2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 * This file implements the ULE scheduler. ULE supports independent CPU
31 * run queues and fine grain locking. It has superior interactive
32 * performance under load even on uni-processor systems.
35 * ULE is the last three letters in schedule. It owes its name to a
36 * generic user created for a scheduling system by Paul Mikesell at
37 * Isilon Systems and a general lack of creativity on the part of the author.
40 #include <sys/cdefs.h>
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
44 #include <sys/param.h>
45 #include <sys/systm.h>
47 #include <sys/kernel.h>
49 #include <sys/limits.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>
62 #include <sys/umtxvar.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
68 #include <sys/pmckern.h>
72 #include <sys/dtrace_bsd.h>
73 int __read_mostly 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 int 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 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
110 #define THREAD_CAN_SCHED(td, cpu) \
111 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
113 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
114 sizeof(struct thread0_storage),
115 "increase struct thread0_storage.t0st_sched size");
118 * Priority ranges used for interactive and non-interactive timeshare
119 * threads. The timeshare priorities are split up into four ranges.
120 * The first range handles interactive threads. The last three ranges
121 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
122 * ranges supporting nice values.
124 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
125 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
126 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
128 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
129 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
130 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
131 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
134 * Cpu percentage computation macros and defines.
136 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
137 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
138 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
139 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
140 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
141 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
143 #define SCHED_TICK_SECS 10
144 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
145 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
146 #define SCHED_TICK_SHIFT 10
147 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
148 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 * These macros determine priorities for non-interactive threads. They are
152 * assigned a priority based on their recent cpu utilization as expressed
153 * by the ratio of ticks to the tick total. NHALF priorities at the start
154 * and end of the MIN to MAX timeshare range are only reachable with negative
155 * or positive nice respectively.
157 * PRI_RANGE: Priority range for utilization dependent priorities.
158 * PRI_NRESV: Number of nice values.
159 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
160 * PRI_NICE: Determines the part of the priority inherited from nice.
162 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
163 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
164 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
165 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
166 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
167 #define SCHED_PRI_TICKS(ts) \
168 (SCHED_TICK_HZ((ts)) / \
169 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
170 #define SCHED_PRI_NICE(nice) (nice)
173 * These determine the interactivity of a process. Interactivity differs from
174 * cpu utilization in that it expresses the voluntary time slept vs time ran
175 * while cpu utilization includes all time not running. This more accurately
176 * models the intent of the thread.
178 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
179 * before throttling back.
180 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
181 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
182 * INTERACT_THRESH: Threshold for placement on the current runq.
184 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
185 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
186 #define SCHED_INTERACT_MAX (100)
187 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
188 #define SCHED_INTERACT_THRESH (30)
191 * These parameters determine the slice behavior for batch work.
193 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
194 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
196 /* Flags kept in td_flags. */
197 #define TDF_PICKCPU TDF_SCHED0 /* Thread should pick new CPU. */
198 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
201 * tickincr: Converts a stathz tick into a hz domain scaled by
202 * the shift factor. Without the shift the error rate
203 * due to rounding would be unacceptably high.
204 * realstathz: stathz is sometimes 0 and run off of hz.
205 * sched_slice: Runtime of each thread before rescheduling.
206 * preempt_thresh: Priority threshold for preemption and remote IPIs.
208 static u_int __read_mostly sched_interact = SCHED_INTERACT_THRESH;
209 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
210 static int __read_mostly realstathz = 127; /* reset during boot. */
211 static int __read_mostly sched_slice = 10; /* reset during boot. */
212 static int __read_mostly sched_slice_min = 1; /* reset during boot. */
214 #ifdef FULL_PREEMPTION
215 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
217 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
220 static int __read_mostly preempt_thresh = 0;
222 static int __read_mostly static_boost = PRI_MIN_BATCH;
223 static int __read_mostly sched_idlespins = 10000;
224 static int __read_mostly sched_idlespinthresh = -1;
227 * tdq - per processor runqs and statistics. A mutex synchronizes access to
228 * most fields. Some fields are loaded or modified without the mutex.
231 * (c) constant after initialization
232 * (f) flag, set with the tdq lock held, cleared on local CPU
233 * (l) all accesses are CPU-local
234 * (ls) stores are performed by the local CPU, loads may be lockless
235 * (t) all accesses are protected by the tdq mutex
236 * (ts) stores are serialized by the tdq mutex, loads may be lockless
240 * Ordered to improve efficiency of cpu_search() and switch().
241 * tdq_lock is padded to avoid false sharing with tdq_load and
244 struct mtx_padalign tdq_lock; /* run queue lock. */
245 struct cpu_group *tdq_cg; /* (c) Pointer to cpu topology. */
246 struct thread *tdq_curthread; /* (t) Current executing thread. */
247 int tdq_load; /* (ts) Aggregate load. */
248 int tdq_sysload; /* (ts) For loadavg, !ITHD load. */
249 int tdq_cpu_idle; /* (ls) cpu_idle() is active. */
250 int tdq_transferable; /* (ts) Transferable thread count. */
251 short tdq_switchcnt; /* (l) Switches this tick. */
252 short tdq_oldswitchcnt; /* (l) Switches last tick. */
253 u_char tdq_lowpri; /* (ts) Lowest priority thread. */
254 u_char tdq_owepreempt; /* (f) Remote preemption pending. */
255 u_char tdq_idx; /* (t) Current insert index. */
256 u_char tdq_ridx; /* (t) Current removal index. */
257 int tdq_id; /* (c) cpuid. */
258 struct runq tdq_realtime; /* (t) real-time run queue. */
259 struct runq tdq_timeshare; /* (t) timeshare run queue. */
260 struct runq tdq_idle; /* (t) Queue of IDLE threads. */
261 char tdq_name[TDQ_NAME_LEN];
263 char tdq_loadname[TDQ_LOADNAME_LEN];
267 /* Idle thread states and config. */
268 #define TDQ_RUNNING 1
271 /* Lockless accessors. */
272 #define TDQ_LOAD(tdq) atomic_load_int(&(tdq)->tdq_load)
273 #define TDQ_TRANSFERABLE(tdq) atomic_load_int(&(tdq)->tdq_transferable)
274 #define TDQ_SWITCHCNT(tdq) (atomic_load_short(&(tdq)->tdq_switchcnt) + \
275 atomic_load_short(&(tdq)->tdq_oldswitchcnt))
276 #define TDQ_SWITCHCNT_INC(tdq) (atomic_store_short(&(tdq)->tdq_switchcnt, \
277 atomic_load_short(&(tdq)->tdq_switchcnt) + 1))
280 struct cpu_group __read_mostly *cpu_top; /* CPU topology */
282 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
283 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
288 static int rebalance = 1;
289 static int balance_interval = 128; /* Default set in sched_initticks(). */
290 static int __read_mostly affinity;
291 static int __read_mostly steal_idle = 1;
292 static int __read_mostly steal_thresh = 2;
293 static int __read_mostly always_steal = 0;
294 static int __read_mostly trysteal_limit = 2;
297 * One thread queue per processor.
299 static struct tdq __read_mostly *balance_tdq;
300 static int balance_ticks;
301 DPCPU_DEFINE_STATIC(struct tdq, tdq);
302 DPCPU_DEFINE_STATIC(uint32_t, randomval);
304 #define TDQ_SELF() ((struct tdq *)PCPU_GET(sched))
305 #define TDQ_CPU(x) (DPCPU_ID_PTR((x), tdq))
306 #define TDQ_ID(x) ((x)->tdq_id)
308 static struct tdq tdq_cpu;
310 #define TDQ_ID(x) (0)
311 #define TDQ_SELF() (&tdq_cpu)
312 #define TDQ_CPU(x) (&tdq_cpu)
315 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
316 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
317 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
318 #define TDQ_TRYLOCK(t) mtx_trylock_spin(TDQ_LOCKPTR((t)))
319 #define TDQ_TRYLOCK_FLAGS(t, f) mtx_trylock_spin_flags(TDQ_LOCKPTR((t)), (f))
320 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
321 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
323 static void sched_setpreempt(int);
324 static void sched_priority(struct thread *);
325 static void sched_thread_priority(struct thread *, u_char);
326 static int sched_interact_score(struct thread *);
327 static void sched_interact_update(struct thread *);
328 static void sched_interact_fork(struct thread *);
329 static void sched_pctcpu_update(struct td_sched *, int);
331 /* Operations on per processor queues */
332 static struct thread *tdq_choose(struct tdq *);
333 static void tdq_setup(struct tdq *, int i);
334 static void tdq_load_add(struct tdq *, struct thread *);
335 static void tdq_load_rem(struct tdq *, struct thread *);
336 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
337 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
338 static inline int sched_shouldpreempt(int, int, int);
339 static void tdq_print(int cpu);
340 static void runq_print(struct runq *rq);
341 static int tdq_add(struct tdq *, struct thread *, int);
343 static int tdq_move(struct tdq *, struct tdq *);
344 static int tdq_idled(struct tdq *);
345 static void tdq_notify(struct tdq *, int lowpri);
346 static struct thread *tdq_steal(struct tdq *, int);
347 static struct thread *runq_steal(struct runq *, int);
348 static int sched_pickcpu(struct thread *, int);
349 static void sched_balance(void);
350 static bool sched_balance_pair(struct tdq *, struct tdq *);
351 static inline struct tdq *sched_setcpu(struct thread *, int, int);
352 static inline void thread_unblock_switch(struct thread *, struct mtx *);
353 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
354 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
355 struct cpu_group *cg, int indent);
358 static void sched_setup(void *dummy);
359 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
361 static void sched_initticks(void *dummy);
362 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
365 SDT_PROVIDER_DEFINE(sched);
367 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
368 "struct proc *", "uint8_t");
369 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
370 "struct proc *", "void *");
371 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
372 "struct proc *", "void *", "int");
373 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
374 "struct proc *", "uint8_t", "struct thread *");
375 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
376 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
378 SDT_PROBE_DEFINE(sched, , , on__cpu);
379 SDT_PROBE_DEFINE(sched, , , remain__cpu);
380 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
384 * Print the threads waiting on a run-queue.
387 runq_print(struct runq *rq)
395 for (i = 0; i < RQB_LEN; i++) {
396 printf("\t\trunq bits %d 0x%zx\n",
397 i, rq->rq_status.rqb_bits[i]);
398 for (j = 0; j < RQB_BPW; j++)
399 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
400 pri = j + (i << RQB_L2BPW);
401 rqh = &rq->rq_queues[pri];
402 TAILQ_FOREACH(td, rqh, td_runq) {
403 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
404 td, td->td_name, td->td_priority,
405 td->td_rqindex, pri);
412 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
421 printf("tdq %d:\n", TDQ_ID(tdq));
422 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
423 printf("\tLock name: %s\n", tdq->tdq_name);
424 printf("\tload: %d\n", tdq->tdq_load);
425 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
426 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
427 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
428 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
429 printf("\tload transferable: %d\n", tdq->tdq_transferable);
430 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
431 printf("\trealtime runq:\n");
432 runq_print(&tdq->tdq_realtime);
433 printf("\ttimeshare runq:\n");
434 runq_print(&tdq->tdq_timeshare);
435 printf("\tidle runq:\n");
436 runq_print(&tdq->tdq_idle);
440 sched_shouldpreempt(int pri, int cpri, int remote)
443 * If the new priority is not better than the current priority there is
449 * Always preempt idle.
451 if (cpri >= PRI_MIN_IDLE)
454 * If preemption is disabled don't preempt others.
456 if (preempt_thresh == 0)
459 * Preempt if we exceed the threshold.
461 if (pri <= preempt_thresh)
464 * If we're interactive or better and there is non-interactive
465 * or worse running preempt only remote processors.
467 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
473 * Add a thread to the actual run-queue. Keeps transferable counts up to
474 * date with what is actually on the run-queue. Selects the correct
475 * queue position for timeshare threads.
478 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
483 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
484 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
486 pri = td->td_priority;
487 ts = td_get_sched(td);
489 if (THREAD_CAN_MIGRATE(td)) {
490 tdq->tdq_transferable++;
491 ts->ts_flags |= TSF_XFERABLE;
493 if (pri < PRI_MIN_BATCH) {
494 ts->ts_runq = &tdq->tdq_realtime;
495 } else if (pri <= PRI_MAX_BATCH) {
496 ts->ts_runq = &tdq->tdq_timeshare;
497 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
498 ("Invalid priority %d on timeshare runq", pri));
500 * This queue contains only priorities between MIN and MAX
501 * batch. Use the whole queue to represent these values.
503 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
504 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
505 pri = (pri + tdq->tdq_idx) % RQ_NQS;
507 * This effectively shortens the queue by one so we
508 * can have a one slot difference between idx and
509 * ridx while we wait for threads to drain.
511 if (tdq->tdq_ridx != tdq->tdq_idx &&
512 pri == tdq->tdq_ridx)
513 pri = (unsigned char)(pri - 1) % RQ_NQS;
516 runq_add_pri(ts->ts_runq, td, pri, flags);
519 ts->ts_runq = &tdq->tdq_idle;
520 runq_add(ts->ts_runq, td, flags);
524 * Remove a thread from a run-queue. This typically happens when a thread
525 * is selected to run. Running threads are not on the queue and the
526 * transferable count does not reflect them.
529 tdq_runq_rem(struct tdq *tdq, struct thread *td)
533 ts = td_get_sched(td);
534 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
535 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
536 KASSERT(ts->ts_runq != NULL,
537 ("tdq_runq_remove: thread %p null ts_runq", td));
538 if (ts->ts_flags & TSF_XFERABLE) {
539 tdq->tdq_transferable--;
540 ts->ts_flags &= ~TSF_XFERABLE;
542 if (ts->ts_runq == &tdq->tdq_timeshare) {
543 if (tdq->tdq_idx != tdq->tdq_ridx)
544 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
546 runq_remove_idx(ts->ts_runq, td, NULL);
548 runq_remove(ts->ts_runq, td);
552 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
553 * for this thread to the referenced thread queue.
556 tdq_load_add(struct tdq *tdq, struct thread *td)
559 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
560 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
563 if ((td->td_flags & TDF_NOLOAD) == 0)
565 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
566 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
570 * Remove the load from a thread that is transitioning to a sleep state or
574 tdq_load_rem(struct tdq *tdq, struct thread *td)
577 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
578 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
579 KASSERT(tdq->tdq_load != 0,
580 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
583 if ((td->td_flags & TDF_NOLOAD) == 0)
585 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
586 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
590 * Bound timeshare latency by decreasing slice size as load increases. We
591 * consider the maximum latency as the sum of the threads waiting to run
592 * aside from curthread and target no more than sched_slice latency but
593 * no less than sched_slice_min runtime.
596 tdq_slice(struct tdq *tdq)
601 * It is safe to use sys_load here because this is called from
602 * contexts where timeshare threads are running and so there
603 * cannot be higher priority load in the system.
605 load = tdq->tdq_sysload - 1;
606 if (load >= SCHED_SLICE_MIN_DIVISOR)
607 return (sched_slice_min);
609 return (sched_slice);
610 return (sched_slice / load);
614 * Set lowpri to its exact value by searching the run-queue and
615 * evaluating curthread. curthread may be passed as an optimization.
618 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
622 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
624 ctd = tdq->tdq_curthread;
625 td = tdq_choose(tdq);
626 if (td == NULL || td->td_priority > ctd->td_priority)
627 tdq->tdq_lowpri = ctd->td_priority;
629 tdq->tdq_lowpri = td->td_priority;
634 * We need some randomness. Implement a classic Linear Congruential
635 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
636 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
637 * of the random state (in the low bits of our answer) to keep
638 * the maximum randomness.
645 rndptr = DPCPU_PTR(randomval);
646 *rndptr = *rndptr * 69069 + 5;
648 return (*rndptr >> 16);
652 cpuset_t *cs_mask; /* The mask of allowed CPUs to choose from. */
653 int cs_prefer; /* Prefer this CPU and groups including it. */
654 int cs_running; /* The thread is now running at cs_prefer. */
655 int cs_pri; /* Min priority for low. */
656 int cs_load; /* Max load for low, min load for high. */
657 int cs_trans; /* Min transferable load for high. */
660 struct cpu_search_res {
661 int csr_cpu; /* The best CPU found. */
662 int csr_load; /* The load of cs_cpu. */
666 * Search the tree of cpu_groups for the lowest or highest loaded CPU.
667 * These routines actually compare the load on all paths through the tree
668 * and find the least loaded cpu on the least loaded path, which may differ
669 * from the least loaded cpu in the system. This balances work among caches
673 cpu_search_lowest(const struct cpu_group *cg, const struct cpu_search *s,
674 struct cpu_search_res *r)
676 struct cpu_search_res lr;
678 int c, bload, l, load, p, total;
684 /* Loop through children CPU groups if there are any. */
685 if (cg->cg_children > 0) {
686 for (c = cg->cg_children - 1; c >= 0; c--) {
687 load = cpu_search_lowest(&cg->cg_child[c], s, &lr);
691 * When balancing do not prefer SMT groups with load >1.
692 * It allows round-robin between SMT groups with equal
693 * load within parent group for more fair scheduling.
695 if (__predict_false(s->cs_running) &&
696 (cg->cg_child[c].cg_flags & CG_FLAG_THREAD) &&
697 load >= 128 && (load & 128) != 0)
700 if (lr.csr_cpu >= 0 && (load < bload ||
701 (load == bload && lr.csr_load < r->csr_load))) {
703 r->csr_cpu = lr.csr_cpu;
704 r->csr_load = lr.csr_load;
710 /* Loop through children CPUs otherwise. */
711 for (c = cg->cg_last; c >= cg->cg_first; c--) {
712 if (!CPU_ISSET(c, &cg->cg_mask))
716 if (c == s->cs_prefer) {
717 if (__predict_false(s->cs_running))
726 * Check this CPU is acceptable.
727 * If the threads is already on the CPU, don't look on the TDQ
728 * priority, since it can be the priority of the thread itself.
730 if (l > s->cs_load ||
731 (atomic_load_char(&tdq->tdq_lowpri) <= s->cs_pri &&
732 (!s->cs_running || c != s->cs_prefer)) ||
733 !CPU_ISSET(c, s->cs_mask))
737 * When balancing do not prefer CPUs with load > 1.
738 * It allows round-robin between CPUs with equal load
739 * within the CPU group for more fair scheduling.
741 if (__predict_false(s->cs_running) && l > 0)
744 load -= sched_random() % 128;
745 if (bload > load - p) {
755 cpu_search_highest(const struct cpu_group *cg, const struct cpu_search *s,
756 struct cpu_search_res *r)
758 struct cpu_search_res lr;
760 int c, bload, l, load, total;
766 /* Loop through children CPU groups if there are any. */
767 if (cg->cg_children > 0) {
768 for (c = cg->cg_children - 1; c >= 0; c--) {
769 load = cpu_search_highest(&cg->cg_child[c], s, &lr);
771 if (lr.csr_cpu >= 0 && (load > bload ||
772 (load == bload && lr.csr_load > r->csr_load))) {
774 r->csr_cpu = lr.csr_cpu;
775 r->csr_load = lr.csr_load;
781 /* Loop through children CPUs otherwise. */
782 for (c = cg->cg_last; c >= cg->cg_first; c--) {
783 if (!CPU_ISSET(c, &cg->cg_mask))
791 * Check this CPU is acceptable.
793 if (l < s->cs_load || TDQ_TRANSFERABLE(tdq) < s->cs_trans ||
794 !CPU_ISSET(c, s->cs_mask))
797 load -= sched_random() % 256;
808 * Find the cpu with the least load via the least loaded path that has a
809 * lowpri greater than pri pri. A pri of -1 indicates any priority is
813 sched_lowest(const struct cpu_group *cg, cpuset_t *mask, int pri, int maxload,
814 int prefer, int running)
817 struct cpu_search_res r;
819 s.cs_prefer = prefer;
820 s.cs_running = running;
824 cpu_search_lowest(cg, &s, &r);
829 * Find the cpu with the highest load via the highest loaded path.
832 sched_highest(const struct cpu_group *cg, cpuset_t *mask, int minload,
836 struct cpu_search_res r;
840 s.cs_trans = mintrans;
841 cpu_search_highest(cg, &s, &r);
846 sched_balance_group(struct cpu_group *cg)
850 cpuset_t hmask, lmask;
851 int high, low, anylow;
855 high = sched_highest(cg, &hmask, 1, 0);
856 /* Stop if there is no more CPU with transferrable threads. */
859 CPU_CLR(high, &hmask);
860 CPU_COPY(&hmask, &lmask);
861 /* Stop if there is no more CPU left for low. */
862 if (CPU_EMPTY(&lmask))
865 if (TDQ_LOAD(tdq) == 1) {
867 * There is only one running thread. We can't move
868 * it from here, so tell it to pick new CPU by itself.
871 td = tdq->tdq_curthread;
872 if (td->td_lock == TDQ_LOCKPTR(tdq) &&
873 (td->td_flags & TDF_IDLETD) == 0 &&
874 THREAD_CAN_MIGRATE(td)) {
875 td->td_flags |= TDF_PICKCPU;
876 ast_sched_locked(td, TDA_SCHED);
878 ipi_cpu(high, IPI_AST);
885 if (TDQ_TRANSFERABLE(tdq) == 0)
887 low = sched_lowest(cg, &lmask, -1, TDQ_LOAD(tdq) - 1, high, 1);
888 /* Stop if we looked well and found no less loaded CPU. */
889 if (anylow && low == -1)
891 /* Go to next high if we found no less loaded CPU. */
894 /* Transfer thread from high to low. */
895 if (sched_balance_pair(tdq, TDQ_CPU(low))) {
896 /* CPU that got thread can no longer be a donor. */
897 CPU_CLR(low, &hmask);
900 * If failed, then there is no threads on high
901 * that can run on this low. Drop low from low
902 * mask and look for different one.
904 CPU_CLR(low, &lmask);
916 balance_ticks = max(balance_interval / 2, 1) +
917 (sched_random() % balance_interval);
920 sched_balance_group(cpu_top);
925 * Lock two thread queues using their address to maintain lock order.
928 tdq_lock_pair(struct tdq *one, struct tdq *two)
932 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
935 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
940 * Unlock two thread queues. Order is not important here.
943 tdq_unlock_pair(struct tdq *one, struct tdq *two)
950 * Transfer load between two imbalanced thread queues. Returns true if a thread
951 * was moved between the queues, and false otherwise.
954 sched_balance_pair(struct tdq *high, struct tdq *low)
960 tdq_lock_pair(high, low);
963 * Transfer a thread from high to low.
965 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load) {
966 lowpri = tdq_move(high, low);
969 * In case the target isn't the current CPU notify it of
970 * the new load, possibly sending an IPI to force it to
971 * reschedule. Otherwise maybe schedule a preemption.
974 if (cpu != PCPU_GET(cpuid))
975 tdq_notify(low, lowpri);
977 sched_setpreempt(low->tdq_lowpri);
981 tdq_unlock_pair(high, low);
986 * Move a thread from one thread queue to another. Returns -1 if the source
987 * queue was empty, else returns the maximum priority of all threads in
988 * the destination queue prior to the addition of the new thread. In the latter
989 * case, this priority can be used to determine whether an IPI needs to be
993 tdq_move(struct tdq *from, struct tdq *to)
998 TDQ_LOCK_ASSERT(from, MA_OWNED);
999 TDQ_LOCK_ASSERT(to, MA_OWNED);
1002 td = tdq_steal(from, cpu);
1007 * Although the run queue is locked the thread may be
1008 * blocked. We can not set the lock until it is unblocked.
1010 thread_lock_block_wait(td);
1012 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from));
1013 td->td_lock = TDQ_LOCKPTR(to);
1014 td_get_sched(td)->ts_cpu = cpu;
1015 return (tdq_add(to, td, SRQ_YIELDING));
1019 * This tdq has idled. Try to steal a thread from another cpu and switch
1023 tdq_idled(struct tdq *tdq)
1025 struct cpu_group *cg, *parent;
1028 int cpu, switchcnt, goup;
1030 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
1033 CPU_CLR(PCPU_GET(cpuid), &mask);
1035 switchcnt = TDQ_SWITCHCNT(tdq);
1036 for (cg = tdq->tdq_cg, goup = 0; ; ) {
1037 cpu = sched_highest(cg, &mask, steal_thresh, 1);
1039 * We were assigned a thread but not preempted. Returning
1040 * 0 here will cause our caller to switch to it.
1046 * We found no CPU to steal from in this group. Escalate to
1047 * the parent and repeat. But if parent has only two children
1048 * groups we can avoid searching this group again by searching
1049 * the other one specifically and then escalating two levels.
1056 parent = cg->cg_parent;
1059 if (parent->cg_children == 2) {
1060 if (cg == &parent->cg_child[0])
1061 cg = &parent->cg_child[1];
1063 cg = &parent->cg_child[0];
1069 steal = TDQ_CPU(cpu);
1071 * The data returned by sched_highest() is stale and
1072 * the chosen CPU no longer has an eligible thread.
1074 * Testing this ahead of tdq_lock_pair() only catches
1075 * this situation about 20% of the time on an 8 core
1076 * 16 thread Ryzen 7, but it still helps performance.
1078 if (TDQ_LOAD(steal) < steal_thresh ||
1079 TDQ_TRANSFERABLE(steal) == 0)
1082 * Try to lock both queues. If we are assigned a thread while
1083 * waited for the lock, switch to it now instead of stealing.
1084 * If we can't get the lock, then somebody likely got there
1085 * first so continue searching.
1088 if (tdq->tdq_load > 0) {
1089 mi_switch(SW_VOL | SWT_IDLE);
1092 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) {
1094 CPU_CLR(cpu, &mask);
1098 * The data returned by sched_highest() is stale and
1099 * the chosen CPU no longer has an eligible thread, or
1100 * we were preempted and the CPU loading info may be out
1101 * of date. The latter is rare. In either case restart
1104 if (TDQ_LOAD(steal) < steal_thresh ||
1105 TDQ_TRANSFERABLE(steal) == 0 ||
1106 switchcnt != TDQ_SWITCHCNT(tdq)) {
1107 tdq_unlock_pair(tdq, steal);
1111 * Steal the thread and switch to it.
1113 if (tdq_move(steal, tdq) != -1)
1116 * We failed to acquire a thread even though it looked
1117 * like one was available. This could be due to affinity
1118 * restrictions or for other reasons. Loop again after
1119 * removing this CPU from the set. The restart logic
1120 * above does not restore this CPU to the set due to the
1121 * likelyhood of failing here again.
1123 CPU_CLR(cpu, &mask);
1124 tdq_unlock_pair(tdq, steal);
1127 mi_switch(SW_VOL | SWT_IDLE);
1132 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1134 * "lowpri" is the minimum scheduling priority among all threads on
1135 * the queue prior to the addition of the new thread.
1138 tdq_notify(struct tdq *tdq, int lowpri)
1142 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1143 KASSERT(tdq->tdq_lowpri <= lowpri,
1144 ("tdq_notify: lowpri %d > tdq_lowpri %d", lowpri, tdq->tdq_lowpri));
1146 if (tdq->tdq_owepreempt)
1150 * Check to see if the newly added thread should preempt the one
1151 * currently running.
1153 if (!sched_shouldpreempt(tdq->tdq_lowpri, lowpri, 1))
1157 * Make sure that our caller's earlier update to tdq_load is
1158 * globally visible before we read tdq_cpu_idle. Idle thread
1159 * accesses both of them without locks, and the order is important.
1161 atomic_thread_fence_seq_cst();
1164 * Try to figure out if we can signal the idle thread instead of sending
1165 * an IPI. This check is racy; at worst, we will deliever an IPI
1169 if (TD_IS_IDLETHREAD(tdq->tdq_curthread) &&
1170 (atomic_load_int(&tdq->tdq_cpu_idle) == 0 || cpu_idle_wakeup(cpu)))
1174 * The run queues have been updated, so any switch on the remote CPU
1175 * will satisfy the preemption request.
1177 tdq->tdq_owepreempt = 1;
1178 ipi_cpu(cpu, IPI_PREEMPT);
1182 * Steals load from a timeshare queue. Honors the rotating queue head
1185 static struct thread *
1186 runq_steal_from(struct runq *rq, int cpu, u_char start)
1190 struct thread *td, *first;
1194 rqb = &rq->rq_status;
1195 bit = start & (RQB_BPW -1);
1198 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1199 if (rqb->rqb_bits[i] == 0)
1202 bit = RQB_FFS(rqb->rqb_bits[i]);
1203 for (; bit < RQB_BPW; bit++) {
1204 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1206 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1207 TAILQ_FOREACH(td, rqh, td_runq) {
1209 if (THREAD_CAN_MIGRATE(td) &&
1210 THREAD_CAN_SCHED(td, cpu))
1222 if (first && THREAD_CAN_MIGRATE(first) &&
1223 THREAD_CAN_SCHED(first, cpu))
1229 * Steals load from a standard linear queue.
1231 static struct thread *
1232 runq_steal(struct runq *rq, int cpu)
1240 rqb = &rq->rq_status;
1241 for (word = 0; word < RQB_LEN; word++) {
1242 if (rqb->rqb_bits[word] == 0)
1244 for (bit = 0; bit < RQB_BPW; bit++) {
1245 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1247 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1248 TAILQ_FOREACH(td, rqh, td_runq)
1249 if (THREAD_CAN_MIGRATE(td) &&
1250 THREAD_CAN_SCHED(td, cpu))
1258 * Attempt to steal a thread in priority order from a thread queue.
1260 static struct thread *
1261 tdq_steal(struct tdq *tdq, int cpu)
1265 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1266 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1268 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1269 cpu, tdq->tdq_ridx)) != NULL)
1271 return (runq_steal(&tdq->tdq_idle, cpu));
1275 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1276 * current lock and returns with the assigned queue locked.
1278 static inline struct tdq *
1279 sched_setcpu(struct thread *td, int cpu, int flags)
1285 THREAD_LOCK_ASSERT(td, MA_OWNED);
1287 td_get_sched(td)->ts_cpu = cpu;
1289 * If the lock matches just return the queue.
1291 if (td->td_lock == TDQ_LOCKPTR(tdq)) {
1292 KASSERT((flags & SRQ_HOLD) == 0,
1293 ("sched_setcpu: Invalid lock for SRQ_HOLD"));
1298 * The hard case, migration, we need to block the thread first to
1299 * prevent order reversals with other cpus locks.
1302 mtx = thread_lock_block(td);
1303 if ((flags & SRQ_HOLD) == 0)
1304 mtx_unlock_spin(mtx);
1306 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1311 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1312 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1313 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1314 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1315 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1316 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1319 sched_pickcpu(struct thread *td, int flags)
1321 struct cpu_group *cg, *ccg;
1322 struct td_sched *ts;
1325 int cpu, pri, r, self, intr;
1327 self = PCPU_GET(cpuid);
1328 ts = td_get_sched(td);
1329 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1330 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1331 if (smp_started == 0)
1334 * Don't migrate a running thread from sched_switch().
1336 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1337 return (ts->ts_cpu);
1339 * Prefer to run interrupt threads on the processors that generate
1342 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1343 curthread->td_intr_nesting_level) {
1345 if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1346 SCHED_STAT_INC(pickcpu_idle_affinity);
1355 tdq = TDQ_CPU(ts->ts_cpu);
1359 * If the thread can run on the last cpu and the affinity has not
1360 * expired and it is idle, run it there.
1362 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1363 atomic_load_char(&tdq->tdq_lowpri) >= PRI_MIN_IDLE &&
1364 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1365 if (cg->cg_flags & CG_FLAG_THREAD) {
1366 /* Check all SMT threads for being idle. */
1367 for (cpu = cg->cg_first; cpu <= cg->cg_last; cpu++) {
1369 atomic_load_char(&TDQ_CPU(cpu)->tdq_lowpri);
1370 if (CPU_ISSET(cpu, &cg->cg_mask) &&
1374 if (cpu > cg->cg_last) {
1375 SCHED_STAT_INC(pickcpu_idle_affinity);
1376 return (ts->ts_cpu);
1379 SCHED_STAT_INC(pickcpu_idle_affinity);
1380 return (ts->ts_cpu);
1385 * Search for the last level cache CPU group in the tree.
1386 * Skip SMT, identical groups and caches with expired affinity.
1387 * Interrupt threads affinity is explicit and never expires.
1389 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1390 if (cg->cg_flags & CG_FLAG_THREAD)
1392 if (cg->cg_children == 1 || cg->cg_count == 1)
1394 if (cg->cg_level == CG_SHARE_NONE ||
1395 (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1399 /* Found LLC shared by all CPUs, so do a global search. */
1403 mask = &td->td_cpuset->cs_mask;
1404 pri = td->td_priority;
1405 r = TD_IS_RUNNING(td);
1407 * Try hard to keep interrupts within found LLC. Search the LLC for
1408 * the least loaded CPU we can run now. For NUMA systems it should
1409 * be within target domain, and it also reduces scheduling overhead.
1411 if (ccg != NULL && intr) {
1412 cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu, r);
1414 SCHED_STAT_INC(pickcpu_intrbind);
1416 /* Search the LLC for the least loaded idle CPU we can run now. */
1418 cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1419 INT_MAX, ts->ts_cpu, r);
1421 SCHED_STAT_INC(pickcpu_affinity);
1423 /* Search globally for the least loaded CPU we can run now. */
1425 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu, r);
1427 SCHED_STAT_INC(pickcpu_lowest);
1429 /* Search globally for the least loaded CPU. */
1431 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu, r);
1433 SCHED_STAT_INC(pickcpu_lowest);
1435 KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1436 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1438 * Compare the lowest loaded cpu to current cpu.
1441 if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1442 atomic_load_char(&tdq->tdq_lowpri) < PRI_MIN_IDLE &&
1443 TDQ_LOAD(TDQ_SELF()) <= TDQ_LOAD(tdq) + 1) {
1444 SCHED_STAT_INC(pickcpu_local);
1447 if (cpu != ts->ts_cpu)
1448 SCHED_STAT_INC(pickcpu_migration);
1454 * Pick the highest priority task we have and return it.
1456 static struct thread *
1457 tdq_choose(struct tdq *tdq)
1461 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1462 td = runq_choose(&tdq->tdq_realtime);
1465 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1467 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1468 ("tdq_choose: Invalid priority on timeshare queue %d",
1472 td = runq_choose(&tdq->tdq_idle);
1474 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1475 ("tdq_choose: Invalid priority on idle queue %d",
1484 * Initialize a thread queue.
1487 tdq_setup(struct tdq *tdq, int id)
1491 printf("ULE: setup cpu %d\n", id);
1492 runq_init(&tdq->tdq_realtime);
1493 runq_init(&tdq->tdq_timeshare);
1494 runq_init(&tdq->tdq_idle);
1496 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1497 "sched lock %d", (int)TDQ_ID(tdq));
1498 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN);
1500 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1501 "CPU %d load", (int)TDQ_ID(tdq));
1507 sched_setup_smp(void)
1512 cpu_top = smp_topo();
1514 tdq = DPCPU_ID_PTR(i, tdq);
1516 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1517 if (tdq->tdq_cg == NULL)
1518 panic("Can't find cpu group for %d\n", i);
1519 DPCPU_ID_SET(i, randomval, i * 69069 + 5);
1521 PCPU_SET(sched, DPCPU_PTR(tdq));
1522 balance_tdq = TDQ_SELF();
1527 * Setup the thread queues and initialize the topology based on MD
1531 sched_setup(void *dummy)
1538 tdq_setup(TDQ_SELF(), 0);
1542 /* Add thread0's load since it's running. */
1544 thread0.td_lock = TDQ_LOCKPTR(tdq);
1545 tdq_load_add(tdq, &thread0);
1546 tdq->tdq_curthread = &thread0;
1547 tdq->tdq_lowpri = thread0.td_priority;
1552 * This routine determines time constants after stathz and hz are setup.
1556 sched_initticks(void *dummy)
1560 realstathz = stathz ? stathz : hz;
1561 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1562 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1563 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1567 * tickincr is shifted out by 10 to avoid rounding errors due to
1568 * hz not being evenly divisible by stathz on all platforms.
1570 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1572 * This does not work for values of stathz that are more than
1573 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1580 * Set the default balance interval now that we know
1581 * what realstathz is.
1583 balance_interval = realstathz;
1584 balance_ticks = balance_interval;
1585 affinity = SCHED_AFFINITY_DEFAULT;
1587 if (sched_idlespinthresh < 0)
1588 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1592 * This is the core of the interactivity algorithm. Determines a score based
1593 * on past behavior. It is the ratio of sleep time to run time scaled to
1594 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1595 * differs from the cpu usage because it does not account for time spent
1596 * waiting on a run-queue. Would be prettier if we had floating point.
1598 * When a thread's sleep time is greater than its run time the
1602 * interactivity score = ---------------------
1603 * sleep time / run time
1606 * When a thread's run time is greater than its sleep time the
1610 * interactivity score = 2 * scaling factor - ---------------------
1611 * run time / sleep time
1614 sched_interact_score(struct thread *td)
1616 struct td_sched *ts;
1619 ts = td_get_sched(td);
1621 * The score is only needed if this is likely to be an interactive
1622 * task. Don't go through the expense of computing it if there's
1625 if (sched_interact <= SCHED_INTERACT_HALF &&
1626 ts->ts_runtime >= ts->ts_slptime)
1627 return (SCHED_INTERACT_HALF);
1629 if (ts->ts_runtime > ts->ts_slptime) {
1630 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1631 return (SCHED_INTERACT_HALF +
1632 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1634 if (ts->ts_slptime > ts->ts_runtime) {
1635 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1636 return (ts->ts_runtime / div);
1638 /* runtime == slptime */
1640 return (SCHED_INTERACT_HALF);
1643 * This can happen if slptime and runtime are 0.
1650 * Scale the scheduling priority according to the "interactivity" of this
1654 sched_priority(struct thread *td)
1658 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1661 * If the score is interactive we place the thread in the realtime
1662 * queue with a priority that is less than kernel and interrupt
1663 * priorities. These threads are not subject to nice restrictions.
1665 * Scores greater than this are placed on the normal timeshare queue
1666 * where the priority is partially decided by the most recent cpu
1667 * utilization and the rest is decided by nice value.
1669 * The nice value of the process has a linear effect on the calculated
1670 * score. Negative nice values make it easier for a thread to be
1671 * considered interactive.
1673 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1674 if (score < sched_interact) {
1675 pri = PRI_MIN_INTERACT;
1676 pri += (PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) * score /
1678 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1679 ("sched_priority: invalid interactive priority %u score %u",
1682 pri = SCHED_PRI_MIN;
1683 if (td_get_sched(td)->ts_ticks)
1684 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1685 SCHED_PRI_RANGE - 1);
1686 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1687 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1688 ("sched_priority: invalid priority %u: nice %d, "
1689 "ticks %d ftick %d ltick %d tick pri %d",
1690 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1691 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1692 SCHED_PRI_TICKS(td_get_sched(td))));
1694 sched_user_prio(td, pri);
1700 * This routine enforces a maximum limit on the amount of scheduling history
1701 * kept. It is called after either the slptime or runtime is adjusted. This
1702 * function is ugly due to integer math.
1705 sched_interact_update(struct thread *td)
1707 struct td_sched *ts;
1710 ts = td_get_sched(td);
1711 sum = ts->ts_runtime + ts->ts_slptime;
1712 if (sum < SCHED_SLP_RUN_MAX)
1715 * This only happens from two places:
1716 * 1) We have added an unusual amount of run time from fork_exit.
1717 * 2) We have added an unusual amount of sleep time from sched_sleep().
1719 if (sum > SCHED_SLP_RUN_MAX * 2) {
1720 if (ts->ts_runtime > ts->ts_slptime) {
1721 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1724 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1730 * If we have exceeded by more than 1/5th then the algorithm below
1731 * will not bring us back into range. Dividing by two here forces
1732 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1734 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1735 ts->ts_runtime /= 2;
1736 ts->ts_slptime /= 2;
1739 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1740 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1744 * Scale back the interactivity history when a child thread is created. The
1745 * history is inherited from the parent but the thread may behave totally
1746 * differently. For example, a shell spawning a compiler process. We want
1747 * to learn that the compiler is behaving badly very quickly.
1750 sched_interact_fork(struct thread *td)
1752 struct td_sched *ts;
1756 ts = td_get_sched(td);
1757 sum = ts->ts_runtime + ts->ts_slptime;
1758 if (sum > SCHED_SLP_RUN_FORK) {
1759 ratio = sum / SCHED_SLP_RUN_FORK;
1760 ts->ts_runtime /= ratio;
1761 ts->ts_slptime /= ratio;
1766 * Called from proc0_init() to setup the scheduler fields.
1771 struct td_sched *ts0;
1774 * Set up the scheduler specific parts of thread0.
1776 ts0 = td_get_sched(&thread0);
1777 ts0->ts_ltick = ticks;
1778 ts0->ts_ftick = ticks;
1780 ts0->ts_cpu = curcpu; /* set valid CPU number */
1784 * schedinit_ap() is needed prior to calling sched_throw(NULL) to ensure that
1785 * the pcpu requirements are met for any calls in the period between curthread
1786 * initialization and sched_throw(). One can safely add threads to the queue
1787 * before sched_throw(), for instance, as long as the thread lock is setup
1790 * TDQ_SELF() relies on the below sched pcpu setting; it may be used only
1791 * after schedinit_ap().
1798 PCPU_SET(sched, DPCPU_PTR(tdq));
1800 PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1804 * This is only somewhat accurate since given many processes of the same
1805 * priority they will switch when their slices run out, which will be
1806 * at most sched_slice stathz ticks.
1809 sched_rr_interval(void)
1812 /* Convert sched_slice from stathz to hz. */
1813 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1817 * Update the percent cpu tracking information when it is requested or
1818 * the total history exceeds the maximum. We keep a sliding history of
1819 * tick counts that slowly decays. This is less precise than the 4BSD
1820 * mechanism since it happens with less regular and frequent events.
1823 sched_pctcpu_update(struct td_sched *ts, int run)
1828 * The signed difference may be negative if the thread hasn't run for
1829 * over half of the ticks rollover period.
1831 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1833 ts->ts_ftick = t - SCHED_TICK_TARG;
1834 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1835 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1836 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1837 ts->ts_ftick = t - SCHED_TICK_TARG;
1840 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1845 * Adjust the priority of a thread. Move it to the appropriate run-queue
1846 * if necessary. This is the back-end for several priority related
1850 sched_thread_priority(struct thread *td, u_char prio)
1855 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1856 "prio:%d", td->td_priority, "new prio:%d", prio,
1857 KTR_ATTR_LINKED, sched_tdname(curthread));
1858 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1859 if (td != curthread && prio < td->td_priority) {
1860 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1861 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1862 prio, KTR_ATTR_LINKED, sched_tdname(td));
1863 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1866 THREAD_LOCK_ASSERT(td, MA_OWNED);
1867 if (td->td_priority == prio)
1870 * If the priority has been elevated due to priority
1871 * propagation, we may have to move ourselves to a new
1872 * queue. This could be optimized to not re-add in some
1875 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1877 td->td_priority = prio;
1878 sched_add(td, SRQ_BORROWING | SRQ_HOLDTD);
1882 * If the thread is currently running we may have to adjust the lowpri
1883 * information so other cpus are aware of our current priority.
1885 if (TD_IS_RUNNING(td)) {
1886 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
1887 oldpri = td->td_priority;
1888 td->td_priority = prio;
1889 if (prio < tdq->tdq_lowpri)
1890 tdq->tdq_lowpri = prio;
1891 else if (tdq->tdq_lowpri == oldpri)
1892 tdq_setlowpri(tdq, td);
1895 td->td_priority = prio;
1899 * Update a thread's priority when it is lent another thread's
1903 sched_lend_prio(struct thread *td, u_char prio)
1906 td->td_flags |= TDF_BORROWING;
1907 sched_thread_priority(td, prio);
1911 * Restore a thread's priority when priority propagation is
1912 * over. The prio argument is the minimum priority the thread
1913 * needs to have to satisfy other possible priority lending
1914 * requests. If the thread's regular priority is less
1915 * important than prio, the thread will keep a priority boost
1919 sched_unlend_prio(struct thread *td, u_char prio)
1923 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1924 td->td_base_pri <= PRI_MAX_TIMESHARE)
1925 base_pri = td->td_user_pri;
1927 base_pri = td->td_base_pri;
1928 if (prio >= base_pri) {
1929 td->td_flags &= ~TDF_BORROWING;
1930 sched_thread_priority(td, base_pri);
1932 sched_lend_prio(td, prio);
1936 * Standard entry for setting the priority to an absolute value.
1939 sched_prio(struct thread *td, u_char prio)
1943 /* First, update the base priority. */
1944 td->td_base_pri = prio;
1947 * If the thread is borrowing another thread's priority, don't
1948 * ever lower the priority.
1950 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1953 /* Change the real priority. */
1954 oldprio = td->td_priority;
1955 sched_thread_priority(td, prio);
1958 * If the thread is on a turnstile, then let the turnstile update
1961 if (TD_ON_LOCK(td) && oldprio != prio)
1962 turnstile_adjust(td, oldprio);
1966 * Set the base interrupt thread priority.
1969 sched_ithread_prio(struct thread *td, u_char prio)
1971 THREAD_LOCK_ASSERT(td, MA_OWNED);
1972 MPASS(td->td_pri_class == PRI_ITHD);
1973 td->td_base_ithread_pri = prio;
1974 sched_prio(td, prio);
1978 * Set the base user priority, does not effect current running priority.
1981 sched_user_prio(struct thread *td, u_char prio)
1984 td->td_base_user_pri = prio;
1985 if (td->td_lend_user_pri <= prio)
1987 td->td_user_pri = prio;
1991 sched_lend_user_prio(struct thread *td, u_char prio)
1994 THREAD_LOCK_ASSERT(td, MA_OWNED);
1995 td->td_lend_user_pri = prio;
1996 td->td_user_pri = min(prio, td->td_base_user_pri);
1997 if (td->td_priority > td->td_user_pri)
1998 sched_prio(td, td->td_user_pri);
1999 else if (td->td_priority != td->td_user_pri)
2000 ast_sched_locked(td, TDA_SCHED);
2004 * Like the above but first check if there is anything to do.
2007 sched_lend_user_prio_cond(struct thread *td, u_char prio)
2010 if (td->td_lend_user_pri == prio)
2014 sched_lend_user_prio(td, prio);
2020 * This tdq is about to idle. Try to steal a thread from another CPU before
2021 * choosing the idle thread.
2024 tdq_trysteal(struct tdq *tdq)
2026 struct cpu_group *cg, *parent;
2031 if (smp_started == 0 || steal_idle == 0 || trysteal_limit == 0 ||
2032 tdq->tdq_cg == NULL)
2035 CPU_CLR(PCPU_GET(cpuid), &mask);
2036 /* We don't want to be preempted while we're iterating. */
2039 for (i = 1, cg = tdq->tdq_cg, goup = 0; ; ) {
2040 cpu = sched_highest(cg, &mask, steal_thresh, 1);
2042 * If a thread was added while interrupts were disabled don't
2045 if (TDQ_LOAD(tdq) > 0) {
2051 * We found no CPU to steal from in this group. Escalate to
2052 * the parent and repeat. But if parent has only two children
2053 * groups we can avoid searching this group again by searching
2054 * the other one specifically and then escalating two levels.
2061 if (++i > trysteal_limit) {
2065 parent = cg->cg_parent;
2066 if (parent == NULL) {
2070 if (parent->cg_children == 2) {
2071 if (cg == &parent->cg_child[0])
2072 cg = &parent->cg_child[1];
2074 cg = &parent->cg_child[0];
2080 steal = TDQ_CPU(cpu);
2082 * The data returned by sched_highest() is stale and
2083 * the chosen CPU no longer has an eligible thread.
2084 * At this point unconditionally exit the loop to bound
2085 * the time spent in the critcal section.
2087 if (TDQ_LOAD(steal) < steal_thresh ||
2088 TDQ_TRANSFERABLE(steal) == 0)
2091 * Try to lock both queues. If we are assigned a thread while
2092 * waited for the lock, switch to it now instead of stealing.
2093 * If we can't get the lock, then somebody likely got there
2097 if (tdq->tdq_load > 0)
2099 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0)
2102 * The data returned by sched_highest() is stale and
2103 * the chosen CPU no longer has an eligible thread.
2105 if (TDQ_LOAD(steal) < steal_thresh ||
2106 TDQ_TRANSFERABLE(steal) == 0) {
2111 * If we fail to acquire one due to affinity restrictions,
2112 * bail out and let the idle thread to a more complete search
2113 * outside of a critical section.
2115 if (tdq_move(steal, tdq) == -1) {
2127 * Handle migration from sched_switch(). This happens only for
2131 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2138 KASSERT(THREAD_CAN_MIGRATE(td) ||
2139 (td_get_sched(td)->ts_flags & TSF_BOUND) != 0,
2140 ("Thread %p shouldn't migrate", td));
2141 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2142 "thread %s queued on absent CPU %d.", td->td_name,
2143 td_get_sched(td)->ts_cpu));
2144 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2146 tdq_load_rem(tdq, td);
2148 * Do the lock dance required to avoid LOR. We have an
2149 * extra spinlock nesting from sched_switch() which will
2150 * prevent preemption while we're holding neither run-queue lock.
2154 lowpri = tdq_add(tdn, td, flags);
2155 tdq_notify(tdn, lowpri);
2159 return (TDQ_LOCKPTR(tdn));
2163 * thread_lock_unblock() that does not assume td_lock is blocked.
2166 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2168 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2173 * Switch threads. This function has to handle threads coming in while
2174 * blocked for some reason, running, or idle. It also must deal with
2175 * migrating a thread from one queue to another as running threads may
2176 * be assigned elsewhere via binding.
2179 sched_switch(struct thread *td, int flags)
2181 struct thread *newtd;
2183 struct td_sched *ts;
2186 int cpuid, preempted;
2191 THREAD_LOCK_ASSERT(td, MA_OWNED);
2193 cpuid = PCPU_GET(cpuid);
2195 ts = td_get_sched(td);
2196 sched_pctcpu_update(ts, 1);
2198 pickcpu = (td->td_flags & TDF_PICKCPU) != 0;
2200 ts->ts_rltick = ticks - affinity * MAX_CACHE_LEVELS;
2202 ts->ts_rltick = ticks;
2204 td->td_lastcpu = td->td_oncpu;
2205 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2206 (flags & SW_PREEMPT) != 0;
2207 td->td_flags &= ~(TDF_PICKCPU | TDF_SLICEEND);
2208 ast_unsched_locked(td, TDA_SCHED);
2209 td->td_owepreempt = 0;
2210 atomic_store_char(&tdq->tdq_owepreempt, 0);
2211 if (!TD_IS_IDLETHREAD(td))
2212 TDQ_SWITCHCNT_INC(tdq);
2215 * Always block the thread lock so we can drop the tdq lock early.
2217 mtx = thread_lock_block(td);
2219 if (TD_IS_IDLETHREAD(td)) {
2220 MPASS(mtx == TDQ_LOCKPTR(tdq));
2222 } else if (TD_IS_RUNNING(td)) {
2223 MPASS(mtx == TDQ_LOCKPTR(tdq));
2224 srqflag = SRQ_OURSELF | SRQ_YIELDING |
2225 (preempted ? SRQ_PREEMPTED : 0);
2227 if (THREAD_CAN_MIGRATE(td) && (!THREAD_CAN_SCHED(td, ts->ts_cpu)
2229 ts->ts_cpu = sched_pickcpu(td, 0);
2231 if (ts->ts_cpu == cpuid)
2232 tdq_runq_add(tdq, td, srqflag);
2234 mtx = sched_switch_migrate(tdq, td, srqflag);
2236 /* This thread must be going to sleep. */
2237 if (mtx != TDQ_LOCKPTR(tdq)) {
2238 mtx_unlock_spin(mtx);
2241 tdq_load_rem(tdq, td);
2243 if (tdq->tdq_load == 0)
2248 #if (KTR_COMPILE & KTR_SCHED) != 0
2249 if (TD_IS_IDLETHREAD(td))
2250 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2251 "prio:%d", td->td_priority);
2253 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2254 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2255 "lockname:\"%s\"", td->td_lockname);
2259 * We enter here with the thread blocked and assigned to the
2260 * appropriate cpu run-queue or sleep-queue and with the current
2261 * thread-queue locked.
2263 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2264 MPASS(td == tdq->tdq_curthread);
2265 newtd = choosethread();
2266 sched_pctcpu_update(td_get_sched(newtd), 0);
2270 * Call the MD code to switch contexts if necessary.
2274 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2275 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2277 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2279 #ifdef KDTRACE_HOOKS
2281 * If DTrace has set the active vtime enum to anything
2282 * other than INACTIVE (0), then it should have set the
2285 if (dtrace_vtime_active)
2286 (*dtrace_vtime_switch_func)(newtd);
2288 td->td_oncpu = NOCPU;
2289 cpu_switch(td, newtd, mtx);
2290 cpuid = td->td_oncpu = PCPU_GET(cpuid);
2292 SDT_PROBE0(sched, , , on__cpu);
2294 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2295 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2298 thread_unblock_switch(td, mtx);
2299 SDT_PROBE0(sched, , , remain__cpu);
2301 KASSERT(curthread->td_md.md_spinlock_count == 1,
2302 ("invalid count %d", curthread->td_md.md_spinlock_count));
2304 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2305 "prio:%d", td->td_priority);
2309 * Adjust thread priorities as a result of a nice request.
2312 sched_nice(struct proc *p, int nice)
2316 PROC_LOCK_ASSERT(p, MA_OWNED);
2319 FOREACH_THREAD_IN_PROC(p, td) {
2322 sched_prio(td, td->td_base_user_pri);
2328 * Record the sleep time for the interactivity scorer.
2331 sched_sleep(struct thread *td, int prio)
2334 THREAD_LOCK_ASSERT(td, MA_OWNED);
2336 td->td_slptick = ticks;
2337 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2338 td->td_flags |= TDF_CANSWAP;
2339 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2341 if (static_boost == 1 && prio)
2342 sched_prio(td, prio);
2343 else if (static_boost && td->td_priority > static_boost)
2344 sched_prio(td, static_boost);
2348 * Schedule a thread to resume execution and record how long it voluntarily
2349 * slept. We also update the pctcpu, interactivity, and priority.
2351 * Requires the thread lock on entry, drops on exit.
2354 sched_wakeup(struct thread *td, int srqflags)
2356 struct td_sched *ts;
2359 THREAD_LOCK_ASSERT(td, MA_OWNED);
2360 ts = td_get_sched(td);
2361 td->td_flags &= ~TDF_CANSWAP;
2364 * If we slept for more than a tick update our interactivity and
2367 slptick = td->td_slptick;
2369 if (slptick && slptick != ticks) {
2370 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2371 sched_interact_update(td);
2372 sched_pctcpu_update(ts, 0);
2376 * When resuming an idle ithread, restore its base ithread
2379 if (PRI_BASE(td->td_pri_class) == PRI_ITHD &&
2380 td->td_priority != td->td_base_ithread_pri)
2381 sched_prio(td, td->td_base_ithread_pri);
2384 * Reset the slice value since we slept and advanced the round-robin.
2387 sched_add(td, SRQ_BORING | srqflags);
2391 * Penalize the parent for creating a new child and initialize the child's
2395 sched_fork(struct thread *td, struct thread *child)
2397 THREAD_LOCK_ASSERT(td, MA_OWNED);
2398 sched_pctcpu_update(td_get_sched(td), 1);
2399 sched_fork_thread(td, child);
2401 * Penalize the parent and child for forking.
2403 sched_interact_fork(child);
2404 sched_priority(child);
2405 td_get_sched(td)->ts_runtime += tickincr;
2406 sched_interact_update(td);
2411 * Fork a new thread, may be within the same process.
2414 sched_fork_thread(struct thread *td, struct thread *child)
2416 struct td_sched *ts;
2417 struct td_sched *ts2;
2421 THREAD_LOCK_ASSERT(td, MA_OWNED);
2425 ts = td_get_sched(td);
2426 ts2 = td_get_sched(child);
2427 child->td_oncpu = NOCPU;
2428 child->td_lastcpu = NOCPU;
2429 child->td_lock = TDQ_LOCKPTR(tdq);
2430 child->td_cpuset = cpuset_ref(td->td_cpuset);
2431 child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2432 ts2->ts_cpu = ts->ts_cpu;
2435 * Grab our parents cpu estimation information.
2437 ts2->ts_ticks = ts->ts_ticks;
2438 ts2->ts_ltick = ts->ts_ltick;
2439 ts2->ts_ftick = ts->ts_ftick;
2441 * Do not inherit any borrowed priority from the parent.
2443 child->td_priority = child->td_base_pri;
2445 * And update interactivity score.
2447 ts2->ts_slptime = ts->ts_slptime;
2448 ts2->ts_runtime = ts->ts_runtime;
2449 /* Attempt to quickly learn interactivity. */
2450 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2452 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2457 * Adjust the priority class of a thread.
2460 sched_class(struct thread *td, int class)
2463 THREAD_LOCK_ASSERT(td, MA_OWNED);
2464 if (td->td_pri_class == class)
2466 td->td_pri_class = class;
2470 * Return some of the child's priority and interactivity to the parent.
2473 sched_exit(struct proc *p, struct thread *child)
2477 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2478 "prio:%d", child->td_priority);
2479 PROC_LOCK_ASSERT(p, MA_OWNED);
2480 td = FIRST_THREAD_IN_PROC(p);
2481 sched_exit_thread(td, child);
2485 * Penalize another thread for the time spent on this one. This helps to
2486 * worsen the priority and interactivity of processes which schedule batch
2487 * jobs such as make. This has little effect on the make process itself but
2488 * causes new processes spawned by it to receive worse scores immediately.
2491 sched_exit_thread(struct thread *td, struct thread *child)
2494 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2495 "prio:%d", child->td_priority);
2497 * Give the child's runtime to the parent without returning the
2498 * sleep time as a penalty to the parent. This causes shells that
2499 * launch expensive things to mark their children as expensive.
2502 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2503 sched_interact_update(td);
2509 sched_preempt(struct thread *td)
2514 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2518 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2519 if (td->td_priority > tdq->tdq_lowpri) {
2520 if (td->td_critnest == 1) {
2521 flags = SW_INVOL | SW_PREEMPT;
2522 flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
2525 /* Switch dropped thread lock. */
2528 td->td_owepreempt = 1;
2530 tdq->tdq_owepreempt = 0;
2536 * Fix priorities on return to user-space. Priorities may be elevated due
2537 * to static priorities in msleep() or similar.
2540 sched_userret_slowpath(struct thread *td)
2544 td->td_priority = td->td_user_pri;
2545 td->td_base_pri = td->td_user_pri;
2546 tdq_setlowpri(TDQ_SELF(), td);
2550 SCHED_STAT_DEFINE(ithread_demotions, "Interrupt thread priority demotions");
2551 SCHED_STAT_DEFINE(ithread_preemptions,
2552 "Interrupt thread preemptions due to time-sharing");
2555 * Return time slice for a given thread. For ithreads this is
2556 * sched_slice. For other threads it is tdq_slice(tdq).
2559 td_slice(struct thread *td, struct tdq *tdq)
2561 if (PRI_BASE(td->td_pri_class) == PRI_ITHD)
2562 return (sched_slice);
2563 return (tdq_slice(tdq));
2567 * Handle a stathz tick. This is really only relevant for timeshare
2568 * and interrupt threads.
2571 sched_clock(struct thread *td, int cnt)
2574 struct td_sched *ts;
2576 THREAD_LOCK_ASSERT(td, MA_OWNED);
2580 * We run the long term load balancer infrequently on the first cpu.
2582 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 &&
2583 balance_ticks != 0) {
2584 balance_ticks -= cnt;
2585 if (balance_ticks <= 0)
2590 * Save the old switch count so we have a record of the last ticks
2591 * activity. Initialize the new switch count based on our load.
2592 * If there is some activity seed it to reflect that.
2594 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2595 tdq->tdq_switchcnt = tdq->tdq_load;
2598 * Advance the insert index once for each tick to ensure that all
2599 * threads get a chance to run.
2601 if (tdq->tdq_idx == tdq->tdq_ridx) {
2602 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2603 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2604 tdq->tdq_ridx = tdq->tdq_idx;
2606 ts = td_get_sched(td);
2607 sched_pctcpu_update(ts, 1);
2608 if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td))
2611 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2613 * We used a tick; charge it to the thread so
2614 * that we can compute our interactivity.
2616 td_get_sched(td)->ts_runtime += tickincr * cnt;
2617 sched_interact_update(td);
2622 * Force a context switch if the current thread has used up a full
2623 * time slice (default is 100ms).
2625 ts->ts_slice += cnt;
2626 if (ts->ts_slice >= td_slice(td, tdq)) {
2630 * If an ithread uses a full quantum, demote its
2631 * priority and preempt it.
2633 if (PRI_BASE(td->td_pri_class) == PRI_ITHD) {
2634 SCHED_STAT_INC(ithread_preemptions);
2635 td->td_owepreempt = 1;
2636 if (td->td_base_pri + RQ_PPQ < PRI_MAX_ITHD) {
2637 SCHED_STAT_INC(ithread_demotions);
2638 sched_prio(td, td->td_base_pri + RQ_PPQ);
2641 ast_sched_locked(td, TDA_SCHED);
2642 td->td_flags |= TDF_SLICEEND;
2648 sched_estcpu(struct thread *td __unused)
2655 * Return whether the current CPU has runnable tasks. Used for in-kernel
2656 * cooperative idle threads.
2659 sched_runnable(void)
2667 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2668 if (TDQ_LOAD(tdq) > 0)
2671 if (TDQ_LOAD(tdq) - 1 > 0)
2679 * Choose the highest priority thread to run. The thread is removed from
2680 * the run-queue while running however the load remains.
2689 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2690 td = tdq_choose(tdq);
2692 tdq_runq_rem(tdq, td);
2693 tdq->tdq_lowpri = td->td_priority;
2695 tdq->tdq_lowpri = PRI_MAX_IDLE;
2696 td = PCPU_GET(idlethread);
2698 tdq->tdq_curthread = td;
2703 * Set owepreempt if the currently running thread has lower priority than "pri".
2704 * Preemption never happens directly in ULE, we always request it once we exit a
2708 sched_setpreempt(int pri)
2714 THREAD_LOCK_ASSERT(ctd, MA_OWNED);
2716 cpri = ctd->td_priority;
2718 ast_sched_locked(ctd, TDA_SCHED);
2719 if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2721 if (!sched_shouldpreempt(pri, cpri, 0))
2723 ctd->td_owepreempt = 1;
2727 * Add a thread to a thread queue. Select the appropriate runq and add the
2728 * thread to it. This is the internal function called when the tdq is
2732 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2736 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2737 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
2738 KASSERT((td->td_inhibitors == 0),
2739 ("sched_add: trying to run inhibited thread"));
2740 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2741 ("sched_add: bad thread state"));
2742 KASSERT(td->td_flags & TDF_INMEM,
2743 ("sched_add: thread swapped out"));
2745 lowpri = tdq->tdq_lowpri;
2746 if (td->td_priority < lowpri)
2747 tdq->tdq_lowpri = td->td_priority;
2748 tdq_runq_add(tdq, td, flags);
2749 tdq_load_add(tdq, td);
2754 * Select the target thread queue and add a thread to it. Request
2755 * preemption or IPI a remote processor if required.
2757 * Requires the thread lock on entry, drops on exit.
2760 sched_add(struct thread *td, int flags)
2767 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2768 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2769 sched_tdname(curthread));
2770 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2771 KTR_ATTR_LINKED, sched_tdname(td));
2772 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2773 flags & SRQ_PREEMPTED);
2774 THREAD_LOCK_ASSERT(td, MA_OWNED);
2776 * Recalculate the priority before we select the target cpu or
2779 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2783 * Pick the destination cpu and if it isn't ours transfer to the
2786 cpu = sched_pickcpu(td, flags);
2787 tdq = sched_setcpu(td, cpu, flags);
2788 lowpri = tdq_add(tdq, td, flags);
2789 if (cpu != PCPU_GET(cpuid))
2790 tdq_notify(tdq, lowpri);
2791 else if (!(flags & SRQ_YIELDING))
2792 sched_setpreempt(td->td_priority);
2796 * Now that the thread is moving to the run-queue, set the lock
2797 * to the scheduler's lock.
2799 if (td->td_lock != TDQ_LOCKPTR(tdq)) {
2801 if ((flags & SRQ_HOLD) != 0)
2802 td->td_lock = TDQ_LOCKPTR(tdq);
2804 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2806 (void)tdq_add(tdq, td, flags);
2807 if (!(flags & SRQ_YIELDING))
2808 sched_setpreempt(td->td_priority);
2810 if (!(flags & SRQ_HOLDTD))
2815 * Remove a thread from a run-queue without running it. This is used
2816 * when we're stealing a thread from a remote queue. Otherwise all threads
2817 * exit by calling sched_exit_thread() and sched_throw() themselves.
2820 sched_rem(struct thread *td)
2824 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2825 "prio:%d", td->td_priority);
2826 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2827 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2828 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2829 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2830 KASSERT(TD_ON_RUNQ(td),
2831 ("sched_rem: thread not on run queue"));
2832 tdq_runq_rem(tdq, td);
2833 tdq_load_rem(tdq, td);
2835 if (td->td_priority == tdq->tdq_lowpri)
2836 tdq_setlowpri(tdq, NULL);
2840 * Fetch cpu utilization information. Updates on demand.
2843 sched_pctcpu(struct thread *td)
2846 struct td_sched *ts;
2849 ts = td_get_sched(td);
2851 THREAD_LOCK_ASSERT(td, MA_OWNED);
2852 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2856 /* How many rtick per second ? */
2857 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2858 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2865 * Enforce affinity settings for a thread. Called after adjustments to
2869 sched_affinity(struct thread *td)
2872 struct td_sched *ts;
2874 THREAD_LOCK_ASSERT(td, MA_OWNED);
2875 ts = td_get_sched(td);
2876 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2878 if (TD_ON_RUNQ(td)) {
2880 sched_add(td, SRQ_BORING | SRQ_HOLDTD);
2883 if (!TD_IS_RUNNING(td))
2886 * Force a switch before returning to userspace. If the
2887 * target thread is not running locally send an ipi to force
2890 ast_sched_locked(td, TDA_SCHED);
2891 if (td != curthread)
2892 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2897 * Bind a thread to a target cpu.
2900 sched_bind(struct thread *td, int cpu)
2902 struct td_sched *ts;
2904 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2905 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2906 ts = td_get_sched(td);
2907 if (ts->ts_flags & TSF_BOUND)
2909 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2910 ts->ts_flags |= TSF_BOUND;
2912 if (PCPU_GET(cpuid) == cpu)
2915 /* When we return from mi_switch we'll be on the correct cpu. */
2916 mi_switch(SW_VOL | SWT_BIND);
2921 * Release a bound thread.
2924 sched_unbind(struct thread *td)
2926 struct td_sched *ts;
2928 THREAD_LOCK_ASSERT(td, MA_OWNED);
2929 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2930 ts = td_get_sched(td);
2931 if ((ts->ts_flags & TSF_BOUND) == 0)
2933 ts->ts_flags &= ~TSF_BOUND;
2938 sched_is_bound(struct thread *td)
2940 THREAD_LOCK_ASSERT(td, MA_OWNED);
2941 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2948 sched_relinquish(struct thread *td)
2951 mi_switch(SW_VOL | SWT_RELINQUISH);
2955 * Return the total system load.
2966 total += atomic_load_int(&TDQ_CPU(i)->tdq_sysload);
2969 return (atomic_load_int(&TDQ_SELF()->tdq_sysload));
2974 sched_sizeof_proc(void)
2976 return (sizeof(struct proc));
2980 sched_sizeof_thread(void)
2982 return (sizeof(struct thread) + sizeof(struct td_sched));
2986 #define TDQ_IDLESPIN(tdq) \
2987 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2989 #define TDQ_IDLESPIN(tdq) 1
2993 * The actual idle process.
2996 sched_idletd(void *dummy)
3000 int oldswitchcnt, switchcnt;
3003 mtx_assert(&Giant, MA_NOTOWNED);
3006 THREAD_NO_SLEEPING();
3009 if (TDQ_LOAD(tdq)) {
3011 mi_switch(SW_VOL | SWT_IDLE);
3013 switchcnt = TDQ_SWITCHCNT(tdq);
3015 if (always_steal || switchcnt != oldswitchcnt) {
3016 oldswitchcnt = switchcnt;
3017 if (tdq_idled(tdq) == 0)
3020 switchcnt = TDQ_SWITCHCNT(tdq);
3022 oldswitchcnt = switchcnt;
3025 * If we're switching very frequently, spin while checking
3026 * for load rather than entering a low power state that
3027 * may require an IPI. However, don't do any busy
3028 * loops while on SMT machines as this simply steals
3029 * cycles from cores doing useful work.
3031 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
3032 for (i = 0; i < sched_idlespins; i++) {
3039 /* If there was context switch during spin, restart it. */
3040 switchcnt = TDQ_SWITCHCNT(tdq);
3041 if (TDQ_LOAD(tdq) != 0 || switchcnt != oldswitchcnt)
3044 /* Run main MD idle handler. */
3045 atomic_store_int(&tdq->tdq_cpu_idle, 1);
3047 * Make sure that the tdq_cpu_idle update is globally visible
3048 * before cpu_idle() reads tdq_load. The order is important
3049 * to avoid races with tdq_notify().
3051 atomic_thread_fence_seq_cst();
3053 * Checking for again after the fence picks up assigned
3054 * threads often enough to make it worthwhile to do so in
3055 * order to avoid calling cpu_idle().
3057 if (TDQ_LOAD(tdq) != 0) {
3058 atomic_store_int(&tdq->tdq_cpu_idle, 0);
3061 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
3062 atomic_store_int(&tdq->tdq_cpu_idle, 0);
3065 * Account thread-less hardware interrupts and
3066 * other wakeup reasons equal to context switches.
3068 switchcnt = TDQ_SWITCHCNT(tdq);
3069 if (switchcnt != oldswitchcnt)
3071 TDQ_SWITCHCNT_INC(tdq);
3077 * sched_throw_grab() chooses a thread from the queue to switch to
3078 * next. It returns with the tdq lock dropped in a spinlock section to
3079 * keep interrupts disabled until the CPU is running in a proper threaded
3082 static struct thread *
3083 sched_throw_grab(struct tdq *tdq)
3085 struct thread *newtd;
3087 newtd = choosethread();
3090 KASSERT(curthread->td_md.md_spinlock_count == 1,
3091 ("invalid count %d", curthread->td_md.md_spinlock_count));
3096 * A CPU is entering for the first time.
3099 sched_ap_entry(void)
3101 struct thread *newtd;
3106 /* This should have been setup in schedinit_ap(). */
3107 THREAD_LOCKPTR_ASSERT(curthread, TDQ_LOCKPTR(tdq));
3110 /* Correct spinlock nesting. */
3112 PCPU_SET(switchtime, cpu_ticks());
3113 PCPU_SET(switchticks, ticks);
3115 newtd = sched_throw_grab(tdq);
3117 /* doesn't return */
3118 cpu_throw(NULL, newtd);
3122 * A thread is exiting.
3125 sched_throw(struct thread *td)
3127 struct thread *newtd;
3133 THREAD_LOCK_ASSERT(td, MA_OWNED);
3134 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq));
3136 tdq_load_rem(tdq, td);
3137 td->td_lastcpu = td->td_oncpu;
3138 td->td_oncpu = NOCPU;
3139 thread_lock_block(td);
3141 newtd = sched_throw_grab(tdq);
3143 /* doesn't return */
3144 cpu_switch(td, newtd, TDQ_LOCKPTR(tdq));
3148 * This is called from fork_exit(). Just acquire the correct locks and
3149 * let fork do the rest of the work.
3152 sched_fork_exit(struct thread *td)
3158 * Finish setting up thread glue so that it begins execution in a
3159 * non-nested critical section with the scheduler lock held.
3161 KASSERT(curthread->td_md.md_spinlock_count == 1,
3162 ("invalid count %d", curthread->td_md.md_spinlock_count));
3163 cpuid = PCPU_GET(cpuid);
3167 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
3168 td->td_oncpu = cpuid;
3169 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
3170 "prio:%d", td->td_priority);
3171 SDT_PROBE0(sched, , , on__cpu);
3175 * Create on first use to catch odd startup conditions.
3178 sched_tdname(struct thread *td)
3181 struct td_sched *ts;
3183 ts = td_get_sched(td);
3184 if (ts->ts_name[0] == '\0')
3185 snprintf(ts->ts_name, sizeof(ts->ts_name),
3186 "%s tid %d", td->td_name, td->td_tid);
3187 return (ts->ts_name);
3189 return (td->td_name);
3195 sched_clear_tdname(struct thread *td)
3197 struct td_sched *ts;
3199 ts = td_get_sched(td);
3200 ts->ts_name[0] = '\0';
3207 * Build the CPU topology dump string. Is recursively called to collect
3208 * the topology tree.
3211 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
3214 char cpusetbuf[CPUSETBUFSIZ];
3217 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
3218 "", 1 + indent / 2, cg->cg_level);
3219 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3220 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3222 for (i = cg->cg_first; i <= cg->cg_last; i++) {
3223 if (CPU_ISSET(i, &cg->cg_mask)) {
3228 sbuf_printf(sb, "%d", i);
3231 sbuf_cat(sb, "</cpu>\n");
3233 if (cg->cg_flags != 0) {
3234 sbuf_printf(sb, "%*s <flags>", indent, "");
3235 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3236 sbuf_cat(sb, "<flag name=\"HTT\">HTT group</flag>");
3237 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3238 sbuf_cat(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3239 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3240 sbuf_cat(sb, "<flag name=\"SMT\">SMT group</flag>");
3241 if ((cg->cg_flags & CG_FLAG_NODE) != 0)
3242 sbuf_cat(sb, "<flag name=\"NODE\">NUMA node</flag>");
3243 sbuf_cat(sb, "</flags>\n");
3246 if (cg->cg_children > 0) {
3247 sbuf_printf(sb, "%*s <children>\n", indent, "");
3248 for (i = 0; i < cg->cg_children; i++)
3249 sysctl_kern_sched_topology_spec_internal(sb,
3250 &cg->cg_child[i], indent+2);
3251 sbuf_printf(sb, "%*s </children>\n", indent, "");
3253 sbuf_printf(sb, "%*s</group>\n", indent, "");
3258 * Sysctl handler for retrieving topology dump. It's a wrapper for
3259 * the recursive sysctl_kern_smp_topology_spec_internal().
3262 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3267 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3269 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3273 sbuf_cat(topo, "<groups>\n");
3274 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3275 sbuf_cat(topo, "</groups>\n");
3278 err = sbuf_finish(topo);
3287 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3289 int error, new_val, period;
3291 period = 1000000 / realstathz;
3292 new_val = period * sched_slice;
3293 error = sysctl_handle_int(oidp, &new_val, 0, req);
3294 if (error != 0 || req->newptr == NULL)
3298 sched_slice = imax(1, (new_val + period / 2) / period);
3299 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3300 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3305 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
3307 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3309 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum,
3310 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
3311 sysctl_kern_quantum, "I",
3312 "Quantum for timeshare threads in microseconds");
3313 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3314 "Quantum for timeshare threads in stathz ticks");
3315 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3316 "Interactivity score threshold");
3317 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3319 "Maximal (lowest) priority for preemption");
3320 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3321 "Assign static kernel priorities to sleeping threads");
3322 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3323 "Number of times idle thread will spin waiting for new work");
3324 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3325 &sched_idlespinthresh, 0,
3326 "Threshold before we will permit idle thread spinning");
3328 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3329 "Number of hz ticks to keep thread affinity for");
3330 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3331 "Enables the long-term load balancer");
3332 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3333 &balance_interval, 0,
3334 "Average period in stathz ticks to run the long-term balancer");
3335 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3336 "Attempts to steal work from other cores before idling");
3337 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3338 "Minimum load on remote CPU before we'll steal");
3339 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3340 0, "Topological distance limit for stealing threads in sched_switch()");
3341 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3342 "Always run the stealer from the idle thread");
3343 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3344 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3345 "XML dump of detected CPU topology");
3348 /* ps compat. All cpu percentages from ULE are weighted. */
3349 static int ccpu = 0;
3350 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0,
3351 "Decay factor used for updating %CPU in 4BSD scheduler");