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 __read_mostly 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 __read_mostly sched_interact = SCHED_INTERACT_THRESH;
210 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
211 static int __read_mostly realstathz = 127; /* reset during boot. */
212 static int __read_mostly sched_slice = 10; /* reset during boot. */
213 static int __read_mostly sched_slice_min = 1; /* reset during boot. */
215 #ifdef FULL_PREEMPTION
216 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
218 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
221 static int __read_mostly preempt_thresh = 0;
223 static int __read_mostly static_boost = PRI_MIN_BATCH;
224 static int __read_mostly sched_idlespins = 10000;
225 static int __read_mostly 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_owepreempt; /* Remote preemption 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 __read_mostly *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(). */
275 static int __read_mostly affinity;
276 static int __read_mostly steal_idle = 1;
277 static int __read_mostly steal_thresh = 2;
278 static int __read_mostly always_steal = 0;
279 static int __read_mostly trysteal_limit = 2;
282 * One thread queue per processor.
284 static struct tdq __read_mostly *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 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_BLOCKED_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 THREAD_LOCK_BLOCKED_ASSERT(td, 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_BLOCKED_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 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
560 THREAD_LOCK_BLOCKED_ASSERT(td, 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 static __always_inline int cpu_search(const struct cpu_group *cg,
647 struct cpu_search *low, struct cpu_search *high, const int match);
648 int __noinline cpu_search_lowest(const struct cpu_group *cg,
649 struct cpu_search *low);
650 int __noinline cpu_search_highest(const struct cpu_group *cg,
651 struct cpu_search *high);
652 int __noinline cpu_search_both(const struct cpu_group *cg,
653 struct cpu_search *low, struct cpu_search *high);
656 * Search the tree of cpu_groups for the lowest or highest loaded cpu
657 * according to the match argument. This routine actually compares the
658 * load on all paths through the tree and finds the least loaded cpu on
659 * the least loaded path, which may differ from the least loaded cpu in
660 * the system. This balances work among caches and buses.
662 * This inline is instantiated in three forms below using constants for the
663 * match argument. It is reduced to the minimum set for each case. It is
664 * also recursive to the depth of the tree.
666 static __always_inline int
667 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
668 struct cpu_search *high, const int match)
670 struct cpu_search lgroup;
671 struct cpu_search hgroup;
673 struct cpu_group *child;
675 int cpu, i, hload, lload, load, total, rnd;
678 cpumask = cg->cg_mask;
679 if (match & CPU_SEARCH_LOWEST) {
683 if (match & CPU_SEARCH_HIGHEST) {
688 /* Iterate through the child CPU groups and then remaining CPUs. */
689 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
691 #ifdef HAVE_INLINE_FFSL
692 cpu = CPU_FFS(&cpumask) - 1;
694 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
701 child = &cg->cg_child[i - 1];
703 if (match & CPU_SEARCH_LOWEST)
705 if (match & CPU_SEARCH_HIGHEST)
707 if (child) { /* Handle child CPU group. */
708 CPU_ANDNOT(&cpumask, &child->cg_mask);
710 case CPU_SEARCH_LOWEST:
711 load = cpu_search_lowest(child, &lgroup);
713 case CPU_SEARCH_HIGHEST:
714 load = cpu_search_highest(child, &hgroup);
716 case CPU_SEARCH_BOTH:
717 load = cpu_search_both(child, &lgroup, &hgroup);
720 } else { /* Handle child CPU. */
721 CPU_CLR(cpu, &cpumask);
723 load = tdq->tdq_load * 256;
724 rnd = sched_random() % 32;
725 if (match & CPU_SEARCH_LOWEST) {
726 if (cpu == low->cs_prefer)
728 /* If that CPU is allowed and get data. */
729 if (tdq->tdq_lowpri > lgroup.cs_pri &&
730 tdq->tdq_load <= lgroup.cs_limit &&
731 CPU_ISSET(cpu, &lgroup.cs_mask)) {
733 lgroup.cs_load = load - rnd;
736 if (match & CPU_SEARCH_HIGHEST)
737 if (tdq->tdq_load >= hgroup.cs_limit &&
738 tdq->tdq_transferable &&
739 CPU_ISSET(cpu, &hgroup.cs_mask)) {
741 hgroup.cs_load = load - rnd;
746 /* We have info about child item. Compare it. */
747 if (match & CPU_SEARCH_LOWEST) {
748 if (lgroup.cs_cpu >= 0 &&
750 (load == lload && lgroup.cs_load < low->cs_load))) {
752 low->cs_cpu = lgroup.cs_cpu;
753 low->cs_load = lgroup.cs_load;
756 if (match & CPU_SEARCH_HIGHEST)
757 if (hgroup.cs_cpu >= 0 &&
759 (load == hload && hgroup.cs_load > high->cs_load))) {
761 high->cs_cpu = hgroup.cs_cpu;
762 high->cs_load = hgroup.cs_load;
766 if (i == 0 && CPU_EMPTY(&cpumask))
769 #ifndef HAVE_INLINE_FFSL
778 * cpu_search instantiations must pass constants to maintain the inline
782 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
784 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
788 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
790 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
794 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
795 struct cpu_search *high)
797 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
801 * Find the cpu with the least load via the least loaded path that has a
802 * lowpri greater than pri pri. A pri of -1 indicates any priority is
806 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
809 struct cpu_search low;
812 low.cs_prefer = prefer;
815 low.cs_limit = maxload;
816 cpu_search_lowest(cg, &low);
821 * Find the cpu with the highest load via the highest loaded path.
824 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
826 struct cpu_search high;
830 high.cs_limit = minload;
831 cpu_search_highest(cg, &high);
836 sched_balance_group(struct cpu_group *cg)
839 cpuset_t hmask, lmask;
840 int high, low, anylow;
844 high = sched_highest(cg, hmask, 2);
845 /* Stop if there is no more CPU with transferrable threads. */
848 CPU_CLR(high, &hmask);
849 CPU_COPY(&hmask, &lmask);
850 /* Stop if there is no more CPU left for low. */
851 if (CPU_EMPTY(&lmask))
856 low = sched_lowest(cg, lmask, -1, tdq->tdq_load - 1, high);
857 /* Stop if we looked well and found no less loaded CPU. */
858 if (anylow && low == -1)
860 /* Go to next high if we found no less loaded CPU. */
863 /* Transfer thread from high to low. */
864 if (sched_balance_pair(tdq, TDQ_CPU(low))) {
865 /* CPU that got thread can no longer be a donor. */
866 CPU_CLR(low, &hmask);
869 * If failed, then there is no threads on high
870 * that can run on this low. Drop low from low
871 * mask and look for different one.
873 CPU_CLR(low, &lmask);
885 balance_ticks = max(balance_interval / 2, 1) +
886 (sched_random() % balance_interval);
889 sched_balance_group(cpu_top);
894 * Lock two thread queues using their address to maintain lock order.
897 tdq_lock_pair(struct tdq *one, struct tdq *two)
901 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
904 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
909 * Unlock two thread queues. Order is not important here.
912 tdq_unlock_pair(struct tdq *one, struct tdq *two)
919 * Transfer load between two imbalanced thread queues.
922 sched_balance_pair(struct tdq *high, struct tdq *low)
927 tdq_lock_pair(high, low);
930 * Transfer a thread from high to low.
932 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
933 (td = tdq_move(high, low)) != NULL) {
935 * In case the target isn't the current cpu notify it of the
936 * new load, possibly sending an IPI to force it to reschedule.
939 if (cpu != PCPU_GET(cpuid))
942 tdq_unlock_pair(high, low);
947 * Move a thread from one thread queue to another.
949 static struct thread *
950 tdq_move(struct tdq *from, struct tdq *to)
956 TDQ_LOCK_ASSERT(from, MA_OWNED);
957 TDQ_LOCK_ASSERT(to, MA_OWNED);
961 td = tdq_steal(tdq, cpu);
966 * Although the run queue is locked the thread may be
967 * blocked. We can not set the lock until it is unblocked.
969 thread_lock_block_wait(td);
971 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from));
972 td->td_lock = TDQ_LOCKPTR(to);
973 td_get_sched(td)->ts_cpu = cpu;
974 tdq_add(to, td, SRQ_YIELDING);
980 * This tdq has idled. Try to steal a thread from another cpu and switch
984 tdq_idled(struct tdq *tdq)
986 struct cpu_group *cg;
991 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
994 CPU_CLR(PCPU_GET(cpuid), &mask);
996 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
997 for (cg = tdq->tdq_cg; ; ) {
998 cpu = sched_highest(cg, mask, steal_thresh);
1000 * We were assigned a thread but not preempted. Returning
1001 * 0 here will cause our caller to switch to it.
1011 steal = TDQ_CPU(cpu);
1013 * The data returned by sched_highest() is stale and
1014 * the chosen CPU no longer has an eligible thread.
1016 * Testing this ahead of tdq_lock_pair() only catches
1017 * this situation about 20% of the time on an 8 core
1018 * 16 thread Ryzen 7, but it still helps performance.
1020 if (steal->tdq_load < steal_thresh ||
1021 steal->tdq_transferable == 0)
1023 tdq_lock_pair(tdq, steal);
1025 * We were assigned a thread while waiting for the locks.
1026 * Switch to it now instead of stealing a thread.
1031 * The data returned by sched_highest() is stale and
1032 * the chosen CPU no longer has an eligible thread, or
1033 * we were preempted and the CPU loading info may be out
1034 * of date. The latter is rare. In either case restart
1037 if (steal->tdq_load < steal_thresh ||
1038 steal->tdq_transferable == 0 ||
1039 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1040 tdq_unlock_pair(tdq, steal);
1044 * Steal the thread and switch to it.
1046 if (tdq_move(steal, tdq) != NULL)
1049 * We failed to acquire a thread even though it looked
1050 * like one was available. This could be due to affinity
1051 * restrictions or for other reasons. Loop again after
1052 * removing this CPU from the set. The restart logic
1053 * above does not restore this CPU to the set due to the
1054 * likelyhood of failing here again.
1056 CPU_CLR(cpu, &mask);
1057 tdq_unlock_pair(tdq, steal);
1060 mi_switch(SW_VOL | SWT_IDLE);
1065 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1068 tdq_notify(struct tdq *tdq, struct thread *td)
1074 if (tdq->tdq_owepreempt)
1076 cpu = td_get_sched(td)->ts_cpu;
1077 pri = td->td_priority;
1078 ctd = pcpu_find(cpu)->pc_curthread;
1079 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1083 * Make sure that our caller's earlier update to tdq_load is
1084 * globally visible before we read tdq_cpu_idle. Idle thread
1085 * accesses both of them without locks, and the order is important.
1087 atomic_thread_fence_seq_cst();
1089 if (TD_IS_IDLETHREAD(ctd)) {
1091 * If the MD code has an idle wakeup routine try that before
1092 * falling back to IPI.
1094 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1099 * The run queues have been updated, so any switch on the remote CPU
1100 * will satisfy the preemption request.
1102 tdq->tdq_owepreempt = 1;
1103 ipi_cpu(cpu, IPI_PREEMPT);
1107 * Steals load from a timeshare queue. Honors the rotating queue head
1110 static struct thread *
1111 runq_steal_from(struct runq *rq, int cpu, u_char start)
1115 struct thread *td, *first;
1119 rqb = &rq->rq_status;
1120 bit = start & (RQB_BPW -1);
1123 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1124 if (rqb->rqb_bits[i] == 0)
1127 bit = RQB_FFS(rqb->rqb_bits[i]);
1128 for (; bit < RQB_BPW; bit++) {
1129 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1131 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1132 TAILQ_FOREACH(td, rqh, td_runq) {
1133 if (first && THREAD_CAN_MIGRATE(td) &&
1134 THREAD_CAN_SCHED(td, cpu))
1145 if (first && THREAD_CAN_MIGRATE(first) &&
1146 THREAD_CAN_SCHED(first, cpu))
1152 * Steals load from a standard linear queue.
1154 static struct thread *
1155 runq_steal(struct runq *rq, int cpu)
1163 rqb = &rq->rq_status;
1164 for (word = 0; word < RQB_LEN; word++) {
1165 if (rqb->rqb_bits[word] == 0)
1167 for (bit = 0; bit < RQB_BPW; bit++) {
1168 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1170 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1171 TAILQ_FOREACH(td, rqh, td_runq)
1172 if (THREAD_CAN_MIGRATE(td) &&
1173 THREAD_CAN_SCHED(td, cpu))
1181 * Attempt to steal a thread in priority order from a thread queue.
1183 static struct thread *
1184 tdq_steal(struct tdq *tdq, int cpu)
1188 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1189 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1191 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1192 cpu, tdq->tdq_ridx)) != NULL)
1194 return (runq_steal(&tdq->tdq_idle, cpu));
1198 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1199 * current lock and returns with the assigned queue locked.
1201 static inline struct tdq *
1202 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)) {
1215 KASSERT((flags & SRQ_HOLD) == 0,
1216 ("sched_setcpu: Invalid lock for SRQ_HOLD"));
1221 * The hard case, migration, we need to block the thread first to
1222 * prevent order reversals with other cpus locks.
1225 mtx = thread_lock_block(td);
1226 if ((flags & SRQ_HOLD) == 0)
1227 mtx_unlock_spin(mtx);
1229 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1234 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1235 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1236 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1237 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1238 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1239 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1242 sched_pickcpu(struct thread *td, int flags)
1244 struct cpu_group *cg, *ccg;
1245 struct td_sched *ts;
1248 int cpu, pri, self, intr;
1250 self = PCPU_GET(cpuid);
1251 ts = td_get_sched(td);
1252 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1253 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1254 if (smp_started == 0)
1257 * Don't migrate a running thread from sched_switch().
1259 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1260 return (ts->ts_cpu);
1262 * Prefer to run interrupt threads on the processors that generate
1265 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1266 curthread->td_intr_nesting_level) {
1268 if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1269 SCHED_STAT_INC(pickcpu_idle_affinity);
1278 tdq = TDQ_CPU(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 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 /* Check all SMT threads for being idle. */
1290 for (cpu = CPU_FFS(&cg->cg_mask) - 1; ; cpu++) {
1291 if (CPU_ISSET(cpu, &cg->cg_mask) &&
1292 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1294 if (cpu >= mp_maxid) {
1295 SCHED_STAT_INC(pickcpu_idle_affinity);
1296 return (ts->ts_cpu);
1300 SCHED_STAT_INC(pickcpu_idle_affinity);
1301 return (ts->ts_cpu);
1306 * Search for the last level cache CPU group in the tree.
1307 * Skip SMT, identical groups and caches with expired affinity.
1308 * Interrupt threads affinity is explicit and never expires.
1310 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1311 if (cg->cg_flags & CG_FLAG_THREAD)
1313 if (cg->cg_children == 1 || cg->cg_count == 1)
1315 if (cg->cg_level == CG_SHARE_NONE ||
1316 (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1320 /* Found LLC shared by all CPUs, so do a global search. */
1324 mask = td->td_cpuset->cs_mask;
1325 pri = td->td_priority;
1327 * Try hard to keep interrupts within found LLC. Search the LLC for
1328 * the least loaded CPU we can run now. For NUMA systems it should
1329 * be within target domain, and it also reduces scheduling overhead.
1331 if (ccg != NULL && intr) {
1332 cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu);
1334 SCHED_STAT_INC(pickcpu_intrbind);
1336 /* Search the LLC for the least loaded idle CPU we can run now. */
1338 cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1339 INT_MAX, ts->ts_cpu);
1341 SCHED_STAT_INC(pickcpu_affinity);
1343 /* Search globally for the least loaded CPU we can run now. */
1345 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1347 SCHED_STAT_INC(pickcpu_lowest);
1349 /* Search globally for the least loaded CPU. */
1351 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1353 SCHED_STAT_INC(pickcpu_lowest);
1355 KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1356 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1358 * Compare the lowest loaded cpu to current cpu.
1361 if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1362 tdq->tdq_lowpri < PRI_MIN_IDLE &&
1363 TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1364 SCHED_STAT_INC(pickcpu_local);
1367 if (cpu != ts->ts_cpu)
1368 SCHED_STAT_INC(pickcpu_migration);
1374 * Pick the highest priority task we have and return it.
1376 static struct thread *
1377 tdq_choose(struct tdq *tdq)
1381 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1382 td = runq_choose(&tdq->tdq_realtime);
1385 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1387 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1388 ("tdq_choose: Invalid priority on timeshare queue %d",
1392 td = runq_choose(&tdq->tdq_idle);
1394 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1395 ("tdq_choose: Invalid priority on idle queue %d",
1404 * Initialize a thread queue.
1407 tdq_setup(struct tdq *tdq, int id)
1411 printf("ULE: setup cpu %d\n", id);
1412 runq_init(&tdq->tdq_realtime);
1413 runq_init(&tdq->tdq_timeshare);
1414 runq_init(&tdq->tdq_idle);
1416 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1417 "sched lock %d", (int)TDQ_ID(tdq));
1418 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN);
1420 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1421 "CPU %d load", (int)TDQ_ID(tdq));
1427 sched_setup_smp(void)
1432 cpu_top = smp_topo();
1434 tdq = DPCPU_ID_PTR(i, tdq);
1436 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1437 if (tdq->tdq_cg == NULL)
1438 panic("Can't find cpu group for %d\n", i);
1440 PCPU_SET(sched, DPCPU_PTR(tdq));
1441 balance_tdq = TDQ_SELF();
1446 * Setup the thread queues and initialize the topology based on MD
1450 sched_setup(void *dummy)
1457 tdq_setup(TDQ_SELF(), 0);
1461 /* Add thread0's load since it's running. */
1463 thread0.td_lock = TDQ_LOCKPTR(tdq);
1464 tdq_load_add(tdq, &thread0);
1465 tdq->tdq_lowpri = thread0.td_priority;
1470 * This routine determines time constants after stathz and hz are setup.
1474 sched_initticks(void *dummy)
1478 realstathz = stathz ? stathz : hz;
1479 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1480 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1481 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1485 * tickincr is shifted out by 10 to avoid rounding errors due to
1486 * hz not being evenly divisible by stathz on all platforms.
1488 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1490 * This does not work for values of stathz that are more than
1491 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1498 * Set the default balance interval now that we know
1499 * what realstathz is.
1501 balance_interval = realstathz;
1502 balance_ticks = balance_interval;
1503 affinity = SCHED_AFFINITY_DEFAULT;
1505 if (sched_idlespinthresh < 0)
1506 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1510 * This is the core of the interactivity algorithm. Determines a score based
1511 * on past behavior. It is the ratio of sleep time to run time scaled to
1512 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1513 * differs from the cpu usage because it does not account for time spent
1514 * waiting on a run-queue. Would be prettier if we had floating point.
1516 * When a thread's sleep time is greater than its run time the
1520 * interactivity score = ---------------------
1521 * sleep time / run time
1524 * When a thread's run time is greater than its sleep time the
1528 * interactivity score = --------------------- + scaling factor
1529 * run time / sleep time
1532 sched_interact_score(struct thread *td)
1534 struct td_sched *ts;
1537 ts = td_get_sched(td);
1539 * The score is only needed if this is likely to be an interactive
1540 * task. Don't go through the expense of computing it if there's
1543 if (sched_interact <= SCHED_INTERACT_HALF &&
1544 ts->ts_runtime >= ts->ts_slptime)
1545 return (SCHED_INTERACT_HALF);
1547 if (ts->ts_runtime > ts->ts_slptime) {
1548 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1549 return (SCHED_INTERACT_HALF +
1550 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1552 if (ts->ts_slptime > ts->ts_runtime) {
1553 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1554 return (ts->ts_runtime / div);
1556 /* runtime == slptime */
1558 return (SCHED_INTERACT_HALF);
1561 * This can happen if slptime and runtime are 0.
1568 * Scale the scheduling priority according to the "interactivity" of this
1572 sched_priority(struct thread *td)
1577 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1580 * If the score is interactive we place the thread in the realtime
1581 * queue with a priority that is less than kernel and interrupt
1582 * priorities. These threads are not subject to nice restrictions.
1584 * Scores greater than this are placed on the normal timeshare queue
1585 * where the priority is partially decided by the most recent cpu
1586 * utilization and the rest is decided by nice value.
1588 * The nice value of the process has a linear effect on the calculated
1589 * score. Negative nice values make it easier for a thread to be
1590 * considered interactive.
1592 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1593 if (score < sched_interact) {
1594 pri = PRI_MIN_INTERACT;
1595 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1596 sched_interact) * score;
1597 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1598 ("sched_priority: invalid interactive priority %d score %d",
1601 pri = SCHED_PRI_MIN;
1602 if (td_get_sched(td)->ts_ticks)
1603 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1604 SCHED_PRI_RANGE - 1);
1605 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1606 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1607 ("sched_priority: invalid priority %d: nice %d, "
1608 "ticks %d ftick %d ltick %d tick pri %d",
1609 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1610 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1611 SCHED_PRI_TICKS(td_get_sched(td))));
1613 sched_user_prio(td, pri);
1619 * This routine enforces a maximum limit on the amount of scheduling history
1620 * kept. It is called after either the slptime or runtime is adjusted. This
1621 * function is ugly due to integer math.
1624 sched_interact_update(struct thread *td)
1626 struct td_sched *ts;
1629 ts = td_get_sched(td);
1630 sum = ts->ts_runtime + ts->ts_slptime;
1631 if (sum < SCHED_SLP_RUN_MAX)
1634 * This only happens from two places:
1635 * 1) We have added an unusual amount of run time from fork_exit.
1636 * 2) We have added an unusual amount of sleep time from sched_sleep().
1638 if (sum > SCHED_SLP_RUN_MAX * 2) {
1639 if (ts->ts_runtime > ts->ts_slptime) {
1640 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1643 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1649 * If we have exceeded by more than 1/5th then the algorithm below
1650 * will not bring us back into range. Dividing by two here forces
1651 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1653 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1654 ts->ts_runtime /= 2;
1655 ts->ts_slptime /= 2;
1658 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1659 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1663 * Scale back the interactivity history when a child thread is created. The
1664 * history is inherited from the parent but the thread may behave totally
1665 * differently. For example, a shell spawning a compiler process. We want
1666 * to learn that the compiler is behaving badly very quickly.
1669 sched_interact_fork(struct thread *td)
1671 struct td_sched *ts;
1675 ts = td_get_sched(td);
1676 sum = ts->ts_runtime + ts->ts_slptime;
1677 if (sum > SCHED_SLP_RUN_FORK) {
1678 ratio = sum / SCHED_SLP_RUN_FORK;
1679 ts->ts_runtime /= ratio;
1680 ts->ts_slptime /= ratio;
1685 * Called from proc0_init() to setup the scheduler fields.
1690 struct td_sched *ts0;
1693 * Set up the scheduler specific parts of thread0.
1695 ts0 = td_get_sched(&thread0);
1696 ts0->ts_ltick = ticks;
1697 ts0->ts_ftick = ticks;
1699 ts0->ts_cpu = curcpu; /* set valid CPU number */
1703 * This is only somewhat accurate since given many processes of the same
1704 * priority they will switch when their slices run out, which will be
1705 * at most sched_slice stathz ticks.
1708 sched_rr_interval(void)
1711 /* Convert sched_slice from stathz to hz. */
1712 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1716 * Update the percent cpu tracking information when it is requested or
1717 * the total history exceeds the maximum. We keep a sliding history of
1718 * tick counts that slowly decays. This is less precise than the 4BSD
1719 * mechanism since it happens with less regular and frequent events.
1722 sched_pctcpu_update(struct td_sched *ts, int run)
1727 * The signed difference may be negative if the thread hasn't run for
1728 * over half of the ticks rollover period.
1730 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1732 ts->ts_ftick = t - SCHED_TICK_TARG;
1733 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1734 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1735 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1736 ts->ts_ftick = t - SCHED_TICK_TARG;
1739 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1744 * Adjust the priority of a thread. Move it to the appropriate run-queue
1745 * if necessary. This is the back-end for several priority related
1749 sched_thread_priority(struct thread *td, u_char prio)
1751 struct td_sched *ts;
1755 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1756 "prio:%d", td->td_priority, "new prio:%d", prio,
1757 KTR_ATTR_LINKED, sched_tdname(curthread));
1758 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1759 if (td != curthread && prio < td->td_priority) {
1760 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1761 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1762 prio, KTR_ATTR_LINKED, sched_tdname(td));
1763 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1766 ts = td_get_sched(td);
1767 THREAD_LOCK_ASSERT(td, MA_OWNED);
1768 if (td->td_priority == prio)
1771 * If the priority has been elevated due to priority
1772 * propagation, we may have to move ourselves to a new
1773 * queue. This could be optimized to not re-add in some
1776 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1778 td->td_priority = prio;
1779 sched_add(td, SRQ_BORROWING | SRQ_HOLDTD);
1783 * If the thread is currently running we may have to adjust the lowpri
1784 * information so other cpus are aware of our current priority.
1786 if (TD_IS_RUNNING(td)) {
1787 tdq = TDQ_CPU(ts->ts_cpu);
1788 oldpri = td->td_priority;
1789 td->td_priority = prio;
1790 if (prio < tdq->tdq_lowpri)
1791 tdq->tdq_lowpri = prio;
1792 else if (tdq->tdq_lowpri == oldpri)
1793 tdq_setlowpri(tdq, td);
1796 td->td_priority = prio;
1800 * Update a thread's priority when it is lent another thread's
1804 sched_lend_prio(struct thread *td, u_char prio)
1807 td->td_flags |= TDF_BORROWING;
1808 sched_thread_priority(td, prio);
1812 * Restore a thread's priority when priority propagation is
1813 * over. The prio argument is the minimum priority the thread
1814 * needs to have to satisfy other possible priority lending
1815 * requests. If the thread's regular priority is less
1816 * important than prio, the thread will keep a priority boost
1820 sched_unlend_prio(struct thread *td, u_char prio)
1824 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1825 td->td_base_pri <= PRI_MAX_TIMESHARE)
1826 base_pri = td->td_user_pri;
1828 base_pri = td->td_base_pri;
1829 if (prio >= base_pri) {
1830 td->td_flags &= ~TDF_BORROWING;
1831 sched_thread_priority(td, base_pri);
1833 sched_lend_prio(td, prio);
1837 * Standard entry for setting the priority to an absolute value.
1840 sched_prio(struct thread *td, u_char prio)
1844 /* First, update the base priority. */
1845 td->td_base_pri = prio;
1848 * If the thread is borrowing another thread's priority, don't
1849 * ever lower the priority.
1851 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1854 /* Change the real priority. */
1855 oldprio = td->td_priority;
1856 sched_thread_priority(td, prio);
1859 * If the thread is on a turnstile, then let the turnstile update
1862 if (TD_ON_LOCK(td) && oldprio != prio)
1863 turnstile_adjust(td, oldprio);
1867 * Set the base user priority, does not effect current running priority.
1870 sched_user_prio(struct thread *td, u_char prio)
1873 td->td_base_user_pri = prio;
1874 if (td->td_lend_user_pri <= prio)
1876 td->td_user_pri = prio;
1880 sched_lend_user_prio(struct thread *td, u_char prio)
1883 THREAD_LOCK_ASSERT(td, MA_OWNED);
1884 td->td_lend_user_pri = prio;
1885 td->td_user_pri = min(prio, td->td_base_user_pri);
1886 if (td->td_priority > td->td_user_pri)
1887 sched_prio(td, td->td_user_pri);
1888 else if (td->td_priority != td->td_user_pri)
1889 td->td_flags |= TDF_NEEDRESCHED;
1893 * Like the above but first check if there is anything to do.
1896 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1899 if (td->td_lend_user_pri != prio)
1901 if (td->td_user_pri != min(prio, td->td_base_user_pri))
1903 if (td->td_priority >= td->td_user_pri)
1909 sched_lend_user_prio(td, prio);
1915 * This tdq is about to idle. Try to steal a thread from another CPU before
1916 * choosing the idle thread.
1919 tdq_trysteal(struct tdq *tdq)
1921 struct cpu_group *cg;
1926 if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1929 CPU_CLR(PCPU_GET(cpuid), &mask);
1930 /* We don't want to be preempted while we're iterating. */
1933 for (i = 1, cg = tdq->tdq_cg; ; ) {
1934 cpu = sched_highest(cg, mask, steal_thresh);
1936 * If a thread was added while interrupts were disabled don't
1939 if (tdq->tdq_load > 0) {
1946 if (cg == NULL || i > trysteal_limit) {
1952 steal = TDQ_CPU(cpu);
1954 * The data returned by sched_highest() is stale and
1955 * the chosen CPU no longer has an eligible thread.
1957 if (steal->tdq_load < steal_thresh ||
1958 steal->tdq_transferable == 0)
1960 tdq_lock_pair(tdq, steal);
1962 * If we get to this point, unconditonally exit the loop
1963 * to bound the time spent in the critcal section.
1965 * If a thread was added while interrupts were disabled don't
1968 if (tdq->tdq_load > 0) {
1973 * The data returned by sched_highest() is stale and
1974 * the chosen CPU no longer has an eligible thread.
1976 if (steal->tdq_load < steal_thresh ||
1977 steal->tdq_transferable == 0) {
1982 * If we fail to acquire one due to affinity restrictions,
1983 * bail out and let the idle thread to a more complete search
1984 * outside of a critical section.
1986 if (tdq_move(steal, tdq) == NULL) {
1998 * Handle migration from sched_switch(). This happens only for
2002 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2006 KASSERT(THREAD_CAN_MIGRATE(td) ||
2007 (td_get_sched(td)->ts_flags & TSF_BOUND) != 0,
2008 ("Thread %p shouldn't migrate", td));
2009 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2010 "thread %s queued on absent CPU %d.", td->td_name,
2011 td_get_sched(td)->ts_cpu));
2012 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2014 tdq_load_rem(tdq, td);
2016 * Do the lock dance required to avoid LOR. We have an
2017 * extra spinlock nesting from sched_switch() which will
2018 * prevent preemption while we're holding neither run-queue lock.
2022 tdq_add(tdn, td, flags);
2023 tdq_notify(tdn, td);
2027 return (TDQ_LOCKPTR(tdn));
2031 * thread_lock_unblock() that does not assume td_lock is blocked.
2034 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2036 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2041 * Switch threads. This function has to handle threads coming in while
2042 * blocked for some reason, running, or idle. It also must deal with
2043 * migrating a thread from one queue to another as running threads may
2044 * be assigned elsewhere via binding.
2047 sched_switch(struct thread *td, int flags)
2049 struct thread *newtd;
2051 struct td_sched *ts;
2054 int cpuid, preempted;
2056 THREAD_LOCK_ASSERT(td, MA_OWNED);
2058 cpuid = PCPU_GET(cpuid);
2060 ts = td_get_sched(td);
2061 sched_pctcpu_update(ts, 1);
2062 ts->ts_rltick = ticks;
2063 td->td_lastcpu = td->td_oncpu;
2064 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2065 (flags & SW_PREEMPT) != 0;
2066 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2067 td->td_owepreempt = 0;
2068 tdq->tdq_owepreempt = 0;
2069 if (!TD_IS_IDLETHREAD(td))
2070 tdq->tdq_switchcnt++;
2073 * Always block the thread lock so we can drop the tdq lock early.
2075 mtx = thread_lock_block(td);
2077 if (TD_IS_IDLETHREAD(td)) {
2078 MPASS(mtx == TDQ_LOCKPTR(tdq));
2080 } else if (TD_IS_RUNNING(td)) {
2081 MPASS(mtx == TDQ_LOCKPTR(tdq));
2082 srqflag = preempted ?
2083 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2084 SRQ_OURSELF|SRQ_YIELDING;
2086 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2087 ts->ts_cpu = sched_pickcpu(td, 0);
2089 if (ts->ts_cpu == cpuid)
2090 tdq_runq_add(tdq, td, srqflag);
2092 mtx = sched_switch_migrate(tdq, td, srqflag);
2094 /* This thread must be going to sleep. */
2095 if (mtx != TDQ_LOCKPTR(tdq)) {
2096 mtx_unlock_spin(mtx);
2099 tdq_load_rem(tdq, td);
2101 if (tdq->tdq_load == 0)
2106 #if (KTR_COMPILE & KTR_SCHED) != 0
2107 if (TD_IS_IDLETHREAD(td))
2108 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2109 "prio:%d", td->td_priority);
2111 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2112 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2113 "lockname:\"%s\"", td->td_lockname);
2117 * We enter here with the thread blocked and assigned to the
2118 * appropriate cpu run-queue or sleep-queue and with the current
2119 * thread-queue locked.
2121 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2122 newtd = choosethread();
2123 sched_pctcpu_update(td_get_sched(newtd), 0);
2127 * Call the MD code to switch contexts if necessary.
2131 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2132 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2134 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2136 #ifdef KDTRACE_HOOKS
2138 * If DTrace has set the active vtime enum to anything
2139 * other than INACTIVE (0), then it should have set the
2142 if (dtrace_vtime_active)
2143 (*dtrace_vtime_switch_func)(newtd);
2145 td->td_oncpu = NOCPU;
2146 cpu_switch(td, newtd, mtx);
2147 cpuid = td->td_oncpu = PCPU_GET(cpuid);
2149 SDT_PROBE0(sched, , , on__cpu);
2151 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2152 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2155 thread_unblock_switch(td, mtx);
2156 SDT_PROBE0(sched, , , remain__cpu);
2158 KASSERT(curthread->td_md.md_spinlock_count == 1,
2159 ("invalid count %d", curthread->td_md.md_spinlock_count));
2161 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2162 "prio:%d", td->td_priority);
2166 * Adjust thread priorities as a result of a nice request.
2169 sched_nice(struct proc *p, int nice)
2173 PROC_LOCK_ASSERT(p, MA_OWNED);
2176 FOREACH_THREAD_IN_PROC(p, td) {
2179 sched_prio(td, td->td_base_user_pri);
2185 * Record the sleep time for the interactivity scorer.
2188 sched_sleep(struct thread *td, int prio)
2191 THREAD_LOCK_ASSERT(td, MA_OWNED);
2193 td->td_slptick = ticks;
2194 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2195 td->td_flags |= TDF_CANSWAP;
2196 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2198 if (static_boost == 1 && prio)
2199 sched_prio(td, prio);
2200 else if (static_boost && td->td_priority > static_boost)
2201 sched_prio(td, static_boost);
2205 * Schedule a thread to resume execution and record how long it voluntarily
2206 * slept. We also update the pctcpu, interactivity, and priority.
2208 * Requires the thread lock on entry, drops on exit.
2211 sched_wakeup(struct thread *td, int srqflags)
2213 struct td_sched *ts;
2216 THREAD_LOCK_ASSERT(td, MA_OWNED);
2217 ts = td_get_sched(td);
2218 td->td_flags &= ~TDF_CANSWAP;
2221 * If we slept for more than a tick update our interactivity and
2224 slptick = td->td_slptick;
2226 if (slptick && slptick != ticks) {
2227 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2228 sched_interact_update(td);
2229 sched_pctcpu_update(ts, 0);
2232 * Reset the slice value since we slept and advanced the round-robin.
2235 sched_add(td, SRQ_BORING | srqflags);
2239 * Penalize the parent for creating a new child and initialize the child's
2243 sched_fork(struct thread *td, struct thread *child)
2245 THREAD_LOCK_ASSERT(td, MA_OWNED);
2246 sched_pctcpu_update(td_get_sched(td), 1);
2247 sched_fork_thread(td, child);
2249 * Penalize the parent and child for forking.
2251 sched_interact_fork(child);
2252 sched_priority(child);
2253 td_get_sched(td)->ts_runtime += tickincr;
2254 sched_interact_update(td);
2259 * Fork a new thread, may be within the same process.
2262 sched_fork_thread(struct thread *td, struct thread *child)
2264 struct td_sched *ts;
2265 struct td_sched *ts2;
2269 THREAD_LOCK_ASSERT(td, MA_OWNED);
2273 ts = td_get_sched(td);
2274 ts2 = td_get_sched(child);
2275 child->td_oncpu = NOCPU;
2276 child->td_lastcpu = NOCPU;
2277 child->td_lock = TDQ_LOCKPTR(tdq);
2278 child->td_cpuset = cpuset_ref(td->td_cpuset);
2279 child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2280 ts2->ts_cpu = ts->ts_cpu;
2283 * Grab our parents cpu estimation information.
2285 ts2->ts_ticks = ts->ts_ticks;
2286 ts2->ts_ltick = ts->ts_ltick;
2287 ts2->ts_ftick = ts->ts_ftick;
2289 * Do not inherit any borrowed priority from the parent.
2291 child->td_priority = child->td_base_pri;
2293 * And update interactivity score.
2295 ts2->ts_slptime = ts->ts_slptime;
2296 ts2->ts_runtime = ts->ts_runtime;
2297 /* Attempt to quickly learn interactivity. */
2298 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2300 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2305 * Adjust the priority class of a thread.
2308 sched_class(struct thread *td, int class)
2311 THREAD_LOCK_ASSERT(td, MA_OWNED);
2312 if (td->td_pri_class == class)
2314 td->td_pri_class = class;
2318 * Return some of the child's priority and interactivity to the parent.
2321 sched_exit(struct proc *p, struct thread *child)
2325 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2326 "prio:%d", child->td_priority);
2327 PROC_LOCK_ASSERT(p, MA_OWNED);
2328 td = FIRST_THREAD_IN_PROC(p);
2329 sched_exit_thread(td, child);
2333 * Penalize another thread for the time spent on this one. This helps to
2334 * worsen the priority and interactivity of processes which schedule batch
2335 * jobs such as make. This has little effect on the make process itself but
2336 * causes new processes spawned by it to receive worse scores immediately.
2339 sched_exit_thread(struct thread *td, struct thread *child)
2342 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2343 "prio:%d", child->td_priority);
2345 * Give the child's runtime to the parent without returning the
2346 * sleep time as a penalty to the parent. This causes shells that
2347 * launch expensive things to mark their children as expensive.
2350 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2351 sched_interact_update(td);
2357 sched_preempt(struct thread *td)
2362 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2366 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2367 if (td->td_priority > tdq->tdq_lowpri) {
2368 if (td->td_critnest == 1) {
2369 flags = SW_INVOL | SW_PREEMPT;
2370 flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
2373 /* Switch dropped thread lock. */
2376 td->td_owepreempt = 1;
2378 tdq->tdq_owepreempt = 0;
2384 * Fix priorities on return to user-space. Priorities may be elevated due
2385 * to static priorities in msleep() or similar.
2388 sched_userret_slowpath(struct thread *td)
2392 td->td_priority = td->td_user_pri;
2393 td->td_base_pri = td->td_user_pri;
2394 tdq_setlowpri(TDQ_SELF(), td);
2399 * Handle a stathz tick. This is really only relevant for timeshare
2403 sched_clock(struct thread *td, int cnt)
2406 struct td_sched *ts;
2408 THREAD_LOCK_ASSERT(td, MA_OWNED);
2412 * We run the long term load balancer infrequently on the first cpu.
2414 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 &&
2415 balance_ticks != 0) {
2416 balance_ticks -= cnt;
2417 if (balance_ticks <= 0)
2422 * Save the old switch count so we have a record of the last ticks
2423 * activity. Initialize the new switch count based on our load.
2424 * If there is some activity seed it to reflect that.
2426 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2427 tdq->tdq_switchcnt = tdq->tdq_load;
2429 * Advance the insert index once for each tick to ensure that all
2430 * threads get a chance to run.
2432 if (tdq->tdq_idx == tdq->tdq_ridx) {
2433 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2434 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2435 tdq->tdq_ridx = tdq->tdq_idx;
2437 ts = td_get_sched(td);
2438 sched_pctcpu_update(ts, 1);
2439 if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td))
2442 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2444 * We used a tick; charge it to the thread so
2445 * that we can compute our interactivity.
2447 td_get_sched(td)->ts_runtime += tickincr * cnt;
2448 sched_interact_update(td);
2453 * Force a context switch if the current thread has used up a full
2454 * time slice (default is 100ms).
2456 ts->ts_slice += cnt;
2457 if (ts->ts_slice >= tdq_slice(tdq)) {
2459 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2464 sched_estcpu(struct thread *td __unused)
2471 * Return whether the current CPU has runnable tasks. Used for in-kernel
2472 * cooperative idle threads.
2475 sched_runnable(void)
2483 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2484 if (tdq->tdq_load > 0)
2487 if (tdq->tdq_load - 1 > 0)
2495 * Choose the highest priority thread to run. The thread is removed from
2496 * the run-queue while running however the load remains. For SMP we set
2497 * the tdq in the global idle bitmask if it idles here.
2506 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2507 td = tdq_choose(tdq);
2509 tdq_runq_rem(tdq, td);
2510 tdq->tdq_lowpri = td->td_priority;
2513 tdq->tdq_lowpri = PRI_MAX_IDLE;
2514 return (PCPU_GET(idlethread));
2518 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2519 * we always request it once we exit a critical section.
2522 sched_setpreempt(struct thread *td)
2528 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2531 pri = td->td_priority;
2532 cpri = ctd->td_priority;
2534 ctd->td_flags |= TDF_NEEDRESCHED;
2535 if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2537 if (!sched_shouldpreempt(pri, cpri, 0))
2539 ctd->td_owepreempt = 1;
2543 * Add a thread to a thread queue. Select the appropriate runq and add the
2544 * thread to it. This is the internal function called when the tdq is
2548 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2551 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2552 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
2553 KASSERT((td->td_inhibitors == 0),
2554 ("sched_add: trying to run inhibited thread"));
2555 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2556 ("sched_add: bad thread state"));
2557 KASSERT(td->td_flags & TDF_INMEM,
2558 ("sched_add: thread swapped out"));
2560 if (td->td_priority < tdq->tdq_lowpri)
2561 tdq->tdq_lowpri = td->td_priority;
2562 tdq_runq_add(tdq, td, flags);
2563 tdq_load_add(tdq, td);
2567 * Select the target thread queue and add a thread to it. Request
2568 * preemption or IPI a remote processor if required.
2570 * Requires the thread lock on entry, drops on exit.
2573 sched_add(struct thread *td, int flags)
2580 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2581 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2582 sched_tdname(curthread));
2583 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2584 KTR_ATTR_LINKED, sched_tdname(td));
2585 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2586 flags & SRQ_PREEMPTED);
2587 THREAD_LOCK_ASSERT(td, MA_OWNED);
2589 * Recalculate the priority before we select the target cpu or
2592 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2596 * Pick the destination cpu and if it isn't ours transfer to the
2599 cpu = sched_pickcpu(td, flags);
2600 tdq = sched_setcpu(td, cpu, flags);
2601 tdq_add(tdq, td, flags);
2602 if (cpu != PCPU_GET(cpuid))
2603 tdq_notify(tdq, td);
2604 else if (!(flags & SRQ_YIELDING))
2605 sched_setpreempt(td);
2610 * Now that the thread is moving to the run-queue, set the lock
2611 * to the scheduler's lock.
2613 if ((flags & SRQ_HOLD) != 0)
2614 td->td_lock = TDQ_LOCKPTR(tdq);
2616 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2617 tdq_add(tdq, td, flags);
2618 if (!(flags & SRQ_YIELDING))
2619 sched_setpreempt(td);
2621 if (!(flags & SRQ_HOLDTD))
2626 * Remove a thread from a run-queue without running it. This is used
2627 * when we're stealing a thread from a remote queue. Otherwise all threads
2628 * exit by calling sched_exit_thread() and sched_throw() themselves.
2631 sched_rem(struct thread *td)
2635 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2636 "prio:%d", td->td_priority);
2637 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2638 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2639 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2640 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2641 KASSERT(TD_ON_RUNQ(td),
2642 ("sched_rem: thread not on run queue"));
2643 tdq_runq_rem(tdq, td);
2644 tdq_load_rem(tdq, td);
2646 if (td->td_priority == tdq->tdq_lowpri)
2647 tdq_setlowpri(tdq, NULL);
2651 * Fetch cpu utilization information. Updates on demand.
2654 sched_pctcpu(struct thread *td)
2657 struct td_sched *ts;
2660 ts = td_get_sched(td);
2662 THREAD_LOCK_ASSERT(td, MA_OWNED);
2663 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2667 /* How many rtick per second ? */
2668 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2669 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2676 * Enforce affinity settings for a thread. Called after adjustments to
2680 sched_affinity(struct thread *td)
2683 struct td_sched *ts;
2685 THREAD_LOCK_ASSERT(td, MA_OWNED);
2686 ts = td_get_sched(td);
2687 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2689 if (TD_ON_RUNQ(td)) {
2691 sched_add(td, SRQ_BORING | SRQ_HOLDTD);
2694 if (!TD_IS_RUNNING(td))
2697 * Force a switch before returning to userspace. If the
2698 * target thread is not running locally send an ipi to force
2701 td->td_flags |= TDF_NEEDRESCHED;
2702 if (td != curthread)
2703 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2708 * Bind a thread to a target cpu.
2711 sched_bind(struct thread *td, int cpu)
2713 struct td_sched *ts;
2715 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2716 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2717 ts = td_get_sched(td);
2718 if (ts->ts_flags & TSF_BOUND)
2720 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2721 ts->ts_flags |= TSF_BOUND;
2723 if (PCPU_GET(cpuid) == cpu)
2726 /* When we return from mi_switch we'll be on the correct cpu. */
2732 * Release a bound thread.
2735 sched_unbind(struct thread *td)
2737 struct td_sched *ts;
2739 THREAD_LOCK_ASSERT(td, MA_OWNED);
2740 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2741 ts = td_get_sched(td);
2742 if ((ts->ts_flags & TSF_BOUND) == 0)
2744 ts->ts_flags &= ~TSF_BOUND;
2749 sched_is_bound(struct thread *td)
2751 THREAD_LOCK_ASSERT(td, MA_OWNED);
2752 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2759 sched_relinquish(struct thread *td)
2762 mi_switch(SW_VOL | SWT_RELINQUISH);
2766 * Return the total system load.
2777 total += TDQ_CPU(i)->tdq_sysload;
2780 return (TDQ_SELF()->tdq_sysload);
2785 sched_sizeof_proc(void)
2787 return (sizeof(struct proc));
2791 sched_sizeof_thread(void)
2793 return (sizeof(struct thread) + sizeof(struct td_sched));
2797 #define TDQ_IDLESPIN(tdq) \
2798 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2800 #define TDQ_IDLESPIN(tdq) 1
2804 * The actual idle process.
2807 sched_idletd(void *dummy)
2811 int oldswitchcnt, switchcnt;
2814 mtx_assert(&Giant, MA_NOTOWNED);
2817 THREAD_NO_SLEEPING();
2820 if (tdq->tdq_load) {
2822 mi_switch(SW_VOL | SWT_IDLE);
2824 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2826 if (always_steal || switchcnt != oldswitchcnt) {
2827 oldswitchcnt = switchcnt;
2828 if (tdq_idled(tdq) == 0)
2831 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2833 oldswitchcnt = switchcnt;
2836 * If we're switching very frequently, spin while checking
2837 * for load rather than entering a low power state that
2838 * may require an IPI. However, don't do any busy
2839 * loops while on SMT machines as this simply steals
2840 * cycles from cores doing useful work.
2842 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2843 for (i = 0; i < sched_idlespins; i++) {
2850 /* If there was context switch during spin, restart it. */
2851 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2852 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2855 /* Run main MD idle handler. */
2856 tdq->tdq_cpu_idle = 1;
2858 * Make sure that tdq_cpu_idle update is globally visible
2859 * before cpu_idle() read tdq_load. The order is important
2860 * to avoid race with tdq_notify.
2862 atomic_thread_fence_seq_cst();
2864 * Checking for again after the fence picks up assigned
2865 * threads often enough to make it worthwhile to do so in
2866 * order to avoid calling cpu_idle().
2868 if (tdq->tdq_load != 0) {
2869 tdq->tdq_cpu_idle = 0;
2872 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2873 tdq->tdq_cpu_idle = 0;
2876 * Account thread-less hardware interrupts and
2877 * other wakeup reasons equal to context switches.
2879 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2880 if (switchcnt != oldswitchcnt)
2882 tdq->tdq_switchcnt++;
2888 * A CPU is entering for the first time or a thread is exiting.
2891 sched_throw(struct thread *td)
2893 struct thread *newtd;
2896 if (__predict_false(td == NULL)) {
2898 PCPU_SET(sched, DPCPU_PTR(tdq));
2900 /* Correct spinlock nesting and acquire the correct lock. */
2904 PCPU_SET(switchtime, cpu_ticks());
2905 PCPU_SET(switchticks, ticks);
2906 PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(tdq);
2909 THREAD_LOCK_ASSERT(td, MA_OWNED);
2910 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq));
2911 tdq_load_rem(tdq, td);
2912 td->td_lastcpu = td->td_oncpu;
2913 td->td_oncpu = NOCPU;
2914 thread_lock_block(td);
2916 newtd = choosethread();
2919 KASSERT(curthread->td_md.md_spinlock_count == 1,
2920 ("invalid count %d", curthread->td_md.md_spinlock_count));
2921 /* doesn't return */
2922 if (__predict_false(td == NULL))
2923 cpu_throw(td, newtd); /* doesn't return */
2925 cpu_switch(td, newtd, TDQ_LOCKPTR(tdq));
2929 * This is called from fork_exit(). Just acquire the correct locks and
2930 * let fork do the rest of the work.
2933 sched_fork_exit(struct thread *td)
2939 * Finish setting up thread glue so that it begins execution in a
2940 * non-nested critical section with the scheduler lock held.
2942 KASSERT(curthread->td_md.md_spinlock_count == 1,
2943 ("invalid count %d", curthread->td_md.md_spinlock_count));
2944 cpuid = PCPU_GET(cpuid);
2948 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2949 td->td_oncpu = cpuid;
2950 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2951 "prio:%d", td->td_priority);
2952 SDT_PROBE0(sched, , , on__cpu);
2956 * Create on first use to catch odd startup conditons.
2959 sched_tdname(struct thread *td)
2962 struct td_sched *ts;
2964 ts = td_get_sched(td);
2965 if (ts->ts_name[0] == '\0')
2966 snprintf(ts->ts_name, sizeof(ts->ts_name),
2967 "%s tid %d", td->td_name, td->td_tid);
2968 return (ts->ts_name);
2970 return (td->td_name);
2976 sched_clear_tdname(struct thread *td)
2978 struct td_sched *ts;
2980 ts = td_get_sched(td);
2981 ts->ts_name[0] = '\0';
2988 * Build the CPU topology dump string. Is recursively called to collect
2989 * the topology tree.
2992 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2995 char cpusetbuf[CPUSETBUFSIZ];
2998 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2999 "", 1 + indent / 2, cg->cg_level);
3000 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3001 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3003 for (i = 0; i < MAXCPU; i++) {
3004 if (CPU_ISSET(i, &cg->cg_mask)) {
3006 sbuf_printf(sb, ", ");
3009 sbuf_printf(sb, "%d", i);
3012 sbuf_printf(sb, "</cpu>\n");
3014 if (cg->cg_flags != 0) {
3015 sbuf_printf(sb, "%*s <flags>", indent, "");
3016 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3017 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3018 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3019 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3020 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3021 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3022 sbuf_printf(sb, "</flags>\n");
3025 if (cg->cg_children > 0) {
3026 sbuf_printf(sb, "%*s <children>\n", indent, "");
3027 for (i = 0; i < cg->cg_children; i++)
3028 sysctl_kern_sched_topology_spec_internal(sb,
3029 &cg->cg_child[i], indent+2);
3030 sbuf_printf(sb, "%*s </children>\n", indent, "");
3032 sbuf_printf(sb, "%*s</group>\n", indent, "");
3037 * Sysctl handler for retrieving topology dump. It's a wrapper for
3038 * the recursive sysctl_kern_smp_topology_spec_internal().
3041 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3046 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3048 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3052 sbuf_printf(topo, "<groups>\n");
3053 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3054 sbuf_printf(topo, "</groups>\n");
3057 err = sbuf_finish(topo);
3066 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3068 int error, new_val, period;
3070 period = 1000000 / realstathz;
3071 new_val = period * sched_slice;
3072 error = sysctl_handle_int(oidp, &new_val, 0, req);
3073 if (error != 0 || req->newptr == NULL)
3077 sched_slice = imax(1, (new_val + period / 2) / period);
3078 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3079 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3084 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3085 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3087 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3088 NULL, 0, sysctl_kern_quantum, "I",
3089 "Quantum for timeshare threads in microseconds");
3090 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3091 "Quantum for timeshare threads in stathz ticks");
3092 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3093 "Interactivity score threshold");
3094 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3096 "Maximal (lowest) priority for preemption");
3097 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3098 "Assign static kernel priorities to sleeping threads");
3099 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3100 "Number of times idle thread will spin waiting for new work");
3101 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3102 &sched_idlespinthresh, 0,
3103 "Threshold before we will permit idle thread spinning");
3105 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3106 "Number of hz ticks to keep thread affinity for");
3107 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3108 "Enables the long-term load balancer");
3109 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3110 &balance_interval, 0,
3111 "Average period in stathz ticks to run the long-term balancer");
3112 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3113 "Attempts to steal work from other cores before idling");
3114 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3115 "Minimum load on remote CPU before we'll steal");
3116 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3117 0, "Topological distance limit for stealing threads in sched_switch()");
3118 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3119 "Always run the stealer from the idle thread");
3120 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3121 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3122 "XML dump of detected CPU topology");
3125 /* ps compat. All cpu percentages from ULE are weighted. */
3126 static int ccpu = 0;
3127 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");