2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
46 #include <sys/param.h>
47 #include <sys/systm.h>
49 #include <sys/kernel.h>
51 #include <sys/limits.h>
53 #include <sys/mutex.h>
55 #include <sys/resource.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/turnstile.h>
65 #include <sys/vmmeter.h>
66 #include <sys/cpuset.h>
70 #include <sys/pmckern.h>
74 #include <sys/dtrace_bsd.h>
75 int dtrace_vtime_active;
76 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
79 #include <machine/cpu.h>
80 #include <machine/smp.h>
84 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
85 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
86 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
89 * Thread scheduler specific section. All fields are protected
93 struct runq *ts_runq; /* Run-queue we're queued on. */
94 short ts_flags; /* TSF_* flags. */
95 int ts_cpu; /* CPU that we have affinity for. */
96 int ts_rltick; /* Real last tick, for affinity. */
97 int ts_slice; /* Ticks of slice remaining. */
98 u_int ts_slptime; /* Number of ticks we vol. slept */
99 u_int ts_runtime; /* Number of ticks we were running */
100 int ts_ltick; /* Last tick that we were running on */
101 int ts_ftick; /* First tick that we were running on */
102 int ts_ticks; /* Tick count */
104 char ts_name[TS_NAME_LEN];
107 /* flags kept in ts_flags */
108 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
109 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
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)
115 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
116 sizeof(struct thread0_storage),
117 "increase struct thread0_storage.t0st_sched size");
120 * Priority ranges used for interactive and non-interactive timeshare
121 * threads. The timeshare priorities are split up into four ranges.
122 * The first range handles interactive threads. The last three ranges
123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124 * ranges supporting nice values.
126 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
130 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
131 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
136 * Cpu percentage computation macros and defines.
138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
145 #define SCHED_TICK_SECS 10
146 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
147 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
148 #define SCHED_TICK_SHIFT 10
149 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
153 * These macros determine priorities for non-interactive threads. They are
154 * assigned a priority based on their recent cpu utilization as expressed
155 * by the ratio of ticks to the tick total. NHALF priorities at the start
156 * and end of the MIN to MAX timeshare range are only reachable with negative
157 * or positive nice respectively.
159 * PRI_RANGE: Priority range for utilization dependent priorities.
160 * PRI_NRESV: Number of nice values.
161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
162 * PRI_NICE: Determines the part of the priority inherited from nice.
164 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
165 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
166 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define SCHED_PRI_TICKS(ts) \
170 (SCHED_TICK_HZ((ts)) / \
171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define SCHED_PRI_NICE(nice) (nice)
175 * These determine the interactivity of a process. Interactivity differs from
176 * cpu utilization in that it expresses the voluntary time slept vs time ran
177 * while cpu utilization includes all time not running. This more accurately
178 * models the intent of the thread.
180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
181 * before throttling back.
182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
183 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
184 * INTERACT_THRESH: Threshold for placement on the current runq.
186 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
187 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
188 #define SCHED_INTERACT_MAX (100)
189 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
190 #define SCHED_INTERACT_THRESH (30)
193 * These parameters determine the slice behavior for batch work.
195 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
196 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
198 /* Flags kept in td_flags. */
199 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
202 * tickincr: Converts a stathz tick into a hz domain scaled by
203 * the shift factor. Without the shift the error rate
204 * due to rounding would be unacceptably high.
205 * realstathz: stathz is sometimes 0 and run off of hz.
206 * sched_slice: Runtime of each thread before rescheduling.
207 * preempt_thresh: Priority threshold for preemption and remote IPIs.
209 static int sched_interact = SCHED_INTERACT_THRESH;
210 static int tickincr = 8 << SCHED_TICK_SHIFT;
211 static int realstathz = 127; /* reset during boot. */
212 static int sched_slice = 10; /* reset during boot. */
213 static int sched_slice_min = 1; /* reset during boot. */
215 #ifdef FULL_PREEMPTION
216 static int preempt_thresh = PRI_MAX_IDLE;
218 static int preempt_thresh = PRI_MIN_KERN;
221 static int preempt_thresh = 0;
223 static int static_boost = PRI_MIN_BATCH;
224 static int sched_idlespins = 10000;
225 static int sched_idlespinthresh = -1;
228 * tdq - per processor runqs and statistics. All fields are protected by the
229 * tdq_lock. The load and lowpri may be accessed without to avoid excess
230 * locking in sched_pickcpu();
234 * Ordered to improve efficiency of cpu_search() and switch().
235 * tdq_lock is padded to avoid false sharing with tdq_load and
238 struct mtx_padalign tdq_lock; /* run queue lock. */
239 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
240 volatile int tdq_load; /* Aggregate load. */
241 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
242 int tdq_sysload; /* For loadavg, !ITHD load. */
243 volatile int tdq_transferable; /* Transferable thread count. */
244 volatile short tdq_switchcnt; /* Switches this tick. */
245 volatile short tdq_oldswitchcnt; /* Switches last tick. */
246 u_char tdq_lowpri; /* Lowest priority thread. */
247 u_char tdq_ipipending; /* IPI pending. */
248 u_char tdq_idx; /* Current insert index. */
249 u_char tdq_ridx; /* Current removal index. */
250 struct runq tdq_realtime; /* real-time run queue. */
251 struct runq tdq_timeshare; /* timeshare run queue. */
252 struct runq tdq_idle; /* Queue of IDLE threads. */
253 char tdq_name[TDQ_NAME_LEN];
255 char tdq_loadname[TDQ_LOADNAME_LEN];
259 /* Idle thread states and config. */
260 #define TDQ_RUNNING 1
264 struct cpu_group *cpu_top; /* CPU topology */
266 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
267 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
272 static int rebalance = 1;
273 static int balance_interval = 128; /* Default set in sched_initticks(). */
275 static int steal_idle = 1;
276 static int steal_thresh = 2;
277 static int always_steal = 0;
278 static int trysteal_limit = 2;
281 * One thread queue per processor.
283 static struct tdq tdq_cpu[MAXCPU];
284 static struct tdq *balance_tdq;
285 static int balance_ticks;
286 DPCPU_DEFINE_STATIC(uint32_t, randomval);
288 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
289 #define TDQ_CPU(x) (&tdq_cpu[(x)])
290 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
292 static struct tdq tdq_cpu;
294 #define TDQ_ID(x) (0)
295 #define TDQ_SELF() (&tdq_cpu)
296 #define TDQ_CPU(x) (&tdq_cpu)
299 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
300 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
301 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
302 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
303 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
305 static void sched_priority(struct thread *);
306 static void sched_thread_priority(struct thread *, u_char);
307 static int sched_interact_score(struct thread *);
308 static void sched_interact_update(struct thread *);
309 static void sched_interact_fork(struct thread *);
310 static void sched_pctcpu_update(struct td_sched *, int);
312 /* Operations on per processor queues */
313 static struct thread *tdq_choose(struct tdq *);
314 static void tdq_setup(struct tdq *);
315 static void tdq_load_add(struct tdq *, struct thread *);
316 static void tdq_load_rem(struct tdq *, struct thread *);
317 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
318 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
319 static inline int sched_shouldpreempt(int, int, int);
320 void tdq_print(int cpu);
321 static void runq_print(struct runq *rq);
322 static void tdq_add(struct tdq *, struct thread *, int);
324 static struct thread *tdq_move(struct tdq *, struct tdq *);
325 static int tdq_idled(struct tdq *);
326 static void tdq_notify(struct tdq *, struct thread *);
327 static struct thread *tdq_steal(struct tdq *, int);
328 static struct thread *runq_steal(struct runq *, int);
329 static int sched_pickcpu(struct thread *, int);
330 static void sched_balance(void);
331 static int sched_balance_pair(struct tdq *, struct tdq *);
332 static inline struct tdq *sched_setcpu(struct thread *, int, int);
333 static inline void thread_unblock_switch(struct thread *, struct mtx *);
334 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
335 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
336 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
337 struct cpu_group *cg, int indent);
340 static void sched_setup(void *dummy);
341 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
343 static void sched_initticks(void *dummy);
344 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
347 SDT_PROVIDER_DEFINE(sched);
349 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
350 "struct proc *", "uint8_t");
351 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
352 "struct proc *", "void *");
353 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
354 "struct proc *", "void *", "int");
355 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
356 "struct proc *", "uint8_t", "struct thread *");
357 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
358 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
360 SDT_PROBE_DEFINE(sched, , , on__cpu);
361 SDT_PROBE_DEFINE(sched, , , remain__cpu);
362 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
366 * Print the threads waiting on a run-queue.
369 runq_print(struct runq *rq)
377 for (i = 0; i < RQB_LEN; i++) {
378 printf("\t\trunq bits %d 0x%zx\n",
379 i, rq->rq_status.rqb_bits[i]);
380 for (j = 0; j < RQB_BPW; j++)
381 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
382 pri = j + (i << RQB_L2BPW);
383 rqh = &rq->rq_queues[pri];
384 TAILQ_FOREACH(td, rqh, td_runq) {
385 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
386 td, td->td_name, td->td_priority,
387 td->td_rqindex, pri);
394 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
403 printf("tdq %d:\n", TDQ_ID(tdq));
404 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
405 printf("\tLock name: %s\n", tdq->tdq_name);
406 printf("\tload: %d\n", tdq->tdq_load);
407 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
408 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
409 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
410 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
411 printf("\tload transferable: %d\n", tdq->tdq_transferable);
412 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
413 printf("\trealtime runq:\n");
414 runq_print(&tdq->tdq_realtime);
415 printf("\ttimeshare runq:\n");
416 runq_print(&tdq->tdq_timeshare);
417 printf("\tidle runq:\n");
418 runq_print(&tdq->tdq_idle);
422 sched_shouldpreempt(int pri, int cpri, int remote)
425 * If the new priority is not better than the current priority there is
431 * Always preempt idle.
433 if (cpri >= PRI_MIN_IDLE)
436 * If preemption is disabled don't preempt others.
438 if (preempt_thresh == 0)
441 * Preempt if we exceed the threshold.
443 if (pri <= preempt_thresh)
446 * If we're interactive or better and there is non-interactive
447 * or worse running preempt only remote processors.
449 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
455 * Add a thread to the actual run-queue. Keeps transferable counts up to
456 * date with what is actually on the run-queue. Selects the correct
457 * queue position for timeshare threads.
460 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
465 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
466 THREAD_LOCK_ASSERT(td, MA_OWNED);
468 pri = td->td_priority;
469 ts = td_get_sched(td);
471 if (THREAD_CAN_MIGRATE(td)) {
472 tdq->tdq_transferable++;
473 ts->ts_flags |= TSF_XFERABLE;
475 if (pri < PRI_MIN_BATCH) {
476 ts->ts_runq = &tdq->tdq_realtime;
477 } else if (pri <= PRI_MAX_BATCH) {
478 ts->ts_runq = &tdq->tdq_timeshare;
479 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
480 ("Invalid priority %d on timeshare runq", pri));
482 * This queue contains only priorities between MIN and MAX
483 * realtime. Use the whole queue to represent these values.
485 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
486 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
487 pri = (pri + tdq->tdq_idx) % RQ_NQS;
489 * This effectively shortens the queue by one so we
490 * can have a one slot difference between idx and
491 * ridx while we wait for threads to drain.
493 if (tdq->tdq_ridx != tdq->tdq_idx &&
494 pri == tdq->tdq_ridx)
495 pri = (unsigned char)(pri - 1) % RQ_NQS;
498 runq_add_pri(ts->ts_runq, td, pri, flags);
501 ts->ts_runq = &tdq->tdq_idle;
502 runq_add(ts->ts_runq, td, flags);
506 * Remove a thread from a run-queue. This typically happens when a thread
507 * is selected to run. Running threads are not on the queue and the
508 * transferable count does not reflect them.
511 tdq_runq_rem(struct tdq *tdq, struct thread *td)
515 ts = td_get_sched(td);
516 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
517 KASSERT(ts->ts_runq != NULL,
518 ("tdq_runq_remove: thread %p null ts_runq", td));
519 if (ts->ts_flags & TSF_XFERABLE) {
520 tdq->tdq_transferable--;
521 ts->ts_flags &= ~TSF_XFERABLE;
523 if (ts->ts_runq == &tdq->tdq_timeshare) {
524 if (tdq->tdq_idx != tdq->tdq_ridx)
525 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
527 runq_remove_idx(ts->ts_runq, td, NULL);
529 runq_remove(ts->ts_runq, td);
533 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
534 * for this thread to the referenced thread queue.
537 tdq_load_add(struct tdq *tdq, struct thread *td)
540 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
541 THREAD_LOCK_ASSERT(td, MA_OWNED);
544 if ((td->td_flags & TDF_NOLOAD) == 0)
546 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
547 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
551 * Remove the load from a thread that is transitioning to a sleep state or
555 tdq_load_rem(struct tdq *tdq, struct thread *td)
558 THREAD_LOCK_ASSERT(td, MA_OWNED);
559 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
560 KASSERT(tdq->tdq_load != 0,
561 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
564 if ((td->td_flags & TDF_NOLOAD) == 0)
566 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
567 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
571 * Bound timeshare latency by decreasing slice size as load increases. We
572 * consider the maximum latency as the sum of the threads waiting to run
573 * aside from curthread and target no more than sched_slice latency but
574 * no less than sched_slice_min runtime.
577 tdq_slice(struct tdq *tdq)
582 * It is safe to use sys_load here because this is called from
583 * contexts where timeshare threads are running and so there
584 * cannot be higher priority load in the system.
586 load = tdq->tdq_sysload - 1;
587 if (load >= SCHED_SLICE_MIN_DIVISOR)
588 return (sched_slice_min);
590 return (sched_slice);
591 return (sched_slice / load);
595 * Set lowpri to its exact value by searching the run-queue and
596 * evaluating curthread. curthread may be passed as an optimization.
599 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
603 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
605 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
606 td = tdq_choose(tdq);
607 if (td == NULL || td->td_priority > ctd->td_priority)
608 tdq->tdq_lowpri = ctd->td_priority;
610 tdq->tdq_lowpri = td->td_priority;
615 * We need some randomness. Implement a classic Linear Congruential
616 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
617 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
618 * of the random state (in the low bits of our answer) to keep
619 * the maximum randomness.
626 rndptr = DPCPU_PTR(randomval);
627 *rndptr = *rndptr * 69069 + 5;
629 return (*rndptr >> 16);
635 int cs_pri; /* Min priority for low. */
636 int cs_limit; /* Max load for low, min load for high. */
641 #define CPU_SEARCH_LOWEST 0x1
642 #define CPU_SEARCH_HIGHEST 0x2
643 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
645 #define CPUSET_FOREACH(cpu, mask) \
646 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
647 if (CPU_ISSET(cpu, &mask))
649 static __always_inline int cpu_search(const struct cpu_group *cg,
650 struct cpu_search *low, struct cpu_search *high, const int match);
651 int __noinline cpu_search_lowest(const struct cpu_group *cg,
652 struct cpu_search *low);
653 int __noinline cpu_search_highest(const struct cpu_group *cg,
654 struct cpu_search *high);
655 int __noinline cpu_search_both(const struct cpu_group *cg,
656 struct cpu_search *low, struct cpu_search *high);
659 * Search the tree of cpu_groups for the lowest or highest loaded cpu
660 * according to the match argument. This routine actually compares the
661 * load on all paths through the tree and finds the least loaded cpu on
662 * the least loaded path, which may differ from the least loaded cpu in
663 * the system. This balances work among caches and buses.
665 * This inline is instantiated in three forms below using constants for the
666 * match argument. It is reduced to the minimum set for each case. It is
667 * also recursive to the depth of the tree.
669 static __always_inline int
670 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
671 struct cpu_search *high, const int match)
673 struct cpu_search lgroup;
674 struct cpu_search hgroup;
676 struct cpu_group *child;
678 int cpu, i, hload, lload, load, total, rnd;
681 cpumask = cg->cg_mask;
682 if (match & CPU_SEARCH_LOWEST) {
686 if (match & CPU_SEARCH_HIGHEST) {
691 /* Iterate through the child CPU groups and then remaining CPUs. */
692 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
694 #ifdef HAVE_INLINE_FFSL
695 cpu = CPU_FFS(&cpumask) - 1;
697 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
704 child = &cg->cg_child[i - 1];
706 if (match & CPU_SEARCH_LOWEST)
708 if (match & CPU_SEARCH_HIGHEST)
710 if (child) { /* Handle child CPU group. */
711 CPU_NAND(&cpumask, &child->cg_mask);
713 case CPU_SEARCH_LOWEST:
714 load = cpu_search_lowest(child, &lgroup);
716 case CPU_SEARCH_HIGHEST:
717 load = cpu_search_highest(child, &hgroup);
719 case CPU_SEARCH_BOTH:
720 load = cpu_search_both(child, &lgroup, &hgroup);
723 } else { /* Handle child CPU. */
724 CPU_CLR(cpu, &cpumask);
726 load = tdq->tdq_load * 256;
727 rnd = sched_random() % 32;
728 if (match & CPU_SEARCH_LOWEST) {
729 if (cpu == low->cs_prefer)
731 /* If that CPU is allowed and get data. */
732 if (tdq->tdq_lowpri > lgroup.cs_pri &&
733 tdq->tdq_load <= lgroup.cs_limit &&
734 CPU_ISSET(cpu, &lgroup.cs_mask)) {
736 lgroup.cs_load = load - rnd;
739 if (match & CPU_SEARCH_HIGHEST)
740 if (tdq->tdq_load >= hgroup.cs_limit &&
741 tdq->tdq_transferable &&
742 CPU_ISSET(cpu, &hgroup.cs_mask)) {
744 hgroup.cs_load = load - rnd;
749 /* We have info about child item. Compare it. */
750 if (match & CPU_SEARCH_LOWEST) {
751 if (lgroup.cs_cpu >= 0 &&
753 (load == lload && lgroup.cs_load < low->cs_load))) {
755 low->cs_cpu = lgroup.cs_cpu;
756 low->cs_load = lgroup.cs_load;
759 if (match & CPU_SEARCH_HIGHEST)
760 if (hgroup.cs_cpu >= 0 &&
762 (load == hload && hgroup.cs_load > high->cs_load))) {
764 high->cs_cpu = hgroup.cs_cpu;
765 high->cs_load = hgroup.cs_load;
769 if (i == 0 && CPU_EMPTY(&cpumask))
772 #ifndef HAVE_INLINE_FFSL
781 * cpu_search instantiations must pass constants to maintain the inline
785 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
787 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
791 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
793 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
797 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
798 struct cpu_search *high)
800 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
804 * Find the cpu with the least load via the least loaded path that has a
805 * lowpri greater than pri pri. A pri of -1 indicates any priority is
809 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
812 struct cpu_search low;
815 low.cs_prefer = prefer;
818 low.cs_limit = maxload;
819 cpu_search_lowest(cg, &low);
824 * Find the cpu with the highest load via the highest loaded path.
827 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
829 struct cpu_search high;
833 high.cs_limit = minload;
834 cpu_search_highest(cg, &high);
839 sched_balance_group(struct cpu_group *cg)
841 cpuset_t hmask, lmask;
842 int high, low, anylow;
846 high = sched_highest(cg, hmask, 2);
847 /* Stop if there is no more CPU with transferrable threads. */
850 CPU_CLR(high, &hmask);
851 CPU_COPY(&hmask, &lmask);
852 /* Stop if there is no more CPU left for low. */
853 if (CPU_EMPTY(&lmask))
857 low = sched_lowest(cg, lmask, -1,
858 TDQ_CPU(high)->tdq_load - 1, high);
859 /* Stop if we looked well and found no less loaded CPU. */
860 if (anylow && low == -1)
862 /* Go to next high if we found no less loaded CPU. */
865 /* Transfer thread from high to low. */
866 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
867 /* CPU that got thread can no longer be a donor. */
868 CPU_CLR(low, &hmask);
871 * If failed, then there is no threads on high
872 * that can run on this low. Drop low from low
873 * mask and look for different one.
875 CPU_CLR(low, &lmask);
887 balance_ticks = max(balance_interval / 2, 1) +
888 (sched_random() % balance_interval);
891 sched_balance_group(cpu_top);
896 * Lock two thread queues using their address to maintain lock order.
899 tdq_lock_pair(struct tdq *one, struct tdq *two)
903 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
906 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
911 * Unlock two thread queues. Order is not important here.
914 tdq_unlock_pair(struct tdq *one, struct tdq *two)
921 * Transfer load between two imbalanced thread queues.
924 sched_balance_pair(struct tdq *high, struct tdq *low)
929 tdq_lock_pair(high, low);
932 * Transfer a thread from high to low.
934 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
935 (td = tdq_move(high, low)) != NULL) {
937 * In case the target isn't the current cpu notify it of the
938 * new load, possibly sending an IPI to force it to reschedule.
941 if (cpu != PCPU_GET(cpuid))
944 tdq_unlock_pair(high, low);
949 * Move a thread from one thread queue to another.
951 static struct thread *
952 tdq_move(struct tdq *from, struct tdq *to)
959 TDQ_LOCK_ASSERT(from, MA_OWNED);
960 TDQ_LOCK_ASSERT(to, MA_OWNED);
964 td = tdq_steal(tdq, cpu);
967 ts = td_get_sched(td);
969 * Although the run queue is locked the thread may be blocked. Lock
970 * it to clear this and acquire the run-queue lock.
973 /* Drop recursive lock on from acquired via thread_lock(). */
977 td->td_lock = TDQ_LOCKPTR(to);
978 tdq_add(to, td, SRQ_YIELDING);
983 * This tdq has idled. Try to steal a thread from another cpu and switch
987 tdq_idled(struct tdq *tdq)
989 struct cpu_group *cg;
994 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
997 CPU_CLR(PCPU_GET(cpuid), &mask);
999 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
1000 for (cg = tdq->tdq_cg; ; ) {
1001 cpu = sched_highest(cg, mask, steal_thresh);
1003 * We were assigned a thread but not preempted. Returning
1004 * 0 here will cause our caller to switch to it.
1014 steal = TDQ_CPU(cpu);
1016 * The data returned by sched_highest() is stale and
1017 * the chosen CPU no longer has an eligible thread.
1019 * Testing this ahead of tdq_lock_pair() only catches
1020 * this situation about 20% of the time on an 8 core
1021 * 16 thread Ryzen 7, but it still helps performance.
1023 if (steal->tdq_load < steal_thresh ||
1024 steal->tdq_transferable == 0)
1026 tdq_lock_pair(tdq, steal);
1028 * We were assigned a thread while waiting for the locks.
1029 * Switch to it now instead of stealing a thread.
1034 * The data returned by sched_highest() is stale and
1035 * the chosen CPU no longer has an eligible thread, or
1036 * we were preempted and the CPU loading info may be out
1037 * of date. The latter is rare. In either case restart
1040 if (steal->tdq_load < steal_thresh ||
1041 steal->tdq_transferable == 0 ||
1042 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1043 tdq_unlock_pair(tdq, steal);
1047 * Steal the thread and switch to it.
1049 if (tdq_move(steal, tdq) != NULL)
1052 * We failed to acquire a thread even though it looked
1053 * like one was available. This could be due to affinity
1054 * restrictions or for other reasons. Loop again after
1055 * removing this CPU from the set. The restart logic
1056 * above does not restore this CPU to the set due to the
1057 * likelyhood of failing here again.
1059 CPU_CLR(cpu, &mask);
1060 tdq_unlock_pair(tdq, steal);
1063 mi_switch(SW_VOL | SWT_IDLE, NULL);
1064 thread_unlock(curthread);
1069 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1072 tdq_notify(struct tdq *tdq, struct thread *td)
1078 if (tdq->tdq_ipipending)
1080 cpu = td_get_sched(td)->ts_cpu;
1081 pri = td->td_priority;
1082 ctd = pcpu_find(cpu)->pc_curthread;
1083 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1087 * Make sure that our caller's earlier update to tdq_load is
1088 * globally visible before we read tdq_cpu_idle. Idle thread
1089 * accesses both of them without locks, and the order is important.
1091 atomic_thread_fence_seq_cst();
1093 if (TD_IS_IDLETHREAD(ctd)) {
1095 * If the MD code has an idle wakeup routine try that before
1096 * falling back to IPI.
1098 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1101 tdq->tdq_ipipending = 1;
1102 ipi_cpu(cpu, IPI_PREEMPT);
1106 * Steals load from a timeshare queue. Honors the rotating queue head
1109 static struct thread *
1110 runq_steal_from(struct runq *rq, int cpu, u_char start)
1114 struct thread *td, *first;
1118 rqb = &rq->rq_status;
1119 bit = start & (RQB_BPW -1);
1122 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1123 if (rqb->rqb_bits[i] == 0)
1126 bit = RQB_FFS(rqb->rqb_bits[i]);
1127 for (; bit < RQB_BPW; bit++) {
1128 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1130 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1131 TAILQ_FOREACH(td, rqh, td_runq) {
1132 if (first && THREAD_CAN_MIGRATE(td) &&
1133 THREAD_CAN_SCHED(td, cpu))
1144 if (first && THREAD_CAN_MIGRATE(first) &&
1145 THREAD_CAN_SCHED(first, cpu))
1151 * Steals load from a standard linear queue.
1153 static struct thread *
1154 runq_steal(struct runq *rq, int cpu)
1162 rqb = &rq->rq_status;
1163 for (word = 0; word < RQB_LEN; word++) {
1164 if (rqb->rqb_bits[word] == 0)
1166 for (bit = 0; bit < RQB_BPW; bit++) {
1167 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1169 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1170 TAILQ_FOREACH(td, rqh, td_runq)
1171 if (THREAD_CAN_MIGRATE(td) &&
1172 THREAD_CAN_SCHED(td, cpu))
1180 * Attempt to steal a thread in priority order from a thread queue.
1182 static struct thread *
1183 tdq_steal(struct tdq *tdq, int cpu)
1187 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1188 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1190 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1191 cpu, tdq->tdq_ridx)) != NULL)
1193 return (runq_steal(&tdq->tdq_idle, cpu));
1197 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1198 * current lock and returns with the assigned queue locked.
1200 static inline struct tdq *
1201 sched_setcpu(struct thread *td, int cpu, int flags)
1206 THREAD_LOCK_ASSERT(td, MA_OWNED);
1208 td_get_sched(td)->ts_cpu = cpu;
1210 * If the lock matches just return the queue.
1212 if (td->td_lock == TDQ_LOCKPTR(tdq))
1216 * If the thread isn't running its lockptr is a
1217 * turnstile or a sleepqueue. We can just lock_set without
1220 if (TD_CAN_RUN(td)) {
1222 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1227 * The hard case, migration, we need to block the thread first to
1228 * prevent order reversals with other cpus locks.
1231 thread_lock_block(td);
1233 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1238 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1239 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1240 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1241 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1242 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1243 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1246 sched_pickcpu(struct thread *td, int flags)
1248 struct cpu_group *cg, *ccg;
1249 struct td_sched *ts;
1254 self = PCPU_GET(cpuid);
1255 ts = td_get_sched(td);
1256 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1257 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1258 if (smp_started == 0)
1261 * Don't migrate a running thread from sched_switch().
1263 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1264 return (ts->ts_cpu);
1266 * Prefer to run interrupt threads on the processors that generate
1269 pri = td->td_priority;
1270 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1271 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1272 SCHED_STAT_INC(pickcpu_intrbind);
1274 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1275 SCHED_STAT_INC(pickcpu_affinity);
1276 return (ts->ts_cpu);
1280 * If the thread can run on the last cpu and the affinity has not
1281 * expired and it is idle, run it there.
1283 tdq = TDQ_CPU(ts->ts_cpu);
1285 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1286 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1287 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1288 if (cg->cg_flags & CG_FLAG_THREAD) {
1289 CPUSET_FOREACH(cpu, cg->cg_mask) {
1290 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1295 if (cpu > mp_maxid) {
1296 SCHED_STAT_INC(pickcpu_idle_affinity);
1297 return (ts->ts_cpu);
1301 * Search for the last level cache CPU group in the tree.
1302 * Skip caches with expired affinity time and SMT groups.
1303 * Affinity to higher level caches will be handled less aggressively.
1305 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1306 if (cg->cg_flags & CG_FLAG_THREAD)
1308 if (!SCHED_AFFINITY(ts, cg->cg_level))
1315 /* Search the group for the less loaded idle CPU we can run now. */
1316 mask = td->td_cpuset->cs_mask;
1317 if (cg != NULL && cg != cpu_top &&
1318 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1319 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1320 INT_MAX, ts->ts_cpu);
1321 /* Search globally for the less loaded CPU we can run now. */
1323 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1324 /* Search globally for the less loaded CPU. */
1326 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1327 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1328 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1330 * Compare the lowest loaded cpu to current cpu.
1332 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1333 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1334 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1335 SCHED_STAT_INC(pickcpu_local);
1338 SCHED_STAT_INC(pickcpu_lowest);
1339 if (cpu != ts->ts_cpu)
1340 SCHED_STAT_INC(pickcpu_migration);
1346 * Pick the highest priority task we have and return it.
1348 static struct thread *
1349 tdq_choose(struct tdq *tdq)
1353 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1354 td = runq_choose(&tdq->tdq_realtime);
1357 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1359 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1360 ("tdq_choose: Invalid priority on timeshare queue %d",
1364 td = runq_choose(&tdq->tdq_idle);
1366 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1367 ("tdq_choose: Invalid priority on idle queue %d",
1376 * Initialize a thread queue.
1379 tdq_setup(struct tdq *tdq)
1383 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1384 runq_init(&tdq->tdq_realtime);
1385 runq_init(&tdq->tdq_timeshare);
1386 runq_init(&tdq->tdq_idle);
1387 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1388 "sched lock %d", (int)TDQ_ID(tdq));
1389 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1390 MTX_SPIN | MTX_RECURSE);
1392 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1393 "CPU %d load", (int)TDQ_ID(tdq));
1399 sched_setup_smp(void)
1404 cpu_top = smp_topo();
1408 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1409 if (tdq->tdq_cg == NULL)
1410 panic("Can't find cpu group for %d\n", i);
1412 balance_tdq = TDQ_SELF();
1417 * Setup the thread queues and initialize the topology based on MD
1421 sched_setup(void *dummy)
1432 /* Add thread0's load since it's running. */
1434 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1435 tdq_load_add(tdq, &thread0);
1436 tdq->tdq_lowpri = thread0.td_priority;
1441 * This routine determines time constants after stathz and hz are setup.
1445 sched_initticks(void *dummy)
1449 realstathz = stathz ? stathz : hz;
1450 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1451 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1452 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1456 * tickincr is shifted out by 10 to avoid rounding errors due to
1457 * hz not being evenly divisible by stathz on all platforms.
1459 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1461 * This does not work for values of stathz that are more than
1462 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1469 * Set the default balance interval now that we know
1470 * what realstathz is.
1472 balance_interval = realstathz;
1473 balance_ticks = balance_interval;
1474 affinity = SCHED_AFFINITY_DEFAULT;
1476 if (sched_idlespinthresh < 0)
1477 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1482 * This is the core of the interactivity algorithm. Determines a score based
1483 * on past behavior. It is the ratio of sleep time to run time scaled to
1484 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1485 * differs from the cpu usage because it does not account for time spent
1486 * waiting on a run-queue. Would be prettier if we had floating point.
1488 * When a thread's sleep time is greater than its run time the
1492 * interactivity score = ---------------------
1493 * sleep time / run time
1496 * When a thread's run time is greater than its sleep time the
1500 * interactivity score = --------------------- + scaling factor
1501 * run time / sleep time
1504 sched_interact_score(struct thread *td)
1506 struct td_sched *ts;
1509 ts = td_get_sched(td);
1511 * The score is only needed if this is likely to be an interactive
1512 * task. Don't go through the expense of computing it if there's
1515 if (sched_interact <= SCHED_INTERACT_HALF &&
1516 ts->ts_runtime >= ts->ts_slptime)
1517 return (SCHED_INTERACT_HALF);
1519 if (ts->ts_runtime > ts->ts_slptime) {
1520 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1521 return (SCHED_INTERACT_HALF +
1522 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1524 if (ts->ts_slptime > ts->ts_runtime) {
1525 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1526 return (ts->ts_runtime / div);
1528 /* runtime == slptime */
1530 return (SCHED_INTERACT_HALF);
1533 * This can happen if slptime and runtime are 0.
1540 * Scale the scheduling priority according to the "interactivity" of this
1544 sched_priority(struct thread *td)
1549 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1552 * If the score is interactive we place the thread in the realtime
1553 * queue with a priority that is less than kernel and interrupt
1554 * priorities. These threads are not subject to nice restrictions.
1556 * Scores greater than this are placed on the normal timeshare queue
1557 * where the priority is partially decided by the most recent cpu
1558 * utilization and the rest is decided by nice value.
1560 * The nice value of the process has a linear effect on the calculated
1561 * score. Negative nice values make it easier for a thread to be
1562 * considered interactive.
1564 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1565 if (score < sched_interact) {
1566 pri = PRI_MIN_INTERACT;
1567 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1568 sched_interact) * score;
1569 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1570 ("sched_priority: invalid interactive priority %d score %d",
1573 pri = SCHED_PRI_MIN;
1574 if (td_get_sched(td)->ts_ticks)
1575 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1576 SCHED_PRI_RANGE - 1);
1577 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1578 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1579 ("sched_priority: invalid priority %d: nice %d, "
1580 "ticks %d ftick %d ltick %d tick pri %d",
1581 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1582 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1583 SCHED_PRI_TICKS(td_get_sched(td))));
1585 sched_user_prio(td, pri);
1591 * This routine enforces a maximum limit on the amount of scheduling history
1592 * kept. It is called after either the slptime or runtime is adjusted. This
1593 * function is ugly due to integer math.
1596 sched_interact_update(struct thread *td)
1598 struct td_sched *ts;
1601 ts = td_get_sched(td);
1602 sum = ts->ts_runtime + ts->ts_slptime;
1603 if (sum < SCHED_SLP_RUN_MAX)
1606 * This only happens from two places:
1607 * 1) We have added an unusual amount of run time from fork_exit.
1608 * 2) We have added an unusual amount of sleep time from sched_sleep().
1610 if (sum > SCHED_SLP_RUN_MAX * 2) {
1611 if (ts->ts_runtime > ts->ts_slptime) {
1612 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1615 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1621 * If we have exceeded by more than 1/5th then the algorithm below
1622 * will not bring us back into range. Dividing by two here forces
1623 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1625 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1626 ts->ts_runtime /= 2;
1627 ts->ts_slptime /= 2;
1630 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1631 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1635 * Scale back the interactivity history when a child thread is created. The
1636 * history is inherited from the parent but the thread may behave totally
1637 * differently. For example, a shell spawning a compiler process. We want
1638 * to learn that the compiler is behaving badly very quickly.
1641 sched_interact_fork(struct thread *td)
1643 struct td_sched *ts;
1647 ts = td_get_sched(td);
1648 sum = ts->ts_runtime + ts->ts_slptime;
1649 if (sum > SCHED_SLP_RUN_FORK) {
1650 ratio = sum / SCHED_SLP_RUN_FORK;
1651 ts->ts_runtime /= ratio;
1652 ts->ts_slptime /= ratio;
1657 * Called from proc0_init() to setup the scheduler fields.
1662 struct td_sched *ts0;
1665 * Set up the scheduler specific parts of thread0.
1667 ts0 = td_get_sched(&thread0);
1668 ts0->ts_ltick = ticks;
1669 ts0->ts_ftick = ticks;
1671 ts0->ts_cpu = curcpu; /* set valid CPU number */
1675 * This is only somewhat accurate since given many processes of the same
1676 * priority they will switch when their slices run out, which will be
1677 * at most sched_slice stathz ticks.
1680 sched_rr_interval(void)
1683 /* Convert sched_slice from stathz to hz. */
1684 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1688 * Update the percent cpu tracking information when it is requested or
1689 * the total history exceeds the maximum. We keep a sliding history of
1690 * tick counts that slowly decays. This is less precise than the 4BSD
1691 * mechanism since it happens with less regular and frequent events.
1694 sched_pctcpu_update(struct td_sched *ts, int run)
1699 * The signed difference may be negative if the thread hasn't run for
1700 * over half of the ticks rollover period.
1702 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1704 ts->ts_ftick = t - SCHED_TICK_TARG;
1705 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1706 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1707 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1708 ts->ts_ftick = t - SCHED_TICK_TARG;
1711 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1716 * Adjust the priority of a thread. Move it to the appropriate run-queue
1717 * if necessary. This is the back-end for several priority related
1721 sched_thread_priority(struct thread *td, u_char prio)
1723 struct td_sched *ts;
1727 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1728 "prio:%d", td->td_priority, "new prio:%d", prio,
1729 KTR_ATTR_LINKED, sched_tdname(curthread));
1730 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1731 if (td != curthread && prio < td->td_priority) {
1732 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1733 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1734 prio, KTR_ATTR_LINKED, sched_tdname(td));
1735 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1738 ts = td_get_sched(td);
1739 THREAD_LOCK_ASSERT(td, MA_OWNED);
1740 if (td->td_priority == prio)
1743 * If the priority has been elevated due to priority
1744 * propagation, we may have to move ourselves to a new
1745 * queue. This could be optimized to not re-add in some
1748 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1750 td->td_priority = prio;
1751 sched_add(td, SRQ_BORROWING);
1755 * If the thread is currently running we may have to adjust the lowpri
1756 * information so other cpus are aware of our current priority.
1758 if (TD_IS_RUNNING(td)) {
1759 tdq = TDQ_CPU(ts->ts_cpu);
1760 oldpri = td->td_priority;
1761 td->td_priority = prio;
1762 if (prio < tdq->tdq_lowpri)
1763 tdq->tdq_lowpri = prio;
1764 else if (tdq->tdq_lowpri == oldpri)
1765 tdq_setlowpri(tdq, td);
1768 td->td_priority = prio;
1772 * Update a thread's priority when it is lent another thread's
1776 sched_lend_prio(struct thread *td, u_char prio)
1779 td->td_flags |= TDF_BORROWING;
1780 sched_thread_priority(td, prio);
1784 * Restore a thread's priority when priority propagation is
1785 * over. The prio argument is the minimum priority the thread
1786 * needs to have to satisfy other possible priority lending
1787 * requests. If the thread's regular priority is less
1788 * important than prio, the thread will keep a priority boost
1792 sched_unlend_prio(struct thread *td, u_char prio)
1796 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1797 td->td_base_pri <= PRI_MAX_TIMESHARE)
1798 base_pri = td->td_user_pri;
1800 base_pri = td->td_base_pri;
1801 if (prio >= base_pri) {
1802 td->td_flags &= ~TDF_BORROWING;
1803 sched_thread_priority(td, base_pri);
1805 sched_lend_prio(td, prio);
1809 * Standard entry for setting the priority to an absolute value.
1812 sched_prio(struct thread *td, u_char prio)
1816 /* First, update the base priority. */
1817 td->td_base_pri = prio;
1820 * If the thread is borrowing another thread's priority, don't
1821 * ever lower the priority.
1823 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1826 /* Change the real priority. */
1827 oldprio = td->td_priority;
1828 sched_thread_priority(td, prio);
1831 * If the thread is on a turnstile, then let the turnstile update
1834 if (TD_ON_LOCK(td) && oldprio != prio)
1835 turnstile_adjust(td, oldprio);
1839 * Set the base user priority, does not effect current running priority.
1842 sched_user_prio(struct thread *td, u_char prio)
1845 td->td_base_user_pri = prio;
1846 if (td->td_lend_user_pri <= prio)
1848 td->td_user_pri = prio;
1852 sched_lend_user_prio(struct thread *td, u_char prio)
1855 THREAD_LOCK_ASSERT(td, MA_OWNED);
1856 td->td_lend_user_pri = prio;
1857 td->td_user_pri = min(prio, td->td_base_user_pri);
1858 if (td->td_priority > td->td_user_pri)
1859 sched_prio(td, td->td_user_pri);
1860 else if (td->td_priority != td->td_user_pri)
1861 td->td_flags |= TDF_NEEDRESCHED;
1865 * Like the above but first check if there is anything to do.
1868 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1871 if (td->td_lend_user_pri != prio)
1873 if (td->td_user_pri != min(prio, td->td_base_user_pri))
1875 if (td->td_priority >= td->td_user_pri)
1881 sched_lend_user_prio(td, prio);
1887 * This tdq is about to idle. Try to steal a thread from another CPU before
1888 * choosing the idle thread.
1891 tdq_trysteal(struct tdq *tdq)
1893 struct cpu_group *cg;
1898 if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1901 CPU_CLR(PCPU_GET(cpuid), &mask);
1902 /* We don't want to be preempted while we're iterating. */
1905 for (i = 1, cg = tdq->tdq_cg; ; ) {
1906 cpu = sched_highest(cg, mask, steal_thresh);
1908 * If a thread was added while interrupts were disabled don't
1911 if (tdq->tdq_load > 0) {
1918 if (cg == NULL || i > trysteal_limit) {
1924 steal = TDQ_CPU(cpu);
1926 * The data returned by sched_highest() is stale and
1927 * the chosen CPU no longer has an eligible thread.
1929 if (steal->tdq_load < steal_thresh ||
1930 steal->tdq_transferable == 0)
1932 tdq_lock_pair(tdq, steal);
1934 * If we get to this point, unconditonally exit the loop
1935 * to bound the time spent in the critcal section.
1937 * If a thread was added while interrupts were disabled don't
1940 if (tdq->tdq_load > 0) {
1945 * The data returned by sched_highest() is stale and
1946 * the chosen CPU no longer has an eligible thread.
1948 if (steal->tdq_load < steal_thresh ||
1949 steal->tdq_transferable == 0) {
1954 * If we fail to acquire one due to affinity restrictions,
1955 * bail out and let the idle thread to a more complete search
1956 * outside of a critical section.
1958 if (tdq_move(steal, tdq) == NULL) {
1970 * Handle migration from sched_switch(). This happens only for
1974 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1978 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
1979 "thread %s queued on absent CPU %d.", td->td_name,
1980 td_get_sched(td)->ts_cpu));
1981 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1983 tdq_load_rem(tdq, td);
1985 * Do the lock dance required to avoid LOR. We grab an extra
1986 * spinlock nesting to prevent preemption while we're
1987 * not holding either run-queue lock.
1990 thread_lock_block(td); /* This releases the lock on tdq. */
1993 * Acquire both run-queue locks before placing the thread on the new
1994 * run-queue to avoid deadlocks created by placing a thread with a
1995 * blocked lock on the run-queue of a remote processor. The deadlock
1996 * occurs when a third processor attempts to lock the two queues in
1997 * question while the target processor is spinning with its own
1998 * run-queue lock held while waiting for the blocked lock to clear.
2000 tdq_lock_pair(tdn, tdq);
2001 tdq_add(tdn, td, flags);
2002 tdq_notify(tdn, td);
2006 return (TDQ_LOCKPTR(tdn));
2010 * Variadic version of thread_lock_unblock() that does not assume td_lock
2014 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2016 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2021 * Switch threads. This function has to handle threads coming in while
2022 * blocked for some reason, running, or idle. It also must deal with
2023 * migrating a thread from one queue to another as running threads may
2024 * be assigned elsewhere via binding.
2027 sched_switch(struct thread *td, struct thread *newtd, int flags)
2030 struct td_sched *ts;
2033 int cpuid, preempted;
2035 THREAD_LOCK_ASSERT(td, MA_OWNED);
2036 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2038 cpuid = PCPU_GET(cpuid);
2039 tdq = TDQ_CPU(cpuid);
2040 ts = td_get_sched(td);
2042 sched_pctcpu_update(ts, 1);
2043 ts->ts_rltick = ticks;
2044 td->td_lastcpu = td->td_oncpu;
2045 td->td_oncpu = NOCPU;
2046 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2047 (flags & SW_PREEMPT) != 0;
2048 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2049 td->td_owepreempt = 0;
2050 if (!TD_IS_IDLETHREAD(td))
2051 tdq->tdq_switchcnt++;
2053 * The lock pointer in an idle thread should never change. Reset it
2054 * to CAN_RUN as well.
2056 if (TD_IS_IDLETHREAD(td)) {
2057 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2059 } else if (TD_IS_RUNNING(td)) {
2060 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2061 srqflag = preempted ?
2062 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2063 SRQ_OURSELF|SRQ_YIELDING;
2065 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2066 ts->ts_cpu = sched_pickcpu(td, 0);
2068 if (ts->ts_cpu == cpuid)
2069 tdq_runq_add(tdq, td, srqflag);
2071 KASSERT(THREAD_CAN_MIGRATE(td) ||
2072 (ts->ts_flags & TSF_BOUND) != 0,
2073 ("Thread %p shouldn't migrate", td));
2074 mtx = sched_switch_migrate(tdq, td, srqflag);
2077 /* This thread must be going to sleep. */
2079 mtx = thread_lock_block(td);
2080 tdq_load_rem(tdq, td);
2082 if (tdq->tdq_load == 0)
2087 #if (KTR_COMPILE & KTR_SCHED) != 0
2088 if (TD_IS_IDLETHREAD(td))
2089 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2090 "prio:%d", td->td_priority);
2092 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2093 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2094 "lockname:\"%s\"", td->td_lockname);
2098 * We enter here with the thread blocked and assigned to the
2099 * appropriate cpu run-queue or sleep-queue and with the current
2100 * thread-queue locked.
2102 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2103 newtd = choosethread();
2105 * Call the MD code to switch contexts if necessary.
2109 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2110 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2112 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2113 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2114 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2115 sched_pctcpu_update(td_get_sched(newtd), 0);
2117 #ifdef KDTRACE_HOOKS
2119 * If DTrace has set the active vtime enum to anything
2120 * other than INACTIVE (0), then it should have set the
2123 if (dtrace_vtime_active)
2124 (*dtrace_vtime_switch_func)(newtd);
2127 cpu_switch(td, newtd, mtx);
2129 * We may return from cpu_switch on a different cpu. However,
2130 * we always return with td_lock pointing to the current cpu's
2133 cpuid = PCPU_GET(cpuid);
2134 tdq = TDQ_CPU(cpuid);
2135 lock_profile_obtain_lock_success(
2136 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2138 SDT_PROBE0(sched, , , on__cpu);
2140 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2141 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2144 thread_unblock_switch(td, mtx);
2145 SDT_PROBE0(sched, , , remain__cpu);
2148 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2149 "prio:%d", td->td_priority);
2152 * Assert that all went well and return.
2154 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2155 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2156 td->td_oncpu = cpuid;
2160 * Adjust thread priorities as a result of a nice request.
2163 sched_nice(struct proc *p, int nice)
2167 PROC_LOCK_ASSERT(p, MA_OWNED);
2170 FOREACH_THREAD_IN_PROC(p, td) {
2173 sched_prio(td, td->td_base_user_pri);
2179 * Record the sleep time for the interactivity scorer.
2182 sched_sleep(struct thread *td, int prio)
2185 THREAD_LOCK_ASSERT(td, MA_OWNED);
2187 td->td_slptick = ticks;
2188 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2189 td->td_flags |= TDF_CANSWAP;
2190 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2192 if (static_boost == 1 && prio)
2193 sched_prio(td, prio);
2194 else if (static_boost && td->td_priority > static_boost)
2195 sched_prio(td, static_boost);
2199 * Schedule a thread to resume execution and record how long it voluntarily
2200 * slept. We also update the pctcpu, interactivity, and priority.
2203 sched_wakeup(struct thread *td)
2205 struct td_sched *ts;
2208 THREAD_LOCK_ASSERT(td, MA_OWNED);
2209 ts = td_get_sched(td);
2210 td->td_flags &= ~TDF_CANSWAP;
2212 * If we slept for more than a tick update our interactivity and
2215 slptick = td->td_slptick;
2217 if (slptick && slptick != ticks) {
2218 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2219 sched_interact_update(td);
2220 sched_pctcpu_update(ts, 0);
2223 * Reset the slice value since we slept and advanced the round-robin.
2226 sched_add(td, SRQ_BORING);
2230 * Penalize the parent for creating a new child and initialize the child's
2234 sched_fork(struct thread *td, struct thread *child)
2236 THREAD_LOCK_ASSERT(td, MA_OWNED);
2237 sched_pctcpu_update(td_get_sched(td), 1);
2238 sched_fork_thread(td, child);
2240 * Penalize the parent and child for forking.
2242 sched_interact_fork(child);
2243 sched_priority(child);
2244 td_get_sched(td)->ts_runtime += tickincr;
2245 sched_interact_update(td);
2250 * Fork a new thread, may be within the same process.
2253 sched_fork_thread(struct thread *td, struct thread *child)
2255 struct td_sched *ts;
2256 struct td_sched *ts2;
2260 THREAD_LOCK_ASSERT(td, MA_OWNED);
2264 ts = td_get_sched(td);
2265 ts2 = td_get_sched(child);
2266 child->td_oncpu = NOCPU;
2267 child->td_lastcpu = NOCPU;
2268 child->td_lock = TDQ_LOCKPTR(tdq);
2269 child->td_cpuset = cpuset_ref(td->td_cpuset);
2270 child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2271 ts2->ts_cpu = ts->ts_cpu;
2274 * Grab our parents cpu estimation information.
2276 ts2->ts_ticks = ts->ts_ticks;
2277 ts2->ts_ltick = ts->ts_ltick;
2278 ts2->ts_ftick = ts->ts_ftick;
2280 * Do not inherit any borrowed priority from the parent.
2282 child->td_priority = child->td_base_pri;
2284 * And update interactivity score.
2286 ts2->ts_slptime = ts->ts_slptime;
2287 ts2->ts_runtime = ts->ts_runtime;
2288 /* Attempt to quickly learn interactivity. */
2289 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2291 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2296 * Adjust the priority class of a thread.
2299 sched_class(struct thread *td, int class)
2302 THREAD_LOCK_ASSERT(td, MA_OWNED);
2303 if (td->td_pri_class == class)
2305 td->td_pri_class = class;
2309 * Return some of the child's priority and interactivity to the parent.
2312 sched_exit(struct proc *p, struct thread *child)
2316 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2317 "prio:%d", child->td_priority);
2318 PROC_LOCK_ASSERT(p, MA_OWNED);
2319 td = FIRST_THREAD_IN_PROC(p);
2320 sched_exit_thread(td, child);
2324 * Penalize another thread for the time spent on this one. This helps to
2325 * worsen the priority and interactivity of processes which schedule batch
2326 * jobs such as make. This has little effect on the make process itself but
2327 * causes new processes spawned by it to receive worse scores immediately.
2330 sched_exit_thread(struct thread *td, struct thread *child)
2333 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2334 "prio:%d", child->td_priority);
2336 * Give the child's runtime to the parent without returning the
2337 * sleep time as a penalty to the parent. This causes shells that
2338 * launch expensive things to mark their children as expensive.
2341 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2342 sched_interact_update(td);
2348 sched_preempt(struct thread *td)
2352 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2356 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2357 tdq->tdq_ipipending = 0;
2358 if (td->td_priority > tdq->tdq_lowpri) {
2361 flags = SW_INVOL | SW_PREEMPT;
2362 if (td->td_critnest > 1)
2363 td->td_owepreempt = 1;
2364 else if (TD_IS_IDLETHREAD(td))
2365 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2367 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2373 * Fix priorities on return to user-space. Priorities may be elevated due
2374 * to static priorities in msleep() or similar.
2377 sched_userret_slowpath(struct thread *td)
2381 td->td_priority = td->td_user_pri;
2382 td->td_base_pri = td->td_user_pri;
2383 tdq_setlowpri(TDQ_SELF(), td);
2388 * Handle a stathz tick. This is really only relevant for timeshare
2392 sched_clock(struct thread *td)
2395 struct td_sched *ts;
2397 THREAD_LOCK_ASSERT(td, MA_OWNED);
2401 * We run the long term load balancer infrequently on the first cpu.
2403 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2404 if (balance_ticks && --balance_ticks == 0)
2409 * Save the old switch count so we have a record of the last ticks
2410 * activity. Initialize the new switch count based on our load.
2411 * If there is some activity seed it to reflect that.
2413 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2414 tdq->tdq_switchcnt = tdq->tdq_load;
2416 * Advance the insert index once for each tick to ensure that all
2417 * threads get a chance to run.
2419 if (tdq->tdq_idx == tdq->tdq_ridx) {
2420 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2421 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2422 tdq->tdq_ridx = tdq->tdq_idx;
2424 ts = td_get_sched(td);
2425 sched_pctcpu_update(ts, 1);
2426 if (td->td_pri_class & PRI_FIFO_BIT)
2428 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2430 * We used a tick; charge it to the thread so
2431 * that we can compute our interactivity.
2433 td_get_sched(td)->ts_runtime += tickincr;
2434 sched_interact_update(td);
2439 * Force a context switch if the current thread has used up a full
2440 * time slice (default is 100ms).
2442 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2444 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2449 sched_estcpu(struct thread *td __unused)
2456 * Return whether the current CPU has runnable tasks. Used for in-kernel
2457 * cooperative idle threads.
2460 sched_runnable(void)
2468 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2469 if (tdq->tdq_load > 0)
2472 if (tdq->tdq_load - 1 > 0)
2480 * Choose the highest priority thread to run. The thread is removed from
2481 * the run-queue while running however the load remains. For SMP we set
2482 * the tdq in the global idle bitmask if it idles here.
2491 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2492 td = tdq_choose(tdq);
2494 tdq_runq_rem(tdq, td);
2495 tdq->tdq_lowpri = td->td_priority;
2498 tdq->tdq_lowpri = PRI_MAX_IDLE;
2499 return (PCPU_GET(idlethread));
2503 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2504 * we always request it once we exit a critical section.
2507 sched_setpreempt(struct thread *td)
2513 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2516 pri = td->td_priority;
2517 cpri = ctd->td_priority;
2519 ctd->td_flags |= TDF_NEEDRESCHED;
2520 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2522 if (!sched_shouldpreempt(pri, cpri, 0))
2524 ctd->td_owepreempt = 1;
2528 * Add a thread to a thread queue. Select the appropriate runq and add the
2529 * thread to it. This is the internal function called when the tdq is
2533 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2536 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2537 KASSERT((td->td_inhibitors == 0),
2538 ("sched_add: trying to run inhibited thread"));
2539 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2540 ("sched_add: bad thread state"));
2541 KASSERT(td->td_flags & TDF_INMEM,
2542 ("sched_add: thread swapped out"));
2544 if (td->td_priority < tdq->tdq_lowpri)
2545 tdq->tdq_lowpri = td->td_priority;
2546 tdq_runq_add(tdq, td, flags);
2547 tdq_load_add(tdq, td);
2551 * Select the target thread queue and add a thread to it. Request
2552 * preemption or IPI a remote processor if required.
2555 sched_add(struct thread *td, int flags)
2562 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2563 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2564 sched_tdname(curthread));
2565 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2566 KTR_ATTR_LINKED, sched_tdname(td));
2567 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2568 flags & SRQ_PREEMPTED);
2569 THREAD_LOCK_ASSERT(td, MA_OWNED);
2571 * Recalculate the priority before we select the target cpu or
2574 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2578 * Pick the destination cpu and if it isn't ours transfer to the
2581 cpu = sched_pickcpu(td, flags);
2582 tdq = sched_setcpu(td, cpu, flags);
2583 tdq_add(tdq, td, flags);
2584 if (cpu != PCPU_GET(cpuid)) {
2585 tdq_notify(tdq, td);
2592 * Now that the thread is moving to the run-queue, set the lock
2593 * to the scheduler's lock.
2595 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2596 tdq_add(tdq, td, flags);
2598 if (!(flags & SRQ_YIELDING))
2599 sched_setpreempt(td);
2603 * Remove a thread from a run-queue without running it. This is used
2604 * when we're stealing a thread from a remote queue. Otherwise all threads
2605 * exit by calling sched_exit_thread() and sched_throw() themselves.
2608 sched_rem(struct thread *td)
2612 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2613 "prio:%d", td->td_priority);
2614 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2615 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2616 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2617 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2618 KASSERT(TD_ON_RUNQ(td),
2619 ("sched_rem: thread not on run queue"));
2620 tdq_runq_rem(tdq, td);
2621 tdq_load_rem(tdq, td);
2623 if (td->td_priority == tdq->tdq_lowpri)
2624 tdq_setlowpri(tdq, NULL);
2628 * Fetch cpu utilization information. Updates on demand.
2631 sched_pctcpu(struct thread *td)
2634 struct td_sched *ts;
2637 ts = td_get_sched(td);
2639 THREAD_LOCK_ASSERT(td, MA_OWNED);
2640 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2644 /* How many rtick per second ? */
2645 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2646 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2653 * Enforce affinity settings for a thread. Called after adjustments to
2657 sched_affinity(struct thread *td)
2660 struct td_sched *ts;
2662 THREAD_LOCK_ASSERT(td, MA_OWNED);
2663 ts = td_get_sched(td);
2664 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2666 if (TD_ON_RUNQ(td)) {
2668 sched_add(td, SRQ_BORING);
2671 if (!TD_IS_RUNNING(td))
2674 * Force a switch before returning to userspace. If the
2675 * target thread is not running locally send an ipi to force
2678 td->td_flags |= TDF_NEEDRESCHED;
2679 if (td != curthread)
2680 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2685 * Bind a thread to a target cpu.
2688 sched_bind(struct thread *td, int cpu)
2690 struct td_sched *ts;
2692 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2693 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2694 ts = td_get_sched(td);
2695 if (ts->ts_flags & TSF_BOUND)
2697 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2698 ts->ts_flags |= TSF_BOUND;
2700 if (PCPU_GET(cpuid) == cpu)
2703 /* When we return from mi_switch we'll be on the correct cpu. */
2704 mi_switch(SW_VOL, NULL);
2708 * Release a bound thread.
2711 sched_unbind(struct thread *td)
2713 struct td_sched *ts;
2715 THREAD_LOCK_ASSERT(td, MA_OWNED);
2716 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2717 ts = td_get_sched(td);
2718 if ((ts->ts_flags & TSF_BOUND) == 0)
2720 ts->ts_flags &= ~TSF_BOUND;
2725 sched_is_bound(struct thread *td)
2727 THREAD_LOCK_ASSERT(td, MA_OWNED);
2728 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2735 sched_relinquish(struct thread *td)
2738 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2743 * Return the total system load.
2754 total += TDQ_CPU(i)->tdq_sysload;
2757 return (TDQ_SELF()->tdq_sysload);
2762 sched_sizeof_proc(void)
2764 return (sizeof(struct proc));
2768 sched_sizeof_thread(void)
2770 return (sizeof(struct thread) + sizeof(struct td_sched));
2774 #define TDQ_IDLESPIN(tdq) \
2775 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2777 #define TDQ_IDLESPIN(tdq) 1
2781 * The actual idle process.
2784 sched_idletd(void *dummy)
2788 int oldswitchcnt, switchcnt;
2791 mtx_assert(&Giant, MA_NOTOWNED);
2794 THREAD_NO_SLEEPING();
2797 if (tdq->tdq_load) {
2799 mi_switch(SW_VOL | SWT_IDLE, NULL);
2802 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2804 if (always_steal || switchcnt != oldswitchcnt) {
2805 oldswitchcnt = switchcnt;
2806 if (tdq_idled(tdq) == 0)
2809 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2811 oldswitchcnt = switchcnt;
2814 * If we're switching very frequently, spin while checking
2815 * for load rather than entering a low power state that
2816 * may require an IPI. However, don't do any busy
2817 * loops while on SMT machines as this simply steals
2818 * cycles from cores doing useful work.
2820 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2821 for (i = 0; i < sched_idlespins; i++) {
2828 /* If there was context switch during spin, restart it. */
2829 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2830 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2833 /* Run main MD idle handler. */
2834 tdq->tdq_cpu_idle = 1;
2836 * Make sure that tdq_cpu_idle update is globally visible
2837 * before cpu_idle() read tdq_load. The order is important
2838 * to avoid race with tdq_notify.
2840 atomic_thread_fence_seq_cst();
2842 * Checking for again after the fence picks up assigned
2843 * threads often enough to make it worthwhile to do so in
2844 * order to avoid calling cpu_idle().
2846 if (tdq->tdq_load != 0) {
2847 tdq->tdq_cpu_idle = 0;
2850 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2851 tdq->tdq_cpu_idle = 0;
2854 * Account thread-less hardware interrupts and
2855 * other wakeup reasons equal to context switches.
2857 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2858 if (switchcnt != oldswitchcnt)
2860 tdq->tdq_switchcnt++;
2866 * A CPU is entering for the first time or a thread is exiting.
2869 sched_throw(struct thread *td)
2871 struct thread *newtd;
2876 /* Correct spinlock nesting and acquire the correct lock. */
2879 PCPU_SET(switchtime, cpu_ticks());
2880 PCPU_SET(switchticks, ticks);
2882 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2883 tdq_load_rem(tdq, td);
2884 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2885 td->td_lastcpu = td->td_oncpu;
2886 td->td_oncpu = NOCPU;
2888 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2889 newtd = choosethread();
2890 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2891 cpu_throw(td, newtd); /* doesn't return */
2895 * This is called from fork_exit(). Just acquire the correct locks and
2896 * let fork do the rest of the work.
2899 sched_fork_exit(struct thread *td)
2905 * Finish setting up thread glue so that it begins execution in a
2906 * non-nested critical section with the scheduler lock held.
2908 cpuid = PCPU_GET(cpuid);
2909 tdq = TDQ_CPU(cpuid);
2910 if (TD_IS_IDLETHREAD(td))
2911 td->td_lock = TDQ_LOCKPTR(tdq);
2912 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2913 td->td_oncpu = cpuid;
2914 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2915 lock_profile_obtain_lock_success(
2916 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2918 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2919 "prio:%d", td->td_priority);
2920 SDT_PROBE0(sched, , , on__cpu);
2924 * Create on first use to catch odd startup conditons.
2927 sched_tdname(struct thread *td)
2930 struct td_sched *ts;
2932 ts = td_get_sched(td);
2933 if (ts->ts_name[0] == '\0')
2934 snprintf(ts->ts_name, sizeof(ts->ts_name),
2935 "%s tid %d", td->td_name, td->td_tid);
2936 return (ts->ts_name);
2938 return (td->td_name);
2944 sched_clear_tdname(struct thread *td)
2946 struct td_sched *ts;
2948 ts = td_get_sched(td);
2949 ts->ts_name[0] = '\0';
2956 * Build the CPU topology dump string. Is recursively called to collect
2957 * the topology tree.
2960 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2963 char cpusetbuf[CPUSETBUFSIZ];
2966 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2967 "", 1 + indent / 2, cg->cg_level);
2968 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2969 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2971 for (i = 0; i < MAXCPU; i++) {
2972 if (CPU_ISSET(i, &cg->cg_mask)) {
2974 sbuf_printf(sb, ", ");
2977 sbuf_printf(sb, "%d", i);
2980 sbuf_printf(sb, "</cpu>\n");
2982 if (cg->cg_flags != 0) {
2983 sbuf_printf(sb, "%*s <flags>", indent, "");
2984 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2985 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2986 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2987 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2988 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2989 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2990 sbuf_printf(sb, "</flags>\n");
2993 if (cg->cg_children > 0) {
2994 sbuf_printf(sb, "%*s <children>\n", indent, "");
2995 for (i = 0; i < cg->cg_children; i++)
2996 sysctl_kern_sched_topology_spec_internal(sb,
2997 &cg->cg_child[i], indent+2);
2998 sbuf_printf(sb, "%*s </children>\n", indent, "");
3000 sbuf_printf(sb, "%*s</group>\n", indent, "");
3005 * Sysctl handler for retrieving topology dump. It's a wrapper for
3006 * the recursive sysctl_kern_smp_topology_spec_internal().
3009 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3014 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3016 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3020 sbuf_printf(topo, "<groups>\n");
3021 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3022 sbuf_printf(topo, "</groups>\n");
3025 err = sbuf_finish(topo);
3034 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3036 int error, new_val, period;
3038 period = 1000000 / realstathz;
3039 new_val = period * sched_slice;
3040 error = sysctl_handle_int(oidp, &new_val, 0, req);
3041 if (error != 0 || req->newptr == NULL)
3045 sched_slice = imax(1, (new_val + period / 2) / period);
3046 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3047 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3052 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3053 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3055 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3056 NULL, 0, sysctl_kern_quantum, "I",
3057 "Quantum for timeshare threads in microseconds");
3058 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3059 "Quantum for timeshare threads in stathz ticks");
3060 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3061 "Interactivity score threshold");
3062 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3064 "Maximal (lowest) priority for preemption");
3065 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3066 "Assign static kernel priorities to sleeping threads");
3067 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3068 "Number of times idle thread will spin waiting for new work");
3069 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3070 &sched_idlespinthresh, 0,
3071 "Threshold before we will permit idle thread spinning");
3073 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3074 "Number of hz ticks to keep thread affinity for");
3075 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3076 "Enables the long-term load balancer");
3077 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3078 &balance_interval, 0,
3079 "Average period in stathz ticks to run the long-term balancer");
3080 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3081 "Attempts to steal work from other cores before idling");
3082 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3083 "Minimum load on remote CPU before we'll steal");
3084 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3085 0, "Topological distance limit for stealing threads in sched_switch()");
3086 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3087 "Always run the stealer from the idle thread");
3088 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3089 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3090 "XML dump of detected CPU topology");
3093 /* ps compat. All cpu percentages from ULE are weighted. */
3094 static int ccpu = 0;
3095 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");