2 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice unmodified, this list of conditions, and the following
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 * This file implements the ULE scheduler. ULE supports independent CPU
29 * run queues and fine grain locking. It has superior interactive
30 * performance under load even on uni-processor systems.
33 * ULE is the last three letters in schedule. It owes its name to a
34 * generic user created for a scheduling system by Paul Mikesell at
35 * Isilon Systems and a general lack of creativity on the part of the author.
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_kdtrace.h"
43 #include "opt_sched.h"
45 #include <sys/param.h>
46 #include <sys/systm.h>
48 #include <sys/kernel.h>
51 #include <sys/mutex.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
68 #include <sys/pmckern.h>
72 #include <sys/dtrace_bsd.h>
73 int dtrace_vtime_active;
74 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
80 #if defined(__powerpc__) && defined(BOOKE_E500)
81 #error "This architecture is not currently compatible with ULE"
86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
87 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
88 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
91 * Thread scheduler specific section. All fields are protected
95 struct runq *ts_runq; /* Run-queue we're queued on. */
96 short ts_flags; /* TSF_* flags. */
97 u_char ts_cpu; /* CPU that we have affinity for. */
98 int ts_rltick; /* Real last tick, for affinity. */
99 int ts_slice; /* Ticks of slice remaining. */
100 u_int ts_slptime; /* Number of ticks we vol. slept */
101 u_int ts_runtime; /* Number of ticks we were running */
102 int ts_ltick; /* Last tick that we were running on */
103 int ts_ftick; /* First tick that we were running on */
104 int ts_ticks; /* Tick count */
106 char ts_name[TS_NAME_LEN];
109 /* flags kept in ts_flags */
110 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
111 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
113 static struct td_sched td_sched0;
115 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
116 #define THREAD_CAN_SCHED(td, cpu) \
117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
120 * Priority ranges used for interactive and non-interactive timeshare
121 * threads. The timeshare priorities are split up into four ranges.
122 * The first range handles interactive threads. The last three ranges
123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124 * ranges supporting nice values.
126 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
130 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
131 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
136 * Cpu percentage computation macros and defines.
138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
145 #define SCHED_TICK_SECS 10
146 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
147 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
148 #define SCHED_TICK_SHIFT 10
149 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
153 * These macros determine priorities for non-interactive threads. They are
154 * assigned a priority based on their recent cpu utilization as expressed
155 * by the ratio of ticks to the tick total. NHALF priorities at the start
156 * and end of the MIN to MAX timeshare range are only reachable with negative
157 * or positive nice respectively.
159 * PRI_RANGE: Priority range for utilization dependent priorities.
160 * PRI_NRESV: Number of nice values.
161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
162 * PRI_NICE: Determines the part of the priority inherited from nice.
164 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
165 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
166 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define SCHED_PRI_TICKS(ts) \
170 (SCHED_TICK_HZ((ts)) / \
171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define SCHED_PRI_NICE(nice) (nice)
175 * These determine the interactivity of a process. Interactivity differs from
176 * cpu utilization in that it expresses the voluntary time slept vs time ran
177 * while cpu utilization includes all time not running. This more accurately
178 * models the intent of the thread.
180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
181 * before throttling back.
182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
183 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
184 * INTERACT_THRESH: Threshold for placement on the current runq.
186 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
187 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
188 #define SCHED_INTERACT_MAX (100)
189 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
190 #define SCHED_INTERACT_THRESH (30)
193 * These parameters determine the slice behavior for batch work.
195 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
196 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
198 /* Flags kept in td_flags. */
199 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
202 * tickincr: Converts a stathz tick into a hz domain scaled by
203 * the shift factor. Without the shift the error rate
204 * due to rounding would be unacceptably high.
205 * realstathz: stathz is sometimes 0 and run off of hz.
206 * sched_slice: Runtime of each thread before rescheduling.
207 * preempt_thresh: Priority threshold for preemption and remote IPIs.
209 static int sched_interact = SCHED_INTERACT_THRESH;
210 static int tickincr = 8 << SCHED_TICK_SHIFT;
211 static int realstathz = 127; /* reset during boot. */
212 static int sched_slice = 10; /* reset during boot. */
213 static int sched_slice_min = 1; /* reset during boot. */
215 #ifdef FULL_PREEMPTION
216 static int preempt_thresh = PRI_MAX_IDLE;
218 static int preempt_thresh = PRI_MIN_KERN;
221 static int preempt_thresh = 0;
223 static int static_boost = PRI_MIN_BATCH;
224 static int sched_idlespins = 10000;
225 static int sched_idlespinthresh = -1;
228 * tdq - per processor runqs and statistics. All fields are protected by the
229 * tdq_lock. The load and lowpri may be accessed without to avoid excess
230 * locking in sched_pickcpu();
234 * Ordered to improve efficiency of cpu_search() and switch().
235 * tdq_lock is padded to avoid false sharing with tdq_load and
238 struct mtx_padalign tdq_lock; /* run queue lock. */
239 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
240 volatile int tdq_load; /* Aggregate load. */
241 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
242 int tdq_sysload; /* For loadavg, !ITHD load. */
243 int tdq_transferable; /* Transferable thread count. */
244 short tdq_switchcnt; /* Switches this tick. */
245 short tdq_oldswitchcnt; /* Switches last tick. */
246 u_char tdq_lowpri; /* Lowest priority thread. */
247 u_char tdq_ipipending; /* IPI pending. */
248 u_char tdq_idx; /* Current insert index. */
249 u_char tdq_ridx; /* Current removal index. */
250 struct runq tdq_realtime; /* real-time run queue. */
251 struct runq tdq_timeshare; /* timeshare run queue. */
252 struct runq tdq_idle; /* Queue of IDLE threads. */
253 char tdq_name[TDQ_NAME_LEN];
255 char tdq_loadname[TDQ_LOADNAME_LEN];
259 /* Idle thread states and config. */
260 #define TDQ_RUNNING 1
264 struct cpu_group *cpu_top; /* CPU topology */
266 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
267 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
272 static int rebalance = 1;
273 static int balance_interval = 128; /* Default set in sched_initticks(). */
275 static int steal_idle = 1;
276 static int steal_thresh = 2;
279 * One thread queue per processor.
281 static struct tdq tdq_cpu[MAXCPU];
282 static struct tdq *balance_tdq;
283 static int balance_ticks;
284 static DPCPU_DEFINE(uint32_t, randomval);
286 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
287 #define TDQ_CPU(x) (&tdq_cpu[(x)])
288 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
290 static struct tdq tdq_cpu;
292 #define TDQ_ID(x) (0)
293 #define TDQ_SELF() (&tdq_cpu)
294 #define TDQ_CPU(x) (&tdq_cpu)
297 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
298 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
299 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
300 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
301 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
303 static void sched_priority(struct thread *);
304 static void sched_thread_priority(struct thread *, u_char);
305 static int sched_interact_score(struct thread *);
306 static void sched_interact_update(struct thread *);
307 static void sched_interact_fork(struct thread *);
308 static void sched_pctcpu_update(struct td_sched *, int);
310 /* Operations on per processor queues */
311 static struct thread *tdq_choose(struct tdq *);
312 static void tdq_setup(struct tdq *);
313 static void tdq_load_add(struct tdq *, struct thread *);
314 static void tdq_load_rem(struct tdq *, struct thread *);
315 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
316 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
317 static inline int sched_shouldpreempt(int, int, int);
318 void tdq_print(int cpu);
319 static void runq_print(struct runq *rq);
320 static void tdq_add(struct tdq *, struct thread *, int);
322 static int tdq_move(struct tdq *, struct tdq *);
323 static int tdq_idled(struct tdq *);
324 static void tdq_notify(struct tdq *, struct thread *);
325 static struct thread *tdq_steal(struct tdq *, int);
326 static struct thread *runq_steal(struct runq *, int);
327 static int sched_pickcpu(struct thread *, int);
328 static void sched_balance(void);
329 static int sched_balance_pair(struct tdq *, struct tdq *);
330 static inline struct tdq *sched_setcpu(struct thread *, int, int);
331 static inline void thread_unblock_switch(struct thread *, struct mtx *);
332 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
333 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
334 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
335 struct cpu_group *cg, int indent);
338 static void sched_setup(void *dummy);
339 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
341 static void sched_initticks(void *dummy);
342 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
345 SDT_PROVIDER_DEFINE(sched);
347 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
348 "struct proc *", "uint8_t");
349 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
350 "struct proc *", "void *");
351 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
352 "struct proc *", "void *", "int");
353 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
354 "struct proc *", "uint8_t", "struct thread *");
355 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
356 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
358 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
359 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
360 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
364 * Print the threads waiting on a run-queue.
367 runq_print(struct runq *rq)
375 for (i = 0; i < RQB_LEN; i++) {
376 printf("\t\trunq bits %d 0x%zx\n",
377 i, rq->rq_status.rqb_bits[i]);
378 for (j = 0; j < RQB_BPW; j++)
379 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
380 pri = j + (i << RQB_L2BPW);
381 rqh = &rq->rq_queues[pri];
382 TAILQ_FOREACH(td, rqh, td_runq) {
383 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
384 td, td->td_name, td->td_priority,
385 td->td_rqindex, pri);
392 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
401 printf("tdq %d:\n", TDQ_ID(tdq));
402 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
403 printf("\tLock name: %s\n", tdq->tdq_name);
404 printf("\tload: %d\n", tdq->tdq_load);
405 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
406 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
407 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
408 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
409 printf("\tload transferable: %d\n", tdq->tdq_transferable);
410 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
411 printf("\trealtime runq:\n");
412 runq_print(&tdq->tdq_realtime);
413 printf("\ttimeshare runq:\n");
414 runq_print(&tdq->tdq_timeshare);
415 printf("\tidle runq:\n");
416 runq_print(&tdq->tdq_idle);
420 sched_shouldpreempt(int pri, int cpri, int remote)
423 * If the new priority is not better than the current priority there is
429 * Always preempt idle.
431 if (cpri >= PRI_MIN_IDLE)
434 * If preemption is disabled don't preempt others.
436 if (preempt_thresh == 0)
439 * Preempt if we exceed the threshold.
441 if (pri <= preempt_thresh)
444 * If we're interactive or better and there is non-interactive
445 * or worse running preempt only remote processors.
447 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
453 * Add a thread to the actual run-queue. Keeps transferable counts up to
454 * date with what is actually on the run-queue. Selects the correct
455 * queue position for timeshare threads.
458 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
463 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
464 THREAD_LOCK_ASSERT(td, MA_OWNED);
466 pri = td->td_priority;
469 if (THREAD_CAN_MIGRATE(td)) {
470 tdq->tdq_transferable++;
471 ts->ts_flags |= TSF_XFERABLE;
473 if (pri < PRI_MIN_BATCH) {
474 ts->ts_runq = &tdq->tdq_realtime;
475 } else if (pri <= PRI_MAX_BATCH) {
476 ts->ts_runq = &tdq->tdq_timeshare;
477 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
478 ("Invalid priority %d on timeshare runq", pri));
480 * This queue contains only priorities between MIN and MAX
481 * realtime. Use the whole queue to represent these values.
483 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
484 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
485 pri = (pri + tdq->tdq_idx) % RQ_NQS;
487 * This effectively shortens the queue by one so we
488 * can have a one slot difference between idx and
489 * ridx while we wait for threads to drain.
491 if (tdq->tdq_ridx != tdq->tdq_idx &&
492 pri == tdq->tdq_ridx)
493 pri = (unsigned char)(pri - 1) % RQ_NQS;
496 runq_add_pri(ts->ts_runq, td, pri, flags);
499 ts->ts_runq = &tdq->tdq_idle;
500 runq_add(ts->ts_runq, td, flags);
504 * Remove a thread from a run-queue. This typically happens when a thread
505 * is selected to run. Running threads are not on the queue and the
506 * transferable count does not reflect them.
509 tdq_runq_rem(struct tdq *tdq, struct thread *td)
514 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
515 KASSERT(ts->ts_runq != NULL,
516 ("tdq_runq_remove: thread %p null ts_runq", td));
517 if (ts->ts_flags & TSF_XFERABLE) {
518 tdq->tdq_transferable--;
519 ts->ts_flags &= ~TSF_XFERABLE;
521 if (ts->ts_runq == &tdq->tdq_timeshare) {
522 if (tdq->tdq_idx != tdq->tdq_ridx)
523 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
525 runq_remove_idx(ts->ts_runq, td, NULL);
527 runq_remove(ts->ts_runq, td);
531 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
532 * for this thread to the referenced thread queue.
535 tdq_load_add(struct tdq *tdq, struct thread *td)
538 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
539 THREAD_LOCK_ASSERT(td, MA_OWNED);
542 if ((td->td_flags & TDF_NOLOAD) == 0)
544 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
545 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
549 * Remove the load from a thread that is transitioning to a sleep state or
553 tdq_load_rem(struct tdq *tdq, struct thread *td)
556 THREAD_LOCK_ASSERT(td, MA_OWNED);
557 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
558 KASSERT(tdq->tdq_load != 0,
559 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
562 if ((td->td_flags & TDF_NOLOAD) == 0)
564 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
565 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
569 * Bound timeshare latency by decreasing slice size as load increases. We
570 * consider the maximum latency as the sum of the threads waiting to run
571 * aside from curthread and target no more than sched_slice latency but
572 * no less than sched_slice_min runtime.
575 tdq_slice(struct tdq *tdq)
580 * It is safe to use sys_load here because this is called from
581 * contexts where timeshare threads are running and so there
582 * cannot be higher priority load in the system.
584 load = tdq->tdq_sysload - 1;
585 if (load >= SCHED_SLICE_MIN_DIVISOR)
586 return (sched_slice_min);
588 return (sched_slice);
589 return (sched_slice / load);
593 * Set lowpri to its exact value by searching the run-queue and
594 * evaluating curthread. curthread may be passed as an optimization.
597 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
601 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
603 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
604 td = tdq_choose(tdq);
605 if (td == NULL || td->td_priority > ctd->td_priority)
606 tdq->tdq_lowpri = ctd->td_priority;
608 tdq->tdq_lowpri = td->td_priority;
615 int cs_pri; /* Min priority for low. */
616 int cs_limit; /* Max load for low, min load for high. */
621 #define CPU_SEARCH_LOWEST 0x1
622 #define CPU_SEARCH_HIGHEST 0x2
623 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
625 #define CPUSET_FOREACH(cpu, mask) \
626 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
627 if (CPU_ISSET(cpu, &mask))
629 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
630 struct cpu_search *high, const int match);
631 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
632 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
633 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
634 struct cpu_search *high);
637 * Search the tree of cpu_groups for the lowest or highest loaded cpu
638 * according to the match argument. This routine actually compares the
639 * load on all paths through the tree and finds the least loaded cpu on
640 * the least loaded path, which may differ from the least loaded cpu in
641 * the system. This balances work among caches and busses.
643 * This inline is instantiated in three forms below using constants for the
644 * match argument. It is reduced to the minimum set for each case. It is
645 * also recursive to the depth of the tree.
648 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
649 struct cpu_search *high, const int match)
651 struct cpu_search lgroup;
652 struct cpu_search hgroup;
654 struct cpu_group *child;
656 int cpu, i, hload, lload, load, total, rnd, *rndptr;
659 cpumask = cg->cg_mask;
660 if (match & CPU_SEARCH_LOWEST) {
664 if (match & CPU_SEARCH_HIGHEST) {
669 /* Iterate through the child CPU groups and then remaining CPUs. */
670 for (i = cg->cg_children, cpu = mp_maxid; i >= 0; ) {
672 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
678 child = &cg->cg_child[i - 1];
680 if (match & CPU_SEARCH_LOWEST)
682 if (match & CPU_SEARCH_HIGHEST)
684 if (child) { /* Handle child CPU group. */
685 CPU_NAND(&cpumask, &child->cg_mask);
687 case CPU_SEARCH_LOWEST:
688 load = cpu_search_lowest(child, &lgroup);
690 case CPU_SEARCH_HIGHEST:
691 load = cpu_search_highest(child, &hgroup);
693 case CPU_SEARCH_BOTH:
694 load = cpu_search_both(child, &lgroup, &hgroup);
697 } else { /* Handle child CPU. */
699 load = tdq->tdq_load * 256;
700 rndptr = DPCPU_PTR(randomval);
701 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
702 if (match & CPU_SEARCH_LOWEST) {
703 if (cpu == low->cs_prefer)
705 /* If that CPU is allowed and get data. */
706 if (tdq->tdq_lowpri > lgroup.cs_pri &&
707 tdq->tdq_load <= lgroup.cs_limit &&
708 CPU_ISSET(cpu, &lgroup.cs_mask)) {
710 lgroup.cs_load = load - rnd;
713 if (match & CPU_SEARCH_HIGHEST)
714 if (tdq->tdq_load >= hgroup.cs_limit &&
715 tdq->tdq_transferable &&
716 CPU_ISSET(cpu, &hgroup.cs_mask)) {
718 hgroup.cs_load = load - rnd;
723 /* We have info about child item. Compare it. */
724 if (match & CPU_SEARCH_LOWEST) {
725 if (lgroup.cs_cpu >= 0 &&
727 (load == lload && lgroup.cs_load < low->cs_load))) {
729 low->cs_cpu = lgroup.cs_cpu;
730 low->cs_load = lgroup.cs_load;
733 if (match & CPU_SEARCH_HIGHEST)
734 if (hgroup.cs_cpu >= 0 &&
736 (load == hload && hgroup.cs_load > high->cs_load))) {
738 high->cs_cpu = hgroup.cs_cpu;
739 high->cs_load = hgroup.cs_load;
743 if (i == 0 && CPU_EMPTY(&cpumask))
752 * cpu_search instantiations must pass constants to maintain the inline
756 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
758 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
762 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
764 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
768 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
769 struct cpu_search *high)
771 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
775 * Find the cpu with the least load via the least loaded path that has a
776 * lowpri greater than pri pri. A pri of -1 indicates any priority is
780 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
783 struct cpu_search low;
786 low.cs_prefer = prefer;
789 low.cs_limit = maxload;
790 cpu_search_lowest(cg, &low);
795 * Find the cpu with the highest load via the highest loaded path.
798 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
800 struct cpu_search high;
804 high.cs_limit = minload;
805 cpu_search_highest(cg, &high);
810 * Simultaneously find the highest and lowest loaded cpu reachable via
814 sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
816 struct cpu_search high;
817 struct cpu_search low;
822 low.cs_limit = INT_MAX;
827 cpu_search_both(cg, &low, &high);
828 *lowcpu = low.cs_cpu;
829 *highcpu = high.cs_cpu;
834 sched_balance_group(struct cpu_group *cg)
836 cpuset_t hmask, lmask;
837 int high, low, anylow;
841 high = sched_highest(cg, hmask, 1);
842 /* Stop if there is no more CPU with transferrable threads. */
845 CPU_CLR(high, &hmask);
846 CPU_COPY(&hmask, &lmask);
847 /* Stop if there is no more CPU left for low. */
848 if (CPU_EMPTY(&lmask))
852 low = sched_lowest(cg, lmask, -1,
853 TDQ_CPU(high)->tdq_load - 1, high);
854 /* Stop if we looked well and found no less loaded CPU. */
855 if (anylow && low == -1)
857 /* Go to next high if we found no less loaded CPU. */
860 /* Transfer thread from high to low. */
861 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
862 /* CPU that got thread can no longer be a donor. */
863 CPU_CLR(low, &hmask);
866 * If failed, then there is no threads on high
867 * that can run on this low. Drop low from low
868 * mask and look for different one.
870 CPU_CLR(low, &lmask);
883 * Select a random time between .5 * balance_interval and
884 * 1.5 * balance_interval.
886 balance_ticks = max(balance_interval / 2, 1);
887 balance_ticks += random() % balance_interval;
888 if (smp_started == 0 || rebalance == 0)
892 sched_balance_group(cpu_top);
897 * Lock two thread queues using their address to maintain lock order.
900 tdq_lock_pair(struct tdq *one, struct tdq *two)
904 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
907 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
912 * Unlock two thread queues. Order is not important here.
915 tdq_unlock_pair(struct tdq *one, struct tdq *two)
922 * Transfer load between two imbalanced thread queues.
925 sched_balance_pair(struct tdq *high, struct tdq *low)
930 tdq_lock_pair(high, low);
933 * Determine what the imbalance is and then adjust that to how many
934 * threads we actually have to give up (transferable).
936 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
937 (moved = tdq_move(high, low)) > 0) {
939 * In case the target isn't the current cpu IPI it to force a
940 * reschedule with the new workload.
943 if (cpu != PCPU_GET(cpuid))
944 ipi_cpu(cpu, IPI_PREEMPT);
946 tdq_unlock_pair(high, low);
951 * Move a thread from one thread queue to another.
954 tdq_move(struct tdq *from, struct tdq *to)
961 TDQ_LOCK_ASSERT(from, MA_OWNED);
962 TDQ_LOCK_ASSERT(to, MA_OWNED);
966 td = tdq_steal(tdq, cpu);
971 * Although the run queue is locked the thread may be blocked. Lock
972 * it to clear this and acquire the run-queue lock.
975 /* Drop recursive lock on from acquired via thread_lock(). */
979 td->td_lock = TDQ_LOCKPTR(to);
980 tdq_add(to, td, SRQ_YIELDING);
985 * This tdq has idled. Try to steal a thread from another cpu and switch
989 tdq_idled(struct tdq *tdq)
991 struct cpu_group *cg;
997 if (smp_started == 0 || steal_idle == 0)
1000 CPU_CLR(PCPU_GET(cpuid), &mask);
1001 /* We don't want to be preempted while we're iterating. */
1003 for (cg = tdq->tdq_cg; cg != NULL; ) {
1004 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1005 thresh = steal_thresh;
1008 cpu = sched_highest(cg, mask, thresh);
1013 steal = TDQ_CPU(cpu);
1014 CPU_CLR(cpu, &mask);
1015 tdq_lock_pair(tdq, steal);
1016 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1017 tdq_unlock_pair(tdq, steal);
1021 * If a thread was added while interrupts were disabled don't
1022 * steal one here. If we fail to acquire one due to affinity
1023 * restrictions loop again with this cpu removed from the
1026 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1027 tdq_unlock_pair(tdq, steal);
1032 mi_switch(SW_VOL | SWT_IDLE, NULL);
1033 thread_unlock(curthread);
1042 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1045 tdq_notify(struct tdq *tdq, struct thread *td)
1051 if (tdq->tdq_ipipending)
1053 cpu = td->td_sched->ts_cpu;
1054 pri = td->td_priority;
1055 ctd = pcpu_find(cpu)->pc_curthread;
1056 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1058 if (TD_IS_IDLETHREAD(ctd)) {
1060 * If the MD code has an idle wakeup routine try that before
1061 * falling back to IPI.
1063 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1066 tdq->tdq_ipipending = 1;
1067 ipi_cpu(cpu, IPI_PREEMPT);
1071 * Steals load from a timeshare queue. Honors the rotating queue head
1074 static struct thread *
1075 runq_steal_from(struct runq *rq, int cpu, u_char start)
1079 struct thread *td, *first;
1084 rqb = &rq->rq_status;
1085 bit = start & (RQB_BPW -1);
1089 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1090 if (rqb->rqb_bits[i] == 0)
1093 for (pri = bit; pri < RQB_BPW; pri++)
1094 if (rqb->rqb_bits[i] & (1ul << pri))
1099 pri = RQB_FFS(rqb->rqb_bits[i]);
1100 pri += (i << RQB_L2BPW);
1101 rqh = &rq->rq_queues[pri];
1102 TAILQ_FOREACH(td, rqh, td_runq) {
1103 if (first && THREAD_CAN_MIGRATE(td) &&
1104 THREAD_CAN_SCHED(td, cpu))
1114 if (first && THREAD_CAN_MIGRATE(first) &&
1115 THREAD_CAN_SCHED(first, cpu))
1121 * Steals load from a standard linear queue.
1123 static struct thread *
1124 runq_steal(struct runq *rq, int cpu)
1132 rqb = &rq->rq_status;
1133 for (word = 0; word < RQB_LEN; word++) {
1134 if (rqb->rqb_bits[word] == 0)
1136 for (bit = 0; bit < RQB_BPW; bit++) {
1137 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1139 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1140 TAILQ_FOREACH(td, rqh, td_runq)
1141 if (THREAD_CAN_MIGRATE(td) &&
1142 THREAD_CAN_SCHED(td, cpu))
1150 * Attempt to steal a thread in priority order from a thread queue.
1152 static struct thread *
1153 tdq_steal(struct tdq *tdq, int cpu)
1157 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1158 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1160 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1161 cpu, tdq->tdq_ridx)) != NULL)
1163 return (runq_steal(&tdq->tdq_idle, cpu));
1167 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1168 * current lock and returns with the assigned queue locked.
1170 static inline struct tdq *
1171 sched_setcpu(struct thread *td, int cpu, int flags)
1176 THREAD_LOCK_ASSERT(td, MA_OWNED);
1178 td->td_sched->ts_cpu = cpu;
1180 * If the lock matches just return the queue.
1182 if (td->td_lock == TDQ_LOCKPTR(tdq))
1186 * If the thread isn't running its lockptr is a
1187 * turnstile or a sleepqueue. We can just lock_set without
1190 if (TD_CAN_RUN(td)) {
1192 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1197 * The hard case, migration, we need to block the thread first to
1198 * prevent order reversals with other cpus locks.
1201 thread_lock_block(td);
1203 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1208 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1209 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1210 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1211 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1212 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1213 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1216 sched_pickcpu(struct thread *td, int flags)
1218 struct cpu_group *cg, *ccg;
1219 struct td_sched *ts;
1224 self = PCPU_GET(cpuid);
1226 if (smp_started == 0)
1229 * Don't migrate a running thread from sched_switch().
1231 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1232 return (ts->ts_cpu);
1234 * Prefer to run interrupt threads on the processors that generate
1237 pri = td->td_priority;
1238 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1239 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1240 SCHED_STAT_INC(pickcpu_intrbind);
1242 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1243 SCHED_STAT_INC(pickcpu_affinity);
1244 return (ts->ts_cpu);
1248 * If the thread can run on the last cpu and the affinity has not
1249 * expired or it is idle run it there.
1251 tdq = TDQ_CPU(ts->ts_cpu);
1253 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1254 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1255 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1256 if (cg->cg_flags & CG_FLAG_THREAD) {
1257 CPUSET_FOREACH(cpu, cg->cg_mask) {
1258 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1263 if (cpu > mp_maxid) {
1264 SCHED_STAT_INC(pickcpu_idle_affinity);
1265 return (ts->ts_cpu);
1269 * Search for the last level cache CPU group in the tree.
1270 * Skip caches with expired affinity time and SMT groups.
1271 * Affinity to higher level caches will be handled less aggressively.
1273 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1274 if (cg->cg_flags & CG_FLAG_THREAD)
1276 if (!SCHED_AFFINITY(ts, cg->cg_level))
1283 /* Search the group for the less loaded idle CPU we can run now. */
1284 mask = td->td_cpuset->cs_mask;
1285 if (cg != NULL && cg != cpu_top &&
1286 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1287 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1288 INT_MAX, ts->ts_cpu);
1289 /* Search globally for the less loaded CPU we can run now. */
1291 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1292 /* Search globally for the less loaded CPU. */
1294 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1295 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1297 * Compare the lowest loaded cpu to current cpu.
1299 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1300 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1301 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1302 SCHED_STAT_INC(pickcpu_local);
1305 SCHED_STAT_INC(pickcpu_lowest);
1306 if (cpu != ts->ts_cpu)
1307 SCHED_STAT_INC(pickcpu_migration);
1313 * Pick the highest priority task we have and return it.
1315 static struct thread *
1316 tdq_choose(struct tdq *tdq)
1320 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1321 td = runq_choose(&tdq->tdq_realtime);
1324 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1326 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1327 ("tdq_choose: Invalid priority on timeshare queue %d",
1331 td = runq_choose(&tdq->tdq_idle);
1333 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1334 ("tdq_choose: Invalid priority on idle queue %d",
1343 * Initialize a thread queue.
1346 tdq_setup(struct tdq *tdq)
1350 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1351 runq_init(&tdq->tdq_realtime);
1352 runq_init(&tdq->tdq_timeshare);
1353 runq_init(&tdq->tdq_idle);
1354 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1355 "sched lock %d", (int)TDQ_ID(tdq));
1356 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1357 MTX_SPIN | MTX_RECURSE);
1359 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1360 "CPU %d load", (int)TDQ_ID(tdq));
1366 sched_setup_smp(void)
1371 cpu_top = smp_topo();
1375 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1376 if (tdq->tdq_cg == NULL)
1377 panic("Can't find cpu group for %d\n", i);
1379 balance_tdq = TDQ_SELF();
1385 * Setup the thread queues and initialize the topology based on MD
1389 sched_setup(void *dummy)
1400 /* Add thread0's load since it's running. */
1402 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1403 tdq_load_add(tdq, &thread0);
1404 tdq->tdq_lowpri = thread0.td_priority;
1409 * This routine determines time constants after stathz and hz are setup.
1413 sched_initticks(void *dummy)
1417 realstathz = stathz ? stathz : hz;
1418 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1419 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1420 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1424 * tickincr is shifted out by 10 to avoid rounding errors due to
1425 * hz not being evenly divisible by stathz on all platforms.
1427 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1429 * This does not work for values of stathz that are more than
1430 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1437 * Set the default balance interval now that we know
1438 * what realstathz is.
1440 balance_interval = realstathz;
1441 affinity = SCHED_AFFINITY_DEFAULT;
1443 if (sched_idlespinthresh < 0)
1444 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1449 * This is the core of the interactivity algorithm. Determines a score based
1450 * on past behavior. It is the ratio of sleep time to run time scaled to
1451 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1452 * differs from the cpu usage because it does not account for time spent
1453 * waiting on a run-queue. Would be prettier if we had floating point.
1456 sched_interact_score(struct thread *td)
1458 struct td_sched *ts;
1463 * The score is only needed if this is likely to be an interactive
1464 * task. Don't go through the expense of computing it if there's
1467 if (sched_interact <= SCHED_INTERACT_HALF &&
1468 ts->ts_runtime >= ts->ts_slptime)
1469 return (SCHED_INTERACT_HALF);
1471 if (ts->ts_runtime > ts->ts_slptime) {
1472 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1473 return (SCHED_INTERACT_HALF +
1474 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1476 if (ts->ts_slptime > ts->ts_runtime) {
1477 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1478 return (ts->ts_runtime / div);
1480 /* runtime == slptime */
1482 return (SCHED_INTERACT_HALF);
1485 * This can happen if slptime and runtime are 0.
1492 * Scale the scheduling priority according to the "interactivity" of this
1496 sched_priority(struct thread *td)
1501 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1504 * If the score is interactive we place the thread in the realtime
1505 * queue with a priority that is less than kernel and interrupt
1506 * priorities. These threads are not subject to nice restrictions.
1508 * Scores greater than this are placed on the normal timeshare queue
1509 * where the priority is partially decided by the most recent cpu
1510 * utilization and the rest is decided by nice value.
1512 * The nice value of the process has a linear effect on the calculated
1513 * score. Negative nice values make it easier for a thread to be
1514 * considered interactive.
1516 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1517 if (score < sched_interact) {
1518 pri = PRI_MIN_INTERACT;
1519 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1520 sched_interact) * score;
1521 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1522 ("sched_priority: invalid interactive priority %d score %d",
1525 pri = SCHED_PRI_MIN;
1526 if (td->td_sched->ts_ticks)
1527 pri += min(SCHED_PRI_TICKS(td->td_sched),
1529 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1530 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1531 ("sched_priority: invalid priority %d: nice %d, "
1532 "ticks %d ftick %d ltick %d tick pri %d",
1533 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1534 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1535 SCHED_PRI_TICKS(td->td_sched)));
1537 sched_user_prio(td, pri);
1543 * This routine enforces a maximum limit on the amount of scheduling history
1544 * kept. It is called after either the slptime or runtime is adjusted. This
1545 * function is ugly due to integer math.
1548 sched_interact_update(struct thread *td)
1550 struct td_sched *ts;
1554 sum = ts->ts_runtime + ts->ts_slptime;
1555 if (sum < SCHED_SLP_RUN_MAX)
1558 * This only happens from two places:
1559 * 1) We have added an unusual amount of run time from fork_exit.
1560 * 2) We have added an unusual amount of sleep time from sched_sleep().
1562 if (sum > SCHED_SLP_RUN_MAX * 2) {
1563 if (ts->ts_runtime > ts->ts_slptime) {
1564 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1567 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1573 * If we have exceeded by more than 1/5th then the algorithm below
1574 * will not bring us back into range. Dividing by two here forces
1575 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1577 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1578 ts->ts_runtime /= 2;
1579 ts->ts_slptime /= 2;
1582 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1583 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1587 * Scale back the interactivity history when a child thread is created. The
1588 * history is inherited from the parent but the thread may behave totally
1589 * differently. For example, a shell spawning a compiler process. We want
1590 * to learn that the compiler is behaving badly very quickly.
1593 sched_interact_fork(struct thread *td)
1598 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1599 if (sum > SCHED_SLP_RUN_FORK) {
1600 ratio = sum / SCHED_SLP_RUN_FORK;
1601 td->td_sched->ts_runtime /= ratio;
1602 td->td_sched->ts_slptime /= ratio;
1607 * Called from proc0_init() to setup the scheduler fields.
1614 * Set up the scheduler specific parts of proc0.
1616 proc0.p_sched = NULL; /* XXX */
1617 thread0.td_sched = &td_sched0;
1618 td_sched0.ts_ltick = ticks;
1619 td_sched0.ts_ftick = ticks;
1620 td_sched0.ts_slice = 0;
1624 * This is only somewhat accurate since given many processes of the same
1625 * priority they will switch when their slices run out, which will be
1626 * at most sched_slice stathz ticks.
1629 sched_rr_interval(void)
1632 /* Convert sched_slice from stathz to hz. */
1633 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1637 * Update the percent cpu tracking information when it is requested or
1638 * the total history exceeds the maximum. We keep a sliding history of
1639 * tick counts that slowly decays. This is less precise than the 4BSD
1640 * mechanism since it happens with less regular and frequent events.
1643 sched_pctcpu_update(struct td_sched *ts, int run)
1647 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1649 ts->ts_ftick = t - SCHED_TICK_TARG;
1650 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1651 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1652 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1653 ts->ts_ftick = t - SCHED_TICK_TARG;
1656 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1661 * Adjust the priority of a thread. Move it to the appropriate run-queue
1662 * if necessary. This is the back-end for several priority related
1666 sched_thread_priority(struct thread *td, u_char prio)
1668 struct td_sched *ts;
1672 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1673 "prio:%d", td->td_priority, "new prio:%d", prio,
1674 KTR_ATTR_LINKED, sched_tdname(curthread));
1675 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
1676 if (td != curthread && prio < td->td_priority) {
1677 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1678 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1679 prio, KTR_ATTR_LINKED, sched_tdname(td));
1680 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
1684 THREAD_LOCK_ASSERT(td, MA_OWNED);
1685 if (td->td_priority == prio)
1688 * If the priority has been elevated due to priority
1689 * propagation, we may have to move ourselves to a new
1690 * queue. This could be optimized to not re-add in some
1693 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1695 td->td_priority = prio;
1696 sched_add(td, SRQ_BORROWING);
1700 * If the thread is currently running we may have to adjust the lowpri
1701 * information so other cpus are aware of our current priority.
1703 if (TD_IS_RUNNING(td)) {
1704 tdq = TDQ_CPU(ts->ts_cpu);
1705 oldpri = td->td_priority;
1706 td->td_priority = prio;
1707 if (prio < tdq->tdq_lowpri)
1708 tdq->tdq_lowpri = prio;
1709 else if (tdq->tdq_lowpri == oldpri)
1710 tdq_setlowpri(tdq, td);
1713 td->td_priority = prio;
1717 * Update a thread's priority when it is lent another thread's
1721 sched_lend_prio(struct thread *td, u_char prio)
1724 td->td_flags |= TDF_BORROWING;
1725 sched_thread_priority(td, prio);
1729 * Restore a thread's priority when priority propagation is
1730 * over. The prio argument is the minimum priority the thread
1731 * needs to have to satisfy other possible priority lending
1732 * requests. If the thread's regular priority is less
1733 * important than prio, the thread will keep a priority boost
1737 sched_unlend_prio(struct thread *td, u_char prio)
1741 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1742 td->td_base_pri <= PRI_MAX_TIMESHARE)
1743 base_pri = td->td_user_pri;
1745 base_pri = td->td_base_pri;
1746 if (prio >= base_pri) {
1747 td->td_flags &= ~TDF_BORROWING;
1748 sched_thread_priority(td, base_pri);
1750 sched_lend_prio(td, prio);
1754 * Standard entry for setting the priority to an absolute value.
1757 sched_prio(struct thread *td, u_char prio)
1761 /* First, update the base priority. */
1762 td->td_base_pri = prio;
1765 * If the thread is borrowing another thread's priority, don't
1766 * ever lower the priority.
1768 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1771 /* Change the real priority. */
1772 oldprio = td->td_priority;
1773 sched_thread_priority(td, prio);
1776 * If the thread is on a turnstile, then let the turnstile update
1779 if (TD_ON_LOCK(td) && oldprio != prio)
1780 turnstile_adjust(td, oldprio);
1784 * Set the base user priority, does not effect current running priority.
1787 sched_user_prio(struct thread *td, u_char prio)
1790 td->td_base_user_pri = prio;
1791 if (td->td_lend_user_pri <= prio)
1793 td->td_user_pri = prio;
1797 sched_lend_user_prio(struct thread *td, u_char prio)
1800 THREAD_LOCK_ASSERT(td, MA_OWNED);
1801 td->td_lend_user_pri = prio;
1802 td->td_user_pri = min(prio, td->td_base_user_pri);
1803 if (td->td_priority > td->td_user_pri)
1804 sched_prio(td, td->td_user_pri);
1805 else if (td->td_priority != td->td_user_pri)
1806 td->td_flags |= TDF_NEEDRESCHED;
1810 * Handle migration from sched_switch(). This happens only for
1814 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1818 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1820 tdq_load_rem(tdq, td);
1822 * Do the lock dance required to avoid LOR. We grab an extra
1823 * spinlock nesting to prevent preemption while we're
1824 * not holding either run-queue lock.
1827 thread_lock_block(td); /* This releases the lock on tdq. */
1830 * Acquire both run-queue locks before placing the thread on the new
1831 * run-queue to avoid deadlocks created by placing a thread with a
1832 * blocked lock on the run-queue of a remote processor. The deadlock
1833 * occurs when a third processor attempts to lock the two queues in
1834 * question while the target processor is spinning with its own
1835 * run-queue lock held while waiting for the blocked lock to clear.
1837 tdq_lock_pair(tdn, tdq);
1838 tdq_add(tdn, td, flags);
1839 tdq_notify(tdn, td);
1843 return (TDQ_LOCKPTR(tdn));
1847 * Variadic version of thread_lock_unblock() that does not assume td_lock
1851 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1853 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1858 * Switch threads. This function has to handle threads coming in while
1859 * blocked for some reason, running, or idle. It also must deal with
1860 * migrating a thread from one queue to another as running threads may
1861 * be assigned elsewhere via binding.
1864 sched_switch(struct thread *td, struct thread *newtd, int flags)
1867 struct td_sched *ts;
1870 int cpuid, preempted;
1872 THREAD_LOCK_ASSERT(td, MA_OWNED);
1873 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1875 cpuid = PCPU_GET(cpuid);
1876 tdq = TDQ_CPU(cpuid);
1879 sched_pctcpu_update(ts, 1);
1880 ts->ts_rltick = ticks;
1881 td->td_lastcpu = td->td_oncpu;
1882 td->td_oncpu = NOCPU;
1883 preempted = !(td->td_flags & TDF_SLICEEND);
1884 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1885 td->td_owepreempt = 0;
1886 if (!TD_IS_IDLETHREAD(td))
1887 tdq->tdq_switchcnt++;
1889 * The lock pointer in an idle thread should never change. Reset it
1890 * to CAN_RUN as well.
1892 if (TD_IS_IDLETHREAD(td)) {
1893 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1895 } else if (TD_IS_RUNNING(td)) {
1896 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1897 srqflag = preempted ?
1898 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1899 SRQ_OURSELF|SRQ_YIELDING;
1901 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1902 ts->ts_cpu = sched_pickcpu(td, 0);
1904 if (ts->ts_cpu == cpuid)
1905 tdq_runq_add(tdq, td, srqflag);
1907 KASSERT(THREAD_CAN_MIGRATE(td) ||
1908 (ts->ts_flags & TSF_BOUND) != 0,
1909 ("Thread %p shouldn't migrate", td));
1910 mtx = sched_switch_migrate(tdq, td, srqflag);
1913 /* This thread must be going to sleep. */
1915 mtx = thread_lock_block(td);
1916 tdq_load_rem(tdq, td);
1919 * We enter here with the thread blocked and assigned to the
1920 * appropriate cpu run-queue or sleep-queue and with the current
1921 * thread-queue locked.
1923 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1924 newtd = choosethread();
1926 * Call the MD code to switch contexts if necessary.
1930 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1931 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1933 SDT_PROBE2(sched, , , off_cpu, newtd, newtd->td_proc);
1934 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1935 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1936 sched_pctcpu_update(newtd->td_sched, 0);
1938 #ifdef KDTRACE_HOOKS
1940 * If DTrace has set the active vtime enum to anything
1941 * other than INACTIVE (0), then it should have set the
1944 if (dtrace_vtime_active)
1945 (*dtrace_vtime_switch_func)(newtd);
1948 cpu_switch(td, newtd, mtx);
1950 * We may return from cpu_switch on a different cpu. However,
1951 * we always return with td_lock pointing to the current cpu's
1954 cpuid = PCPU_GET(cpuid);
1955 tdq = TDQ_CPU(cpuid);
1956 lock_profile_obtain_lock_success(
1957 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1959 SDT_PROBE0(sched, , , on_cpu);
1961 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1962 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1965 thread_unblock_switch(td, mtx);
1966 SDT_PROBE0(sched, , , remain_cpu);
1969 * Assert that all went well and return.
1971 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1972 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1973 td->td_oncpu = cpuid;
1977 * Adjust thread priorities as a result of a nice request.
1980 sched_nice(struct proc *p, int nice)
1984 PROC_LOCK_ASSERT(p, MA_OWNED);
1987 FOREACH_THREAD_IN_PROC(p, td) {
1990 sched_prio(td, td->td_base_user_pri);
1996 * Record the sleep time for the interactivity scorer.
1999 sched_sleep(struct thread *td, int prio)
2002 THREAD_LOCK_ASSERT(td, MA_OWNED);
2004 td->td_slptick = ticks;
2005 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2006 td->td_flags |= TDF_CANSWAP;
2007 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2009 if (static_boost == 1 && prio)
2010 sched_prio(td, prio);
2011 else if (static_boost && td->td_priority > static_boost)
2012 sched_prio(td, static_boost);
2016 * Schedule a thread to resume execution and record how long it voluntarily
2017 * slept. We also update the pctcpu, interactivity, and priority.
2020 sched_wakeup(struct thread *td)
2022 struct td_sched *ts;
2025 THREAD_LOCK_ASSERT(td, MA_OWNED);
2027 td->td_flags &= ~TDF_CANSWAP;
2029 * If we slept for more than a tick update our interactivity and
2032 slptick = td->td_slptick;
2034 if (slptick && slptick != ticks) {
2035 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2036 sched_interact_update(td);
2037 sched_pctcpu_update(ts, 0);
2040 * Reset the slice value since we slept and advanced the round-robin.
2043 sched_add(td, SRQ_BORING);
2047 * Penalize the parent for creating a new child and initialize the child's
2051 sched_fork(struct thread *td, struct thread *child)
2053 THREAD_LOCK_ASSERT(td, MA_OWNED);
2054 sched_pctcpu_update(td->td_sched, 1);
2055 sched_fork_thread(td, child);
2057 * Penalize the parent and child for forking.
2059 sched_interact_fork(child);
2060 sched_priority(child);
2061 td->td_sched->ts_runtime += tickincr;
2062 sched_interact_update(td);
2067 * Fork a new thread, may be within the same process.
2070 sched_fork_thread(struct thread *td, struct thread *child)
2072 struct td_sched *ts;
2073 struct td_sched *ts2;
2077 THREAD_LOCK_ASSERT(td, MA_OWNED);
2082 ts2 = child->td_sched;
2083 child->td_lock = TDQ_LOCKPTR(tdq);
2084 child->td_cpuset = cpuset_ref(td->td_cpuset);
2085 ts2->ts_cpu = ts->ts_cpu;
2088 * Grab our parents cpu estimation information.
2090 ts2->ts_ticks = ts->ts_ticks;
2091 ts2->ts_ltick = ts->ts_ltick;
2092 ts2->ts_ftick = ts->ts_ftick;
2094 * Do not inherit any borrowed priority from the parent.
2096 child->td_priority = child->td_base_pri;
2098 * And update interactivity score.
2100 ts2->ts_slptime = ts->ts_slptime;
2101 ts2->ts_runtime = ts->ts_runtime;
2102 /* Attempt to quickly learn interactivity. */
2103 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2105 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2110 * Adjust the priority class of a thread.
2113 sched_class(struct thread *td, int class)
2116 THREAD_LOCK_ASSERT(td, MA_OWNED);
2117 if (td->td_pri_class == class)
2119 td->td_pri_class = class;
2123 * Return some of the child's priority and interactivity to the parent.
2126 sched_exit(struct proc *p, struct thread *child)
2130 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2131 "prio:%d", child->td_priority);
2132 PROC_LOCK_ASSERT(p, MA_OWNED);
2133 td = FIRST_THREAD_IN_PROC(p);
2134 sched_exit_thread(td, child);
2138 * Penalize another thread for the time spent on this one. This helps to
2139 * worsen the priority and interactivity of processes which schedule batch
2140 * jobs such as make. This has little effect on the make process itself but
2141 * causes new processes spawned by it to receive worse scores immediately.
2144 sched_exit_thread(struct thread *td, struct thread *child)
2147 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2148 "prio:%d", child->td_priority);
2150 * Give the child's runtime to the parent without returning the
2151 * sleep time as a penalty to the parent. This causes shells that
2152 * launch expensive things to mark their children as expensive.
2155 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2156 sched_interact_update(td);
2162 sched_preempt(struct thread *td)
2166 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2170 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2171 tdq->tdq_ipipending = 0;
2172 if (td->td_priority > tdq->tdq_lowpri) {
2175 flags = SW_INVOL | SW_PREEMPT;
2176 if (td->td_critnest > 1)
2177 td->td_owepreempt = 1;
2178 else if (TD_IS_IDLETHREAD(td))
2179 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2181 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2187 * Fix priorities on return to user-space. Priorities may be elevated due
2188 * to static priorities in msleep() or similar.
2191 sched_userret(struct thread *td)
2194 * XXX we cheat slightly on the locking here to avoid locking in
2195 * the usual case. Setting td_priority here is essentially an
2196 * incomplete workaround for not setting it properly elsewhere.
2197 * Now that some interrupt handlers are threads, not setting it
2198 * properly elsewhere can clobber it in the window between setting
2199 * it here and returning to user mode, so don't waste time setting
2200 * it perfectly here.
2202 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2203 ("thread with borrowed priority returning to userland"));
2204 if (td->td_priority != td->td_user_pri) {
2206 td->td_priority = td->td_user_pri;
2207 td->td_base_pri = td->td_user_pri;
2208 tdq_setlowpri(TDQ_SELF(), td);
2214 * Handle a stathz tick. This is really only relevant for timeshare
2218 sched_clock(struct thread *td)
2221 struct td_sched *ts;
2223 THREAD_LOCK_ASSERT(td, MA_OWNED);
2227 * We run the long term load balancer infrequently on the first cpu.
2229 if (balance_tdq == tdq) {
2230 if (balance_ticks && --balance_ticks == 0)
2235 * Save the old switch count so we have a record of the last ticks
2236 * activity. Initialize the new switch count based on our load.
2237 * If there is some activity seed it to reflect that.
2239 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2240 tdq->tdq_switchcnt = tdq->tdq_load;
2242 * Advance the insert index once for each tick to ensure that all
2243 * threads get a chance to run.
2245 if (tdq->tdq_idx == tdq->tdq_ridx) {
2246 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2247 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2248 tdq->tdq_ridx = tdq->tdq_idx;
2251 sched_pctcpu_update(ts, 1);
2252 if (td->td_pri_class & PRI_FIFO_BIT)
2254 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2256 * We used a tick; charge it to the thread so
2257 * that we can compute our interactivity.
2259 td->td_sched->ts_runtime += tickincr;
2260 sched_interact_update(td);
2265 * Force a context switch if the current thread has used up a full
2266 * time slice (default is 100ms).
2268 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2270 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2275 * Called once per hz tick.
2284 * Return whether the current CPU has runnable tasks. Used for in-kernel
2285 * cooperative idle threads.
2288 sched_runnable(void)
2296 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2297 if (tdq->tdq_load > 0)
2300 if (tdq->tdq_load - 1 > 0)
2308 * Choose the highest priority thread to run. The thread is removed from
2309 * the run-queue while running however the load remains. For SMP we set
2310 * the tdq in the global idle bitmask if it idles here.
2319 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2320 td = tdq_choose(tdq);
2322 tdq_runq_rem(tdq, td);
2323 tdq->tdq_lowpri = td->td_priority;
2326 tdq->tdq_lowpri = PRI_MAX_IDLE;
2327 return (PCPU_GET(idlethread));
2331 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2332 * we always request it once we exit a critical section.
2335 sched_setpreempt(struct thread *td)
2341 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2344 pri = td->td_priority;
2345 cpri = ctd->td_priority;
2347 ctd->td_flags |= TDF_NEEDRESCHED;
2348 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2350 if (!sched_shouldpreempt(pri, cpri, 0))
2352 ctd->td_owepreempt = 1;
2356 * Add a thread to a thread queue. Select the appropriate runq and add the
2357 * thread to it. This is the internal function called when the tdq is
2361 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2364 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2365 KASSERT((td->td_inhibitors == 0),
2366 ("sched_add: trying to run inhibited thread"));
2367 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2368 ("sched_add: bad thread state"));
2369 KASSERT(td->td_flags & TDF_INMEM,
2370 ("sched_add: thread swapped out"));
2372 if (td->td_priority < tdq->tdq_lowpri)
2373 tdq->tdq_lowpri = td->td_priority;
2374 tdq_runq_add(tdq, td, flags);
2375 tdq_load_add(tdq, td);
2379 * Select the target thread queue and add a thread to it. Request
2380 * preemption or IPI a remote processor if required.
2383 sched_add(struct thread *td, int flags)
2390 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2391 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2392 sched_tdname(curthread));
2393 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2394 KTR_ATTR_LINKED, sched_tdname(td));
2395 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2396 flags & SRQ_PREEMPTED);
2397 THREAD_LOCK_ASSERT(td, MA_OWNED);
2399 * Recalculate the priority before we select the target cpu or
2402 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2406 * Pick the destination cpu and if it isn't ours transfer to the
2409 cpu = sched_pickcpu(td, flags);
2410 tdq = sched_setcpu(td, cpu, flags);
2411 tdq_add(tdq, td, flags);
2412 if (cpu != PCPU_GET(cpuid)) {
2413 tdq_notify(tdq, td);
2420 * Now that the thread is moving to the run-queue, set the lock
2421 * to the scheduler's lock.
2423 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2424 tdq_add(tdq, td, flags);
2426 if (!(flags & SRQ_YIELDING))
2427 sched_setpreempt(td);
2431 * Remove a thread from a run-queue without running it. This is used
2432 * when we're stealing a thread from a remote queue. Otherwise all threads
2433 * exit by calling sched_exit_thread() and sched_throw() themselves.
2436 sched_rem(struct thread *td)
2440 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2441 "prio:%d", td->td_priority);
2442 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2443 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2444 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2445 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2446 KASSERT(TD_ON_RUNQ(td),
2447 ("sched_rem: thread not on run queue"));
2448 tdq_runq_rem(tdq, td);
2449 tdq_load_rem(tdq, td);
2451 if (td->td_priority == tdq->tdq_lowpri)
2452 tdq_setlowpri(tdq, NULL);
2456 * Fetch cpu utilization information. Updates on demand.
2459 sched_pctcpu(struct thread *td)
2462 struct td_sched *ts;
2469 THREAD_LOCK_ASSERT(td, MA_OWNED);
2470 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2474 /* How many rtick per second ? */
2475 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2476 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2483 * Enforce affinity settings for a thread. Called after adjustments to
2487 sched_affinity(struct thread *td)
2490 struct td_sched *ts;
2492 THREAD_LOCK_ASSERT(td, MA_OWNED);
2494 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2496 if (TD_ON_RUNQ(td)) {
2498 sched_add(td, SRQ_BORING);
2501 if (!TD_IS_RUNNING(td))
2504 * Force a switch before returning to userspace. If the
2505 * target thread is not running locally send an ipi to force
2508 td->td_flags |= TDF_NEEDRESCHED;
2509 if (td != curthread)
2510 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2515 * Bind a thread to a target cpu.
2518 sched_bind(struct thread *td, int cpu)
2520 struct td_sched *ts;
2522 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2523 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2525 if (ts->ts_flags & TSF_BOUND)
2527 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2528 ts->ts_flags |= TSF_BOUND;
2530 if (PCPU_GET(cpuid) == cpu)
2533 /* When we return from mi_switch we'll be on the correct cpu. */
2534 mi_switch(SW_VOL, NULL);
2538 * Release a bound thread.
2541 sched_unbind(struct thread *td)
2543 struct td_sched *ts;
2545 THREAD_LOCK_ASSERT(td, MA_OWNED);
2546 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2548 if ((ts->ts_flags & TSF_BOUND) == 0)
2550 ts->ts_flags &= ~TSF_BOUND;
2555 sched_is_bound(struct thread *td)
2557 THREAD_LOCK_ASSERT(td, MA_OWNED);
2558 return (td->td_sched->ts_flags & TSF_BOUND);
2565 sched_relinquish(struct thread *td)
2568 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2573 * Return the total system load.
2584 total += TDQ_CPU(i)->tdq_sysload;
2587 return (TDQ_SELF()->tdq_sysload);
2592 sched_sizeof_proc(void)
2594 return (sizeof(struct proc));
2598 sched_sizeof_thread(void)
2600 return (sizeof(struct thread) + sizeof(struct td_sched));
2604 #define TDQ_IDLESPIN(tdq) \
2605 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2607 #define TDQ_IDLESPIN(tdq) 1
2611 * The actual idle process.
2614 sched_idletd(void *dummy)
2618 int oldswitchcnt, switchcnt;
2621 mtx_assert(&Giant, MA_NOTOWNED);
2624 THREAD_NO_SLEEPING();
2627 if (tdq->tdq_load) {
2629 mi_switch(SW_VOL | SWT_IDLE, NULL);
2632 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2634 if (switchcnt != oldswitchcnt) {
2635 oldswitchcnt = switchcnt;
2636 if (tdq_idled(tdq) == 0)
2639 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2641 oldswitchcnt = switchcnt;
2644 * If we're switching very frequently, spin while checking
2645 * for load rather than entering a low power state that
2646 * may require an IPI. However, don't do any busy
2647 * loops while on SMT machines as this simply steals
2648 * cycles from cores doing useful work.
2650 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2651 for (i = 0; i < sched_idlespins; i++) {
2658 /* If there was context switch during spin, restart it. */
2659 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2660 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2663 /* Run main MD idle handler. */
2664 tdq->tdq_cpu_idle = 1;
2665 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2666 tdq->tdq_cpu_idle = 0;
2669 * Account thread-less hardware interrupts and
2670 * other wakeup reasons equal to context switches.
2672 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2673 if (switchcnt != oldswitchcnt)
2675 tdq->tdq_switchcnt++;
2681 * A CPU is entering for the first time or a thread is exiting.
2684 sched_throw(struct thread *td)
2686 struct thread *newtd;
2691 /* Correct spinlock nesting and acquire the correct lock. */
2694 PCPU_SET(switchtime, cpu_ticks());
2695 PCPU_SET(switchticks, ticks);
2697 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2698 tdq_load_rem(tdq, td);
2699 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2701 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2702 newtd = choosethread();
2703 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2704 cpu_throw(td, newtd); /* doesn't return */
2708 * This is called from fork_exit(). Just acquire the correct locks and
2709 * let fork do the rest of the work.
2712 sched_fork_exit(struct thread *td)
2714 struct td_sched *ts;
2719 * Finish setting up thread glue so that it begins execution in a
2720 * non-nested critical section with the scheduler lock held.
2722 cpuid = PCPU_GET(cpuid);
2723 tdq = TDQ_CPU(cpuid);
2725 if (TD_IS_IDLETHREAD(td))
2726 td->td_lock = TDQ_LOCKPTR(tdq);
2727 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2728 td->td_oncpu = cpuid;
2729 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2730 lock_profile_obtain_lock_success(
2731 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2735 * Create on first use to catch odd startup conditons.
2738 sched_tdname(struct thread *td)
2741 struct td_sched *ts;
2744 if (ts->ts_name[0] == '\0')
2745 snprintf(ts->ts_name, sizeof(ts->ts_name),
2746 "%s tid %d", td->td_name, td->td_tid);
2747 return (ts->ts_name);
2749 return (td->td_name);
2755 sched_clear_tdname(struct thread *td)
2757 struct td_sched *ts;
2760 ts->ts_name[0] = '\0';
2767 * Build the CPU topology dump string. Is recursively called to collect
2768 * the topology tree.
2771 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2774 char cpusetbuf[CPUSETBUFSIZ];
2777 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2778 "", 1 + indent / 2, cg->cg_level);
2779 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2780 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2782 for (i = 0; i < MAXCPU; i++) {
2783 if (CPU_ISSET(i, &cg->cg_mask)) {
2785 sbuf_printf(sb, ", ");
2788 sbuf_printf(sb, "%d", i);
2791 sbuf_printf(sb, "</cpu>\n");
2793 if (cg->cg_flags != 0) {
2794 sbuf_printf(sb, "%*s <flags>", indent, "");
2795 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2796 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2797 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2798 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2799 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2800 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2801 sbuf_printf(sb, "</flags>\n");
2804 if (cg->cg_children > 0) {
2805 sbuf_printf(sb, "%*s <children>\n", indent, "");
2806 for (i = 0; i < cg->cg_children; i++)
2807 sysctl_kern_sched_topology_spec_internal(sb,
2808 &cg->cg_child[i], indent+2);
2809 sbuf_printf(sb, "%*s </children>\n", indent, "");
2811 sbuf_printf(sb, "%*s</group>\n", indent, "");
2816 * Sysctl handler for retrieving topology dump. It's a wrapper for
2817 * the recursive sysctl_kern_smp_topology_spec_internal().
2820 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2825 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2827 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2831 sbuf_printf(topo, "<groups>\n");
2832 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2833 sbuf_printf(topo, "</groups>\n");
2837 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2846 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2848 int error, new_val, period;
2850 period = 1000000 / realstathz;
2851 new_val = period * sched_slice;
2852 error = sysctl_handle_int(oidp, &new_val, 0, req);
2853 if (error != 0 || req->newptr == NULL)
2857 sched_slice = imax(1, (new_val + period / 2) / period);
2858 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2859 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2864 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2865 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2867 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2868 NULL, 0, sysctl_kern_quantum, "I",
2869 "Quantum for timeshare threads in microseconds");
2870 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2871 "Quantum for timeshare threads in stathz ticks");
2872 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2873 "Interactivity score threshold");
2874 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2876 "Maximal (lowest) priority for preemption");
2877 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2878 "Assign static kernel priorities to sleeping threads");
2879 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2880 "Number of times idle thread will spin waiting for new work");
2881 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2882 &sched_idlespinthresh, 0,
2883 "Threshold before we will permit idle thread spinning");
2885 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2886 "Number of hz ticks to keep thread affinity for");
2887 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2888 "Enables the long-term load balancer");
2889 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2890 &balance_interval, 0,
2891 "Average period in stathz ticks to run the long-term balancer");
2892 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2893 "Attempts to steal work from other cores before idling");
2894 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2895 "Minimum load on remote CPU before we'll steal");
2896 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2897 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2898 "XML dump of detected CPU topology");
2901 /* ps compat. All cpu percentages from ULE are weighted. */
2902 static int ccpu = 0;
2903 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");