched_rem(struct thread *td)
2361 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2362 "prio:%d", td->td_priority);
2363 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2364 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2365 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2366 KASSERT(TD_ON_RUNQ(td),
2367 ("sched_rem: thread not on run queue"));
2368 tdq_runq_rem(tdq, td);
2369 tdq_load_rem(tdq, td);
2371 if (td->td_priority == tdq->tdq_lowpri)
2372 tdq_setlowpri(tdq, NULL);
2376 * Fetch cpu utilization information. Updates on demand.
2379 sched_pctcpu(struct thread *td)
2382 struct td_sched *ts;
2389 THREAD_LOCK_ASSERT(td, MA_OWNED);
2393 sched_pctcpu_update(ts);
2394 /* How many rtick per second ? */
2395 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2396 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2403 * Enforce affinity settings for a thread. Called after adjustments to
2407 sched_affinity(struct thread *td)
2410 struct td_sched *ts;
2412 THREAD_LOCK_ASSERT(td, MA_OWNED);
2414 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2416 if (TD_ON_RUNQ(td)) {
2418 sched_add(td, SRQ_BORING);
2421 if (!TD_IS_RUNNING(td))
2424 * Force a switch before returning to userspace. If the
2425 * target thread is not running locally send an ipi to force
2428 td->td_flags |= TDF_NEEDRESCHED;
2429 if (td != curthread)
2430 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2435 * Bind a thread to a target cpu.
2438 sched_bind(struct thread *td, int cpu)
2440 struct td_sched *ts;
2442 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2443 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2445 if (ts->ts_flags & TSF_BOUND)
2447 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
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);
2466 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2468 if ((ts->ts_flags & TSF_BOUND) == 0)
2470 ts->ts_flags &= ~TSF_BOUND;
2475 sched_is_bound(struct thread *td)
2477 THREAD_LOCK_ASSERT(td, MA_OWNED);
2478 return (td->td_sched->ts_flags & TSF_BOUND);
2485 sched_relinquish(struct thread *td)
2488 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2493 * Return the total system load.
2504 total += TDQ_CPU(i)->tdq_sysload;
2507 return (TDQ_SELF()->tdq_sysload);
2512 sched_sizeof_proc(void)
2514 return (sizeof(struct proc));
2518 sched_sizeof_thread(void)
2520 return (sizeof(struct thread) + sizeof(struct td_sched));
2524 #define TDQ_IDLESPIN(tdq) \
2525 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2527 #define TDQ_IDLESPIN(tdq) 1
2531 * The actual idle process.
2534 sched_idletd(void *dummy)
2541 mtx_assert(&Giant, MA_NOTOWNED);
2546 if (tdq_idled(tdq) == 0)
2549 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2551 * If we're switching very frequently, spin while checking
2552 * for load rather than entering a low power state that
2553 * may require an IPI. However, don't do any busy
2554 * loops while on SMT machines as this simply steals
2555 * cycles from cores doing useful work.
2557 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2558 for (i = 0; i < sched_idlespins; i++) {
2564 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2565 if (tdq->tdq_load == 0)
2566 cpu_idle(switchcnt > 1);
2567 if (tdq->tdq_load) {
2569 mi_switch(SW_VOL | SWT_IDLE, NULL);
2576 * A CPU is entering for the first time or a thread is exiting.
2579 sched_throw(struct thread *td)
2581 struct thread *newtd;
2586 /* Correct spinlock nesting and acquire the correct lock. */
2589 PCPU_SET(switchtime, cpu_ticks());
2590 PCPU_SET(switchticks, ticks);
2592 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2593 tdq_load_rem(tdq, td);
2594 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2596 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2597 newtd = choosethread();
2598 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2599 cpu_throw(td, newtd); /* doesn't return */
2603 * This is called from fork_exit(). Just acquire the correct locks and
2604 * let fork do the rest of the work.
2607 sched_fork_exit(struct thread *td)
2609 struct td_sched *ts;
2614 * Finish setting up thread glue so that it begins execution in a
2615 * non-nested critical section with the scheduler lock held.
2617 cpuid = PCPU_GET(cpuid);
2618 tdq = TDQ_CPU(cpuid);
2620 if (TD_IS_IDLETHREAD(td))
2621 td->td_lock = TDQ_LOCKPTR(tdq);
2622 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2623 td->td_oncpu = cpuid;
2624 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2625 lock_profile_obtain_lock_success(
2626 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2630 * Create on first use to catch odd startup conditons.
2633 sched_tdname(struct thread *td)
2636 struct td_sched *ts;
2639 if (ts->ts_name[0] == '\0')
2640 snprintf(ts->ts_name, sizeof(ts->ts_name),
2641 "%s tid %d", td->td_name, td->td_tid);
2642 return (ts->ts_name);
2644 return (td->td_name);
2651 * Build the CPU topology dump string. Is recursively called to collect
2652 * the topology tree.
2655 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2660 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2661 "", 1 + indent / 2, cg->cg_level);
2662 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2663 cg->cg_count, cg->cg_mask);
2665 for (i = 0; i < MAXCPU; i++) {
2666 if ((cg->cg_mask & (1 << i)) != 0) {
2668 sbuf_printf(sb, ", ");
2671 sbuf_printf(sb, "%d", i);
2674 sbuf_printf(sb, "</cpu>\n");
2676 if (cg->cg_flags != 0) {
2677 sbuf_printf(sb, "%*s <flags>", indent, "");
2678 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2679 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2680 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2681 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2682 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2683 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2684 sbuf_printf(sb, "</flags>\n");
2687 if (cg->cg_children > 0) {
2688 sbuf_printf(sb, "%*s <children>\n", indent, "");
2689 for (i = 0; i < cg->cg_children; i++)
2690 sysctl_kern_sched_topology_spec_internal(sb,
2691 &cg->cg_child[i], indent+2);
2692 sbuf_printf(sb, "%*s </children>\n", indent, "");
2694 sbuf_printf(sb, "%*s</group>\n", indent, "");
2699 * Sysctl handler for retrieving topology dump. It's a wrapper for
2700 * the recursive sysctl_kern_smp_topology_spec_internal().
2703 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2708 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2710 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2714 sbuf_printf(topo, "<groups>\n");
2715 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2716 sbuf_printf(topo, "</groups>\n");
2720 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2727 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2728 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2730 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2731 "Slice size for timeshare threads");
2732 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2733 "Interactivity score threshold");
2734 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2735 0,"Min priority for preemption, lower priorities have greater precedence");
2736 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2737 0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2738 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2739 0,"Number of times idle will spin waiting for new work.");
2740 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2741 0,"Threshold before we will permit idle spinning.");
2743 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2744 "Number of hz ticks to keep thread affinity for");
2745 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2746 "Enables the long-term load balancer");
2747 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2748 &balance_interval, 0,
2749 "Average frequency in stathz ticks to run the long-term balancer");
2750 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2751 "Steals work from another hyper-threaded core on idle");
2752 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2753 "Attempts to steal work from other cores before idling");
2754 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2755 "Minimum load on remote cpu before we'll steal");
2757 /* Retrieve SMP topology */
2758 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2759 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2760 "XML dump of detected CPU topology");
2763 /* ps compat. All cpu percentages from ULE are weighted. */
2764 static int ccpu = 0;
2765 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");