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_PICKCPU TDF_SCHED0 /* Thread should pick new CPU. */
200 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
203 * tickincr: Converts a stathz tick into a hz domain scaled by
204 * the shift factor. Without the shift the error rate
205 * due to rounding would be unacceptably high.
206 * realstathz: stathz is sometimes 0 and run off of hz.
207 * sched_slice: Runtime of each thread before rescheduling.
208 * preempt_thresh: Priority threshold for preemption and remote IPIs.
210 static u_int __read_mostly sched_interact = SCHED_INTERACT_THRESH;
211 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
212 static int __read_mostly realstathz = 127; /* reset during boot. */
213 static int __read_mostly sched_slice = 10; /* reset during boot. */
214 static int __read_mostly sched_slice_min = 1; /* reset during boot. */
216 #ifdef FULL_PREEMPTION
217 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
219 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
222 static int __read_mostly preempt_thresh = 0;
224 static int __read_mostly static_boost = PRI_MIN_BATCH;
225 static int __read_mostly sched_idlespins = 10000;
226 static int __read_mostly sched_idlespinthresh = -1;
229 * tdq - per processor runqs and statistics. All fields are protected by the
230 * tdq_lock. The load and lowpri may be accessed without to avoid excess
231 * locking in sched_pickcpu();
235 * Ordered to improve efficiency of cpu_search() and switch().
236 * tdq_lock is padded to avoid false sharing with tdq_load and
239 struct mtx_padalign tdq_lock; /* run queue lock. */
240 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
241 volatile int tdq_load; /* Aggregate load. */
242 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
243 int tdq_sysload; /* For loadavg, !ITHD load. */
244 volatile int tdq_transferable; /* Transferable thread count. */
245 volatile short tdq_switchcnt; /* Switches this tick. */
246 volatile short tdq_oldswitchcnt; /* Switches last tick. */
247 u_char tdq_lowpri; /* Lowest priority thread. */
248 u_char tdq_owepreempt; /* Remote preemption pending. */
249 u_char tdq_idx; /* Current insert index. */
250 u_char tdq_ridx; /* Current removal index. */
251 int tdq_id; /* cpuid. */
252 struct runq tdq_realtime; /* real-time run queue. */
253 struct runq tdq_timeshare; /* timeshare run queue. */
254 struct runq tdq_idle; /* Queue of IDLE threads. */
255 char tdq_name[TDQ_NAME_LEN];
257 char tdq_loadname[TDQ_LOADNAME_LEN];
261 /* Idle thread states and config. */
262 #define TDQ_RUNNING 1
266 struct cpu_group __read_mostly *cpu_top; /* CPU topology */
268 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
269 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
274 static int rebalance = 1;
275 static int balance_interval = 128; /* Default set in sched_initticks(). */
276 static int __read_mostly affinity;
277 static int __read_mostly steal_idle = 1;
278 static int __read_mostly steal_thresh = 2;
279 static int __read_mostly always_steal = 0;
280 static int __read_mostly trysteal_limit = 2;
283 * One thread queue per processor.
285 static struct tdq __read_mostly *balance_tdq;
286 static int balance_ticks;
287 DPCPU_DEFINE_STATIC(struct tdq, tdq);
288 DPCPU_DEFINE_STATIC(uint32_t, randomval);
290 #define TDQ_SELF() ((struct tdq *)PCPU_GET(sched))
291 #define TDQ_CPU(x) (DPCPU_ID_PTR((x), tdq))
292 #define TDQ_ID(x) ((x)->tdq_id)
294 static struct tdq tdq_cpu;
296 #define TDQ_ID(x) (0)
297 #define TDQ_SELF() (&tdq_cpu)
298 #define TDQ_CPU(x) (&tdq_cpu)
301 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
302 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
303 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
304 #define TDQ_TRYLOCK(t) mtx_trylock_spin(TDQ_LOCKPTR((t)))
305 #define TDQ_TRYLOCK_FLAGS(t, f) mtx_trylock_spin_flags(TDQ_LOCKPTR((t)), (f))
306 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
307 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
309 static void sched_priority(struct thread *);
310 static void sched_thread_priority(struct thread *, u_char);
311 static int sched_interact_score(struct thread *);
312 static void sched_interact_update(struct thread *);
313 static void sched_interact_fork(struct thread *);
314 static void sched_pctcpu_update(struct td_sched *, int);
316 /* Operations on per processor queues */
317 static struct thread *tdq_choose(struct tdq *);
318 static void tdq_setup(struct tdq *, int i);
319 static void tdq_load_add(struct tdq *, struct thread *);
320 static void tdq_load_rem(struct tdq *, struct thread *);
321 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
322 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
323 static inline int sched_shouldpreempt(int, int, int);
324 void tdq_print(int cpu);
325 static void runq_print(struct runq *rq);
326 static void tdq_add(struct tdq *, struct thread *, int);
328 static struct thread *tdq_move(struct tdq *, struct tdq *);
329 static int tdq_idled(struct tdq *);
330 static void tdq_notify(struct tdq *, struct thread *);
331 static struct thread *tdq_steal(struct tdq *, int);
332 static struct thread *runq_steal(struct runq *, int);
333 static int sched_pickcpu(struct thread *, int);
334 static void sched_balance(void);
335 static int sched_balance_pair(struct tdq *, struct tdq *);
336 static inline struct tdq *sched_setcpu(struct thread *, int, int);
337 static inline void thread_unblock_switch(struct thread *, struct mtx *);
338 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
339 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
340 struct cpu_group *cg, int indent);
343 static void sched_setup(void *dummy);
344 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
346 static void sched_initticks(void *dummy);
347 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
350 SDT_PROVIDER_DEFINE(sched);
352 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
353 "struct proc *", "uint8_t");
354 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
355 "struct proc *", "void *");
356 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
357 "struct proc *", "void *", "int");
358 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
359 "struct proc *", "uint8_t", "struct thread *");
360 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
361 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
363 SDT_PROBE_DEFINE(sched, , , on__cpu);
364 SDT_PROBE_DEFINE(sched, , , remain__cpu);
365 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
369 * Print the threads waiting on a run-queue.
372 runq_print(struct runq *rq)
380 for (i = 0; i < RQB_LEN; i++) {
381 printf("\t\trunq bits %d 0x%zx\n",
382 i, rq->rq_status.rqb_bits[i]);
383 for (j = 0; j < RQB_BPW; j++)
384 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
385 pri = j + (i << RQB_L2BPW);
386 rqh = &rq->rq_queues[pri];
387 TAILQ_FOREACH(td, rqh, td_runq) {
388 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
389 td, td->td_name, td->td_priority,
390 td->td_rqindex, pri);
397 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
406 printf("tdq %d:\n", TDQ_ID(tdq));
407 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
408 printf("\tLock name: %s\n", tdq->tdq_name);
409 printf("\tload: %d\n", tdq->tdq_load);
410 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
411 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
412 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
413 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
414 printf("\tload transferable: %d\n", tdq->tdq_transferable);
415 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
416 printf("\trealtime runq:\n");
417 runq_print(&tdq->tdq_realtime);
418 printf("\ttimeshare runq:\n");
419 runq_print(&tdq->tdq_timeshare);
420 printf("\tidle runq:\n");
421 runq_print(&tdq->tdq_idle);
425 sched_shouldpreempt(int pri, int cpri, int remote)
428 * If the new priority is not better than the current priority there is
434 * Always preempt idle.
436 if (cpri >= PRI_MIN_IDLE)
439 * If preemption is disabled don't preempt others.
441 if (preempt_thresh == 0)
444 * Preempt if we exceed the threshold.
446 if (pri <= preempt_thresh)
449 * If we're interactive or better and there is non-interactive
450 * or worse running preempt only remote processors.
452 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
458 * Add a thread to the actual run-queue. Keeps transferable counts up to
459 * date with what is actually on the run-queue. Selects the correct
460 * queue position for timeshare threads.
463 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
468 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
469 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
471 pri = td->td_priority;
472 ts = td_get_sched(td);
474 if (THREAD_CAN_MIGRATE(td)) {
475 tdq->tdq_transferable++;
476 ts->ts_flags |= TSF_XFERABLE;
478 if (pri < PRI_MIN_BATCH) {
479 ts->ts_runq = &tdq->tdq_realtime;
480 } else if (pri <= PRI_MAX_BATCH) {
481 ts->ts_runq = &tdq->tdq_timeshare;
482 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
483 ("Invalid priority %d on timeshare runq", pri));
485 * This queue contains only priorities between MIN and MAX
486 * realtime. Use the whole queue to represent these values.
488 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
489 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
490 pri = (pri + tdq->tdq_idx) % RQ_NQS;
492 * This effectively shortens the queue by one so we
493 * can have a one slot difference between idx and
494 * ridx while we wait for threads to drain.
496 if (tdq->tdq_ridx != tdq->tdq_idx &&
497 pri == tdq->tdq_ridx)
498 pri = (unsigned char)(pri - 1) % RQ_NQS;
501 runq_add_pri(ts->ts_runq, td, pri, flags);
504 ts->ts_runq = &tdq->tdq_idle;
505 runq_add(ts->ts_runq, td, flags);
509 * Remove a thread from a run-queue. This typically happens when a thread
510 * is selected to run. Running threads are not on the queue and the
511 * transferable count does not reflect them.
514 tdq_runq_rem(struct tdq *tdq, struct thread *td)
518 ts = td_get_sched(td);
519 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
520 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
521 KASSERT(ts->ts_runq != NULL,
522 ("tdq_runq_remove: thread %p null ts_runq", td));
523 if (ts->ts_flags & TSF_XFERABLE) {
524 tdq->tdq_transferable--;
525 ts->ts_flags &= ~TSF_XFERABLE;
527 if (ts->ts_runq == &tdq->tdq_timeshare) {
528 if (tdq->tdq_idx != tdq->tdq_ridx)
529 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
531 runq_remove_idx(ts->ts_runq, td, NULL);
533 runq_remove(ts->ts_runq, td);
537 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
538 * for this thread to the referenced thread queue.
541 tdq_load_add(struct tdq *tdq, struct thread *td)
544 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
545 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
548 if ((td->td_flags & TDF_NOLOAD) == 0)
550 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
551 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
555 * Remove the load from a thread that is transitioning to a sleep state or
559 tdq_load_rem(struct tdq *tdq, struct thread *td)
562 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
563 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
564 KASSERT(tdq->tdq_load != 0,
565 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
568 if ((td->td_flags & TDF_NOLOAD) == 0)
570 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
571 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
575 * Bound timeshare latency by decreasing slice size as load increases. We
576 * consider the maximum latency as the sum of the threads waiting to run
577 * aside from curthread and target no more than sched_slice latency but
578 * no less than sched_slice_min runtime.
581 tdq_slice(struct tdq *tdq)
586 * It is safe to use sys_load here because this is called from
587 * contexts where timeshare threads are running and so there
588 * cannot be higher priority load in the system.
590 load = tdq->tdq_sysload - 1;
591 if (load >= SCHED_SLICE_MIN_DIVISOR)
592 return (sched_slice_min);
594 return (sched_slice);
595 return (sched_slice / load);
599 * Set lowpri to its exact value by searching the run-queue and
600 * evaluating curthread. curthread may be passed as an optimization.
603 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
607 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
609 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
610 td = tdq_choose(tdq);
611 if (td == NULL || td->td_priority > ctd->td_priority)
612 tdq->tdq_lowpri = ctd->td_priority;
614 tdq->tdq_lowpri = td->td_priority;
619 * We need some randomness. Implement a classic Linear Congruential
620 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
621 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
622 * of the random state (in the low bits of our answer) to keep
623 * the maximum randomness.
630 rndptr = DPCPU_PTR(randomval);
631 *rndptr = *rndptr * 69069 + 5;
633 return (*rndptr >> 16);
637 cpuset_t *cs_mask; /* The mask of allowed CPUs to choose from. */
638 int cs_prefer; /* Prefer this CPU and groups including it. */
639 int cs_running; /* The thread is now running at cs_prefer. */
640 int cs_pri; /* Min priority for low. */
641 int cs_load; /* Max load for low, min load for high. */
642 int cs_trans; /* Min transferable load for high. */
645 struct cpu_search_res {
646 int csr_cpu; /* The best CPU found. */
647 int csr_load; /* The load of cs_cpu. */
651 * Search the tree of cpu_groups for the lowest or highest loaded CPU.
652 * These routines actually compare the load on all paths through the tree
653 * and find the least loaded cpu on the least loaded path, which may differ
654 * from the least loaded cpu in the system. This balances work among caches
658 cpu_search_lowest(const struct cpu_group *cg, const struct cpu_search *s,
659 struct cpu_search_res *r)
661 struct cpu_search_res lr;
663 int c, bload, l, load, p, total;
669 /* Loop through children CPU groups if there are any. */
670 if (cg->cg_children > 0) {
671 for (c = cg->cg_children - 1; c >= 0; c--) {
672 load = cpu_search_lowest(&cg->cg_child[c], s, &lr);
676 * When balancing do not prefer SMT groups with load >1.
677 * It allows round-robin between SMT groups with equal
678 * load within parent group for more fair scheduling.
680 if (__predict_false(s->cs_running) &&
681 (cg->cg_child[c].cg_flags & CG_FLAG_THREAD) &&
682 load >= 128 && (load & 128) != 0)
685 if (lr.csr_cpu >= 0 && (load < bload ||
686 (load == bload && lr.csr_load < r->csr_load))) {
688 r->csr_cpu = lr.csr_cpu;
689 r->csr_load = lr.csr_load;
695 /* Loop through children CPUs otherwise. */
696 for (c = cg->cg_last; c >= cg->cg_first; c--) {
697 if (!CPU_ISSET(c, &cg->cg_mask))
701 if (c == s->cs_prefer) {
702 if (__predict_false(s->cs_running))
711 * Check this CPU is acceptable.
712 * If the threads is already on the CPU, don't look on the TDQ
713 * priority, since it can be the priority of the thread itself.
715 if (l > s->cs_load || (tdq->tdq_lowpri <= s->cs_pri &&
716 (!s->cs_running || c != s->cs_prefer)) ||
717 !CPU_ISSET(c, s->cs_mask))
721 * When balancing do not prefer CPUs with load > 1.
722 * It allows round-robin between CPUs with equal load
723 * within the CPU group for more fair scheduling.
725 if (__predict_false(s->cs_running) && l > 0)
728 load -= sched_random() % 128;
729 if (bload > load - p) {
739 cpu_search_highest(const struct cpu_group *cg, const struct cpu_search *s,
740 struct cpu_search_res *r)
742 struct cpu_search_res lr;
744 int c, bload, l, load, total;
750 /* Loop through children CPU groups if there are any. */
751 if (cg->cg_children > 0) {
752 for (c = cg->cg_children - 1; c >= 0; c--) {
753 load = cpu_search_highest(&cg->cg_child[c], s, &lr);
755 if (lr.csr_cpu >= 0 && (load > bload ||
756 (load == bload && lr.csr_load > r->csr_load))) {
758 r->csr_cpu = lr.csr_cpu;
759 r->csr_load = lr.csr_load;
765 /* Loop through children CPUs otherwise. */
766 for (c = cg->cg_last; c >= cg->cg_first; c--) {
767 if (!CPU_ISSET(c, &cg->cg_mask))
775 * Check this CPU is acceptable.
777 if (l < s->cs_load || (tdq->tdq_transferable < s->cs_trans) ||
778 !CPU_ISSET(c, s->cs_mask))
781 load -= sched_random() % 256;
792 * Find the cpu with the least load via the least loaded path that has a
793 * lowpri greater than pri pri. A pri of -1 indicates any priority is
797 sched_lowest(const struct cpu_group *cg, cpuset_t *mask, int pri, int maxload,
798 int prefer, int running)
801 struct cpu_search_res r;
803 s.cs_prefer = prefer;
804 s.cs_running = running;
808 cpu_search_lowest(cg, &s, &r);
813 * Find the cpu with the highest load via the highest loaded path.
816 sched_highest(const struct cpu_group *cg, cpuset_t *mask, int minload,
820 struct cpu_search_res r;
824 s.cs_trans = mintrans;
825 cpu_search_highest(cg, &s, &r);
830 sched_balance_group(struct cpu_group *cg)
834 cpuset_t hmask, lmask;
835 int high, low, anylow;
839 high = sched_highest(cg, &hmask, 1, 0);
840 /* Stop if there is no more CPU with transferrable threads. */
843 CPU_CLR(high, &hmask);
844 CPU_COPY(&hmask, &lmask);
845 /* Stop if there is no more CPU left for low. */
846 if (CPU_EMPTY(&lmask))
849 if (tdq->tdq_load == 1) {
851 * There is only one running thread. We can't move
852 * it from here, so tell it to pick new CPU by itself.
855 td = pcpu_find(high)->pc_curthread;
856 if ((td->td_flags & TDF_IDLETD) == 0 &&
857 THREAD_CAN_MIGRATE(td)) {
858 td->td_flags |= TDF_NEEDRESCHED | TDF_PICKCPU;
860 ipi_cpu(high, IPI_AST);
867 if (tdq->tdq_transferable == 0)
869 low = sched_lowest(cg, &lmask, -1, tdq->tdq_load - 1, high, 1);
870 /* Stop if we looked well and found no less loaded CPU. */
871 if (anylow && low == -1)
873 /* Go to next high if we found no less loaded CPU. */
876 /* Transfer thread from high to low. */
877 if (sched_balance_pair(tdq, TDQ_CPU(low))) {
878 /* CPU that got thread can no longer be a donor. */
879 CPU_CLR(low, &hmask);
882 * If failed, then there is no threads on high
883 * that can run on this low. Drop low from low
884 * mask and look for different one.
886 CPU_CLR(low, &lmask);
898 balance_ticks = max(balance_interval / 2, 1) +
899 (sched_random() % balance_interval);
902 sched_balance_group(cpu_top);
907 * Lock two thread queues using their address to maintain lock order.
910 tdq_lock_pair(struct tdq *one, struct tdq *two)
914 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
917 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
922 * Unlock two thread queues. Order is not important here.
925 tdq_unlock_pair(struct tdq *one, struct tdq *two)
932 * Transfer load between two imbalanced thread queues.
935 sched_balance_pair(struct tdq *high, struct tdq *low)
940 tdq_lock_pair(high, low);
943 * Transfer a thread from high to low.
945 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
946 (td = tdq_move(high, low)) != NULL) {
948 * In case the target isn't the current cpu notify it of the
949 * new load, possibly sending an IPI to force it to reschedule.
952 if (cpu != PCPU_GET(cpuid))
955 tdq_unlock_pair(high, low);
960 * Move a thread from one thread queue to another.
962 static struct thread *
963 tdq_move(struct tdq *from, struct tdq *to)
969 TDQ_LOCK_ASSERT(from, MA_OWNED);
970 TDQ_LOCK_ASSERT(to, MA_OWNED);
974 td = tdq_steal(tdq, cpu);
979 * Although the run queue is locked the thread may be
980 * blocked. We can not set the lock until it is unblocked.
982 thread_lock_block_wait(td);
984 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from));
985 td->td_lock = TDQ_LOCKPTR(to);
986 td_get_sched(td)->ts_cpu = cpu;
987 tdq_add(to, td, SRQ_YIELDING);
993 * This tdq has idled. Try to steal a thread from another cpu and switch
997 tdq_idled(struct tdq *tdq)
999 struct cpu_group *cg, *parent;
1002 int cpu, switchcnt, goup;
1004 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
1007 CPU_CLR(PCPU_GET(cpuid), &mask);
1009 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
1010 for (cg = tdq->tdq_cg, goup = 0; ; ) {
1011 cpu = sched_highest(cg, &mask, steal_thresh, 1);
1013 * We were assigned a thread but not preempted. Returning
1014 * 0 here will cause our caller to switch to it.
1020 * We found no CPU to steal from in this group. Escalate to
1021 * the parent and repeat. But if parent has only two children
1022 * groups we can avoid searching this group again by searching
1023 * the other one specifically and then escalating two levels.
1030 parent = cg->cg_parent;
1033 if (parent->cg_children == 2) {
1034 if (cg == &parent->cg_child[0])
1035 cg = &parent->cg_child[1];
1037 cg = &parent->cg_child[0];
1043 steal = TDQ_CPU(cpu);
1045 * The data returned by sched_highest() is stale and
1046 * the chosen CPU no longer has an eligible thread.
1048 * Testing this ahead of tdq_lock_pair() only catches
1049 * this situation about 20% of the time on an 8 core
1050 * 16 thread Ryzen 7, but it still helps performance.
1052 if (steal->tdq_load < steal_thresh ||
1053 steal->tdq_transferable == 0)
1056 * Try to lock both queues. If we are assigned a thread while
1057 * waited for the lock, switch to it now instead of stealing.
1058 * If we can't get the lock, then somebody likely got there
1059 * first so continue searching.
1062 if (tdq->tdq_load > 0) {
1063 mi_switch(SW_VOL | SWT_IDLE);
1066 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) {
1068 CPU_CLR(cpu, &mask);
1072 * The data returned by sched_highest() is stale and
1073 * the chosen CPU no longer has an eligible thread, or
1074 * we were preempted and the CPU loading info may be out
1075 * of date. The latter is rare. In either case restart
1078 if (steal->tdq_load < steal_thresh ||
1079 steal->tdq_transferable == 0 ||
1080 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1081 tdq_unlock_pair(tdq, steal);
1085 * Steal the thread and switch to it.
1087 if (tdq_move(steal, tdq) != NULL)
1090 * We failed to acquire a thread even though it looked
1091 * like one was available. This could be due to affinity
1092 * restrictions or for other reasons. Loop again after
1093 * removing this CPU from the set. The restart logic
1094 * above does not restore this CPU to the set due to the
1095 * likelyhood of failing here again.
1097 CPU_CLR(cpu, &mask);
1098 tdq_unlock_pair(tdq, steal);
1101 mi_switch(SW_VOL | SWT_IDLE);
1106 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1109 tdq_notify(struct tdq *tdq, struct thread *td)
1115 if (tdq->tdq_owepreempt)
1117 cpu = td_get_sched(td)->ts_cpu;
1118 pri = td->td_priority;
1119 ctd = pcpu_find(cpu)->pc_curthread;
1120 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1124 * Make sure that our caller's earlier update to tdq_load is
1125 * globally visible before we read tdq_cpu_idle. Idle thread
1126 * accesses both of them without locks, and the order is important.
1128 atomic_thread_fence_seq_cst();
1130 if (TD_IS_IDLETHREAD(ctd)) {
1132 * If the MD code has an idle wakeup routine try that before
1133 * falling back to IPI.
1135 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1140 * The run queues have been updated, so any switch on the remote CPU
1141 * will satisfy the preemption request.
1143 tdq->tdq_owepreempt = 1;
1144 ipi_cpu(cpu, IPI_PREEMPT);
1148 * Steals load from a timeshare queue. Honors the rotating queue head
1151 static struct thread *
1152 runq_steal_from(struct runq *rq, int cpu, u_char start)
1156 struct thread *td, *first;
1160 rqb = &rq->rq_status;
1161 bit = start & (RQB_BPW -1);
1164 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1165 if (rqb->rqb_bits[i] == 0)
1168 bit = RQB_FFS(rqb->rqb_bits[i]);
1169 for (; bit < RQB_BPW; bit++) {
1170 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1172 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1173 TAILQ_FOREACH(td, rqh, td_runq) {
1175 if (THREAD_CAN_MIGRATE(td) &&
1176 THREAD_CAN_SCHED(td, cpu))
1188 if (first && THREAD_CAN_MIGRATE(first) &&
1189 THREAD_CAN_SCHED(first, cpu))
1195 * Steals load from a standard linear queue.
1197 static struct thread *
1198 runq_steal(struct runq *rq, int cpu)
1206 rqb = &rq->rq_status;
1207 for (word = 0; word < RQB_LEN; word++) {
1208 if (rqb->rqb_bits[word] == 0)
1210 for (bit = 0; bit < RQB_BPW; bit++) {
1211 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1213 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1214 TAILQ_FOREACH(td, rqh, td_runq)
1215 if (THREAD_CAN_MIGRATE(td) &&
1216 THREAD_CAN_SCHED(td, cpu))
1224 * Attempt to steal a thread in priority order from a thread queue.
1226 static struct thread *
1227 tdq_steal(struct tdq *tdq, int cpu)
1231 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1232 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1234 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1235 cpu, tdq->tdq_ridx)) != NULL)
1237 return (runq_steal(&tdq->tdq_idle, cpu));
1241 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1242 * current lock and returns with the assigned queue locked.
1244 static inline struct tdq *
1245 sched_setcpu(struct thread *td, int cpu, int flags)
1251 THREAD_LOCK_ASSERT(td, MA_OWNED);
1253 td_get_sched(td)->ts_cpu = cpu;
1255 * If the lock matches just return the queue.
1257 if (td->td_lock == TDQ_LOCKPTR(tdq)) {
1258 KASSERT((flags & SRQ_HOLD) == 0,
1259 ("sched_setcpu: Invalid lock for SRQ_HOLD"));
1264 * The hard case, migration, we need to block the thread first to
1265 * prevent order reversals with other cpus locks.
1268 mtx = thread_lock_block(td);
1269 if ((flags & SRQ_HOLD) == 0)
1270 mtx_unlock_spin(mtx);
1272 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1277 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1278 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1279 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1280 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1281 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1282 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1285 sched_pickcpu(struct thread *td, int flags)
1287 struct cpu_group *cg, *ccg;
1288 struct td_sched *ts;
1291 int cpu, pri, r, self, intr;
1293 self = PCPU_GET(cpuid);
1294 ts = td_get_sched(td);
1295 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1296 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1297 if (smp_started == 0)
1300 * Don't migrate a running thread from sched_switch().
1302 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1303 return (ts->ts_cpu);
1305 * Prefer to run interrupt threads on the processors that generate
1308 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1309 curthread->td_intr_nesting_level) {
1311 if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1312 SCHED_STAT_INC(pickcpu_idle_affinity);
1321 tdq = TDQ_CPU(ts->ts_cpu);
1325 * If the thread can run on the last cpu and the affinity has not
1326 * expired and it is idle, run it there.
1328 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1329 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1330 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1331 if (cg->cg_flags & CG_FLAG_THREAD) {
1332 /* Check all SMT threads for being idle. */
1333 for (cpu = cg->cg_first; cpu <= cg->cg_last; cpu++) {
1334 if (CPU_ISSET(cpu, &cg->cg_mask) &&
1335 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1338 if (cpu > cg->cg_last) {
1339 SCHED_STAT_INC(pickcpu_idle_affinity);
1340 return (ts->ts_cpu);
1343 SCHED_STAT_INC(pickcpu_idle_affinity);
1344 return (ts->ts_cpu);
1349 * Search for the last level cache CPU group in the tree.
1350 * Skip SMT, identical groups and caches with expired affinity.
1351 * Interrupt threads affinity is explicit and never expires.
1353 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1354 if (cg->cg_flags & CG_FLAG_THREAD)
1356 if (cg->cg_children == 1 || cg->cg_count == 1)
1358 if (cg->cg_level == CG_SHARE_NONE ||
1359 (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1363 /* Found LLC shared by all CPUs, so do a global search. */
1367 mask = &td->td_cpuset->cs_mask;
1368 pri = td->td_priority;
1369 r = TD_IS_RUNNING(td);
1371 * Try hard to keep interrupts within found LLC. Search the LLC for
1372 * the least loaded CPU we can run now. For NUMA systems it should
1373 * be within target domain, and it also reduces scheduling overhead.
1375 if (ccg != NULL && intr) {
1376 cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu, r);
1378 SCHED_STAT_INC(pickcpu_intrbind);
1380 /* Search the LLC for the least loaded idle CPU we can run now. */
1382 cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1383 INT_MAX, ts->ts_cpu, r);
1385 SCHED_STAT_INC(pickcpu_affinity);
1387 /* Search globally for the least loaded CPU we can run now. */
1389 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu, r);
1391 SCHED_STAT_INC(pickcpu_lowest);
1393 /* Search globally for the least loaded CPU. */
1395 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu, r);
1397 SCHED_STAT_INC(pickcpu_lowest);
1399 KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1400 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1402 * Compare the lowest loaded cpu to current cpu.
1405 if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1406 tdq->tdq_lowpri < PRI_MIN_IDLE &&
1407 TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1408 SCHED_STAT_INC(pickcpu_local);
1411 if (cpu != ts->ts_cpu)
1412 SCHED_STAT_INC(pickcpu_migration);
1418 * Pick the highest priority task we have and return it.
1420 static struct thread *
1421 tdq_choose(struct tdq *tdq)
1425 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1426 td = runq_choose(&tdq->tdq_realtime);
1429 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1431 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1432 ("tdq_choose: Invalid priority on timeshare queue %d",
1436 td = runq_choose(&tdq->tdq_idle);
1438 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1439 ("tdq_choose: Invalid priority on idle queue %d",
1448 * Initialize a thread queue.
1451 tdq_setup(struct tdq *tdq, int id)
1455 printf("ULE: setup cpu %d\n", id);
1456 runq_init(&tdq->tdq_realtime);
1457 runq_init(&tdq->tdq_timeshare);
1458 runq_init(&tdq->tdq_idle);
1460 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1461 "sched lock %d", (int)TDQ_ID(tdq));
1462 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN);
1464 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1465 "CPU %d load", (int)TDQ_ID(tdq));
1471 sched_setup_smp(void)
1476 cpu_top = smp_topo();
1478 tdq = DPCPU_ID_PTR(i, tdq);
1480 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1481 if (tdq->tdq_cg == NULL)
1482 panic("Can't find cpu group for %d\n", i);
1483 DPCPU_ID_SET(i, randomval, i * 69069 + 5);
1485 PCPU_SET(sched, DPCPU_PTR(tdq));
1486 balance_tdq = TDQ_SELF();
1491 * Setup the thread queues and initialize the topology based on MD
1495 sched_setup(void *dummy)
1502 tdq_setup(TDQ_SELF(), 0);
1506 /* Add thread0's load since it's running. */
1508 thread0.td_lock = TDQ_LOCKPTR(tdq);
1509 tdq_load_add(tdq, &thread0);
1510 tdq->tdq_lowpri = thread0.td_priority;
1515 * This routine determines time constants after stathz and hz are setup.
1519 sched_initticks(void *dummy)
1523 realstathz = stathz ? stathz : hz;
1524 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1525 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1526 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1530 * tickincr is shifted out by 10 to avoid rounding errors due to
1531 * hz not being evenly divisible by stathz on all platforms.
1533 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1535 * This does not work for values of stathz that are more than
1536 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1543 * Set the default balance interval now that we know
1544 * what realstathz is.
1546 balance_interval = realstathz;
1547 balance_ticks = balance_interval;
1548 affinity = SCHED_AFFINITY_DEFAULT;
1550 if (sched_idlespinthresh < 0)
1551 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1555 * This is the core of the interactivity algorithm. Determines a score based
1556 * on past behavior. It is the ratio of sleep time to run time scaled to
1557 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1558 * differs from the cpu usage because it does not account for time spent
1559 * waiting on a run-queue. Would be prettier if we had floating point.
1561 * When a thread's sleep time is greater than its run time the
1565 * interactivity score = ---------------------
1566 * sleep time / run time
1569 * When a thread's run time is greater than its sleep time the
1573 * interactivity score = --------------------- + scaling factor
1574 * run time / sleep time
1577 sched_interact_score(struct thread *td)
1579 struct td_sched *ts;
1582 ts = td_get_sched(td);
1584 * The score is only needed if this is likely to be an interactive
1585 * task. Don't go through the expense of computing it if there's
1588 if (sched_interact <= SCHED_INTERACT_HALF &&
1589 ts->ts_runtime >= ts->ts_slptime)
1590 return (SCHED_INTERACT_HALF);
1592 if (ts->ts_runtime > ts->ts_slptime) {
1593 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1594 return (SCHED_INTERACT_HALF +
1595 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1597 if (ts->ts_slptime > ts->ts_runtime) {
1598 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1599 return (ts->ts_runtime / div);
1601 /* runtime == slptime */
1603 return (SCHED_INTERACT_HALF);
1606 * This can happen if slptime and runtime are 0.
1613 * Scale the scheduling priority according to the "interactivity" of this
1617 sched_priority(struct thread *td)
1621 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1624 * If the score is interactive we place the thread in the realtime
1625 * queue with a priority that is less than kernel and interrupt
1626 * priorities. These threads are not subject to nice restrictions.
1628 * Scores greater than this are placed on the normal timeshare queue
1629 * where the priority is partially decided by the most recent cpu
1630 * utilization and the rest is decided by nice value.
1632 * The nice value of the process has a linear effect on the calculated
1633 * score. Negative nice values make it easier for a thread to be
1634 * considered interactive.
1636 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1637 if (score < sched_interact) {
1638 pri = PRI_MIN_INTERACT;
1639 pri += (PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) * score /
1641 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1642 ("sched_priority: invalid interactive priority %u score %u",
1645 pri = SCHED_PRI_MIN;
1646 if (td_get_sched(td)->ts_ticks)
1647 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1648 SCHED_PRI_RANGE - 1);
1649 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1650 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1651 ("sched_priority: invalid priority %u: nice %d, "
1652 "ticks %d ftick %d ltick %d tick pri %d",
1653 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1654 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1655 SCHED_PRI_TICKS(td_get_sched(td))));
1657 sched_user_prio(td, pri);
1663 * This routine enforces a maximum limit on the amount of scheduling history
1664 * kept. It is called after either the slptime or runtime is adjusted. This
1665 * function is ugly due to integer math.
1668 sched_interact_update(struct thread *td)
1670 struct td_sched *ts;
1673 ts = td_get_sched(td);
1674 sum = ts->ts_runtime + ts->ts_slptime;
1675 if (sum < SCHED_SLP_RUN_MAX)
1678 * This only happens from two places:
1679 * 1) We have added an unusual amount of run time from fork_exit.
1680 * 2) We have added an unusual amount of sleep time from sched_sleep().
1682 if (sum > SCHED_SLP_RUN_MAX * 2) {
1683 if (ts->ts_runtime > ts->ts_slptime) {
1684 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1687 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1693 * If we have exceeded by more than 1/5th then the algorithm below
1694 * will not bring us back into range. Dividing by two here forces
1695 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1697 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1698 ts->ts_runtime /= 2;
1699 ts->ts_slptime /= 2;
1702 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1703 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1707 * Scale back the interactivity history when a child thread is created. The
1708 * history is inherited from the parent but the thread may behave totally
1709 * differently. For example, a shell spawning a compiler process. We want
1710 * to learn that the compiler is behaving badly very quickly.
1713 sched_interact_fork(struct thread *td)
1715 struct td_sched *ts;
1719 ts = td_get_sched(td);
1720 sum = ts->ts_runtime + ts->ts_slptime;
1721 if (sum > SCHED_SLP_RUN_FORK) {
1722 ratio = sum / SCHED_SLP_RUN_FORK;
1723 ts->ts_runtime /= ratio;
1724 ts->ts_slptime /= ratio;
1729 * Called from proc0_init() to setup the scheduler fields.
1734 struct td_sched *ts0;
1737 * Set up the scheduler specific parts of thread0.
1739 ts0 = td_get_sched(&thread0);
1740 ts0->ts_ltick = ticks;
1741 ts0->ts_ftick = ticks;
1743 ts0->ts_cpu = curcpu; /* set valid CPU number */
1747 * This is only somewhat accurate since given many processes of the same
1748 * priority they will switch when their slices run out, which will be
1749 * at most sched_slice stathz ticks.
1752 sched_rr_interval(void)
1755 /* Convert sched_slice from stathz to hz. */
1756 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1760 * Update the percent cpu tracking information when it is requested or
1761 * the total history exceeds the maximum. We keep a sliding history of
1762 * tick counts that slowly decays. This is less precise than the 4BSD
1763 * mechanism since it happens with less regular and frequent events.
1766 sched_pctcpu_update(struct td_sched *ts, int run)
1771 * The signed difference may be negative if the thread hasn't run for
1772 * over half of the ticks rollover period.
1774 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1776 ts->ts_ftick = t - SCHED_TICK_TARG;
1777 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1778 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1779 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1780 ts->ts_ftick = t - SCHED_TICK_TARG;
1783 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1788 * Adjust the priority of a thread. Move it to the appropriate run-queue
1789 * if necessary. This is the back-end for several priority related
1793 sched_thread_priority(struct thread *td, u_char prio)
1795 struct td_sched *ts;
1799 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1800 "prio:%d", td->td_priority, "new prio:%d", prio,
1801 KTR_ATTR_LINKED, sched_tdname(curthread));
1802 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1803 if (td != curthread && prio < td->td_priority) {
1804 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1805 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1806 prio, KTR_ATTR_LINKED, sched_tdname(td));
1807 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1810 ts = td_get_sched(td);
1811 THREAD_LOCK_ASSERT(td, MA_OWNED);
1812 if (td->td_priority == prio)
1815 * If the priority has been elevated due to priority
1816 * propagation, we may have to move ourselves to a new
1817 * queue. This could be optimized to not re-add in some
1820 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1822 td->td_priority = prio;
1823 sched_add(td, SRQ_BORROWING | SRQ_HOLDTD);
1827 * If the thread is currently running we may have to adjust the lowpri
1828 * information so other cpus are aware of our current priority.
1830 if (TD_IS_RUNNING(td)) {
1831 tdq = TDQ_CPU(ts->ts_cpu);
1832 oldpri = td->td_priority;
1833 td->td_priority = prio;
1834 if (prio < tdq->tdq_lowpri)
1835 tdq->tdq_lowpri = prio;
1836 else if (tdq->tdq_lowpri == oldpri)
1837 tdq_setlowpri(tdq, td);
1840 td->td_priority = prio;
1844 * Update a thread's priority when it is lent another thread's
1848 sched_lend_prio(struct thread *td, u_char prio)
1851 td->td_flags |= TDF_BORROWING;
1852 sched_thread_priority(td, prio);
1856 * Restore a thread's priority when priority propagation is
1857 * over. The prio argument is the minimum priority the thread
1858 * needs to have to satisfy other possible priority lending
1859 * requests. If the thread's regular priority is less
1860 * important than prio, the thread will keep a priority boost
1864 sched_unlend_prio(struct thread *td, u_char prio)
1868 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1869 td->td_base_pri <= PRI_MAX_TIMESHARE)
1870 base_pri = td->td_user_pri;
1872 base_pri = td->td_base_pri;
1873 if (prio >= base_pri) {
1874 td->td_flags &= ~TDF_BORROWING;
1875 sched_thread_priority(td, base_pri);
1877 sched_lend_prio(td, prio);
1881 * Standard entry for setting the priority to an absolute value.
1884 sched_prio(struct thread *td, u_char prio)
1888 /* First, update the base priority. */
1889 td->td_base_pri = prio;
1892 * If the thread is borrowing another thread's priority, don't
1893 * ever lower the priority.
1895 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1898 /* Change the real priority. */
1899 oldprio = td->td_priority;
1900 sched_thread_priority(td, prio);
1903 * If the thread is on a turnstile, then let the turnstile update
1906 if (TD_ON_LOCK(td) && oldprio != prio)
1907 turnstile_adjust(td, oldprio);
1911 * Set the base user priority, does not effect current running priority.
1914 sched_user_prio(struct thread *td, u_char prio)
1917 td->td_base_user_pri = prio;
1918 if (td->td_lend_user_pri <= prio)
1920 td->td_user_pri = prio;
1924 sched_lend_user_prio(struct thread *td, u_char prio)
1927 THREAD_LOCK_ASSERT(td, MA_OWNED);
1928 td->td_lend_user_pri = prio;
1929 td->td_user_pri = min(prio, td->td_base_user_pri);
1930 if (td->td_priority > td->td_user_pri)
1931 sched_prio(td, td->td_user_pri);
1932 else if (td->td_priority != td->td_user_pri)
1933 td->td_flags |= TDF_NEEDRESCHED;
1937 * Like the above but first check if there is anything to do.
1940 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1943 if (td->td_lend_user_pri != prio)
1945 if (td->td_user_pri != min(prio, td->td_base_user_pri))
1947 if (td->td_priority != td->td_user_pri)
1953 sched_lend_user_prio(td, prio);
1959 * This tdq is about to idle. Try to steal a thread from another CPU before
1960 * choosing the idle thread.
1963 tdq_trysteal(struct tdq *tdq)
1965 struct cpu_group *cg, *parent;
1970 if (smp_started == 0 || steal_idle == 0 || trysteal_limit == 0 ||
1971 tdq->tdq_cg == NULL)
1974 CPU_CLR(PCPU_GET(cpuid), &mask);
1975 /* We don't want to be preempted while we're iterating. */
1978 for (i = 1, cg = tdq->tdq_cg, goup = 0; ; ) {
1979 cpu = sched_highest(cg, &mask, steal_thresh, 1);
1981 * If a thread was added while interrupts were disabled don't
1984 if (tdq->tdq_load > 0) {
1990 * We found no CPU to steal from in this group. Escalate to
1991 * the parent and repeat. But if parent has only two children
1992 * groups we can avoid searching this group again by searching
1993 * the other one specifically and then escalating two levels.
2000 if (++i > trysteal_limit) {
2004 parent = cg->cg_parent;
2005 if (parent == NULL) {
2009 if (parent->cg_children == 2) {
2010 if (cg == &parent->cg_child[0])
2011 cg = &parent->cg_child[1];
2013 cg = &parent->cg_child[0];
2019 steal = TDQ_CPU(cpu);
2021 * The data returned by sched_highest() is stale and
2022 * the chosen CPU no longer has an eligible thread.
2023 * At this point unconditonally exit the loop to bound
2024 * the time spent in the critcal section.
2026 if (steal->tdq_load < steal_thresh ||
2027 steal->tdq_transferable == 0)
2030 * Try to lock both queues. If we are assigned a thread while
2031 * waited for the lock, switch to it now instead of stealing.
2032 * If we can't get the lock, then somebody likely got there
2036 if (tdq->tdq_load > 0)
2038 if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0)
2041 * The data returned by sched_highest() is stale and
2042 * the chosen CPU no longer has an eligible thread.
2044 if (steal->tdq_load < steal_thresh ||
2045 steal->tdq_transferable == 0) {
2050 * If we fail to acquire one due to affinity restrictions,
2051 * bail out and let the idle thread to a more complete search
2052 * outside of a critical section.
2054 if (tdq_move(steal, tdq) == NULL) {
2066 * Handle migration from sched_switch(). This happens only for
2070 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2074 KASSERT(THREAD_CAN_MIGRATE(td) ||
2075 (td_get_sched(td)->ts_flags & TSF_BOUND) != 0,
2076 ("Thread %p shouldn't migrate", td));
2077 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2078 "thread %s queued on absent CPU %d.", td->td_name,
2079 td_get_sched(td)->ts_cpu));
2080 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2082 tdq_load_rem(tdq, td);
2084 * Do the lock dance required to avoid LOR. We have an
2085 * extra spinlock nesting from sched_switch() which will
2086 * prevent preemption while we're holding neither run-queue lock.
2090 tdq_add(tdn, td, flags);
2091 tdq_notify(tdn, td);
2095 return (TDQ_LOCKPTR(tdn));
2099 * thread_lock_unblock() that does not assume td_lock is blocked.
2102 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2104 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2109 * Switch threads. This function has to handle threads coming in while
2110 * blocked for some reason, running, or idle. It also must deal with
2111 * migrating a thread from one queue to another as running threads may
2112 * be assigned elsewhere via binding.
2115 sched_switch(struct thread *td, int flags)
2117 struct thread *newtd;
2119 struct td_sched *ts;
2122 int cpuid, preempted;
2127 THREAD_LOCK_ASSERT(td, MA_OWNED);
2129 cpuid = PCPU_GET(cpuid);
2131 ts = td_get_sched(td);
2132 sched_pctcpu_update(ts, 1);
2134 pickcpu = (td->td_flags & TDF_PICKCPU) != 0;
2136 ts->ts_rltick = ticks - affinity * MAX_CACHE_LEVELS;
2138 ts->ts_rltick = ticks;
2140 td->td_lastcpu = td->td_oncpu;
2141 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2142 (flags & SW_PREEMPT) != 0;
2143 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_PICKCPU | TDF_SLICEEND);
2144 td->td_owepreempt = 0;
2145 tdq->tdq_owepreempt = 0;
2146 if (!TD_IS_IDLETHREAD(td))
2147 tdq->tdq_switchcnt++;
2150 * Always block the thread lock so we can drop the tdq lock early.
2152 mtx = thread_lock_block(td);
2154 if (TD_IS_IDLETHREAD(td)) {
2155 MPASS(mtx == TDQ_LOCKPTR(tdq));
2157 } else if (TD_IS_RUNNING(td)) {
2158 MPASS(mtx == TDQ_LOCKPTR(tdq));
2159 srqflag = preempted ?
2160 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2161 SRQ_OURSELF|SRQ_YIELDING;
2163 if (THREAD_CAN_MIGRATE(td) && (!THREAD_CAN_SCHED(td, ts->ts_cpu)
2165 ts->ts_cpu = sched_pickcpu(td, 0);
2167 if (ts->ts_cpu == cpuid)
2168 tdq_runq_add(tdq, td, srqflag);
2170 mtx = sched_switch_migrate(tdq, td, srqflag);
2172 /* This thread must be going to sleep. */
2173 if (mtx != TDQ_LOCKPTR(tdq)) {
2174 mtx_unlock_spin(mtx);
2177 tdq_load_rem(tdq, td);
2179 if (tdq->tdq_load == 0)
2184 #if (KTR_COMPILE & KTR_SCHED) != 0
2185 if (TD_IS_IDLETHREAD(td))
2186 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2187 "prio:%d", td->td_priority);
2189 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2190 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2191 "lockname:\"%s\"", td->td_lockname);
2195 * We enter here with the thread blocked and assigned to the
2196 * appropriate cpu run-queue or sleep-queue and with the current
2197 * thread-queue locked.
2199 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2200 newtd = choosethread();
2201 sched_pctcpu_update(td_get_sched(newtd), 0);
2205 * Call the MD code to switch contexts if necessary.
2209 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2210 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2212 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2214 #ifdef KDTRACE_HOOKS
2216 * If DTrace has set the active vtime enum to anything
2217 * other than INACTIVE (0), then it should have set the
2220 if (dtrace_vtime_active)
2221 (*dtrace_vtime_switch_func)(newtd);
2223 td->td_oncpu = NOCPU;
2224 cpu_switch(td, newtd, mtx);
2225 cpuid = td->td_oncpu = PCPU_GET(cpuid);
2227 SDT_PROBE0(sched, , , on__cpu);
2229 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2230 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2233 thread_unblock_switch(td, mtx);
2234 SDT_PROBE0(sched, , , remain__cpu);
2236 KASSERT(curthread->td_md.md_spinlock_count == 1,
2237 ("invalid count %d", curthread->td_md.md_spinlock_count));
2239 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2240 "prio:%d", td->td_priority);
2244 * Adjust thread priorities as a result of a nice request.
2247 sched_nice(struct proc *p, int nice)
2251 PROC_LOCK_ASSERT(p, MA_OWNED);
2254 FOREACH_THREAD_IN_PROC(p, td) {
2257 sched_prio(td, td->td_base_user_pri);
2263 * Record the sleep time for the interactivity scorer.
2266 sched_sleep(struct thread *td, int prio)
2269 THREAD_LOCK_ASSERT(td, MA_OWNED);
2271 td->td_slptick = ticks;
2272 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2273 td->td_flags |= TDF_CANSWAP;
2274 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2276 if (static_boost == 1 && prio)
2277 sched_prio(td, prio);
2278 else if (static_boost && td->td_priority > static_boost)
2279 sched_prio(td, static_boost);
2283 * Schedule a thread to resume execution and record how long it voluntarily
2284 * slept. We also update the pctcpu, interactivity, and priority.
2286 * Requires the thread lock on entry, drops on exit.
2289 sched_wakeup(struct thread *td, int srqflags)
2291 struct td_sched *ts;
2294 THREAD_LOCK_ASSERT(td, MA_OWNED);
2295 ts = td_get_sched(td);
2296 td->td_flags &= ~TDF_CANSWAP;
2299 * If we slept for more than a tick update our interactivity and
2302 slptick = td->td_slptick;
2304 if (slptick && slptick != ticks) {
2305 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2306 sched_interact_update(td);
2307 sched_pctcpu_update(ts, 0);
2310 * Reset the slice value since we slept and advanced the round-robin.
2313 sched_add(td, SRQ_BORING | srqflags);
2317 * Penalize the parent for creating a new child and initialize the child's
2321 sched_fork(struct thread *td, struct thread *child)
2323 THREAD_LOCK_ASSERT(td, MA_OWNED);
2324 sched_pctcpu_update(td_get_sched(td), 1);
2325 sched_fork_thread(td, child);
2327 * Penalize the parent and child for forking.
2329 sched_interact_fork(child);
2330 sched_priority(child);
2331 td_get_sched(td)->ts_runtime += tickincr;
2332 sched_interact_update(td);
2337 * Fork a new thread, may be within the same process.
2340 sched_fork_thread(struct thread *td, struct thread *child)
2342 struct td_sched *ts;
2343 struct td_sched *ts2;
2347 THREAD_LOCK_ASSERT(td, MA_OWNED);
2351 ts = td_get_sched(td);
2352 ts2 = td_get_sched(child);
2353 child->td_oncpu = NOCPU;
2354 child->td_lastcpu = NOCPU;
2355 child->td_lock = TDQ_LOCKPTR(tdq);
2356 child->td_cpuset = cpuset_ref(td->td_cpuset);
2357 child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2358 ts2->ts_cpu = ts->ts_cpu;
2361 * Grab our parents cpu estimation information.
2363 ts2->ts_ticks = ts->ts_ticks;
2364 ts2->ts_ltick = ts->ts_ltick;
2365 ts2->ts_ftick = ts->ts_ftick;
2367 * Do not inherit any borrowed priority from the parent.
2369 child->td_priority = child->td_base_pri;
2371 * And update interactivity score.
2373 ts2->ts_slptime = ts->ts_slptime;
2374 ts2->ts_runtime = ts->ts_runtime;
2375 /* Attempt to quickly learn interactivity. */
2376 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2378 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2383 * Adjust the priority class of a thread.
2386 sched_class(struct thread *td, int class)
2389 THREAD_LOCK_ASSERT(td, MA_OWNED);
2390 if (td->td_pri_class == class)
2392 td->td_pri_class = class;
2396 * Return some of the child's priority and interactivity to the parent.
2399 sched_exit(struct proc *p, struct thread *child)
2403 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2404 "prio:%d", child->td_priority);
2405 PROC_LOCK_ASSERT(p, MA_OWNED);
2406 td = FIRST_THREAD_IN_PROC(p);
2407 sched_exit_thread(td, child);
2411 * Penalize another thread for the time spent on this one. This helps to
2412 * worsen the priority and interactivity of processes which schedule batch
2413 * jobs such as make. This has little effect on the make process itself but
2414 * causes new processes spawned by it to receive worse scores immediately.
2417 sched_exit_thread(struct thread *td, struct thread *child)
2420 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2421 "prio:%d", child->td_priority);
2423 * Give the child's runtime to the parent without returning the
2424 * sleep time as a penalty to the parent. This causes shells that
2425 * launch expensive things to mark their children as expensive.
2428 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2429 sched_interact_update(td);
2435 sched_preempt(struct thread *td)
2440 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2444 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2445 if (td->td_priority > tdq->tdq_lowpri) {
2446 if (td->td_critnest == 1) {
2447 flags = SW_INVOL | SW_PREEMPT;
2448 flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
2451 /* Switch dropped thread lock. */
2454 td->td_owepreempt = 1;
2456 tdq->tdq_owepreempt = 0;
2462 * Fix priorities on return to user-space. Priorities may be elevated due
2463 * to static priorities in msleep() or similar.
2466 sched_userret_slowpath(struct thread *td)
2470 td->td_priority = td->td_user_pri;
2471 td->td_base_pri = td->td_user_pri;
2472 tdq_setlowpri(TDQ_SELF(), td);
2477 * Handle a stathz tick. This is really only relevant for timeshare
2481 sched_clock(struct thread *td, int cnt)
2484 struct td_sched *ts;
2486 THREAD_LOCK_ASSERT(td, MA_OWNED);
2490 * We run the long term load balancer infrequently on the first cpu.
2492 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 &&
2493 balance_ticks != 0) {
2494 balance_ticks -= cnt;
2495 if (balance_ticks <= 0)
2500 * Save the old switch count so we have a record of the last ticks
2501 * activity. Initialize the new switch count based on our load.
2502 * If there is some activity seed it to reflect that.
2504 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2505 tdq->tdq_switchcnt = tdq->tdq_load;
2507 * Advance the insert index once for each tick to ensure that all
2508 * threads get a chance to run.
2510 if (tdq->tdq_idx == tdq->tdq_ridx) {
2511 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2512 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2513 tdq->tdq_ridx = tdq->tdq_idx;
2515 ts = td_get_sched(td);
2516 sched_pctcpu_update(ts, 1);
2517 if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td))
2520 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2522 * We used a tick; charge it to the thread so
2523 * that we can compute our interactivity.
2525 td_get_sched(td)->ts_runtime += tickincr * cnt;
2526 sched_interact_update(td);
2531 * Force a context switch if the current thread has used up a full
2532 * time slice (default is 100ms).
2534 ts->ts_slice += cnt;
2535 if (ts->ts_slice >= tdq_slice(tdq)) {
2537 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2542 sched_estcpu(struct thread *td __unused)
2549 * Return whether the current CPU has runnable tasks. Used for in-kernel
2550 * cooperative idle threads.
2553 sched_runnable(void)
2561 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2562 if (tdq->tdq_load > 0)
2565 if (tdq->tdq_load - 1 > 0)
2573 * Choose the highest priority thread to run. The thread is removed from
2574 * the run-queue while running however the load remains. For SMP we set
2575 * the tdq in the global idle bitmask if it idles here.
2584 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2585 td = tdq_choose(tdq);
2587 tdq_runq_rem(tdq, td);
2588 tdq->tdq_lowpri = td->td_priority;
2591 tdq->tdq_lowpri = PRI_MAX_IDLE;
2592 return (PCPU_GET(idlethread));
2596 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2597 * we always request it once we exit a critical section.
2600 sched_setpreempt(struct thread *td)
2606 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2609 pri = td->td_priority;
2610 cpri = ctd->td_priority;
2612 ctd->td_flags |= TDF_NEEDRESCHED;
2613 if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2615 if (!sched_shouldpreempt(pri, cpri, 0))
2617 ctd->td_owepreempt = 1;
2621 * Add a thread to a thread queue. Select the appropriate runq and add the
2622 * thread to it. This is the internal function called when the tdq is
2626 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2629 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2630 THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
2631 KASSERT((td->td_inhibitors == 0),
2632 ("sched_add: trying to run inhibited thread"));
2633 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2634 ("sched_add: bad thread state"));
2635 KASSERT(td->td_flags & TDF_INMEM,
2636 ("sched_add: thread swapped out"));
2638 if (td->td_priority < tdq->tdq_lowpri)
2639 tdq->tdq_lowpri = td->td_priority;
2640 tdq_runq_add(tdq, td, flags);
2641 tdq_load_add(tdq, td);
2645 * Select the target thread queue and add a thread to it. Request
2646 * preemption or IPI a remote processor if required.
2648 * Requires the thread lock on entry, drops on exit.
2651 sched_add(struct thread *td, int flags)
2658 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2659 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2660 sched_tdname(curthread));
2661 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2662 KTR_ATTR_LINKED, sched_tdname(td));
2663 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2664 flags & SRQ_PREEMPTED);
2665 THREAD_LOCK_ASSERT(td, MA_OWNED);
2667 * Recalculate the priority before we select the target cpu or
2670 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2674 * Pick the destination cpu and if it isn't ours transfer to the
2677 cpu = sched_pickcpu(td, flags);
2678 tdq = sched_setcpu(td, cpu, flags);
2679 tdq_add(tdq, td, flags);
2680 if (cpu != PCPU_GET(cpuid))
2681 tdq_notify(tdq, td);
2682 else if (!(flags & SRQ_YIELDING))
2683 sched_setpreempt(td);
2687 * Now that the thread is moving to the run-queue, set the lock
2688 * to the scheduler's lock.
2690 if (td->td_lock != TDQ_LOCKPTR(tdq)) {
2692 if ((flags & SRQ_HOLD) != 0)
2693 td->td_lock = TDQ_LOCKPTR(tdq);
2695 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2697 tdq_add(tdq, td, flags);
2698 if (!(flags & SRQ_YIELDING))
2699 sched_setpreempt(td);
2701 if (!(flags & SRQ_HOLDTD))
2706 * Remove a thread from a run-queue without running it. This is used
2707 * when we're stealing a thread from a remote queue. Otherwise all threads
2708 * exit by calling sched_exit_thread() and sched_throw() themselves.
2711 sched_rem(struct thread *td)
2715 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2716 "prio:%d", td->td_priority);
2717 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2718 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2719 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2720 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2721 KASSERT(TD_ON_RUNQ(td),
2722 ("sched_rem: thread not on run queue"));
2723 tdq_runq_rem(tdq, td);
2724 tdq_load_rem(tdq, td);
2726 if (td->td_priority == tdq->tdq_lowpri)
2727 tdq_setlowpri(tdq, NULL);
2731 * Fetch cpu utilization information. Updates on demand.
2734 sched_pctcpu(struct thread *td)
2737 struct td_sched *ts;
2740 ts = td_get_sched(td);
2742 THREAD_LOCK_ASSERT(td, MA_OWNED);
2743 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2747 /* How many rtick per second ? */
2748 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2749 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2756 * Enforce affinity settings for a thread. Called after adjustments to
2760 sched_affinity(struct thread *td)
2763 struct td_sched *ts;
2765 THREAD_LOCK_ASSERT(td, MA_OWNED);
2766 ts = td_get_sched(td);
2767 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2769 if (TD_ON_RUNQ(td)) {
2771 sched_add(td, SRQ_BORING | SRQ_HOLDTD);
2774 if (!TD_IS_RUNNING(td))
2777 * Force a switch before returning to userspace. If the
2778 * target thread is not running locally send an ipi to force
2781 td->td_flags |= TDF_NEEDRESCHED;
2782 if (td != curthread)
2783 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2788 * Bind a thread to a target cpu.
2791 sched_bind(struct thread *td, int cpu)
2793 struct td_sched *ts;
2795 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2796 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2797 ts = td_get_sched(td);
2798 if (ts->ts_flags & TSF_BOUND)
2800 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2801 ts->ts_flags |= TSF_BOUND;
2803 if (PCPU_GET(cpuid) == cpu)
2806 /* When we return from mi_switch we'll be on the correct cpu. */
2812 * Release a bound thread.
2815 sched_unbind(struct thread *td)
2817 struct td_sched *ts;
2819 THREAD_LOCK_ASSERT(td, MA_OWNED);
2820 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2821 ts = td_get_sched(td);
2822 if ((ts->ts_flags & TSF_BOUND) == 0)
2824 ts->ts_flags &= ~TSF_BOUND;
2829 sched_is_bound(struct thread *td)
2831 THREAD_LOCK_ASSERT(td, MA_OWNED);
2832 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2839 sched_relinquish(struct thread *td)
2842 mi_switch(SW_VOL | SWT_RELINQUISH);
2846 * Return the total system load.
2857 total += TDQ_CPU(i)->tdq_sysload;
2860 return (TDQ_SELF()->tdq_sysload);
2865 sched_sizeof_proc(void)
2867 return (sizeof(struct proc));
2871 sched_sizeof_thread(void)
2873 return (sizeof(struct thread) + sizeof(struct td_sched));
2877 #define TDQ_IDLESPIN(tdq) \
2878 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2880 #define TDQ_IDLESPIN(tdq) 1
2884 * The actual idle process.
2887 sched_idletd(void *dummy)
2891 int oldswitchcnt, switchcnt;
2894 mtx_assert(&Giant, MA_NOTOWNED);
2897 THREAD_NO_SLEEPING();
2900 if (tdq->tdq_load) {
2902 mi_switch(SW_VOL | SWT_IDLE);
2904 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2906 if (always_steal || switchcnt != oldswitchcnt) {
2907 oldswitchcnt = switchcnt;
2908 if (tdq_idled(tdq) == 0)
2911 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2913 oldswitchcnt = switchcnt;
2916 * If we're switching very frequently, spin while checking
2917 * for load rather than entering a low power state that
2918 * may require an IPI. However, don't do any busy
2919 * loops while on SMT machines as this simply steals
2920 * cycles from cores doing useful work.
2922 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2923 for (i = 0; i < sched_idlespins; i++) {
2930 /* If there was context switch during spin, restart it. */
2931 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2932 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2935 /* Run main MD idle handler. */
2936 tdq->tdq_cpu_idle = 1;
2938 * Make sure that tdq_cpu_idle update is globally visible
2939 * before cpu_idle() read tdq_load. The order is important
2940 * to avoid race with tdq_notify.
2942 atomic_thread_fence_seq_cst();
2944 * Checking for again after the fence picks up assigned
2945 * threads often enough to make it worthwhile to do so in
2946 * order to avoid calling cpu_idle().
2948 if (tdq->tdq_load != 0) {
2949 tdq->tdq_cpu_idle = 0;
2952 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2953 tdq->tdq_cpu_idle = 0;
2956 * Account thread-less hardware interrupts and
2957 * other wakeup reasons equal to context switches.
2959 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2960 if (switchcnt != oldswitchcnt)
2962 tdq->tdq_switchcnt++;
2968 * A CPU is entering for the first time or a thread is exiting.
2971 sched_throw(struct thread *td)
2973 struct thread *newtd;
2976 if (__predict_false(td == NULL)) {
2978 PCPU_SET(sched, DPCPU_PTR(tdq));
2980 /* Correct spinlock nesting and acquire the correct lock. */
2984 PCPU_SET(switchtime, cpu_ticks());
2985 PCPU_SET(switchticks, ticks);
2986 PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(tdq);
2989 THREAD_LOCK_ASSERT(td, MA_OWNED);
2990 THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq));
2991 tdq_load_rem(tdq, td);
2992 td->td_lastcpu = td->td_oncpu;
2993 td->td_oncpu = NOCPU;
2994 thread_lock_block(td);
2996 newtd = choosethread();
2999 KASSERT(curthread->td_md.md_spinlock_count == 1,
3000 ("invalid count %d", curthread->td_md.md_spinlock_count));
3001 /* doesn't return */
3002 if (__predict_false(td == NULL))
3003 cpu_throw(td, newtd); /* doesn't return */
3005 cpu_switch(td, newtd, TDQ_LOCKPTR(tdq));
3009 * This is called from fork_exit(). Just acquire the correct locks and
3010 * let fork do the rest of the work.
3013 sched_fork_exit(struct thread *td)
3019 * Finish setting up thread glue so that it begins execution in a
3020 * non-nested critical section with the scheduler lock held.
3022 KASSERT(curthread->td_md.md_spinlock_count == 1,
3023 ("invalid count %d", curthread->td_md.md_spinlock_count));
3024 cpuid = PCPU_GET(cpuid);
3028 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
3029 td->td_oncpu = cpuid;
3030 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
3031 "prio:%d", td->td_priority);
3032 SDT_PROBE0(sched, , , on__cpu);
3036 * Create on first use to catch odd startup conditons.
3039 sched_tdname(struct thread *td)
3042 struct td_sched *ts;
3044 ts = td_get_sched(td);
3045 if (ts->ts_name[0] == '\0')
3046 snprintf(ts->ts_name, sizeof(ts->ts_name),
3047 "%s tid %d", td->td_name, td->td_tid);
3048 return (ts->ts_name);
3050 return (td->td_name);
3056 sched_clear_tdname(struct thread *td)
3058 struct td_sched *ts;
3060 ts = td_get_sched(td);
3061 ts->ts_name[0] = '\0';
3068 * Build the CPU topology dump string. Is recursively called to collect
3069 * the topology tree.
3072 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
3075 char cpusetbuf[CPUSETBUFSIZ];
3078 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
3079 "", 1 + indent / 2, cg->cg_level);
3080 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3081 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3083 for (i = cg->cg_first; i <= cg->cg_last; i++) {
3084 if (CPU_ISSET(i, &cg->cg_mask)) {
3086 sbuf_printf(sb, ", ");
3089 sbuf_printf(sb, "%d", i);
3092 sbuf_printf(sb, "</cpu>\n");
3094 if (cg->cg_flags != 0) {
3095 sbuf_printf(sb, "%*s <flags>", indent, "");
3096 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3097 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3098 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3099 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3100 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3101 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3102 if ((cg->cg_flags & CG_FLAG_NODE) != 0)
3103 sbuf_printf(sb, "<flag name=\"NODE\">NUMA node</flag>");
3104 sbuf_printf(sb, "</flags>\n");
3107 if (cg->cg_children > 0) {
3108 sbuf_printf(sb, "%*s <children>\n", indent, "");
3109 for (i = 0; i < cg->cg_children; i++)
3110 sysctl_kern_sched_topology_spec_internal(sb,
3111 &cg->cg_child[i], indent+2);
3112 sbuf_printf(sb, "%*s </children>\n", indent, "");
3114 sbuf_printf(sb, "%*s</group>\n", indent, "");
3119 * Sysctl handler for retrieving topology dump. It's a wrapper for
3120 * the recursive sysctl_kern_smp_topology_spec_internal().
3123 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3128 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3130 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3134 sbuf_printf(topo, "<groups>\n");
3135 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3136 sbuf_printf(topo, "</groups>\n");
3139 err = sbuf_finish(topo);
3148 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3150 int error, new_val, period;
3152 period = 1000000 / realstathz;
3153 new_val = period * sched_slice;
3154 error = sysctl_handle_int(oidp, &new_val, 0, req);
3155 if (error != 0 || req->newptr == NULL)
3159 sched_slice = imax(1, (new_val + period / 2) / period);
3160 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3161 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3166 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
3168 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3170 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum,
3171 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
3172 sysctl_kern_quantum, "I",
3173 "Quantum for timeshare threads in microseconds");
3174 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3175 "Quantum for timeshare threads in stathz ticks");
3176 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3177 "Interactivity score threshold");
3178 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3180 "Maximal (lowest) priority for preemption");
3181 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3182 "Assign static kernel priorities to sleeping threads");
3183 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3184 "Number of times idle thread will spin waiting for new work");
3185 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3186 &sched_idlespinthresh, 0,
3187 "Threshold before we will permit idle thread spinning");
3189 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3190 "Number of hz ticks to keep thread affinity for");
3191 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3192 "Enables the long-term load balancer");
3193 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3194 &balance_interval, 0,
3195 "Average period in stathz ticks to run the long-term balancer");
3196 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3197 "Attempts to steal work from other cores before idling");
3198 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3199 "Minimum load on remote CPU before we'll steal");
3200 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3201 0, "Topological distance limit for stealing threads in sched_switch()");
3202 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3203 "Always run the stealer from the idle thread");
3204 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3205 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3206 "XML dump of detected CPU topology");
3209 /* ps compat. All cpu percentages from ULE are weighted. */
3210 static int ccpu = 0;
3211 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0,
3212 "Decay factor used for updating %CPU in 4BSD scheduler");