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 int tdq_id; /* cpuid. */
251 struct runq tdq_realtime; /* real-time run queue. */
252 struct runq tdq_timeshare; /* timeshare run queue. */
253 struct runq tdq_idle; /* Queue of IDLE threads. */
254 char tdq_name[TDQ_NAME_LEN];
256 char tdq_loadname[TDQ_LOADNAME_LEN];
260 /* Idle thread states and config. */
261 #define TDQ_RUNNING 1
265 struct cpu_group *cpu_top; /* CPU topology */
267 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
268 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
273 static int rebalance = 1;
274 static int balance_interval = 128; /* Default set in sched_initticks(). */
276 static int steal_idle = 1;
277 static int steal_thresh = 2;
278 static int always_steal = 0;
279 static int trysteal_limit = 2;
282 * One thread queue per processor.
284 static struct tdq *balance_tdq;
285 static int balance_ticks;
286 DPCPU_DEFINE_STATIC(struct tdq, tdq);
287 DPCPU_DEFINE_STATIC(uint32_t, randomval);
289 #define TDQ_SELF() ((struct tdq *)PCPU_GET(sched))
290 #define TDQ_CPU(x) (DPCPU_ID_PTR((x), tdq))
291 #define TDQ_ID(x) ((x)->tdq_id)
293 static struct tdq tdq_cpu;
295 #define TDQ_ID(x) (0)
296 #define TDQ_SELF() (&tdq_cpu)
297 #define TDQ_CPU(x) (&tdq_cpu)
300 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
301 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
302 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
303 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
304 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
306 static void sched_priority(struct thread *);
307 static void sched_thread_priority(struct thread *, u_char);
308 static int sched_interact_score(struct thread *);
309 static void sched_interact_update(struct thread *);
310 static void sched_interact_fork(struct thread *);
311 static void sched_pctcpu_update(struct td_sched *, int);
313 /* Operations on per processor queues */
314 static struct thread *tdq_choose(struct tdq *);
315 static void tdq_setup(struct tdq *, int i);
316 static void tdq_load_add(struct tdq *, struct thread *);
317 static void tdq_load_rem(struct tdq *, struct thread *);
318 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
319 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
320 static inline int sched_shouldpreempt(int, int, int);
321 void tdq_print(int cpu);
322 static void runq_print(struct runq *rq);
323 static void tdq_add(struct tdq *, struct thread *, int);
325 static struct thread *tdq_move(struct tdq *, struct tdq *);
326 static int tdq_idled(struct tdq *);
327 static void tdq_notify(struct tdq *, struct thread *);
328 static struct thread *tdq_steal(struct tdq *, int);
329 static struct thread *runq_steal(struct runq *, int);
330 static int sched_pickcpu(struct thread *, int);
331 static void sched_balance(void);
332 static int sched_balance_pair(struct tdq *, struct tdq *);
333 static inline struct tdq *sched_setcpu(struct thread *, int, int);
334 static inline void thread_unblock_switch(struct thread *, struct mtx *);
335 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
336 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
337 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
338 struct cpu_group *cg, int indent);
341 static void sched_setup(void *dummy);
342 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
344 static void sched_initticks(void *dummy);
345 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
348 SDT_PROVIDER_DEFINE(sched);
350 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
351 "struct proc *", "uint8_t");
352 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
353 "struct proc *", "void *");
354 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
355 "struct proc *", "void *", "int");
356 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
357 "struct proc *", "uint8_t", "struct thread *");
358 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
359 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
361 SDT_PROBE_DEFINE(sched, , , on__cpu);
362 SDT_PROBE_DEFINE(sched, , , remain__cpu);
363 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
367 * Print the threads waiting on a run-queue.
370 runq_print(struct runq *rq)
378 for (i = 0; i < RQB_LEN; i++) {
379 printf("\t\trunq bits %d 0x%zx\n",
380 i, rq->rq_status.rqb_bits[i]);
381 for (j = 0; j < RQB_BPW; j++)
382 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
383 pri = j + (i << RQB_L2BPW);
384 rqh = &rq->rq_queues[pri];
385 TAILQ_FOREACH(td, rqh, td_runq) {
386 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
387 td, td->td_name, td->td_priority,
388 td->td_rqindex, pri);
395 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
404 printf("tdq %d:\n", TDQ_ID(tdq));
405 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
406 printf("\tLock name: %s\n", tdq->tdq_name);
407 printf("\tload: %d\n", tdq->tdq_load);
408 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
409 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
410 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
411 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
412 printf("\tload transferable: %d\n", tdq->tdq_transferable);
413 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
414 printf("\trealtime runq:\n");
415 runq_print(&tdq->tdq_realtime);
416 printf("\ttimeshare runq:\n");
417 runq_print(&tdq->tdq_timeshare);
418 printf("\tidle runq:\n");
419 runq_print(&tdq->tdq_idle);
423 sched_shouldpreempt(int pri, int cpri, int remote)
426 * If the new priority is not better than the current priority there is
432 * Always preempt idle.
434 if (cpri >= PRI_MIN_IDLE)
437 * If preemption is disabled don't preempt others.
439 if (preempt_thresh == 0)
442 * Preempt if we exceed the threshold.
444 if (pri <= preempt_thresh)
447 * If we're interactive or better and there is non-interactive
448 * or worse running preempt only remote processors.
450 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
456 * Add a thread to the actual run-queue. Keeps transferable counts up to
457 * date with what is actually on the run-queue. Selects the correct
458 * queue position for timeshare threads.
461 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
466 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
467 THREAD_LOCK_ASSERT(td, MA_OWNED);
469 pri = td->td_priority;
470 ts = td_get_sched(td);
472 if (THREAD_CAN_MIGRATE(td)) {
473 tdq->tdq_transferable++;
474 ts->ts_flags |= TSF_XFERABLE;
476 if (pri < PRI_MIN_BATCH) {
477 ts->ts_runq = &tdq->tdq_realtime;
478 } else if (pri <= PRI_MAX_BATCH) {
479 ts->ts_runq = &tdq->tdq_timeshare;
480 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
481 ("Invalid priority %d on timeshare runq", pri));
483 * This queue contains only priorities between MIN and MAX
484 * realtime. Use the whole queue to represent these values.
486 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
487 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
488 pri = (pri + tdq->tdq_idx) % RQ_NQS;
490 * This effectively shortens the queue by one so we
491 * can have a one slot difference between idx and
492 * ridx while we wait for threads to drain.
494 if (tdq->tdq_ridx != tdq->tdq_idx &&
495 pri == tdq->tdq_ridx)
496 pri = (unsigned char)(pri - 1) % RQ_NQS;
499 runq_add_pri(ts->ts_runq, td, pri, flags);
502 ts->ts_runq = &tdq->tdq_idle;
503 runq_add(ts->ts_runq, td, flags);
507 * Remove a thread from a run-queue. This typically happens when a thread
508 * is selected to run. Running threads are not on the queue and the
509 * transferable count does not reflect them.
512 tdq_runq_rem(struct tdq *tdq, struct thread *td)
516 ts = td_get_sched(td);
517 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
518 KASSERT(ts->ts_runq != NULL,
519 ("tdq_runq_remove: thread %p null ts_runq", td));
520 if (ts->ts_flags & TSF_XFERABLE) {
521 tdq->tdq_transferable--;
522 ts->ts_flags &= ~TSF_XFERABLE;
524 if (ts->ts_runq == &tdq->tdq_timeshare) {
525 if (tdq->tdq_idx != tdq->tdq_ridx)
526 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
528 runq_remove_idx(ts->ts_runq, td, NULL);
530 runq_remove(ts->ts_runq, td);
534 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
535 * for this thread to the referenced thread queue.
538 tdq_load_add(struct tdq *tdq, struct thread *td)
541 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
542 THREAD_LOCK_ASSERT(td, MA_OWNED);
545 if ((td->td_flags & TDF_NOLOAD) == 0)
547 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
548 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
552 * Remove the load from a thread that is transitioning to a sleep state or
556 tdq_load_rem(struct tdq *tdq, struct thread *td)
559 THREAD_LOCK_ASSERT(td, MA_OWNED);
560 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
561 KASSERT(tdq->tdq_load != 0,
562 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
565 if ((td->td_flags & TDF_NOLOAD) == 0)
567 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
568 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
572 * Bound timeshare latency by decreasing slice size as load increases. We
573 * consider the maximum latency as the sum of the threads waiting to run
574 * aside from curthread and target no more than sched_slice latency but
575 * no less than sched_slice_min runtime.
578 tdq_slice(struct tdq *tdq)
583 * It is safe to use sys_load here because this is called from
584 * contexts where timeshare threads are running and so there
585 * cannot be higher priority load in the system.
587 load = tdq->tdq_sysload - 1;
588 if (load >= SCHED_SLICE_MIN_DIVISOR)
589 return (sched_slice_min);
591 return (sched_slice);
592 return (sched_slice / load);
596 * Set lowpri to its exact value by searching the run-queue and
597 * evaluating curthread. curthread may be passed as an optimization.
600 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
604 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
606 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
607 td = tdq_choose(tdq);
608 if (td == NULL || td->td_priority > ctd->td_priority)
609 tdq->tdq_lowpri = ctd->td_priority;
611 tdq->tdq_lowpri = td->td_priority;
616 * We need some randomness. Implement a classic Linear Congruential
617 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
618 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
619 * of the random state (in the low bits of our answer) to keep
620 * the maximum randomness.
627 rndptr = DPCPU_PTR(randomval);
628 *rndptr = *rndptr * 69069 + 5;
630 return (*rndptr >> 16);
636 int cs_pri; /* Min priority for low. */
637 int cs_limit; /* Max load for low, min load for high. */
642 #define CPU_SEARCH_LOWEST 0x1
643 #define CPU_SEARCH_HIGHEST 0x2
644 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
646 #define CPUSET_FOREACH(cpu, mask) \
647 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
648 if (CPU_ISSET(cpu, &mask))
650 static __always_inline int cpu_search(const struct cpu_group *cg,
651 struct cpu_search *low, struct cpu_search *high, const int match);
652 int __noinline cpu_search_lowest(const struct cpu_group *cg,
653 struct cpu_search *low);
654 int __noinline cpu_search_highest(const struct cpu_group *cg,
655 struct cpu_search *high);
656 int __noinline cpu_search_both(const struct cpu_group *cg,
657 struct cpu_search *low, struct cpu_search *high);
660 * Search the tree of cpu_groups for the lowest or highest loaded cpu
661 * according to the match argument. This routine actually compares the
662 * load on all paths through the tree and finds the least loaded cpu on
663 * the least loaded path, which may differ from the least loaded cpu in
664 * the system. This balances work among caches and buses.
666 * This inline is instantiated in three forms below using constants for the
667 * match argument. It is reduced to the minimum set for each case. It is
668 * also recursive to the depth of the tree.
670 static __always_inline int
671 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
672 struct cpu_search *high, const int match)
674 struct cpu_search lgroup;
675 struct cpu_search hgroup;
677 struct cpu_group *child;
679 int cpu, i, hload, lload, load, total, rnd;
682 cpumask = cg->cg_mask;
683 if (match & CPU_SEARCH_LOWEST) {
687 if (match & CPU_SEARCH_HIGHEST) {
692 /* Iterate through the child CPU groups and then remaining CPUs. */
693 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
695 #ifdef HAVE_INLINE_FFSL
696 cpu = CPU_FFS(&cpumask) - 1;
698 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
705 child = &cg->cg_child[i - 1];
707 if (match & CPU_SEARCH_LOWEST)
709 if (match & CPU_SEARCH_HIGHEST)
711 if (child) { /* Handle child CPU group. */
712 CPU_NAND(&cpumask, &child->cg_mask);
714 case CPU_SEARCH_LOWEST:
715 load = cpu_search_lowest(child, &lgroup);
717 case CPU_SEARCH_HIGHEST:
718 load = cpu_search_highest(child, &hgroup);
720 case CPU_SEARCH_BOTH:
721 load = cpu_search_both(child, &lgroup, &hgroup);
724 } else { /* Handle child CPU. */
725 CPU_CLR(cpu, &cpumask);
727 load = tdq->tdq_load * 256;
728 rnd = sched_random() % 32;
729 if (match & CPU_SEARCH_LOWEST) {
730 if (cpu == low->cs_prefer)
732 /* If that CPU is allowed and get data. */
733 if (tdq->tdq_lowpri > lgroup.cs_pri &&
734 tdq->tdq_load <= lgroup.cs_limit &&
735 CPU_ISSET(cpu, &lgroup.cs_mask)) {
737 lgroup.cs_load = load - rnd;
740 if (match & CPU_SEARCH_HIGHEST)
741 if (tdq->tdq_load >= hgroup.cs_limit &&
742 tdq->tdq_transferable &&
743 CPU_ISSET(cpu, &hgroup.cs_mask)) {
745 hgroup.cs_load = load - rnd;
750 /* We have info about child item. Compare it. */
751 if (match & CPU_SEARCH_LOWEST) {
752 if (lgroup.cs_cpu >= 0 &&
754 (load == lload && lgroup.cs_load < low->cs_load))) {
756 low->cs_cpu = lgroup.cs_cpu;
757 low->cs_load = lgroup.cs_load;
760 if (match & CPU_SEARCH_HIGHEST)
761 if (hgroup.cs_cpu >= 0 &&
763 (load == hload && hgroup.cs_load > high->cs_load))) {
765 high->cs_cpu = hgroup.cs_cpu;
766 high->cs_load = hgroup.cs_load;
770 if (i == 0 && CPU_EMPTY(&cpumask))
773 #ifndef HAVE_INLINE_FFSL
782 * cpu_search instantiations must pass constants to maintain the inline
786 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
788 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
792 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
794 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
798 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
799 struct cpu_search *high)
801 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
805 * Find the cpu with the least load via the least loaded path that has a
806 * lowpri greater than pri pri. A pri of -1 indicates any priority is
810 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
813 struct cpu_search low;
816 low.cs_prefer = prefer;
819 low.cs_limit = maxload;
820 cpu_search_lowest(cg, &low);
825 * Find the cpu with the highest load via the highest loaded path.
828 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
830 struct cpu_search high;
834 high.cs_limit = minload;
835 cpu_search_highest(cg, &high);
840 sched_balance_group(struct cpu_group *cg)
843 cpuset_t hmask, lmask;
844 int high, low, anylow;
848 high = sched_highest(cg, hmask, 2);
849 /* Stop if there is no more CPU with transferrable threads. */
852 CPU_CLR(high, &hmask);
853 CPU_COPY(&hmask, &lmask);
854 /* Stop if there is no more CPU left for low. */
855 if (CPU_EMPTY(&lmask))
860 low = sched_lowest(cg, lmask, -1, tdq->tdq_load - 1, high);
861 /* Stop if we looked well and found no less loaded CPU. */
862 if (anylow && low == -1)
864 /* Go to next high if we found no less loaded CPU. */
867 /* Transfer thread from high to low. */
868 if (sched_balance_pair(tdq, TDQ_CPU(low))) {
869 /* CPU that got thread can no longer be a donor. */
870 CPU_CLR(low, &hmask);
873 * If failed, then there is no threads on high
874 * that can run on this low. Drop low from low
875 * mask and look for different one.
877 CPU_CLR(low, &lmask);
889 balance_ticks = max(balance_interval / 2, 1) +
890 (sched_random() % balance_interval);
893 sched_balance_group(cpu_top);
898 * Lock two thread queues using their address to maintain lock order.
901 tdq_lock_pair(struct tdq *one, struct tdq *two)
905 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
908 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
913 * Unlock two thread queues. Order is not important here.
916 tdq_unlock_pair(struct tdq *one, struct tdq *two)
923 * Transfer load between two imbalanced thread queues.
926 sched_balance_pair(struct tdq *high, struct tdq *low)
931 tdq_lock_pair(high, low);
934 * Transfer a thread from high to low.
936 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
937 (td = tdq_move(high, low)) != NULL) {
939 * In case the target isn't the current cpu notify it of the
940 * new load, possibly sending an IPI to force it to reschedule.
943 if (cpu != PCPU_GET(cpuid))
946 tdq_unlock_pair(high, low);
951 * Move a thread from one thread queue to another.
953 static struct thread *
954 tdq_move(struct tdq *from, struct tdq *to)
961 TDQ_LOCK_ASSERT(from, MA_OWNED);
962 TDQ_LOCK_ASSERT(to, MA_OWNED);
966 td = tdq_steal(tdq, cpu);
969 ts = td_get_sched(td);
971 * Although the run queue is locked the thread may be blocked. Lock
972 * it to clear this and acquire the run-queue lock.
975 /* Drop recursive lock on from acquired via thread_lock(). */
979 td->td_lock = TDQ_LOCKPTR(to);
980 tdq_add(to, td, SRQ_YIELDING);
985 * This tdq has idled. Try to steal a thread from another cpu and switch
989 tdq_idled(struct tdq *tdq)
991 struct cpu_group *cg;
996 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
999 CPU_CLR(PCPU_GET(cpuid), &mask);
1001 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
1002 for (cg = tdq->tdq_cg; ; ) {
1003 cpu = sched_highest(cg, mask, steal_thresh);
1005 * We were assigned a thread but not preempted. Returning
1006 * 0 here will cause our caller to switch to it.
1016 steal = TDQ_CPU(cpu);
1018 * The data returned by sched_highest() is stale and
1019 * the chosen CPU no longer has an eligible thread.
1021 * Testing this ahead of tdq_lock_pair() only catches
1022 * this situation about 20% of the time on an 8 core
1023 * 16 thread Ryzen 7, but it still helps performance.
1025 if (steal->tdq_load < steal_thresh ||
1026 steal->tdq_transferable == 0)
1028 tdq_lock_pair(tdq, steal);
1030 * We were assigned a thread while waiting for the locks.
1031 * Switch to it now instead of stealing a thread.
1036 * The data returned by sched_highest() is stale and
1037 * the chosen CPU no longer has an eligible thread, or
1038 * we were preempted and the CPU loading info may be out
1039 * of date. The latter is rare. In either case restart
1042 if (steal->tdq_load < steal_thresh ||
1043 steal->tdq_transferable == 0 ||
1044 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1045 tdq_unlock_pair(tdq, steal);
1049 * Steal the thread and switch to it.
1051 if (tdq_move(steal, tdq) != NULL)
1054 * We failed to acquire a thread even though it looked
1055 * like one was available. This could be due to affinity
1056 * restrictions or for other reasons. Loop again after
1057 * removing this CPU from the set. The restart logic
1058 * above does not restore this CPU to the set due to the
1059 * likelyhood of failing here again.
1061 CPU_CLR(cpu, &mask);
1062 tdq_unlock_pair(tdq, steal);
1065 mi_switch(SW_VOL | SWT_IDLE, NULL);
1066 thread_unlock(curthread);
1071 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1074 tdq_notify(struct tdq *tdq, struct thread *td)
1080 if (tdq->tdq_ipipending)
1082 cpu = td_get_sched(td)->ts_cpu;
1083 pri = td->td_priority;
1084 ctd = pcpu_find(cpu)->pc_curthread;
1085 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1089 * Make sure that our caller's earlier update to tdq_load is
1090 * globally visible before we read tdq_cpu_idle. Idle thread
1091 * accesses both of them without locks, and the order is important.
1093 atomic_thread_fence_seq_cst();
1095 if (TD_IS_IDLETHREAD(ctd)) {
1097 * If the MD code has an idle wakeup routine try that before
1098 * falling back to IPI.
1100 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1103 tdq->tdq_ipipending = 1;
1104 ipi_cpu(cpu, IPI_PREEMPT);
1108 * Steals load from a timeshare queue. Honors the rotating queue head
1111 static struct thread *
1112 runq_steal_from(struct runq *rq, int cpu, u_char start)
1116 struct thread *td, *first;
1120 rqb = &rq->rq_status;
1121 bit = start & (RQB_BPW -1);
1124 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1125 if (rqb->rqb_bits[i] == 0)
1128 bit = RQB_FFS(rqb->rqb_bits[i]);
1129 for (; bit < RQB_BPW; bit++) {
1130 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1132 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1133 TAILQ_FOREACH(td, rqh, td_runq) {
1134 if (first && THREAD_CAN_MIGRATE(td) &&
1135 THREAD_CAN_SCHED(td, cpu))
1146 if (first && THREAD_CAN_MIGRATE(first) &&
1147 THREAD_CAN_SCHED(first, cpu))
1153 * Steals load from a standard linear queue.
1155 static struct thread *
1156 runq_steal(struct runq *rq, int cpu)
1164 rqb = &rq->rq_status;
1165 for (word = 0; word < RQB_LEN; word++) {
1166 if (rqb->rqb_bits[word] == 0)
1168 for (bit = 0; bit < RQB_BPW; bit++) {
1169 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1171 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1172 TAILQ_FOREACH(td, rqh, td_runq)
1173 if (THREAD_CAN_MIGRATE(td) &&
1174 THREAD_CAN_SCHED(td, cpu))
1182 * Attempt to steal a thread in priority order from a thread queue.
1184 static struct thread *
1185 tdq_steal(struct tdq *tdq, int cpu)
1189 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1190 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1192 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1193 cpu, tdq->tdq_ridx)) != NULL)
1195 return (runq_steal(&tdq->tdq_idle, cpu));
1199 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1200 * current lock and returns with the assigned queue locked.
1202 static inline struct tdq *
1203 sched_setcpu(struct thread *td, int cpu, int flags)
1208 THREAD_LOCK_ASSERT(td, MA_OWNED);
1210 td_get_sched(td)->ts_cpu = cpu;
1212 * If the lock matches just return the queue.
1214 if (td->td_lock == TDQ_LOCKPTR(tdq))
1218 * If the thread isn't running its lockptr is a
1219 * turnstile or a sleepqueue. We can just lock_set without
1222 if (TD_CAN_RUN(td)) {
1224 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1229 * The hard case, migration, we need to block the thread first to
1230 * prevent order reversals with other cpus locks.
1233 thread_lock_block(td);
1235 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1240 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1241 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1242 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1243 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1244 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1245 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1248 sched_pickcpu(struct thread *td, int flags)
1250 struct cpu_group *cg, *ccg;
1251 struct td_sched *ts;
1256 self = PCPU_GET(cpuid);
1257 ts = td_get_sched(td);
1258 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1259 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1260 if (smp_started == 0)
1263 * Don't migrate a running thread from sched_switch().
1265 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1266 return (ts->ts_cpu);
1268 * Prefer to run interrupt threads on the processors that generate
1271 pri = td->td_priority;
1272 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1273 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1274 SCHED_STAT_INC(pickcpu_intrbind);
1276 if (TDQ_SELF()->tdq_lowpri > pri) {
1277 SCHED_STAT_INC(pickcpu_affinity);
1278 return (ts->ts_cpu);
1282 * If the thread can run on the last cpu and the affinity has not
1283 * expired and it is idle, run it there.
1285 tdq = TDQ_CPU(ts->ts_cpu);
1287 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1288 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1289 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1290 if (cg->cg_flags & CG_FLAG_THREAD) {
1291 CPUSET_FOREACH(cpu, cg->cg_mask) {
1292 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1297 if (cpu > mp_maxid) {
1298 SCHED_STAT_INC(pickcpu_idle_affinity);
1299 return (ts->ts_cpu);
1303 * Search for the last level cache CPU group in the tree.
1304 * Skip caches with expired affinity time and SMT groups.
1305 * Affinity to higher level caches will be handled less aggressively.
1307 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1308 if (cg->cg_flags & CG_FLAG_THREAD)
1310 if (!SCHED_AFFINITY(ts, cg->cg_level))
1317 /* Search the group for the less loaded idle CPU we can run now. */
1318 mask = td->td_cpuset->cs_mask;
1319 if (cg != NULL && cg != cpu_top &&
1320 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1321 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1322 INT_MAX, ts->ts_cpu);
1323 /* Search globally for the less loaded CPU we can run now. */
1325 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1326 /* Search globally for the less loaded CPU. */
1328 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1329 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1330 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1332 * Compare the lowest loaded cpu to current cpu.
1335 if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1336 tdq->tdq_lowpri < PRI_MIN_IDLE &&
1337 TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1338 SCHED_STAT_INC(pickcpu_local);
1341 SCHED_STAT_INC(pickcpu_lowest);
1342 if (cpu != ts->ts_cpu)
1343 SCHED_STAT_INC(pickcpu_migration);
1349 * Pick the highest priority task we have and return it.
1351 static struct thread *
1352 tdq_choose(struct tdq *tdq)
1356 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1357 td = runq_choose(&tdq->tdq_realtime);
1360 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1362 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1363 ("tdq_choose: Invalid priority on timeshare queue %d",
1367 td = runq_choose(&tdq->tdq_idle);
1369 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1370 ("tdq_choose: Invalid priority on idle queue %d",
1379 * Initialize a thread queue.
1382 tdq_setup(struct tdq *tdq, int id)
1386 printf("ULE: setup cpu %d\n", id);
1387 runq_init(&tdq->tdq_realtime);
1388 runq_init(&tdq->tdq_timeshare);
1389 runq_init(&tdq->tdq_idle);
1391 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1392 "sched lock %d", (int)TDQ_ID(tdq));
1393 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1394 MTX_SPIN | MTX_RECURSE);
1396 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1397 "CPU %d load", (int)TDQ_ID(tdq));
1403 sched_setup_smp(void)
1408 cpu_top = smp_topo();
1410 tdq = DPCPU_ID_PTR(i, tdq);
1412 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1413 if (tdq->tdq_cg == NULL)
1414 panic("Can't find cpu group for %d\n", i);
1416 PCPU_SET(sched, DPCPU_PTR(tdq));
1417 balance_tdq = TDQ_SELF();
1422 * Setup the thread queues and initialize the topology based on MD
1426 sched_setup(void *dummy)
1433 tdq_setup(TDQ_SELF(), 0);
1437 /* Add thread0's load since it's running. */
1439 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1440 tdq_load_add(tdq, &thread0);
1441 tdq->tdq_lowpri = thread0.td_priority;
1446 * This routine determines time constants after stathz and hz are setup.
1450 sched_initticks(void *dummy)
1454 realstathz = stathz ? stathz : hz;
1455 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1456 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1457 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1461 * tickincr is shifted out by 10 to avoid rounding errors due to
1462 * hz not being evenly divisible by stathz on all platforms.
1464 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1466 * This does not work for values of stathz that are more than
1467 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1474 * Set the default balance interval now that we know
1475 * what realstathz is.
1477 balance_interval = realstathz;
1478 balance_ticks = balance_interval;
1479 affinity = SCHED_AFFINITY_DEFAULT;
1481 if (sched_idlespinthresh < 0)
1482 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1487 * This is the core of the interactivity algorithm. Determines a score based
1488 * on past behavior. It is the ratio of sleep time to run time scaled to
1489 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1490 * differs from the cpu usage because it does not account for time spent
1491 * waiting on a run-queue. Would be prettier if we had floating point.
1493 * When a thread's sleep time is greater than its run time the
1497 * interactivity score = ---------------------
1498 * sleep time / run time
1501 * When a thread's run time is greater than its sleep time the
1505 * interactivity score = --------------------- + scaling factor
1506 * run time / sleep time
1509 sched_interact_score(struct thread *td)
1511 struct td_sched *ts;
1514 ts = td_get_sched(td);
1516 * The score is only needed if this is likely to be an interactive
1517 * task. Don't go through the expense of computing it if there's
1520 if (sched_interact <= SCHED_INTERACT_HALF &&
1521 ts->ts_runtime >= ts->ts_slptime)
1522 return (SCHED_INTERACT_HALF);
1524 if (ts->ts_runtime > ts->ts_slptime) {
1525 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1526 return (SCHED_INTERACT_HALF +
1527 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1529 if (ts->ts_slptime > ts->ts_runtime) {
1530 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1531 return (ts->ts_runtime / div);
1533 /* runtime == slptime */
1535 return (SCHED_INTERACT_HALF);
1538 * This can happen if slptime and runtime are 0.
1545 * Scale the scheduling priority according to the "interactivity" of this
1549 sched_priority(struct thread *td)
1554 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1557 * If the score is interactive we place the thread in the realtime
1558 * queue with a priority that is less than kernel and interrupt
1559 * priorities. These threads are not subject to nice restrictions.
1561 * Scores greater than this are placed on the normal timeshare queue
1562 * where the priority is partially decided by the most recent cpu
1563 * utilization and the rest is decided by nice value.
1565 * The nice value of the process has a linear effect on the calculated
1566 * score. Negative nice values make it easier for a thread to be
1567 * considered interactive.
1569 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1570 if (score < sched_interact) {
1571 pri = PRI_MIN_INTERACT;
1572 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1573 sched_interact) * score;
1574 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1575 ("sched_priority: invalid interactive priority %d score %d",
1578 pri = SCHED_PRI_MIN;
1579 if (td_get_sched(td)->ts_ticks)
1580 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1581 SCHED_PRI_RANGE - 1);
1582 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1583 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1584 ("sched_priority: invalid priority %d: nice %d, "
1585 "ticks %d ftick %d ltick %d tick pri %d",
1586 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1587 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1588 SCHED_PRI_TICKS(td_get_sched(td))));
1590 sched_user_prio(td, pri);
1596 * This routine enforces a maximum limit on the amount of scheduling history
1597 * kept. It is called after either the slptime or runtime is adjusted. This
1598 * function is ugly due to integer math.
1601 sched_interact_update(struct thread *td)
1603 struct td_sched *ts;
1606 ts = td_get_sched(td);
1607 sum = ts->ts_runtime + ts->ts_slptime;
1608 if (sum < SCHED_SLP_RUN_MAX)
1611 * This only happens from two places:
1612 * 1) We have added an unusual amount of run time from fork_exit.
1613 * 2) We have added an unusual amount of sleep time from sched_sleep().
1615 if (sum > SCHED_SLP_RUN_MAX * 2) {
1616 if (ts->ts_runtime > ts->ts_slptime) {
1617 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1620 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1626 * If we have exceeded by more than 1/5th then the algorithm below
1627 * will not bring us back into range. Dividing by two here forces
1628 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1630 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1631 ts->ts_runtime /= 2;
1632 ts->ts_slptime /= 2;
1635 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1636 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1640 * Scale back the interactivity history when a child thread is created. The
1641 * history is inherited from the parent but the thread may behave totally
1642 * differently. For example, a shell spawning a compiler process. We want
1643 * to learn that the compiler is behaving badly very quickly.
1646 sched_interact_fork(struct thread *td)
1648 struct td_sched *ts;
1652 ts = td_get_sched(td);
1653 sum = ts->ts_runtime + ts->ts_slptime;
1654 if (sum > SCHED_SLP_RUN_FORK) {
1655 ratio = sum / SCHED_SLP_RUN_FORK;
1656 ts->ts_runtime /= ratio;
1657 ts->ts_slptime /= ratio;
1662 * Called from proc0_init() to setup the scheduler fields.
1667 struct td_sched *ts0;
1670 * Set up the scheduler specific parts of thread0.
1672 ts0 = td_get_sched(&thread0);
1673 ts0->ts_ltick = ticks;
1674 ts0->ts_ftick = ticks;
1676 ts0->ts_cpu = curcpu; /* set valid CPU number */
1680 * This is only somewhat accurate since given many processes of the same
1681 * priority they will switch when their slices run out, which will be
1682 * at most sched_slice stathz ticks.
1685 sched_rr_interval(void)
1688 /* Convert sched_slice from stathz to hz. */
1689 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1693 * Update the percent cpu tracking information when it is requested or
1694 * the total history exceeds the maximum. We keep a sliding history of
1695 * tick counts that slowly decays. This is less precise than the 4BSD
1696 * mechanism since it happens with less regular and frequent events.
1699 sched_pctcpu_update(struct td_sched *ts, int run)
1704 * The signed difference may be negative if the thread hasn't run for
1705 * over half of the ticks rollover period.
1707 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1709 ts->ts_ftick = t - SCHED_TICK_TARG;
1710 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1711 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1712 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1713 ts->ts_ftick = t - SCHED_TICK_TARG;
1716 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1721 * Adjust the priority of a thread. Move it to the appropriate run-queue
1722 * if necessary. This is the back-end for several priority related
1726 sched_thread_priority(struct thread *td, u_char prio)
1728 struct td_sched *ts;
1732 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1733 "prio:%d", td->td_priority, "new prio:%d", prio,
1734 KTR_ATTR_LINKED, sched_tdname(curthread));
1735 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1736 if (td != curthread && prio < td->td_priority) {
1737 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1738 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1739 prio, KTR_ATTR_LINKED, sched_tdname(td));
1740 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1743 ts = td_get_sched(td);
1744 THREAD_LOCK_ASSERT(td, MA_OWNED);
1745 if (td->td_priority == prio)
1748 * If the priority has been elevated due to priority
1749 * propagation, we may have to move ourselves to a new
1750 * queue. This could be optimized to not re-add in some
1753 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1755 td->td_priority = prio;
1756 sched_add(td, SRQ_BORROWING);
1760 * If the thread is currently running we may have to adjust the lowpri
1761 * information so other cpus are aware of our current priority.
1763 if (TD_IS_RUNNING(td)) {
1764 tdq = TDQ_CPU(ts->ts_cpu);
1765 oldpri = td->td_priority;
1766 td->td_priority = prio;
1767 if (prio < tdq->tdq_lowpri)
1768 tdq->tdq_lowpri = prio;
1769 else if (tdq->tdq_lowpri == oldpri)
1770 tdq_setlowpri(tdq, td);
1773 td->td_priority = prio;
1777 * Update a thread's priority when it is lent another thread's
1781 sched_lend_prio(struct thread *td, u_char prio)
1784 td->td_flags |= TDF_BORROWING;
1785 sched_thread_priority(td, prio);
1789 * Restore a thread's priority when priority propagation is
1790 * over. The prio argument is the minimum priority the thread
1791 * needs to have to satisfy other possible priority lending
1792 * requests. If the thread's regular priority is less
1793 * important than prio, the thread will keep a priority boost
1797 sched_unlend_prio(struct thread *td, u_char prio)
1801 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1802 td->td_base_pri <= PRI_MAX_TIMESHARE)
1803 base_pri = td->td_user_pri;
1805 base_pri = td->td_base_pri;
1806 if (prio >= base_pri) {
1807 td->td_flags &= ~TDF_BORROWING;
1808 sched_thread_priority(td, base_pri);
1810 sched_lend_prio(td, prio);
1814 * Standard entry for setting the priority to an absolute value.
1817 sched_prio(struct thread *td, u_char prio)
1821 /* First, update the base priority. */
1822 td->td_base_pri = prio;
1825 * If the thread is borrowing another thread's priority, don't
1826 * ever lower the priority.
1828 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1831 /* Change the real priority. */
1832 oldprio = td->td_priority;
1833 sched_thread_priority(td, prio);
1836 * If the thread is on a turnstile, then let the turnstile update
1839 if (TD_ON_LOCK(td) && oldprio != prio)
1840 turnstile_adjust(td, oldprio);
1844 * Set the base user priority, does not effect current running priority.
1847 sched_user_prio(struct thread *td, u_char prio)
1850 td->td_base_user_pri = prio;
1851 if (td->td_lend_user_pri <= prio)
1853 td->td_user_pri = prio;
1857 sched_lend_user_prio(struct thread *td, u_char prio)
1860 THREAD_LOCK_ASSERT(td, MA_OWNED);
1861 td->td_lend_user_pri = prio;
1862 td->td_user_pri = min(prio, td->td_base_user_pri);
1863 if (td->td_priority > td->td_user_pri)
1864 sched_prio(td, td->td_user_pri);
1865 else if (td->td_priority != td->td_user_pri)
1866 td->td_flags |= TDF_NEEDRESCHED;
1870 * Like the above but first check if there is anything to do.
1873 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1876 if (td->td_lend_user_pri != prio)
1878 if (td->td_user_pri != min(prio, td->td_base_user_pri))
1880 if (td->td_priority >= td->td_user_pri)
1886 sched_lend_user_prio(td, prio);
1892 * This tdq is about to idle. Try to steal a thread from another CPU before
1893 * choosing the idle thread.
1896 tdq_trysteal(struct tdq *tdq)
1898 struct cpu_group *cg;
1903 if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1906 CPU_CLR(PCPU_GET(cpuid), &mask);
1907 /* We don't want to be preempted while we're iterating. */
1910 for (i = 1, cg = tdq->tdq_cg; ; ) {
1911 cpu = sched_highest(cg, mask, steal_thresh);
1913 * If a thread was added while interrupts were disabled don't
1916 if (tdq->tdq_load > 0) {
1923 if (cg == NULL || i > trysteal_limit) {
1929 steal = TDQ_CPU(cpu);
1931 * The data returned by sched_highest() is stale and
1932 * the chosen CPU no longer has an eligible thread.
1934 if (steal->tdq_load < steal_thresh ||
1935 steal->tdq_transferable == 0)
1937 tdq_lock_pair(tdq, steal);
1939 * If we get to this point, unconditonally exit the loop
1940 * to bound the time spent in the critcal section.
1942 * If a thread was added while interrupts were disabled don't
1945 if (tdq->tdq_load > 0) {
1950 * The data returned by sched_highest() is stale and
1951 * the chosen CPU no longer has an eligible thread.
1953 if (steal->tdq_load < steal_thresh ||
1954 steal->tdq_transferable == 0) {
1959 * If we fail to acquire one due to affinity restrictions,
1960 * bail out and let the idle thread to a more complete search
1961 * outside of a critical section.
1963 if (tdq_move(steal, tdq) == NULL) {
1975 * Handle migration from sched_switch(). This happens only for
1979 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1983 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
1984 "thread %s queued on absent CPU %d.", td->td_name,
1985 td_get_sched(td)->ts_cpu));
1986 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1988 tdq_load_rem(tdq, td);
1990 * Do the lock dance required to avoid LOR. We grab an extra
1991 * spinlock nesting to prevent preemption while we're
1992 * not holding either run-queue lock.
1995 thread_lock_block(td); /* This releases the lock on tdq. */
1998 * Acquire both run-queue locks before placing the thread on the new
1999 * run-queue to avoid deadlocks created by placing a thread with a
2000 * blocked lock on the run-queue of a remote processor. The deadlock
2001 * occurs when a third processor attempts to lock the two queues in
2002 * question while the target processor is spinning with its own
2003 * run-queue lock held while waiting for the blocked lock to clear.
2005 tdq_lock_pair(tdn, tdq);
2006 tdq_add(tdn, td, flags);
2007 tdq_notify(tdn, td);
2011 return (TDQ_LOCKPTR(tdn));
2015 * Variadic version of thread_lock_unblock() that does not assume td_lock
2019 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2021 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2026 * Switch threads. This function has to handle threads coming in while
2027 * blocked for some reason, running, or idle. It also must deal with
2028 * migrating a thread from one queue to another as running threads may
2029 * be assigned elsewhere via binding.
2032 sched_switch(struct thread *td, struct thread *newtd, int flags)
2035 struct td_sched *ts;
2038 int cpuid, preempted;
2040 THREAD_LOCK_ASSERT(td, MA_OWNED);
2041 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2043 cpuid = PCPU_GET(cpuid);
2045 ts = td_get_sched(td);
2047 sched_pctcpu_update(ts, 1);
2048 ts->ts_rltick = ticks;
2049 td->td_lastcpu = td->td_oncpu;
2050 td->td_oncpu = NOCPU;
2051 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2052 (flags & SW_PREEMPT) != 0;
2053 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2054 td->td_owepreempt = 0;
2055 if (!TD_IS_IDLETHREAD(td))
2056 tdq->tdq_switchcnt++;
2058 * The lock pointer in an idle thread should never change. Reset it
2059 * to CAN_RUN as well.
2061 if (TD_IS_IDLETHREAD(td)) {
2062 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2064 } else if (TD_IS_RUNNING(td)) {
2065 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2066 srqflag = preempted ?
2067 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2068 SRQ_OURSELF|SRQ_YIELDING;
2070 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2071 ts->ts_cpu = sched_pickcpu(td, 0);
2073 if (ts->ts_cpu == cpuid)
2074 tdq_runq_add(tdq, td, srqflag);
2076 KASSERT(THREAD_CAN_MIGRATE(td) ||
2077 (ts->ts_flags & TSF_BOUND) != 0,
2078 ("Thread %p shouldn't migrate", td));
2079 mtx = sched_switch_migrate(tdq, td, srqflag);
2082 /* This thread must be going to sleep. */
2084 mtx = thread_lock_block(td);
2085 tdq_load_rem(tdq, td);
2087 if (tdq->tdq_load == 0)
2092 #if (KTR_COMPILE & KTR_SCHED) != 0
2093 if (TD_IS_IDLETHREAD(td))
2094 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2095 "prio:%d", td->td_priority);
2097 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2098 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2099 "lockname:\"%s\"", td->td_lockname);
2103 * We enter here with the thread blocked and assigned to the
2104 * appropriate cpu run-queue or sleep-queue and with the current
2105 * thread-queue locked.
2107 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2108 newtd = choosethread();
2110 * Call the MD code to switch contexts if necessary.
2114 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2115 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2117 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2118 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2119 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2120 sched_pctcpu_update(td_get_sched(newtd), 0);
2122 #ifdef KDTRACE_HOOKS
2124 * If DTrace has set the active vtime enum to anything
2125 * other than INACTIVE (0), then it should have set the
2128 if (dtrace_vtime_active)
2129 (*dtrace_vtime_switch_func)(newtd);
2132 cpu_switch(td, newtd, mtx);
2134 * We may return from cpu_switch on a different cpu. However,
2135 * we always return with td_lock pointing to the current cpu's
2138 cpuid = PCPU_GET(cpuid);
2140 lock_profile_obtain_lock_success(
2141 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2143 SDT_PROBE0(sched, , , on__cpu);
2145 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2146 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2149 thread_unblock_switch(td, mtx);
2150 SDT_PROBE0(sched, , , remain__cpu);
2153 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2154 "prio:%d", td->td_priority);
2157 * Assert that all went well and return.
2159 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2160 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2161 td->td_oncpu = cpuid;
2165 * Adjust thread priorities as a result of a nice request.
2168 sched_nice(struct proc *p, int nice)
2172 PROC_LOCK_ASSERT(p, MA_OWNED);
2175 FOREACH_THREAD_IN_PROC(p, td) {
2178 sched_prio(td, td->td_base_user_pri);
2184 * Record the sleep time for the interactivity scorer.
2187 sched_sleep(struct thread *td, int prio)
2190 THREAD_LOCK_ASSERT(td, MA_OWNED);
2192 td->td_slptick = ticks;
2193 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2194 td->td_flags |= TDF_CANSWAP;
2195 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2197 if (static_boost == 1 && prio)
2198 sched_prio(td, prio);
2199 else if (static_boost && td->td_priority > static_boost)
2200 sched_prio(td, static_boost);
2204 * Schedule a thread to resume execution and record how long it voluntarily
2205 * slept. We also update the pctcpu, interactivity, and priority.
2208 sched_wakeup(struct thread *td)
2210 struct td_sched *ts;
2213 THREAD_LOCK_ASSERT(td, MA_OWNED);
2214 ts = td_get_sched(td);
2215 td->td_flags &= ~TDF_CANSWAP;
2217 * If we slept for more than a tick update our interactivity and
2220 slptick = td->td_slptick;
2222 if (slptick && slptick != ticks) {
2223 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2224 sched_interact_update(td);
2225 sched_pctcpu_update(ts, 0);
2228 * Reset the slice value since we slept and advanced the round-robin.
2231 sched_add(td, SRQ_BORING);
2235 * Penalize the parent for creating a new child and initialize the child's
2239 sched_fork(struct thread *td, struct thread *child)
2241 THREAD_LOCK_ASSERT(td, MA_OWNED);
2242 sched_pctcpu_update(td_get_sched(td), 1);
2243 sched_fork_thread(td, child);
2245 * Penalize the parent and child for forking.
2247 sched_interact_fork(child);
2248 sched_priority(child);
2249 td_get_sched(td)->ts_runtime += tickincr;
2250 sched_interact_update(td);
2255 * Fork a new thread, may be within the same process.
2258 sched_fork_thread(struct thread *td, struct thread *child)
2260 struct td_sched *ts;
2261 struct td_sched *ts2;
2265 THREAD_LOCK_ASSERT(td, MA_OWNED);
2269 ts = td_get_sched(td);
2270 ts2 = td_get_sched(child);
2271 child->td_oncpu = NOCPU;
2272 child->td_lastcpu = NOCPU;
2273 child->td_lock = TDQ_LOCKPTR(tdq);
2274 child->td_cpuset = cpuset_ref(td->td_cpuset);
2275 child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2276 ts2->ts_cpu = ts->ts_cpu;
2279 * Grab our parents cpu estimation information.
2281 ts2->ts_ticks = ts->ts_ticks;
2282 ts2->ts_ltick = ts->ts_ltick;
2283 ts2->ts_ftick = ts->ts_ftick;
2285 * Do not inherit any borrowed priority from the parent.
2287 child->td_priority = child->td_base_pri;
2289 * And update interactivity score.
2291 ts2->ts_slptime = ts->ts_slptime;
2292 ts2->ts_runtime = ts->ts_runtime;
2293 /* Attempt to quickly learn interactivity. */
2294 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2296 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2301 * Adjust the priority class of a thread.
2304 sched_class(struct thread *td, int class)
2307 THREAD_LOCK_ASSERT(td, MA_OWNED);
2308 if (td->td_pri_class == class)
2310 td->td_pri_class = class;
2314 * Return some of the child's priority and interactivity to the parent.
2317 sched_exit(struct proc *p, struct thread *child)
2321 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2322 "prio:%d", child->td_priority);
2323 PROC_LOCK_ASSERT(p, MA_OWNED);
2324 td = FIRST_THREAD_IN_PROC(p);
2325 sched_exit_thread(td, child);
2329 * Penalize another thread for the time spent on this one. This helps to
2330 * worsen the priority and interactivity of processes which schedule batch
2331 * jobs such as make. This has little effect on the make process itself but
2332 * causes new processes spawned by it to receive worse scores immediately.
2335 sched_exit_thread(struct thread *td, struct thread *child)
2338 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2339 "prio:%d", child->td_priority);
2341 * Give the child's runtime to the parent without returning the
2342 * sleep time as a penalty to the parent. This causes shells that
2343 * launch expensive things to mark their children as expensive.
2346 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2347 sched_interact_update(td);
2353 sched_preempt(struct thread *td)
2357 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2361 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2362 tdq->tdq_ipipending = 0;
2363 if (td->td_priority > tdq->tdq_lowpri) {
2366 flags = SW_INVOL | SW_PREEMPT;
2367 if (td->td_critnest > 1)
2368 td->td_owepreempt = 1;
2369 else if (TD_IS_IDLETHREAD(td))
2370 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2372 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2378 * Fix priorities on return to user-space. Priorities may be elevated due
2379 * to static priorities in msleep() or similar.
2382 sched_userret_slowpath(struct thread *td)
2386 td->td_priority = td->td_user_pri;
2387 td->td_base_pri = td->td_user_pri;
2388 tdq_setlowpri(TDQ_SELF(), td);
2393 * Handle a stathz tick. This is really only relevant for timeshare
2397 sched_clock(struct thread *td)
2400 struct td_sched *ts;
2402 THREAD_LOCK_ASSERT(td, MA_OWNED);
2406 * We run the long term load balancer infrequently on the first cpu.
2408 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2409 if (balance_ticks && --balance_ticks == 0)
2414 * Save the old switch count so we have a record of the last ticks
2415 * activity. Initialize the new switch count based on our load.
2416 * If there is some activity seed it to reflect that.
2418 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2419 tdq->tdq_switchcnt = tdq->tdq_load;
2421 * Advance the insert index once for each tick to ensure that all
2422 * threads get a chance to run.
2424 if (tdq->tdq_idx == tdq->tdq_ridx) {
2425 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2426 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2427 tdq->tdq_ridx = tdq->tdq_idx;
2429 ts = td_get_sched(td);
2430 sched_pctcpu_update(ts, 1);
2431 if (td->td_pri_class & PRI_FIFO_BIT)
2433 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2435 * We used a tick; charge it to the thread so
2436 * that we can compute our interactivity.
2438 td_get_sched(td)->ts_runtime += tickincr;
2439 sched_interact_update(td);
2444 * Force a context switch if the current thread has used up a full
2445 * time slice (default is 100ms).
2447 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2449 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2454 sched_estcpu(struct thread *td __unused)
2461 * Return whether the current CPU has runnable tasks. Used for in-kernel
2462 * cooperative idle threads.
2465 sched_runnable(void)
2473 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2474 if (tdq->tdq_load > 0)
2477 if (tdq->tdq_load - 1 > 0)
2485 * Choose the highest priority thread to run. The thread is removed from
2486 * the run-queue while running however the load remains. For SMP we set
2487 * the tdq in the global idle bitmask if it idles here.
2496 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2497 td = tdq_choose(tdq);
2499 tdq_runq_rem(tdq, td);
2500 tdq->tdq_lowpri = td->td_priority;
2503 tdq->tdq_lowpri = PRI_MAX_IDLE;
2504 return (PCPU_GET(idlethread));
2508 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2509 * we always request it once we exit a critical section.
2512 sched_setpreempt(struct thread *td)
2518 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2521 pri = td->td_priority;
2522 cpri = ctd->td_priority;
2524 ctd->td_flags |= TDF_NEEDRESCHED;
2525 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2527 if (!sched_shouldpreempt(pri, cpri, 0))
2529 ctd->td_owepreempt = 1;
2533 * Add a thread to a thread queue. Select the appropriate runq and add the
2534 * thread to it. This is the internal function called when the tdq is
2538 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2541 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2542 KASSERT((td->td_inhibitors == 0),
2543 ("sched_add: trying to run inhibited thread"));
2544 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2545 ("sched_add: bad thread state"));
2546 KASSERT(td->td_flags & TDF_INMEM,
2547 ("sched_add: thread swapped out"));
2549 if (td->td_priority < tdq->tdq_lowpri)
2550 tdq->tdq_lowpri = td->td_priority;
2551 tdq_runq_add(tdq, td, flags);
2552 tdq_load_add(tdq, td);
2556 * Select the target thread queue and add a thread to it. Request
2557 * preemption or IPI a remote processor if required.
2560 sched_add(struct thread *td, int flags)
2567 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2568 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2569 sched_tdname(curthread));
2570 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2571 KTR_ATTR_LINKED, sched_tdname(td));
2572 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2573 flags & SRQ_PREEMPTED);
2574 THREAD_LOCK_ASSERT(td, MA_OWNED);
2576 * Recalculate the priority before we select the target cpu or
2579 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2583 * Pick the destination cpu and if it isn't ours transfer to the
2586 cpu = sched_pickcpu(td, flags);
2587 tdq = sched_setcpu(td, cpu, flags);
2588 tdq_add(tdq, td, flags);
2589 if (cpu != PCPU_GET(cpuid)) {
2590 tdq_notify(tdq, td);
2597 * Now that the thread is moving to the run-queue, set the lock
2598 * to the scheduler's lock.
2600 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2601 tdq_add(tdq, td, flags);
2603 if (!(flags & SRQ_YIELDING))
2604 sched_setpreempt(td);
2608 * Remove a thread from a run-queue without running it. This is used
2609 * when we're stealing a thread from a remote queue. Otherwise all threads
2610 * exit by calling sched_exit_thread() and sched_throw() themselves.
2613 sched_rem(struct thread *td)
2617 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2618 "prio:%d", td->td_priority);
2619 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2620 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2621 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2622 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2623 KASSERT(TD_ON_RUNQ(td),
2624 ("sched_rem: thread not on run queue"));
2625 tdq_runq_rem(tdq, td);
2626 tdq_load_rem(tdq, td);
2628 if (td->td_priority == tdq->tdq_lowpri)
2629 tdq_setlowpri(tdq, NULL);
2633 * Fetch cpu utilization information. Updates on demand.
2636 sched_pctcpu(struct thread *td)
2639 struct td_sched *ts;
2642 ts = td_get_sched(td);
2644 THREAD_LOCK_ASSERT(td, MA_OWNED);
2645 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2649 /* How many rtick per second ? */
2650 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2651 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2658 * Enforce affinity settings for a thread. Called after adjustments to
2662 sched_affinity(struct thread *td)
2665 struct td_sched *ts;
2667 THREAD_LOCK_ASSERT(td, MA_OWNED);
2668 ts = td_get_sched(td);
2669 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2671 if (TD_ON_RUNQ(td)) {
2673 sched_add(td, SRQ_BORING);
2676 if (!TD_IS_RUNNING(td))
2679 * Force a switch before returning to userspace. If the
2680 * target thread is not running locally send an ipi to force
2683 td->td_flags |= TDF_NEEDRESCHED;
2684 if (td != curthread)
2685 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2690 * Bind a thread to a target cpu.
2693 sched_bind(struct thread *td, int cpu)
2695 struct td_sched *ts;
2697 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2698 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2699 ts = td_get_sched(td);
2700 if (ts->ts_flags & TSF_BOUND)
2702 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2703 ts->ts_flags |= TSF_BOUND;
2705 if (PCPU_GET(cpuid) == cpu)
2708 /* When we return from mi_switch we'll be on the correct cpu. */
2709 mi_switch(SW_VOL, NULL);
2713 * Release a bound thread.
2716 sched_unbind(struct thread *td)
2718 struct td_sched *ts;
2720 THREAD_LOCK_ASSERT(td, MA_OWNED);
2721 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2722 ts = td_get_sched(td);
2723 if ((ts->ts_flags & TSF_BOUND) == 0)
2725 ts->ts_flags &= ~TSF_BOUND;
2730 sched_is_bound(struct thread *td)
2732 THREAD_LOCK_ASSERT(td, MA_OWNED);
2733 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2740 sched_relinquish(struct thread *td)
2743 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2748 * Return the total system load.
2759 total += TDQ_CPU(i)->tdq_sysload;
2762 return (TDQ_SELF()->tdq_sysload);
2767 sched_sizeof_proc(void)
2769 return (sizeof(struct proc));
2773 sched_sizeof_thread(void)
2775 return (sizeof(struct thread) + sizeof(struct td_sched));
2779 #define TDQ_IDLESPIN(tdq) \
2780 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2782 #define TDQ_IDLESPIN(tdq) 1
2786 * The actual idle process.
2789 sched_idletd(void *dummy)
2793 int oldswitchcnt, switchcnt;
2796 mtx_assert(&Giant, MA_NOTOWNED);
2799 THREAD_NO_SLEEPING();
2802 if (tdq->tdq_load) {
2804 mi_switch(SW_VOL | SWT_IDLE, NULL);
2807 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2809 if (always_steal || switchcnt != oldswitchcnt) {
2810 oldswitchcnt = switchcnt;
2811 if (tdq_idled(tdq) == 0)
2814 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2816 oldswitchcnt = switchcnt;
2819 * If we're switching very frequently, spin while checking
2820 * for load rather than entering a low power state that
2821 * may require an IPI. However, don't do any busy
2822 * loops while on SMT machines as this simply steals
2823 * cycles from cores doing useful work.
2825 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2826 for (i = 0; i < sched_idlespins; i++) {
2833 /* If there was context switch during spin, restart it. */
2834 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2835 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2838 /* Run main MD idle handler. */
2839 tdq->tdq_cpu_idle = 1;
2841 * Make sure that tdq_cpu_idle update is globally visible
2842 * before cpu_idle() read tdq_load. The order is important
2843 * to avoid race with tdq_notify.
2845 atomic_thread_fence_seq_cst();
2847 * Checking for again after the fence picks up assigned
2848 * threads often enough to make it worthwhile to do so in
2849 * order to avoid calling cpu_idle().
2851 if (tdq->tdq_load != 0) {
2852 tdq->tdq_cpu_idle = 0;
2855 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2856 tdq->tdq_cpu_idle = 0;
2859 * Account thread-less hardware interrupts and
2860 * other wakeup reasons equal to context switches.
2862 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2863 if (switchcnt != oldswitchcnt)
2865 tdq->tdq_switchcnt++;
2871 * A CPU is entering for the first time or a thread is exiting.
2874 sched_throw(struct thread *td)
2876 struct thread *newtd;
2881 PCPU_SET(sched, DPCPU_PTR(tdq));
2883 /* Correct spinlock nesting and acquire the correct lock. */
2887 PCPU_SET(switchtime, cpu_ticks());
2888 PCPU_SET(switchticks, ticks);
2891 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2892 tdq_load_rem(tdq, td);
2893 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2894 td->td_lastcpu = td->td_oncpu;
2895 td->td_oncpu = NOCPU;
2897 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2898 newtd = choosethread();
2899 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2900 cpu_throw(td, newtd); /* doesn't return */
2904 * This is called from fork_exit(). Just acquire the correct locks and
2905 * let fork do the rest of the work.
2908 sched_fork_exit(struct thread *td)
2914 * Finish setting up thread glue so that it begins execution in a
2915 * non-nested critical section with the scheduler lock held.
2917 cpuid = PCPU_GET(cpuid);
2919 if (TD_IS_IDLETHREAD(td))
2920 td->td_lock = TDQ_LOCKPTR(tdq);
2921 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2922 td->td_oncpu = cpuid;
2923 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2924 lock_profile_obtain_lock_success(
2925 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2927 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2928 "prio:%d", td->td_priority);
2929 SDT_PROBE0(sched, , , on__cpu);
2933 * Create on first use to catch odd startup conditons.
2936 sched_tdname(struct thread *td)
2939 struct td_sched *ts;
2941 ts = td_get_sched(td);
2942 if (ts->ts_name[0] == '\0')
2943 snprintf(ts->ts_name, sizeof(ts->ts_name),
2944 "%s tid %d", td->td_name, td->td_tid);
2945 return (ts->ts_name);
2947 return (td->td_name);
2953 sched_clear_tdname(struct thread *td)
2955 struct td_sched *ts;
2957 ts = td_get_sched(td);
2958 ts->ts_name[0] = '\0';
2965 * Build the CPU topology dump string. Is recursively called to collect
2966 * the topology tree.
2969 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2972 char cpusetbuf[CPUSETBUFSIZ];
2975 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2976 "", 1 + indent / 2, cg->cg_level);
2977 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2978 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2980 for (i = 0; i < MAXCPU; i++) {
2981 if (CPU_ISSET(i, &cg->cg_mask)) {
2983 sbuf_printf(sb, ", ");
2986 sbuf_printf(sb, "%d", i);
2989 sbuf_printf(sb, "</cpu>\n");
2991 if (cg->cg_flags != 0) {
2992 sbuf_printf(sb, "%*s <flags>", indent, "");
2993 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2994 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2995 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2996 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2997 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2998 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2999 sbuf_printf(sb, "</flags>\n");
3002 if (cg->cg_children > 0) {
3003 sbuf_printf(sb, "%*s <children>\n", indent, "");
3004 for (i = 0; i < cg->cg_children; i++)
3005 sysctl_kern_sched_topology_spec_internal(sb,
3006 &cg->cg_child[i], indent+2);
3007 sbuf_printf(sb, "%*s </children>\n", indent, "");
3009 sbuf_printf(sb, "%*s</group>\n", indent, "");
3014 * Sysctl handler for retrieving topology dump. It's a wrapper for
3015 * the recursive sysctl_kern_smp_topology_spec_internal().
3018 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3023 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3025 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3029 sbuf_printf(topo, "<groups>\n");
3030 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3031 sbuf_printf(topo, "</groups>\n");
3034 err = sbuf_finish(topo);
3043 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3045 int error, new_val, period;
3047 period = 1000000 / realstathz;
3048 new_val = period * sched_slice;
3049 error = sysctl_handle_int(oidp, &new_val, 0, req);
3050 if (error != 0 || req->newptr == NULL)
3054 sched_slice = imax(1, (new_val + period / 2) / period);
3055 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3056 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3061 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3062 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3064 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3065 NULL, 0, sysctl_kern_quantum, "I",
3066 "Quantum for timeshare threads in microseconds");
3067 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3068 "Quantum for timeshare threads in stathz ticks");
3069 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3070 "Interactivity score threshold");
3071 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3073 "Maximal (lowest) priority for preemption");
3074 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3075 "Assign static kernel priorities to sleeping threads");
3076 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3077 "Number of times idle thread will spin waiting for new work");
3078 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3079 &sched_idlespinthresh, 0,
3080 "Threshold before we will permit idle thread spinning");
3082 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3083 "Number of hz ticks to keep thread affinity for");
3084 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3085 "Enables the long-term load balancer");
3086 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3087 &balance_interval, 0,
3088 "Average period in stathz ticks to run the long-term balancer");
3089 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3090 "Attempts to steal work from other cores before idling");
3091 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3092 "Minimum load on remote CPU before we'll steal");
3093 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3094 0, "Topological distance limit for stealing threads in sched_switch()");
3095 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3096 "Always run the stealer from the idle thread");
3097 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3098 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3099 "XML dump of detected CPU topology");
3102 /* ps compat. All cpu percentages from ULE are weighted. */
3103 static int ccpu = 0;
3104 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");