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>
82 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
83 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
84 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87 * Thread scheduler specific section. All fields are protected
91 struct runq *ts_runq; /* Run-queue we're queued on. */
92 short ts_flags; /* TSF_* flags. */
93 u_char ts_cpu; /* CPU that we have affinity for. */
94 int ts_rltick; /* Real last tick, for affinity. */
95 int ts_slice; /* Ticks of slice remaining. */
96 u_int ts_slptime; /* Number of ticks we vol. slept */
97 u_int ts_runtime; /* Number of ticks we were running */
98 int ts_ltick; /* Last tick that we were running on */
99 int ts_ftick; /* First tick that we were running on */
100 int ts_ticks; /* Tick count */
102 char ts_name[TS_NAME_LEN];
105 /* flags kept in ts_flags */
106 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
107 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
109 static struct td_sched td_sched0;
111 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
112 #define THREAD_CAN_SCHED(td, cpu) \
113 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
116 * Priority ranges used for interactive and non-interactive timeshare
117 * threads. The timeshare priorities are split up into four ranges.
118 * The first range handles interactive threads. The last three ranges
119 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
120 * ranges supporting nice values.
122 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
123 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
124 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
126 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
127 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
128 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
129 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
132 * Cpu percentage computation macros and defines.
134 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
135 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
136 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
137 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
138 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
139 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
141 #define SCHED_TICK_SECS 10
142 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
143 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
144 #define SCHED_TICK_SHIFT 10
145 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
146 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
149 * These macros determine priorities for non-interactive threads. They are
150 * assigned a priority based on their recent cpu utilization as expressed
151 * by the ratio of ticks to the tick total. NHALF priorities at the start
152 * and end of the MIN to MAX timeshare range are only reachable with negative
153 * or positive nice respectively.
155 * PRI_RANGE: Priority range for utilization dependent priorities.
156 * PRI_NRESV: Number of nice values.
157 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
158 * PRI_NICE: Determines the part of the priority inherited from nice.
160 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
161 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
162 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
163 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
164 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
165 #define SCHED_PRI_TICKS(ts) \
166 (SCHED_TICK_HZ((ts)) / \
167 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
168 #define SCHED_PRI_NICE(nice) (nice)
171 * These determine the interactivity of a process. Interactivity differs from
172 * cpu utilization in that it expresses the voluntary time slept vs time ran
173 * while cpu utilization includes all time not running. This more accurately
174 * models the intent of the thread.
176 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
177 * before throttling back.
178 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
179 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
180 * INTERACT_THRESH: Threshold for placement on the current runq.
182 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
183 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
184 #define SCHED_INTERACT_MAX (100)
185 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
186 #define SCHED_INTERACT_THRESH (30)
189 * These parameters determine the slice behavior for batch work.
191 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
192 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
194 /* Flags kept in td_flags. */
195 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
198 * tickincr: Converts a stathz tick into a hz domain scaled by
199 * the shift factor. Without the shift the error rate
200 * due to rounding would be unacceptably high.
201 * realstathz: stathz is sometimes 0 and run off of hz.
202 * sched_slice: Runtime of each thread before rescheduling.
203 * preempt_thresh: Priority threshold for preemption and remote IPIs.
205 static int sched_interact = SCHED_INTERACT_THRESH;
206 static int tickincr = 8 << SCHED_TICK_SHIFT;
207 static int realstathz = 127; /* reset during boot. */
208 static int sched_slice = 10; /* reset during boot. */
209 static int sched_slice_min = 1; /* reset during boot. */
211 #ifdef FULL_PREEMPTION
212 static int preempt_thresh = PRI_MAX_IDLE;
214 static int preempt_thresh = PRI_MIN_KERN;
217 static int preempt_thresh = 0;
219 static int static_boost = PRI_MIN_BATCH;
220 static int sched_idlespins = 10000;
221 static int sched_idlespinthresh = -1;
224 * tdq - per processor runqs and statistics. All fields are protected by the
225 * tdq_lock. The load and lowpri may be accessed without to avoid excess
226 * locking in sched_pickcpu();
230 * Ordered to improve efficiency of cpu_search() and switch().
231 * tdq_lock is padded to avoid false sharing with tdq_load and
234 struct mtx_padalign tdq_lock; /* run queue lock. */
235 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
236 volatile int tdq_load; /* Aggregate load. */
237 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
238 int tdq_sysload; /* For loadavg, !ITHD load. */
239 int tdq_transferable; /* Transferable thread count. */
240 short tdq_switchcnt; /* Switches this tick. */
241 short tdq_oldswitchcnt; /* Switches last tick. */
242 u_char tdq_lowpri; /* Lowest priority thread. */
243 u_char tdq_ipipending; /* IPI pending. */
244 u_char tdq_idx; /* Current insert index. */
245 u_char tdq_ridx; /* Current removal index. */
246 struct runq tdq_realtime; /* real-time run queue. */
247 struct runq tdq_timeshare; /* timeshare run queue. */
248 struct runq tdq_idle; /* Queue of IDLE threads. */
249 char tdq_name[TDQ_NAME_LEN];
251 char tdq_loadname[TDQ_LOADNAME_LEN];
255 /* Idle thread states and config. */
256 #define TDQ_RUNNING 1
260 struct cpu_group *cpu_top; /* CPU topology */
262 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
263 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
268 static int rebalance = 1;
269 static int balance_interval = 128; /* Default set in sched_initticks(). */
271 static int steal_idle = 1;
272 static int steal_thresh = 2;
275 * One thread queue per processor.
277 static struct tdq tdq_cpu[MAXCPU];
278 static struct tdq *balance_tdq;
279 static int balance_ticks;
280 static DPCPU_DEFINE(uint32_t, randomval);
282 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
283 #define TDQ_CPU(x) (&tdq_cpu[(x)])
284 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
286 static struct tdq tdq_cpu;
288 #define TDQ_ID(x) (0)
289 #define TDQ_SELF() (&tdq_cpu)
290 #define TDQ_CPU(x) (&tdq_cpu)
293 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
294 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
295 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
296 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
297 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
299 static void sched_priority(struct thread *);
300 static void sched_thread_priority(struct thread *, u_char);
301 static int sched_interact_score(struct thread *);
302 static void sched_interact_update(struct thread *);
303 static void sched_interact_fork(struct thread *);
304 static void sched_pctcpu_update(struct td_sched *, int);
306 /* Operations on per processor queues */
307 static struct thread *tdq_choose(struct tdq *);
308 static void tdq_setup(struct tdq *);
309 static void tdq_load_add(struct tdq *, struct thread *);
310 static void tdq_load_rem(struct tdq *, struct thread *);
311 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
312 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
313 static inline int sched_shouldpreempt(int, int, int);
314 void tdq_print(int cpu);
315 static void runq_print(struct runq *rq);
316 static void tdq_add(struct tdq *, struct thread *, int);
318 static int tdq_move(struct tdq *, struct tdq *);
319 static int tdq_idled(struct tdq *);
320 static void tdq_notify(struct tdq *, struct thread *);
321 static struct thread *tdq_steal(struct tdq *, int);
322 static struct thread *runq_steal(struct runq *, int);
323 static int sched_pickcpu(struct thread *, int);
324 static void sched_balance(void);
325 static int sched_balance_pair(struct tdq *, struct tdq *);
326 static inline struct tdq *sched_setcpu(struct thread *, int, int);
327 static inline void thread_unblock_switch(struct thread *, struct mtx *);
328 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
329 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
330 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
331 struct cpu_group *cg, int indent);
334 static void sched_setup(void *dummy);
335 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
337 static void sched_initticks(void *dummy);
338 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
341 SDT_PROVIDER_DEFINE(sched);
343 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
344 "struct proc *", "uint8_t");
345 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
346 "struct proc *", "void *");
347 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
348 "struct proc *", "void *", "int");
349 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
350 "struct proc *", "uint8_t", "struct thread *");
351 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
352 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
354 SDT_PROBE_DEFINE(sched, , , on__cpu);
355 SDT_PROBE_DEFINE(sched, , , remain__cpu);
356 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
360 * Print the threads waiting on a run-queue.
363 runq_print(struct runq *rq)
371 for (i = 0; i < RQB_LEN; i++) {
372 printf("\t\trunq bits %d 0x%zx\n",
373 i, rq->rq_status.rqb_bits[i]);
374 for (j = 0; j < RQB_BPW; j++)
375 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
376 pri = j + (i << RQB_L2BPW);
377 rqh = &rq->rq_queues[pri];
378 TAILQ_FOREACH(td, rqh, td_runq) {
379 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
380 td, td->td_name, td->td_priority,
381 td->td_rqindex, pri);
388 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
397 printf("tdq %d:\n", TDQ_ID(tdq));
398 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
399 printf("\tLock name: %s\n", tdq->tdq_name);
400 printf("\tload: %d\n", tdq->tdq_load);
401 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
402 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
403 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
404 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
405 printf("\tload transferable: %d\n", tdq->tdq_transferable);
406 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
407 printf("\trealtime runq:\n");
408 runq_print(&tdq->tdq_realtime);
409 printf("\ttimeshare runq:\n");
410 runq_print(&tdq->tdq_timeshare);
411 printf("\tidle runq:\n");
412 runq_print(&tdq->tdq_idle);
416 sched_shouldpreempt(int pri, int cpri, int remote)
419 * If the new priority is not better than the current priority there is
425 * Always preempt idle.
427 if (cpri >= PRI_MIN_IDLE)
430 * If preemption is disabled don't preempt others.
432 if (preempt_thresh == 0)
435 * Preempt if we exceed the threshold.
437 if (pri <= preempt_thresh)
440 * If we're interactive or better and there is non-interactive
441 * or worse running preempt only remote processors.
443 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
449 * Add a thread to the actual run-queue. Keeps transferable counts up to
450 * date with what is actually on the run-queue. Selects the correct
451 * queue position for timeshare threads.
454 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
459 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
460 THREAD_LOCK_ASSERT(td, MA_OWNED);
462 pri = td->td_priority;
465 if (THREAD_CAN_MIGRATE(td)) {
466 tdq->tdq_transferable++;
467 ts->ts_flags |= TSF_XFERABLE;
469 if (pri < PRI_MIN_BATCH) {
470 ts->ts_runq = &tdq->tdq_realtime;
471 } else if (pri <= PRI_MAX_BATCH) {
472 ts->ts_runq = &tdq->tdq_timeshare;
473 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
474 ("Invalid priority %d on timeshare runq", pri));
476 * This queue contains only priorities between MIN and MAX
477 * realtime. Use the whole queue to represent these values.
479 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
480 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
481 pri = (pri + tdq->tdq_idx) % RQ_NQS;
483 * This effectively shortens the queue by one so we
484 * can have a one slot difference between idx and
485 * ridx while we wait for threads to drain.
487 if (tdq->tdq_ridx != tdq->tdq_idx &&
488 pri == tdq->tdq_ridx)
489 pri = (unsigned char)(pri - 1) % RQ_NQS;
492 runq_add_pri(ts->ts_runq, td, pri, flags);
495 ts->ts_runq = &tdq->tdq_idle;
496 runq_add(ts->ts_runq, td, flags);
500 * Remove a thread from a run-queue. This typically happens when a thread
501 * is selected to run. Running threads are not on the queue and the
502 * transferable count does not reflect them.
505 tdq_runq_rem(struct tdq *tdq, struct thread *td)
510 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
511 KASSERT(ts->ts_runq != NULL,
512 ("tdq_runq_remove: thread %p null ts_runq", td));
513 if (ts->ts_flags & TSF_XFERABLE) {
514 tdq->tdq_transferable--;
515 ts->ts_flags &= ~TSF_XFERABLE;
517 if (ts->ts_runq == &tdq->tdq_timeshare) {
518 if (tdq->tdq_idx != tdq->tdq_ridx)
519 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
521 runq_remove_idx(ts->ts_runq, td, NULL);
523 runq_remove(ts->ts_runq, td);
527 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
528 * for this thread to the referenced thread queue.
531 tdq_load_add(struct tdq *tdq, struct thread *td)
534 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
535 THREAD_LOCK_ASSERT(td, MA_OWNED);
538 if ((td->td_flags & TDF_NOLOAD) == 0)
540 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
541 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
545 * Remove the load from a thread that is transitioning to a sleep state or
549 tdq_load_rem(struct tdq *tdq, struct thread *td)
552 THREAD_LOCK_ASSERT(td, MA_OWNED);
553 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
554 KASSERT(tdq->tdq_load != 0,
555 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
558 if ((td->td_flags & TDF_NOLOAD) == 0)
560 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
561 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
565 * Bound timeshare latency by decreasing slice size as load increases. We
566 * consider the maximum latency as the sum of the threads waiting to run
567 * aside from curthread and target no more than sched_slice latency but
568 * no less than sched_slice_min runtime.
571 tdq_slice(struct tdq *tdq)
576 * It is safe to use sys_load here because this is called from
577 * contexts where timeshare threads are running and so there
578 * cannot be higher priority load in the system.
580 load = tdq->tdq_sysload - 1;
581 if (load >= SCHED_SLICE_MIN_DIVISOR)
582 return (sched_slice_min);
584 return (sched_slice);
585 return (sched_slice / load);
589 * Set lowpri to its exact value by searching the run-queue and
590 * evaluating curthread. curthread may be passed as an optimization.
593 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
597 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
599 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
600 td = tdq_choose(tdq);
601 if (td == NULL || td->td_priority > ctd->td_priority)
602 tdq->tdq_lowpri = ctd->td_priority;
604 tdq->tdq_lowpri = td->td_priority;
611 int cs_pri; /* Min priority for low. */
612 int cs_limit; /* Max load for low, min load for high. */
617 #define CPU_SEARCH_LOWEST 0x1
618 #define CPU_SEARCH_HIGHEST 0x2
619 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
621 #define CPUSET_FOREACH(cpu, mask) \
622 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
623 if (CPU_ISSET(cpu, &mask))
625 static __always_inline int cpu_search(const struct cpu_group *cg,
626 struct cpu_search *low, struct cpu_search *high, const int match);
627 int __noinline cpu_search_lowest(const struct cpu_group *cg,
628 struct cpu_search *low);
629 int __noinline cpu_search_highest(const struct cpu_group *cg,
630 struct cpu_search *high);
631 int __noinline cpu_search_both(const struct cpu_group *cg,
632 struct cpu_search *low, struct cpu_search *high);
635 * Search the tree of cpu_groups for the lowest or highest loaded cpu
636 * according to the match argument. This routine actually compares the
637 * load on all paths through the tree and finds the least loaded cpu on
638 * the least loaded path, which may differ from the least loaded cpu in
639 * the system. This balances work among caches and busses.
641 * This inline is instantiated in three forms below using constants for the
642 * match argument. It is reduced to the minimum set for each case. It is
643 * also recursive to the depth of the tree.
645 static __always_inline int
646 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
647 struct cpu_search *high, const int match)
649 struct cpu_search lgroup;
650 struct cpu_search hgroup;
652 struct cpu_group *child;
654 int cpu, i, hload, lload, load, total, rnd, *rndptr;
657 cpumask = cg->cg_mask;
658 if (match & CPU_SEARCH_LOWEST) {
662 if (match & CPU_SEARCH_HIGHEST) {
667 /* Iterate through the child CPU groups and then remaining CPUs. */
668 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
670 #ifdef HAVE_INLINE_FFSL
671 cpu = CPU_FFS(&cpumask) - 1;
673 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
680 child = &cg->cg_child[i - 1];
682 if (match & CPU_SEARCH_LOWEST)
684 if (match & CPU_SEARCH_HIGHEST)
686 if (child) { /* Handle child CPU group. */
687 CPU_NAND(&cpumask, &child->cg_mask);
689 case CPU_SEARCH_LOWEST:
690 load = cpu_search_lowest(child, &lgroup);
692 case CPU_SEARCH_HIGHEST:
693 load = cpu_search_highest(child, &hgroup);
695 case CPU_SEARCH_BOTH:
696 load = cpu_search_both(child, &lgroup, &hgroup);
699 } else { /* Handle child CPU. */
700 CPU_CLR(cpu, &cpumask);
702 load = tdq->tdq_load * 256;
703 rndptr = DPCPU_PTR(randomval);
704 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
705 if (match & CPU_SEARCH_LOWEST) {
706 if (cpu == low->cs_prefer)
708 /* If that CPU is allowed and get data. */
709 if (tdq->tdq_lowpri > lgroup.cs_pri &&
710 tdq->tdq_load <= lgroup.cs_limit &&
711 CPU_ISSET(cpu, &lgroup.cs_mask)) {
713 lgroup.cs_load = load - rnd;
716 if (match & CPU_SEARCH_HIGHEST)
717 if (tdq->tdq_load >= hgroup.cs_limit &&
718 tdq->tdq_transferable &&
719 CPU_ISSET(cpu, &hgroup.cs_mask)) {
721 hgroup.cs_load = load - rnd;
726 /* We have info about child item. Compare it. */
727 if (match & CPU_SEARCH_LOWEST) {
728 if (lgroup.cs_cpu >= 0 &&
730 (load == lload && lgroup.cs_load < low->cs_load))) {
732 low->cs_cpu = lgroup.cs_cpu;
733 low->cs_load = lgroup.cs_load;
736 if (match & CPU_SEARCH_HIGHEST)
737 if (hgroup.cs_cpu >= 0 &&
739 (load == hload && hgroup.cs_load > high->cs_load))) {
741 high->cs_cpu = hgroup.cs_cpu;
742 high->cs_load = hgroup.cs_load;
746 if (i == 0 && CPU_EMPTY(&cpumask))
749 #ifndef HAVE_INLINE_FFSL
758 * cpu_search instantiations must pass constants to maintain the inline
762 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
764 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
768 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
770 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
774 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
775 struct cpu_search *high)
777 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
781 * Find the cpu with the least load via the least loaded path that has a
782 * lowpri greater than pri pri. A pri of -1 indicates any priority is
786 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
789 struct cpu_search low;
792 low.cs_prefer = prefer;
795 low.cs_limit = maxload;
796 cpu_search_lowest(cg, &low);
801 * Find the cpu with the highest load via the highest loaded path.
804 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
806 struct cpu_search high;
810 high.cs_limit = minload;
811 cpu_search_highest(cg, &high);
816 sched_balance_group(struct cpu_group *cg)
818 cpuset_t hmask, lmask;
819 int high, low, anylow;
823 high = sched_highest(cg, hmask, 1);
824 /* Stop if there is no more CPU with transferrable threads. */
827 CPU_CLR(high, &hmask);
828 CPU_COPY(&hmask, &lmask);
829 /* Stop if there is no more CPU left for low. */
830 if (CPU_EMPTY(&lmask))
834 low = sched_lowest(cg, lmask, -1,
835 TDQ_CPU(high)->tdq_load - 1, high);
836 /* Stop if we looked well and found no less loaded CPU. */
837 if (anylow && low == -1)
839 /* Go to next high if we found no less loaded CPU. */
842 /* Transfer thread from high to low. */
843 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
844 /* CPU that got thread can no longer be a donor. */
845 CPU_CLR(low, &hmask);
848 * If failed, then there is no threads on high
849 * that can run on this low. Drop low from low
850 * mask and look for different one.
852 CPU_CLR(low, &lmask);
865 * Select a random time between .5 * balance_interval and
866 * 1.5 * balance_interval.
868 balance_ticks = max(balance_interval / 2, 1);
869 balance_ticks += random() % balance_interval;
870 if (smp_started == 0 || rebalance == 0)
874 sched_balance_group(cpu_top);
879 * Lock two thread queues using their address to maintain lock order.
882 tdq_lock_pair(struct tdq *one, struct tdq *two)
886 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
889 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
894 * Unlock two thread queues. Order is not important here.
897 tdq_unlock_pair(struct tdq *one, struct tdq *two)
904 * Transfer load between two imbalanced thread queues.
907 sched_balance_pair(struct tdq *high, struct tdq *low)
912 tdq_lock_pair(high, low);
915 * Determine what the imbalance is and then adjust that to how many
916 * threads we actually have to give up (transferable).
918 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
919 (moved = tdq_move(high, low)) > 0) {
921 * In case the target isn't the current cpu IPI it to force a
922 * reschedule with the new workload.
925 if (cpu != PCPU_GET(cpuid))
926 ipi_cpu(cpu, IPI_PREEMPT);
928 tdq_unlock_pair(high, low);
933 * Move a thread from one thread queue to another.
936 tdq_move(struct tdq *from, struct tdq *to)
943 TDQ_LOCK_ASSERT(from, MA_OWNED);
944 TDQ_LOCK_ASSERT(to, MA_OWNED);
948 td = tdq_steal(tdq, cpu);
953 * Although the run queue is locked the thread may be blocked. Lock
954 * it to clear this and acquire the run-queue lock.
957 /* Drop recursive lock on from acquired via thread_lock(). */
961 td->td_lock = TDQ_LOCKPTR(to);
962 tdq_add(to, td, SRQ_YIELDING);
967 * This tdq has idled. Try to steal a thread from another cpu and switch
971 tdq_idled(struct tdq *tdq)
973 struct cpu_group *cg;
979 if (smp_started == 0 || steal_idle == 0)
982 CPU_CLR(PCPU_GET(cpuid), &mask);
983 /* We don't want to be preempted while we're iterating. */
985 for (cg = tdq->tdq_cg; cg != NULL; ) {
986 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
987 thresh = steal_thresh;
990 cpu = sched_highest(cg, mask, thresh);
995 steal = TDQ_CPU(cpu);
997 tdq_lock_pair(tdq, steal);
998 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
999 tdq_unlock_pair(tdq, steal);
1003 * If a thread was added while interrupts were disabled don't
1004 * steal one here. If we fail to acquire one due to affinity
1005 * restrictions loop again with this cpu removed from the
1008 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1009 tdq_unlock_pair(tdq, steal);
1014 mi_switch(SW_VOL | SWT_IDLE, NULL);
1015 thread_unlock(curthread);
1024 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1027 tdq_notify(struct tdq *tdq, struct thread *td)
1033 if (tdq->tdq_ipipending)
1035 cpu = td->td_sched->ts_cpu;
1036 pri = td->td_priority;
1037 ctd = pcpu_find(cpu)->pc_curthread;
1038 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1042 * Make sure that tdq_load updated before calling this function
1043 * is globally visible before we read tdq_cpu_idle. Idle thread
1044 * accesses both of them without locks, and the order is important.
1048 if (TD_IS_IDLETHREAD(ctd)) {
1050 * If the MD code has an idle wakeup routine try that before
1051 * falling back to IPI.
1053 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1056 tdq->tdq_ipipending = 1;
1057 ipi_cpu(cpu, IPI_PREEMPT);
1061 * Steals load from a timeshare queue. Honors the rotating queue head
1064 static struct thread *
1065 runq_steal_from(struct runq *rq, int cpu, u_char start)
1069 struct thread *td, *first;
1074 rqb = &rq->rq_status;
1075 bit = start & (RQB_BPW -1);
1079 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1080 if (rqb->rqb_bits[i] == 0)
1083 for (pri = bit; pri < RQB_BPW; pri++)
1084 if (rqb->rqb_bits[i] & (1ul << pri))
1089 pri = RQB_FFS(rqb->rqb_bits[i]);
1090 pri += (i << RQB_L2BPW);
1091 rqh = &rq->rq_queues[pri];
1092 TAILQ_FOREACH(td, rqh, td_runq) {
1093 if (first && THREAD_CAN_MIGRATE(td) &&
1094 THREAD_CAN_SCHED(td, cpu))
1104 if (first && THREAD_CAN_MIGRATE(first) &&
1105 THREAD_CAN_SCHED(first, cpu))
1111 * Steals load from a standard linear queue.
1113 static struct thread *
1114 runq_steal(struct runq *rq, int cpu)
1122 rqb = &rq->rq_status;
1123 for (word = 0; word < RQB_LEN; word++) {
1124 if (rqb->rqb_bits[word] == 0)
1126 for (bit = 0; bit < RQB_BPW; bit++) {
1127 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1129 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1130 TAILQ_FOREACH(td, rqh, td_runq)
1131 if (THREAD_CAN_MIGRATE(td) &&
1132 THREAD_CAN_SCHED(td, cpu))
1140 * Attempt to steal a thread in priority order from a thread queue.
1142 static struct thread *
1143 tdq_steal(struct tdq *tdq, int cpu)
1147 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1148 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1150 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1151 cpu, tdq->tdq_ridx)) != NULL)
1153 return (runq_steal(&tdq->tdq_idle, cpu));
1157 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1158 * current lock and returns with the assigned queue locked.
1160 static inline struct tdq *
1161 sched_setcpu(struct thread *td, int cpu, int flags)
1166 THREAD_LOCK_ASSERT(td, MA_OWNED);
1168 td->td_sched->ts_cpu = cpu;
1170 * If the lock matches just return the queue.
1172 if (td->td_lock == TDQ_LOCKPTR(tdq))
1176 * If the thread isn't running its lockptr is a
1177 * turnstile or a sleepqueue. We can just lock_set without
1180 if (TD_CAN_RUN(td)) {
1182 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1187 * The hard case, migration, we need to block the thread first to
1188 * prevent order reversals with other cpus locks.
1191 thread_lock_block(td);
1193 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1198 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1199 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1200 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1201 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1202 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1203 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1206 sched_pickcpu(struct thread *td, int flags)
1208 struct cpu_group *cg, *ccg;
1209 struct td_sched *ts;
1214 self = PCPU_GET(cpuid);
1216 if (smp_started == 0)
1219 * Don't migrate a running thread from sched_switch().
1221 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1222 return (ts->ts_cpu);
1224 * Prefer to run interrupt threads on the processors that generate
1227 pri = td->td_priority;
1228 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1229 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1230 SCHED_STAT_INC(pickcpu_intrbind);
1232 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1233 SCHED_STAT_INC(pickcpu_affinity);
1234 return (ts->ts_cpu);
1238 * If the thread can run on the last cpu and the affinity has not
1239 * expired or it is idle run it there.
1241 tdq = TDQ_CPU(ts->ts_cpu);
1243 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1244 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1245 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1246 if (cg->cg_flags & CG_FLAG_THREAD) {
1247 CPUSET_FOREACH(cpu, cg->cg_mask) {
1248 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1253 if (cpu > mp_maxid) {
1254 SCHED_STAT_INC(pickcpu_idle_affinity);
1255 return (ts->ts_cpu);
1259 * Search for the last level cache CPU group in the tree.
1260 * Skip caches with expired affinity time and SMT groups.
1261 * Affinity to higher level caches will be handled less aggressively.
1263 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1264 if (cg->cg_flags & CG_FLAG_THREAD)
1266 if (!SCHED_AFFINITY(ts, cg->cg_level))
1273 /* Search the group for the less loaded idle CPU we can run now. */
1274 mask = td->td_cpuset->cs_mask;
1275 if (cg != NULL && cg != cpu_top &&
1276 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1277 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1278 INT_MAX, ts->ts_cpu);
1279 /* Search globally for the less loaded CPU we can run now. */
1281 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1282 /* Search globally for the less loaded CPU. */
1284 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1285 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1287 * Compare the lowest loaded cpu to current cpu.
1289 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1290 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1291 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1292 SCHED_STAT_INC(pickcpu_local);
1295 SCHED_STAT_INC(pickcpu_lowest);
1296 if (cpu != ts->ts_cpu)
1297 SCHED_STAT_INC(pickcpu_migration);
1303 * Pick the highest priority task we have and return it.
1305 static struct thread *
1306 tdq_choose(struct tdq *tdq)
1310 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1311 td = runq_choose(&tdq->tdq_realtime);
1314 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1316 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1317 ("tdq_choose: Invalid priority on timeshare queue %d",
1321 td = runq_choose(&tdq->tdq_idle);
1323 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1324 ("tdq_choose: Invalid priority on idle queue %d",
1333 * Initialize a thread queue.
1336 tdq_setup(struct tdq *tdq)
1340 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1341 runq_init(&tdq->tdq_realtime);
1342 runq_init(&tdq->tdq_timeshare);
1343 runq_init(&tdq->tdq_idle);
1344 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1345 "sched lock %d", (int)TDQ_ID(tdq));
1346 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1347 MTX_SPIN | MTX_RECURSE);
1349 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1350 "CPU %d load", (int)TDQ_ID(tdq));
1356 sched_setup_smp(void)
1361 cpu_top = smp_topo();
1365 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1366 if (tdq->tdq_cg == NULL)
1367 panic("Can't find cpu group for %d\n", i);
1369 balance_tdq = TDQ_SELF();
1375 * Setup the thread queues and initialize the topology based on MD
1379 sched_setup(void *dummy)
1390 /* Add thread0's load since it's running. */
1392 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1393 tdq_load_add(tdq, &thread0);
1394 tdq->tdq_lowpri = thread0.td_priority;
1399 * This routine determines time constants after stathz and hz are setup.
1403 sched_initticks(void *dummy)
1407 realstathz = stathz ? stathz : hz;
1408 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1409 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1410 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1414 * tickincr is shifted out by 10 to avoid rounding errors due to
1415 * hz not being evenly divisible by stathz on all platforms.
1417 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1419 * This does not work for values of stathz that are more than
1420 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1427 * Set the default balance interval now that we know
1428 * what realstathz is.
1430 balance_interval = realstathz;
1431 affinity = SCHED_AFFINITY_DEFAULT;
1433 if (sched_idlespinthresh < 0)
1434 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1439 * This is the core of the interactivity algorithm. Determines a score based
1440 * on past behavior. It is the ratio of sleep time to run time scaled to
1441 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1442 * differs from the cpu usage because it does not account for time spent
1443 * waiting on a run-queue. Would be prettier if we had floating point.
1446 sched_interact_score(struct thread *td)
1448 struct td_sched *ts;
1453 * The score is only needed if this is likely to be an interactive
1454 * task. Don't go through the expense of computing it if there's
1457 if (sched_interact <= SCHED_INTERACT_HALF &&
1458 ts->ts_runtime >= ts->ts_slptime)
1459 return (SCHED_INTERACT_HALF);
1461 if (ts->ts_runtime > ts->ts_slptime) {
1462 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1463 return (SCHED_INTERACT_HALF +
1464 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1466 if (ts->ts_slptime > ts->ts_runtime) {
1467 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1468 return (ts->ts_runtime / div);
1470 /* runtime == slptime */
1472 return (SCHED_INTERACT_HALF);
1475 * This can happen if slptime and runtime are 0.
1482 * Scale the scheduling priority according to the "interactivity" of this
1486 sched_priority(struct thread *td)
1491 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1494 * If the score is interactive we place the thread in the realtime
1495 * queue with a priority that is less than kernel and interrupt
1496 * priorities. These threads are not subject to nice restrictions.
1498 * Scores greater than this are placed on the normal timeshare queue
1499 * where the priority is partially decided by the most recent cpu
1500 * utilization and the rest is decided by nice value.
1502 * The nice value of the process has a linear effect on the calculated
1503 * score. Negative nice values make it easier for a thread to be
1504 * considered interactive.
1506 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1507 if (score < sched_interact) {
1508 pri = PRI_MIN_INTERACT;
1509 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1510 sched_interact) * score;
1511 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1512 ("sched_priority: invalid interactive priority %d score %d",
1515 pri = SCHED_PRI_MIN;
1516 if (td->td_sched->ts_ticks)
1517 pri += min(SCHED_PRI_TICKS(td->td_sched),
1518 SCHED_PRI_RANGE - 1);
1519 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1520 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1521 ("sched_priority: invalid priority %d: nice %d, "
1522 "ticks %d ftick %d ltick %d tick pri %d",
1523 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1524 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1525 SCHED_PRI_TICKS(td->td_sched)));
1527 sched_user_prio(td, pri);
1533 * This routine enforces a maximum limit on the amount of scheduling history
1534 * kept. It is called after either the slptime or runtime is adjusted. This
1535 * function is ugly due to integer math.
1538 sched_interact_update(struct thread *td)
1540 struct td_sched *ts;
1544 sum = ts->ts_runtime + ts->ts_slptime;
1545 if (sum < SCHED_SLP_RUN_MAX)
1548 * This only happens from two places:
1549 * 1) We have added an unusual amount of run time from fork_exit.
1550 * 2) We have added an unusual amount of sleep time from sched_sleep().
1552 if (sum > SCHED_SLP_RUN_MAX * 2) {
1553 if (ts->ts_runtime > ts->ts_slptime) {
1554 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1557 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1563 * If we have exceeded by more than 1/5th then the algorithm below
1564 * will not bring us back into range. Dividing by two here forces
1565 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1567 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1568 ts->ts_runtime /= 2;
1569 ts->ts_slptime /= 2;
1572 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1573 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1577 * Scale back the interactivity history when a child thread is created. The
1578 * history is inherited from the parent but the thread may behave totally
1579 * differently. For example, a shell spawning a compiler process. We want
1580 * to learn that the compiler is behaving badly very quickly.
1583 sched_interact_fork(struct thread *td)
1588 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1589 if (sum > SCHED_SLP_RUN_FORK) {
1590 ratio = sum / SCHED_SLP_RUN_FORK;
1591 td->td_sched->ts_runtime /= ratio;
1592 td->td_sched->ts_slptime /= ratio;
1597 * Called from proc0_init() to setup the scheduler fields.
1604 * Set up the scheduler specific parts of proc0.
1606 proc0.p_sched = NULL; /* XXX */
1607 thread0.td_sched = &td_sched0;
1608 td_sched0.ts_ltick = ticks;
1609 td_sched0.ts_ftick = ticks;
1610 td_sched0.ts_slice = 0;
1614 * This is only somewhat accurate since given many processes of the same
1615 * priority they will switch when their slices run out, which will be
1616 * at most sched_slice stathz ticks.
1619 sched_rr_interval(void)
1622 /* Convert sched_slice from stathz to hz. */
1623 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1627 * Update the percent cpu tracking information when it is requested or
1628 * the total history exceeds the maximum. We keep a sliding history of
1629 * tick counts that slowly decays. This is less precise than the 4BSD
1630 * mechanism since it happens with less regular and frequent events.
1633 sched_pctcpu_update(struct td_sched *ts, int run)
1637 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1639 ts->ts_ftick = t - SCHED_TICK_TARG;
1640 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1641 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1642 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1643 ts->ts_ftick = t - SCHED_TICK_TARG;
1646 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1651 * Adjust the priority of a thread. Move it to the appropriate run-queue
1652 * if necessary. This is the back-end for several priority related
1656 sched_thread_priority(struct thread *td, u_char prio)
1658 struct td_sched *ts;
1662 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1663 "prio:%d", td->td_priority, "new prio:%d", prio,
1664 KTR_ATTR_LINKED, sched_tdname(curthread));
1665 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1666 if (td != curthread && prio < td->td_priority) {
1667 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1668 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1669 prio, KTR_ATTR_LINKED, sched_tdname(td));
1670 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1674 THREAD_LOCK_ASSERT(td, MA_OWNED);
1675 if (td->td_priority == prio)
1678 * If the priority has been elevated due to priority
1679 * propagation, we may have to move ourselves to a new
1680 * queue. This could be optimized to not re-add in some
1683 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1685 td->td_priority = prio;
1686 sched_add(td, SRQ_BORROWING);
1690 * If the thread is currently running we may have to adjust the lowpri
1691 * information so other cpus are aware of our current priority.
1693 if (TD_IS_RUNNING(td)) {
1694 tdq = TDQ_CPU(ts->ts_cpu);
1695 oldpri = td->td_priority;
1696 td->td_priority = prio;
1697 if (prio < tdq->tdq_lowpri)
1698 tdq->tdq_lowpri = prio;
1699 else if (tdq->tdq_lowpri == oldpri)
1700 tdq_setlowpri(tdq, td);
1703 td->td_priority = prio;
1707 * Update a thread's priority when it is lent another thread's
1711 sched_lend_prio(struct thread *td, u_char prio)
1714 td->td_flags |= TDF_BORROWING;
1715 sched_thread_priority(td, prio);
1719 * Restore a thread's priority when priority propagation is
1720 * over. The prio argument is the minimum priority the thread
1721 * needs to have to satisfy other possible priority lending
1722 * requests. If the thread's regular priority is less
1723 * important than prio, the thread will keep a priority boost
1727 sched_unlend_prio(struct thread *td, u_char prio)
1731 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1732 td->td_base_pri <= PRI_MAX_TIMESHARE)
1733 base_pri = td->td_user_pri;
1735 base_pri = td->td_base_pri;
1736 if (prio >= base_pri) {
1737 td->td_flags &= ~TDF_BORROWING;
1738 sched_thread_priority(td, base_pri);
1740 sched_lend_prio(td, prio);
1744 * Standard entry for setting the priority to an absolute value.
1747 sched_prio(struct thread *td, u_char prio)
1751 /* First, update the base priority. */
1752 td->td_base_pri = prio;
1755 * If the thread is borrowing another thread's priority, don't
1756 * ever lower the priority.
1758 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1761 /* Change the real priority. */
1762 oldprio = td->td_priority;
1763 sched_thread_priority(td, prio);
1766 * If the thread is on a turnstile, then let the turnstile update
1769 if (TD_ON_LOCK(td) && oldprio != prio)
1770 turnstile_adjust(td, oldprio);
1774 * Set the base user priority, does not effect current running priority.
1777 sched_user_prio(struct thread *td, u_char prio)
1780 td->td_base_user_pri = prio;
1781 if (td->td_lend_user_pri <= prio)
1783 td->td_user_pri = prio;
1787 sched_lend_user_prio(struct thread *td, u_char prio)
1790 THREAD_LOCK_ASSERT(td, MA_OWNED);
1791 td->td_lend_user_pri = prio;
1792 td->td_user_pri = min(prio, td->td_base_user_pri);
1793 if (td->td_priority > td->td_user_pri)
1794 sched_prio(td, td->td_user_pri);
1795 else if (td->td_priority != td->td_user_pri)
1796 td->td_flags |= TDF_NEEDRESCHED;
1800 * Handle migration from sched_switch(). This happens only for
1804 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1808 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1810 tdq_load_rem(tdq, td);
1812 * Do the lock dance required to avoid LOR. We grab an extra
1813 * spinlock nesting to prevent preemption while we're
1814 * not holding either run-queue lock.
1817 thread_lock_block(td); /* This releases the lock on tdq. */
1820 * Acquire both run-queue locks before placing the thread on the new
1821 * run-queue to avoid deadlocks created by placing a thread with a
1822 * blocked lock on the run-queue of a remote processor. The deadlock
1823 * occurs when a third processor attempts to lock the two queues in
1824 * question while the target processor is spinning with its own
1825 * run-queue lock held while waiting for the blocked lock to clear.
1827 tdq_lock_pair(tdn, tdq);
1828 tdq_add(tdn, td, flags);
1829 tdq_notify(tdn, td);
1833 return (TDQ_LOCKPTR(tdn));
1837 * Variadic version of thread_lock_unblock() that does not assume td_lock
1841 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1843 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1848 * Switch threads. This function has to handle threads coming in while
1849 * blocked for some reason, running, or idle. It also must deal with
1850 * migrating a thread from one queue to another as running threads may
1851 * be assigned elsewhere via binding.
1854 sched_switch(struct thread *td, struct thread *newtd, int flags)
1857 struct td_sched *ts;
1860 int cpuid, preempted;
1862 THREAD_LOCK_ASSERT(td, MA_OWNED);
1863 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1865 cpuid = PCPU_GET(cpuid);
1866 tdq = TDQ_CPU(cpuid);
1869 sched_pctcpu_update(ts, 1);
1870 ts->ts_rltick = ticks;
1871 td->td_lastcpu = td->td_oncpu;
1872 td->td_oncpu = NOCPU;
1873 preempted = !((td->td_flags & TDF_SLICEEND) ||
1874 (flags & SWT_RELINQUISH));
1875 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1876 td->td_owepreempt = 0;
1877 if (!TD_IS_IDLETHREAD(td))
1878 tdq->tdq_switchcnt++;
1880 * The lock pointer in an idle thread should never change. Reset it
1881 * to CAN_RUN as well.
1883 if (TD_IS_IDLETHREAD(td)) {
1884 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1886 } else if (TD_IS_RUNNING(td)) {
1887 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1888 srqflag = preempted ?
1889 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1890 SRQ_OURSELF|SRQ_YIELDING;
1892 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1893 ts->ts_cpu = sched_pickcpu(td, 0);
1895 if (ts->ts_cpu == cpuid)
1896 tdq_runq_add(tdq, td, srqflag);
1898 KASSERT(THREAD_CAN_MIGRATE(td) ||
1899 (ts->ts_flags & TSF_BOUND) != 0,
1900 ("Thread %p shouldn't migrate", td));
1901 mtx = sched_switch_migrate(tdq, td, srqflag);
1904 /* This thread must be going to sleep. */
1906 mtx = thread_lock_block(td);
1907 tdq_load_rem(tdq, td);
1910 * We enter here with the thread blocked and assigned to the
1911 * appropriate cpu run-queue or sleep-queue and with the current
1912 * thread-queue locked.
1914 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1915 newtd = choosethread();
1917 * Call the MD code to switch contexts if necessary.
1921 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1922 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1924 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1925 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1926 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1927 sched_pctcpu_update(newtd->td_sched, 0);
1929 #ifdef KDTRACE_HOOKS
1931 * If DTrace has set the active vtime enum to anything
1932 * other than INACTIVE (0), then it should have set the
1935 if (dtrace_vtime_active)
1936 (*dtrace_vtime_switch_func)(newtd);
1939 cpu_switch(td, newtd, mtx);
1941 * We may return from cpu_switch on a different cpu. However,
1942 * we always return with td_lock pointing to the current cpu's
1945 cpuid = PCPU_GET(cpuid);
1946 tdq = TDQ_CPU(cpuid);
1947 lock_profile_obtain_lock_success(
1948 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1950 SDT_PROBE0(sched, , , on__cpu);
1952 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1953 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1956 thread_unblock_switch(td, mtx);
1957 SDT_PROBE0(sched, , , remain__cpu);
1960 * Assert that all went well and return.
1962 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1963 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1964 td->td_oncpu = cpuid;
1968 * Adjust thread priorities as a result of a nice request.
1971 sched_nice(struct proc *p, int nice)
1975 PROC_LOCK_ASSERT(p, MA_OWNED);
1978 FOREACH_THREAD_IN_PROC(p, td) {
1981 sched_prio(td, td->td_base_user_pri);
1987 * Record the sleep time for the interactivity scorer.
1990 sched_sleep(struct thread *td, int prio)
1993 THREAD_LOCK_ASSERT(td, MA_OWNED);
1995 td->td_slptick = ticks;
1996 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1997 td->td_flags |= TDF_CANSWAP;
1998 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2000 if (static_boost == 1 && prio)
2001 sched_prio(td, prio);
2002 else if (static_boost && td->td_priority > static_boost)
2003 sched_prio(td, static_boost);
2007 * Schedule a thread to resume execution and record how long it voluntarily
2008 * slept. We also update the pctcpu, interactivity, and priority.
2011 sched_wakeup(struct thread *td)
2013 struct td_sched *ts;
2016 THREAD_LOCK_ASSERT(td, MA_OWNED);
2018 td->td_flags &= ~TDF_CANSWAP;
2020 * If we slept for more than a tick update our interactivity and
2023 slptick = td->td_slptick;
2025 if (slptick && slptick != ticks) {
2026 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2027 sched_interact_update(td);
2028 sched_pctcpu_update(ts, 0);
2031 * Reset the slice value since we slept and advanced the round-robin.
2034 sched_add(td, SRQ_BORING);
2038 * Penalize the parent for creating a new child and initialize the child's
2042 sched_fork(struct thread *td, struct thread *child)
2044 THREAD_LOCK_ASSERT(td, MA_OWNED);
2045 sched_pctcpu_update(td->td_sched, 1);
2046 sched_fork_thread(td, child);
2048 * Penalize the parent and child for forking.
2050 sched_interact_fork(child);
2051 sched_priority(child);
2052 td->td_sched->ts_runtime += tickincr;
2053 sched_interact_update(td);
2058 * Fork a new thread, may be within the same process.
2061 sched_fork_thread(struct thread *td, struct thread *child)
2063 struct td_sched *ts;
2064 struct td_sched *ts2;
2068 THREAD_LOCK_ASSERT(td, MA_OWNED);
2073 ts2 = child->td_sched;
2074 child->td_lock = TDQ_LOCKPTR(tdq);
2075 child->td_cpuset = cpuset_ref(td->td_cpuset);
2076 ts2->ts_cpu = ts->ts_cpu;
2079 * Grab our parents cpu estimation information.
2081 ts2->ts_ticks = ts->ts_ticks;
2082 ts2->ts_ltick = ts->ts_ltick;
2083 ts2->ts_ftick = ts->ts_ftick;
2085 * Do not inherit any borrowed priority from the parent.
2087 child->td_priority = child->td_base_pri;
2089 * And update interactivity score.
2091 ts2->ts_slptime = ts->ts_slptime;
2092 ts2->ts_runtime = ts->ts_runtime;
2093 /* Attempt to quickly learn interactivity. */
2094 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2096 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2101 * Adjust the priority class of a thread.
2104 sched_class(struct thread *td, int class)
2107 THREAD_LOCK_ASSERT(td, MA_OWNED);
2108 if (td->td_pri_class == class)
2110 td->td_pri_class = class;
2114 * Return some of the child's priority and interactivity to the parent.
2117 sched_exit(struct proc *p, struct thread *child)
2121 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2122 "prio:%d", child->td_priority);
2123 PROC_LOCK_ASSERT(p, MA_OWNED);
2124 td = FIRST_THREAD_IN_PROC(p);
2125 sched_exit_thread(td, child);
2129 * Penalize another thread for the time spent on this one. This helps to
2130 * worsen the priority and interactivity of processes which schedule batch
2131 * jobs such as make. This has little effect on the make process itself but
2132 * causes new processes spawned by it to receive worse scores immediately.
2135 sched_exit_thread(struct thread *td, struct thread *child)
2138 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2139 "prio:%d", child->td_priority);
2141 * Give the child's runtime to the parent without returning the
2142 * sleep time as a penalty to the parent. This causes shells that
2143 * launch expensive things to mark their children as expensive.
2146 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2147 sched_interact_update(td);
2153 sched_preempt(struct thread *td)
2157 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2161 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2162 tdq->tdq_ipipending = 0;
2163 if (td->td_priority > tdq->tdq_lowpri) {
2166 flags = SW_INVOL | SW_PREEMPT;
2167 if (td->td_critnest > 1)
2168 td->td_owepreempt = 1;
2169 else if (TD_IS_IDLETHREAD(td))
2170 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2172 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2178 * Fix priorities on return to user-space. Priorities may be elevated due
2179 * to static priorities in msleep() or similar.
2182 sched_userret(struct thread *td)
2185 * XXX we cheat slightly on the locking here to avoid locking in
2186 * the usual case. Setting td_priority here is essentially an
2187 * incomplete workaround for not setting it properly elsewhere.
2188 * Now that some interrupt handlers are threads, not setting it
2189 * properly elsewhere can clobber it in the window between setting
2190 * it here and returning to user mode, so don't waste time setting
2191 * it perfectly here.
2193 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2194 ("thread with borrowed priority returning to userland"));
2195 if (td->td_priority != td->td_user_pri) {
2197 td->td_priority = td->td_user_pri;
2198 td->td_base_pri = td->td_user_pri;
2199 tdq_setlowpri(TDQ_SELF(), td);
2205 * Handle a stathz tick. This is really only relevant for timeshare
2209 sched_clock(struct thread *td)
2212 struct td_sched *ts;
2214 THREAD_LOCK_ASSERT(td, MA_OWNED);
2218 * We run the long term load balancer infrequently on the first cpu.
2220 if (balance_tdq == tdq) {
2221 if (balance_ticks && --balance_ticks == 0)
2226 * Save the old switch count so we have a record of the last ticks
2227 * activity. Initialize the new switch count based on our load.
2228 * If there is some activity seed it to reflect that.
2230 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2231 tdq->tdq_switchcnt = tdq->tdq_load;
2233 * Advance the insert index once for each tick to ensure that all
2234 * threads get a chance to run.
2236 if (tdq->tdq_idx == tdq->tdq_ridx) {
2237 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2238 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2239 tdq->tdq_ridx = tdq->tdq_idx;
2242 sched_pctcpu_update(ts, 1);
2243 if (td->td_pri_class & PRI_FIFO_BIT)
2245 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2247 * We used a tick; charge it to the thread so
2248 * that we can compute our interactivity.
2250 td->td_sched->ts_runtime += tickincr;
2251 sched_interact_update(td);
2256 * Force a context switch if the current thread has used up a full
2257 * time slice (default is 100ms).
2259 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2261 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2266 * Called once per hz tick.
2275 * Return whether the current CPU has runnable tasks. Used for in-kernel
2276 * cooperative idle threads.
2279 sched_runnable(void)
2287 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2288 if (tdq->tdq_load > 0)
2291 if (tdq->tdq_load - 1 > 0)
2299 * Choose the highest priority thread to run. The thread is removed from
2300 * the run-queue while running however the load remains. For SMP we set
2301 * the tdq in the global idle bitmask if it idles here.
2310 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2311 td = tdq_choose(tdq);
2313 tdq_runq_rem(tdq, td);
2314 tdq->tdq_lowpri = td->td_priority;
2317 tdq->tdq_lowpri = PRI_MAX_IDLE;
2318 return (PCPU_GET(idlethread));
2322 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2323 * we always request it once we exit a critical section.
2326 sched_setpreempt(struct thread *td)
2332 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2335 pri = td->td_priority;
2336 cpri = ctd->td_priority;
2338 ctd->td_flags |= TDF_NEEDRESCHED;
2339 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2341 if (!sched_shouldpreempt(pri, cpri, 0))
2343 ctd->td_owepreempt = 1;
2347 * Add a thread to a thread queue. Select the appropriate runq and add the
2348 * thread to it. This is the internal function called when the tdq is
2352 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2355 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2356 KASSERT((td->td_inhibitors == 0),
2357 ("sched_add: trying to run inhibited thread"));
2358 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2359 ("sched_add: bad thread state"));
2360 KASSERT(td->td_flags & TDF_INMEM,
2361 ("sched_add: thread swapped out"));
2363 if (td->td_priority < tdq->tdq_lowpri)
2364 tdq->tdq_lowpri = td->td_priority;
2365 tdq_runq_add(tdq, td, flags);
2366 tdq_load_add(tdq, td);
2370 * Select the target thread queue and add a thread to it. Request
2371 * preemption or IPI a remote processor if required.
2374 sched_add(struct thread *td, int flags)
2381 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2382 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2383 sched_tdname(curthread));
2384 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2385 KTR_ATTR_LINKED, sched_tdname(td));
2386 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2387 flags & SRQ_PREEMPTED);
2388 THREAD_LOCK_ASSERT(td, MA_OWNED);
2390 * Recalculate the priority before we select the target cpu or
2393 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2397 * Pick the destination cpu and if it isn't ours transfer to the
2400 cpu = sched_pickcpu(td, flags);
2401 tdq = sched_setcpu(td, cpu, flags);
2402 tdq_add(tdq, td, flags);
2403 if (cpu != PCPU_GET(cpuid)) {
2404 tdq_notify(tdq, td);
2411 * Now that the thread is moving to the run-queue, set the lock
2412 * to the scheduler's lock.
2414 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2415 tdq_add(tdq, td, flags);
2417 if (!(flags & SRQ_YIELDING))
2418 sched_setpreempt(td);
2422 * Remove a thread from a run-queue without running it. This is used
2423 * when we're stealing a thread from a remote queue. Otherwise all threads
2424 * exit by calling sched_exit_thread() and sched_throw() themselves.
2427 sched_rem(struct thread *td)
2431 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2432 "prio:%d", td->td_priority);
2433 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2434 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2435 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2436 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2437 KASSERT(TD_ON_RUNQ(td),
2438 ("sched_rem: thread not on run queue"));
2439 tdq_runq_rem(tdq, td);
2440 tdq_load_rem(tdq, td);
2442 if (td->td_priority == tdq->tdq_lowpri)
2443 tdq_setlowpri(tdq, NULL);
2447 * Fetch cpu utilization information. Updates on demand.
2450 sched_pctcpu(struct thread *td)
2453 struct td_sched *ts;
2460 THREAD_LOCK_ASSERT(td, MA_OWNED);
2461 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2465 /* How many rtick per second ? */
2466 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2467 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2474 * Enforce affinity settings for a thread. Called after adjustments to
2478 sched_affinity(struct thread *td)
2481 struct td_sched *ts;
2483 THREAD_LOCK_ASSERT(td, MA_OWNED);
2485 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2487 if (TD_ON_RUNQ(td)) {
2489 sched_add(td, SRQ_BORING);
2492 if (!TD_IS_RUNNING(td))
2495 * Force a switch before returning to userspace. If the
2496 * target thread is not running locally send an ipi to force
2499 td->td_flags |= TDF_NEEDRESCHED;
2500 if (td != curthread)
2501 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2506 * Bind a thread to a target cpu.
2509 sched_bind(struct thread *td, int cpu)
2511 struct td_sched *ts;
2513 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2514 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2516 if (ts->ts_flags & TSF_BOUND)
2518 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2519 ts->ts_flags |= TSF_BOUND;
2521 if (PCPU_GET(cpuid) == cpu)
2524 /* When we return from mi_switch we'll be on the correct cpu. */
2525 mi_switch(SW_VOL, NULL);
2529 * Release a bound thread.
2532 sched_unbind(struct thread *td)
2534 struct td_sched *ts;
2536 THREAD_LOCK_ASSERT(td, MA_OWNED);
2537 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2539 if ((ts->ts_flags & TSF_BOUND) == 0)
2541 ts->ts_flags &= ~TSF_BOUND;
2546 sched_is_bound(struct thread *td)
2548 THREAD_LOCK_ASSERT(td, MA_OWNED);
2549 return (td->td_sched->ts_flags & TSF_BOUND);
2556 sched_relinquish(struct thread *td)
2559 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2564 * Return the total system load.
2575 total += TDQ_CPU(i)->tdq_sysload;
2578 return (TDQ_SELF()->tdq_sysload);
2583 sched_sizeof_proc(void)
2585 return (sizeof(struct proc));
2589 sched_sizeof_thread(void)
2591 return (sizeof(struct thread) + sizeof(struct td_sched));
2595 #define TDQ_IDLESPIN(tdq) \
2596 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2598 #define TDQ_IDLESPIN(tdq) 1
2602 * The actual idle process.
2605 sched_idletd(void *dummy)
2609 int oldswitchcnt, switchcnt;
2612 mtx_assert(&Giant, MA_NOTOWNED);
2615 THREAD_NO_SLEEPING();
2618 if (tdq->tdq_load) {
2620 mi_switch(SW_VOL | SWT_IDLE, NULL);
2623 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2625 if (switchcnt != oldswitchcnt) {
2626 oldswitchcnt = switchcnt;
2627 if (tdq_idled(tdq) == 0)
2630 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2632 oldswitchcnt = switchcnt;
2635 * If we're switching very frequently, spin while checking
2636 * for load rather than entering a low power state that
2637 * may require an IPI. However, don't do any busy
2638 * loops while on SMT machines as this simply steals
2639 * cycles from cores doing useful work.
2641 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2642 for (i = 0; i < sched_idlespins; i++) {
2649 /* If there was context switch during spin, restart it. */
2650 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2651 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2654 /* Run main MD idle handler. */
2655 tdq->tdq_cpu_idle = 1;
2657 * Make sure that tdq_cpu_idle update is globally visible
2658 * before cpu_idle() read tdq_load. The order is important
2659 * to avoid race with tdq_notify.
2662 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2663 tdq->tdq_cpu_idle = 0;
2666 * Account thread-less hardware interrupts and
2667 * other wakeup reasons equal to context switches.
2669 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2670 if (switchcnt != oldswitchcnt)
2672 tdq->tdq_switchcnt++;
2678 * A CPU is entering for the first time or a thread is exiting.
2681 sched_throw(struct thread *td)
2683 struct thread *newtd;
2688 /* Correct spinlock nesting and acquire the correct lock. */
2691 PCPU_SET(switchtime, cpu_ticks());
2692 PCPU_SET(switchticks, ticks);
2694 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2695 tdq_load_rem(tdq, td);
2696 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2698 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2699 newtd = choosethread();
2700 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2701 cpu_throw(td, newtd); /* doesn't return */
2705 * This is called from fork_exit(). Just acquire the correct locks and
2706 * let fork do the rest of the work.
2709 sched_fork_exit(struct thread *td)
2715 * Finish setting up thread glue so that it begins execution in a
2716 * non-nested critical section with the scheduler lock held.
2718 cpuid = PCPU_GET(cpuid);
2719 tdq = TDQ_CPU(cpuid);
2720 if (TD_IS_IDLETHREAD(td))
2721 td->td_lock = TDQ_LOCKPTR(tdq);
2722 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2723 td->td_oncpu = cpuid;
2724 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2725 lock_profile_obtain_lock_success(
2726 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2730 * Create on first use to catch odd startup conditons.
2733 sched_tdname(struct thread *td)
2736 struct td_sched *ts;
2739 if (ts->ts_name[0] == '\0')
2740 snprintf(ts->ts_name, sizeof(ts->ts_name),
2741 "%s tid %d", td->td_name, td->td_tid);
2742 return (ts->ts_name);
2744 return (td->td_name);
2750 sched_clear_tdname(struct thread *td)
2752 struct td_sched *ts;
2755 ts->ts_name[0] = '\0';
2762 * Build the CPU topology dump string. Is recursively called to collect
2763 * the topology tree.
2766 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2769 char cpusetbuf[CPUSETBUFSIZ];
2772 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2773 "", 1 + indent / 2, cg->cg_level);
2774 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2775 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2777 for (i = 0; i < MAXCPU; i++) {
2778 if (CPU_ISSET(i, &cg->cg_mask)) {
2780 sbuf_printf(sb, ", ");
2783 sbuf_printf(sb, "%d", i);
2786 sbuf_printf(sb, "</cpu>\n");
2788 if (cg->cg_flags != 0) {
2789 sbuf_printf(sb, "%*s <flags>", indent, "");
2790 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2791 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2792 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2793 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2794 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2795 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2796 sbuf_printf(sb, "</flags>\n");
2799 if (cg->cg_children > 0) {
2800 sbuf_printf(sb, "%*s <children>\n", indent, "");
2801 for (i = 0; i < cg->cg_children; i++)
2802 sysctl_kern_sched_topology_spec_internal(sb,
2803 &cg->cg_child[i], indent+2);
2804 sbuf_printf(sb, "%*s </children>\n", indent, "");
2806 sbuf_printf(sb, "%*s</group>\n", indent, "");
2811 * Sysctl handler for retrieving topology dump. It's a wrapper for
2812 * the recursive sysctl_kern_smp_topology_spec_internal().
2815 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2820 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2822 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2826 sbuf_printf(topo, "<groups>\n");
2827 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2828 sbuf_printf(topo, "</groups>\n");
2832 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2841 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2843 int error, new_val, period;
2845 period = 1000000 / realstathz;
2846 new_val = period * sched_slice;
2847 error = sysctl_handle_int(oidp, &new_val, 0, req);
2848 if (error != 0 || req->newptr == NULL)
2852 sched_slice = imax(1, (new_val + period / 2) / period);
2853 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2854 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2859 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2860 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2862 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2863 NULL, 0, sysctl_kern_quantum, "I",
2864 "Quantum for timeshare threads in microseconds");
2865 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2866 "Quantum for timeshare threads in stathz ticks");
2867 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2868 "Interactivity score threshold");
2869 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2871 "Maximal (lowest) priority for preemption");
2872 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2873 "Assign static kernel priorities to sleeping threads");
2874 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2875 "Number of times idle thread will spin waiting for new work");
2876 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2877 &sched_idlespinthresh, 0,
2878 "Threshold before we will permit idle thread spinning");
2880 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2881 "Number of hz ticks to keep thread affinity for");
2882 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2883 "Enables the long-term load balancer");
2884 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2885 &balance_interval, 0,
2886 "Average period in stathz ticks to run the long-term balancer");
2887 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2888 "Attempts to steal work from other cores before idling");
2889 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2890 "Minimum load on remote CPU before we'll steal");
2891 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2892 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2893 "XML dump of detected CPU topology");
2896 /* ps compat. All cpu percentages from ULE are weighted. */
2897 static int ccpu = 0;
2898 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");