2 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice unmodified, this list of conditions, and the following
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 * This file implements the ULE scheduler. ULE supports independent CPU
29 * run queues and fine grain locking. It has superior interactive
30 * performance under load even on uni-processor systems.
33 * ULE is the last three letters in schedule. It owes its name to a
34 * generic user created for a scheduling system by Paul Mikesell at
35 * Isilon Systems and a general lack of creativity on the part of the author.
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
44 #include <sys/param.h>
45 #include <sys/systm.h>
47 #include <sys/kernel.h>
49 #include <sys/limits.h>
51 #include <sys/mutex.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
68 #include <sys/pmckern.h>
72 #include <sys/dtrace_bsd.h>
73 int dtrace_vtime_active;
74 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
82 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
83 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
84 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87 * Thread scheduler specific section. All fields are protected
91 struct runq *ts_runq; /* Run-queue we're queued on. */
92 short ts_flags; /* TSF_* flags. */
93 int ts_cpu; /* CPU that we have affinity for. */
94 int ts_rltick; /* Real last tick, for affinity. */
95 int ts_slice; /* Ticks of slice remaining. */
96 u_int ts_slptime; /* Number of ticks we vol. slept */
97 u_int ts_runtime; /* Number of ticks we were running */
98 int ts_ltick; /* Last tick that we were running on */
99 int ts_ftick; /* First tick that we were running on */
100 int ts_ticks; /* Tick count */
102 char ts_name[TS_NAME_LEN];
105 /* flags kept in ts_flags */
106 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
107 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
109 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
110 #define THREAD_CAN_SCHED(td, cpu) \
111 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
113 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
114 sizeof(struct thread0_storage),
115 "increase struct thread0_storage.t0st_sched size");
118 * Priority ranges used for interactive and non-interactive timeshare
119 * threads. The timeshare priorities are split up into four ranges.
120 * The first range handles interactive threads. The last three ranges
121 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
122 * ranges supporting nice values.
124 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
125 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
126 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
128 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
129 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
130 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
131 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
134 * Cpu percentage computation macros and defines.
136 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
137 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
138 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
139 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
140 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
141 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
143 #define SCHED_TICK_SECS 10
144 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
145 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
146 #define SCHED_TICK_SHIFT 10
147 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
148 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 * These macros determine priorities for non-interactive threads. They are
152 * assigned a priority based on their recent cpu utilization as expressed
153 * by the ratio of ticks to the tick total. NHALF priorities at the start
154 * and end of the MIN to MAX timeshare range are only reachable with negative
155 * or positive nice respectively.
157 * PRI_RANGE: Priority range for utilization dependent priorities.
158 * PRI_NRESV: Number of nice values.
159 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
160 * PRI_NICE: Determines the part of the priority inherited from nice.
162 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
163 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
164 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
165 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
166 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
167 #define SCHED_PRI_TICKS(ts) \
168 (SCHED_TICK_HZ((ts)) / \
169 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
170 #define SCHED_PRI_NICE(nice) (nice)
173 * These determine the interactivity of a process. Interactivity differs from
174 * cpu utilization in that it expresses the voluntary time slept vs time ran
175 * while cpu utilization includes all time not running. This more accurately
176 * models the intent of the thread.
178 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
179 * before throttling back.
180 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
181 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
182 * INTERACT_THRESH: Threshold for placement on the current runq.
184 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
185 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
186 #define SCHED_INTERACT_MAX (100)
187 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
188 #define SCHED_INTERACT_THRESH (30)
191 * These parameters determine the slice behavior for batch work.
193 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
194 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
196 /* Flags kept in td_flags. */
197 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
200 * tickincr: Converts a stathz tick into a hz domain scaled by
201 * the shift factor. Without the shift the error rate
202 * due to rounding would be unacceptably high.
203 * realstathz: stathz is sometimes 0 and run off of hz.
204 * sched_slice: Runtime of each thread before rescheduling.
205 * preempt_thresh: Priority threshold for preemption and remote IPIs.
207 static int sched_interact = SCHED_INTERACT_THRESH;
208 static int tickincr = 8 << SCHED_TICK_SHIFT;
209 static int realstathz = 127; /* reset during boot. */
210 static int sched_slice = 10; /* reset during boot. */
211 static int sched_slice_min = 1; /* reset during boot. */
213 #ifdef FULL_PREEMPTION
214 static int preempt_thresh = PRI_MAX_IDLE;
216 static int preempt_thresh = PRI_MIN_KERN;
219 static int preempt_thresh = 0;
221 static int static_boost = PRI_MIN_BATCH;
222 static int sched_idlespins = 10000;
223 static int sched_idlespinthresh = -1;
226 * tdq - per processor runqs and statistics. All fields are protected by the
227 * tdq_lock. The load and lowpri may be accessed without to avoid excess
228 * locking in sched_pickcpu();
232 * Ordered to improve efficiency of cpu_search() and switch().
233 * tdq_lock is padded to avoid false sharing with tdq_load and
236 struct mtx_padalign tdq_lock; /* run queue lock. */
237 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
238 volatile int tdq_load; /* Aggregate load. */
239 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
240 int tdq_sysload; /* For loadavg, !ITHD load. */
241 int tdq_transferable; /* Transferable thread count. */
242 short tdq_switchcnt; /* Switches this tick. */
243 short tdq_oldswitchcnt; /* Switches last tick. */
244 u_char tdq_lowpri; /* Lowest priority thread. */
245 u_char tdq_ipipending; /* IPI pending. */
246 u_char tdq_idx; /* Current insert index. */
247 u_char tdq_ridx; /* Current removal index. */
248 struct runq tdq_realtime; /* real-time run queue. */
249 struct runq tdq_timeshare; /* timeshare run queue. */
250 struct runq tdq_idle; /* Queue of IDLE threads. */
251 char tdq_name[TDQ_NAME_LEN];
253 char tdq_loadname[TDQ_LOADNAME_LEN];
257 /* Idle thread states and config. */
258 #define TDQ_RUNNING 1
262 struct cpu_group *cpu_top; /* CPU topology */
264 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
265 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
270 static int rebalance = 1;
271 static int balance_interval = 128; /* Default set in sched_initticks(). */
273 static int steal_idle = 1;
274 static int steal_thresh = 2;
277 * One thread queue per processor.
279 static struct tdq tdq_cpu[MAXCPU];
280 static struct tdq *balance_tdq;
281 static int balance_ticks;
282 static DPCPU_DEFINE(uint32_t, randomval);
284 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
285 #define TDQ_CPU(x) (&tdq_cpu[(x)])
286 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
288 static struct tdq tdq_cpu;
290 #define TDQ_ID(x) (0)
291 #define TDQ_SELF() (&tdq_cpu)
292 #define TDQ_CPU(x) (&tdq_cpu)
295 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
296 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
297 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
298 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
299 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
301 static void sched_priority(struct thread *);
302 static void sched_thread_priority(struct thread *, u_char);
303 static int sched_interact_score(struct thread *);
304 static void sched_interact_update(struct thread *);
305 static void sched_interact_fork(struct thread *);
306 static void sched_pctcpu_update(struct td_sched *, int);
308 /* Operations on per processor queues */
309 static struct thread *tdq_choose(struct tdq *);
310 static void tdq_setup(struct tdq *);
311 static void tdq_load_add(struct tdq *, struct thread *);
312 static void tdq_load_rem(struct tdq *, struct thread *);
313 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
314 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
315 static inline int sched_shouldpreempt(int, int, int);
316 void tdq_print(int cpu);
317 static void runq_print(struct runq *rq);
318 static void tdq_add(struct tdq *, struct thread *, int);
320 static int tdq_move(struct tdq *, struct tdq *);
321 static int tdq_idled(struct tdq *);
322 static void tdq_notify(struct tdq *, struct thread *);
323 static struct thread *tdq_steal(struct tdq *, int);
324 static struct thread *runq_steal(struct runq *, int);
325 static int sched_pickcpu(struct thread *, int);
326 static void sched_balance(void);
327 static int sched_balance_pair(struct tdq *, struct tdq *);
328 static inline struct tdq *sched_setcpu(struct thread *, int, int);
329 static inline void thread_unblock_switch(struct thread *, struct mtx *);
330 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
331 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
332 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
333 struct cpu_group *cg, int indent);
336 static void sched_setup(void *dummy);
337 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
339 static void sched_initticks(void *dummy);
340 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
343 SDT_PROVIDER_DEFINE(sched);
345 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
346 "struct proc *", "uint8_t");
347 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
348 "struct proc *", "void *");
349 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
350 "struct proc *", "void *", "int");
351 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
352 "struct proc *", "uint8_t", "struct thread *");
353 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
354 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
356 SDT_PROBE_DEFINE(sched, , , on__cpu);
357 SDT_PROBE_DEFINE(sched, , , remain__cpu);
358 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
362 * Print the threads waiting on a run-queue.
365 runq_print(struct runq *rq)
373 for (i = 0; i < RQB_LEN; i++) {
374 printf("\t\trunq bits %d 0x%zx\n",
375 i, rq->rq_status.rqb_bits[i]);
376 for (j = 0; j < RQB_BPW; j++)
377 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
378 pri = j + (i << RQB_L2BPW);
379 rqh = &rq->rq_queues[pri];
380 TAILQ_FOREACH(td, rqh, td_runq) {
381 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
382 td, td->td_name, td->td_priority,
383 td->td_rqindex, pri);
390 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
399 printf("tdq %d:\n", TDQ_ID(tdq));
400 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
401 printf("\tLock name: %s\n", tdq->tdq_name);
402 printf("\tload: %d\n", tdq->tdq_load);
403 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
404 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
405 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
406 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
407 printf("\tload transferable: %d\n", tdq->tdq_transferable);
408 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
409 printf("\trealtime runq:\n");
410 runq_print(&tdq->tdq_realtime);
411 printf("\ttimeshare runq:\n");
412 runq_print(&tdq->tdq_timeshare);
413 printf("\tidle runq:\n");
414 runq_print(&tdq->tdq_idle);
418 sched_shouldpreempt(int pri, int cpri, int remote)
421 * If the new priority is not better than the current priority there is
427 * Always preempt idle.
429 if (cpri >= PRI_MIN_IDLE)
432 * If preemption is disabled don't preempt others.
434 if (preempt_thresh == 0)
437 * Preempt if we exceed the threshold.
439 if (pri <= preempt_thresh)
442 * If we're interactive or better and there is non-interactive
443 * or worse running preempt only remote processors.
445 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
451 * Add a thread to the actual run-queue. Keeps transferable counts up to
452 * date with what is actually on the run-queue. Selects the correct
453 * queue position for timeshare threads.
456 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
461 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
462 THREAD_LOCK_ASSERT(td, MA_OWNED);
464 pri = td->td_priority;
465 ts = td_get_sched(td);
467 if (THREAD_CAN_MIGRATE(td)) {
468 tdq->tdq_transferable++;
469 ts->ts_flags |= TSF_XFERABLE;
471 if (pri < PRI_MIN_BATCH) {
472 ts->ts_runq = &tdq->tdq_realtime;
473 } else if (pri <= PRI_MAX_BATCH) {
474 ts->ts_runq = &tdq->tdq_timeshare;
475 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
476 ("Invalid priority %d on timeshare runq", pri));
478 * This queue contains only priorities between MIN and MAX
479 * realtime. Use the whole queue to represent these values.
481 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
482 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
483 pri = (pri + tdq->tdq_idx) % RQ_NQS;
485 * This effectively shortens the queue by one so we
486 * can have a one slot difference between idx and
487 * ridx while we wait for threads to drain.
489 if (tdq->tdq_ridx != tdq->tdq_idx &&
490 pri == tdq->tdq_ridx)
491 pri = (unsigned char)(pri - 1) % RQ_NQS;
494 runq_add_pri(ts->ts_runq, td, pri, flags);
497 ts->ts_runq = &tdq->tdq_idle;
498 runq_add(ts->ts_runq, td, flags);
502 * Remove a thread from a run-queue. This typically happens when a thread
503 * is selected to run. Running threads are not on the queue and the
504 * transferable count does not reflect them.
507 tdq_runq_rem(struct tdq *tdq, struct thread *td)
511 ts = td_get_sched(td);
512 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
513 KASSERT(ts->ts_runq != NULL,
514 ("tdq_runq_remove: thread %p null ts_runq", td));
515 if (ts->ts_flags & TSF_XFERABLE) {
516 tdq->tdq_transferable--;
517 ts->ts_flags &= ~TSF_XFERABLE;
519 if (ts->ts_runq == &tdq->tdq_timeshare) {
520 if (tdq->tdq_idx != tdq->tdq_ridx)
521 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
523 runq_remove_idx(ts->ts_runq, td, NULL);
525 runq_remove(ts->ts_runq, td);
529 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
530 * for this thread to the referenced thread queue.
533 tdq_load_add(struct tdq *tdq, struct thread *td)
536 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
537 THREAD_LOCK_ASSERT(td, MA_OWNED);
540 if ((td->td_flags & TDF_NOLOAD) == 0)
542 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
543 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
547 * Remove the load from a thread that is transitioning to a sleep state or
551 tdq_load_rem(struct tdq *tdq, struct thread *td)
554 THREAD_LOCK_ASSERT(td, MA_OWNED);
555 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
556 KASSERT(tdq->tdq_load != 0,
557 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
560 if ((td->td_flags & TDF_NOLOAD) == 0)
562 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
563 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
567 * Bound timeshare latency by decreasing slice size as load increases. We
568 * consider the maximum latency as the sum of the threads waiting to run
569 * aside from curthread and target no more than sched_slice latency but
570 * no less than sched_slice_min runtime.
573 tdq_slice(struct tdq *tdq)
578 * It is safe to use sys_load here because this is called from
579 * contexts where timeshare threads are running and so there
580 * cannot be higher priority load in the system.
582 load = tdq->tdq_sysload - 1;
583 if (load >= SCHED_SLICE_MIN_DIVISOR)
584 return (sched_slice_min);
586 return (sched_slice);
587 return (sched_slice / load);
591 * Set lowpri to its exact value by searching the run-queue and
592 * evaluating curthread. curthread may be passed as an optimization.
595 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
599 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
601 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
602 td = tdq_choose(tdq);
603 if (td == NULL || td->td_priority > ctd->td_priority)
604 tdq->tdq_lowpri = ctd->td_priority;
606 tdq->tdq_lowpri = td->td_priority;
611 * We need some randomness. Implement a classic Linear Congruential
612 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
613 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
614 * of the random state (in the low bits of our answer) to keep
615 * the maximum randomness.
622 rndptr = DPCPU_PTR(randomval);
623 *rndptr = *rndptr * 69069 + 5;
625 return (*rndptr >> 16);
631 int cs_pri; /* Min priority for low. */
632 int cs_limit; /* Max load for low, min load for high. */
637 #define CPU_SEARCH_LOWEST 0x1
638 #define CPU_SEARCH_HIGHEST 0x2
639 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
641 #define CPUSET_FOREACH(cpu, mask) \
642 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
643 if (CPU_ISSET(cpu, &mask))
645 static __always_inline int cpu_search(const struct cpu_group *cg,
646 struct cpu_search *low, struct cpu_search *high, const int match);
647 int __noinline cpu_search_lowest(const struct cpu_group *cg,
648 struct cpu_search *low);
649 int __noinline cpu_search_highest(const struct cpu_group *cg,
650 struct cpu_search *high);
651 int __noinline cpu_search_both(const struct cpu_group *cg,
652 struct cpu_search *low, struct cpu_search *high);
655 * Search the tree of cpu_groups for the lowest or highest loaded cpu
656 * according to the match argument. This routine actually compares the
657 * load on all paths through the tree and finds the least loaded cpu on
658 * the least loaded path, which may differ from the least loaded cpu in
659 * the system. This balances work among caches and buses.
661 * This inline is instantiated in three forms below using constants for the
662 * match argument. It is reduced to the minimum set for each case. It is
663 * also recursive to the depth of the tree.
665 static __always_inline int
666 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
667 struct cpu_search *high, const int match)
669 struct cpu_search lgroup;
670 struct cpu_search hgroup;
672 struct cpu_group *child;
674 int cpu, i, hload, lload, load, total, rnd;
677 cpumask = cg->cg_mask;
678 if (match & CPU_SEARCH_LOWEST) {
682 if (match & CPU_SEARCH_HIGHEST) {
687 /* Iterate through the child CPU groups and then remaining CPUs. */
688 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
690 #ifdef HAVE_INLINE_FFSL
691 cpu = CPU_FFS(&cpumask) - 1;
693 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
700 child = &cg->cg_child[i - 1];
702 if (match & CPU_SEARCH_LOWEST)
704 if (match & CPU_SEARCH_HIGHEST)
706 if (child) { /* Handle child CPU group. */
707 CPU_NAND(&cpumask, &child->cg_mask);
709 case CPU_SEARCH_LOWEST:
710 load = cpu_search_lowest(child, &lgroup);
712 case CPU_SEARCH_HIGHEST:
713 load = cpu_search_highest(child, &hgroup);
715 case CPU_SEARCH_BOTH:
716 load = cpu_search_both(child, &lgroup, &hgroup);
719 } else { /* Handle child CPU. */
720 CPU_CLR(cpu, &cpumask);
722 load = tdq->tdq_load * 256;
723 rnd = sched_random() % 32;
724 if (match & CPU_SEARCH_LOWEST) {
725 if (cpu == low->cs_prefer)
727 /* If that CPU is allowed and get data. */
728 if (tdq->tdq_lowpri > lgroup.cs_pri &&
729 tdq->tdq_load <= lgroup.cs_limit &&
730 CPU_ISSET(cpu, &lgroup.cs_mask)) {
732 lgroup.cs_load = load - rnd;
735 if (match & CPU_SEARCH_HIGHEST)
736 if (tdq->tdq_load >= hgroup.cs_limit &&
737 tdq->tdq_transferable &&
738 CPU_ISSET(cpu, &hgroup.cs_mask)) {
740 hgroup.cs_load = load - rnd;
745 /* We have info about child item. Compare it. */
746 if (match & CPU_SEARCH_LOWEST) {
747 if (lgroup.cs_cpu >= 0 &&
749 (load == lload && lgroup.cs_load < low->cs_load))) {
751 low->cs_cpu = lgroup.cs_cpu;
752 low->cs_load = lgroup.cs_load;
755 if (match & CPU_SEARCH_HIGHEST)
756 if (hgroup.cs_cpu >= 0 &&
758 (load == hload && hgroup.cs_load > high->cs_load))) {
760 high->cs_cpu = hgroup.cs_cpu;
761 high->cs_load = hgroup.cs_load;
765 if (i == 0 && CPU_EMPTY(&cpumask))
768 #ifndef HAVE_INLINE_FFSL
777 * cpu_search instantiations must pass constants to maintain the inline
781 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
783 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
787 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
789 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
793 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
794 struct cpu_search *high)
796 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
800 * Find the cpu with the least load via the least loaded path that has a
801 * lowpri greater than pri pri. A pri of -1 indicates any priority is
805 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
808 struct cpu_search low;
811 low.cs_prefer = prefer;
814 low.cs_limit = maxload;
815 cpu_search_lowest(cg, &low);
820 * Find the cpu with the highest load via the highest loaded path.
823 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
825 struct cpu_search high;
829 high.cs_limit = minload;
830 cpu_search_highest(cg, &high);
835 sched_balance_group(struct cpu_group *cg)
837 cpuset_t hmask, lmask;
838 int high, low, anylow;
842 high = sched_highest(cg, hmask, 1);
843 /* Stop if there is no more CPU with transferrable threads. */
846 CPU_CLR(high, &hmask);
847 CPU_COPY(&hmask, &lmask);
848 /* Stop if there is no more CPU left for low. */
849 if (CPU_EMPTY(&lmask))
853 low = sched_lowest(cg, lmask, -1,
854 TDQ_CPU(high)->tdq_load - 1, high);
855 /* Stop if we looked well and found no less loaded CPU. */
856 if (anylow && low == -1)
858 /* Go to next high if we found no less loaded CPU. */
861 /* Transfer thread from high to low. */
862 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
863 /* CPU that got thread can no longer be a donor. */
864 CPU_CLR(low, &hmask);
867 * If failed, then there is no threads on high
868 * that can run on this low. Drop low from low
869 * mask and look for different one.
871 CPU_CLR(low, &lmask);
883 if (smp_started == 0 || rebalance == 0)
886 balance_ticks = max(balance_interval / 2, 1) +
887 (sched_random() % balance_interval);
890 sched_balance_group(cpu_top);
895 * Lock two thread queues using their address to maintain lock order.
898 tdq_lock_pair(struct tdq *one, struct tdq *two)
902 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
905 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
910 * Unlock two thread queues. Order is not important here.
913 tdq_unlock_pair(struct tdq *one, struct tdq *two)
920 * Transfer load between two imbalanced thread queues.
923 sched_balance_pair(struct tdq *high, struct tdq *low)
928 tdq_lock_pair(high, low);
931 * Determine what the imbalance is and then adjust that to how many
932 * threads we actually have to give up (transferable).
934 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
935 (moved = tdq_move(high, low)) > 0) {
937 * In case the target isn't the current cpu IPI it to force a
938 * reschedule with the new workload.
941 if (cpu != PCPU_GET(cpuid))
942 ipi_cpu(cpu, IPI_PREEMPT);
944 tdq_unlock_pair(high, low);
949 * Move a thread from one thread queue to another.
952 tdq_move(struct tdq *from, struct tdq *to)
959 TDQ_LOCK_ASSERT(from, MA_OWNED);
960 TDQ_LOCK_ASSERT(to, MA_OWNED);
964 td = tdq_steal(tdq, cpu);
967 ts = td_get_sched(td);
969 * Although the run queue is locked the thread may be blocked. Lock
970 * it to clear this and acquire the run-queue lock.
973 /* Drop recursive lock on from acquired via thread_lock(). */
977 td->td_lock = TDQ_LOCKPTR(to);
978 tdq_add(to, td, SRQ_YIELDING);
983 * This tdq has idled. Try to steal a thread from another cpu and switch
987 tdq_idled(struct tdq *tdq)
989 struct cpu_group *cg;
995 if (smp_started == 0 || steal_idle == 0)
998 CPU_CLR(PCPU_GET(cpuid), &mask);
999 /* We don't want to be preempted while we're iterating. */
1001 for (cg = tdq->tdq_cg; cg != NULL; ) {
1002 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1003 thresh = steal_thresh;
1006 cpu = sched_highest(cg, mask, thresh);
1011 steal = TDQ_CPU(cpu);
1012 CPU_CLR(cpu, &mask);
1013 tdq_lock_pair(tdq, steal);
1014 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1015 tdq_unlock_pair(tdq, steal);
1019 * If a thread was added while interrupts were disabled don't
1020 * steal one here. If we fail to acquire one due to affinity
1021 * restrictions loop again with this cpu removed from the
1024 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1025 tdq_unlock_pair(tdq, steal);
1030 mi_switch(SW_VOL | SWT_IDLE, NULL);
1031 thread_unlock(curthread);
1040 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1043 tdq_notify(struct tdq *tdq, struct thread *td)
1049 if (tdq->tdq_ipipending)
1051 cpu = td_get_sched(td)->ts_cpu;
1052 pri = td->td_priority;
1053 ctd = pcpu_find(cpu)->pc_curthread;
1054 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1058 * Make sure that our caller's earlier update to tdq_load is
1059 * globally visible before we read tdq_cpu_idle. Idle thread
1060 * accesses both of them without locks, and the order is important.
1062 atomic_thread_fence_seq_cst();
1064 if (TD_IS_IDLETHREAD(ctd)) {
1066 * If the MD code has an idle wakeup routine try that before
1067 * falling back to IPI.
1069 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1072 tdq->tdq_ipipending = 1;
1073 ipi_cpu(cpu, IPI_PREEMPT);
1077 * Steals load from a timeshare queue. Honors the rotating queue head
1080 static struct thread *
1081 runq_steal_from(struct runq *rq, int cpu, u_char start)
1085 struct thread *td, *first;
1089 rqb = &rq->rq_status;
1090 bit = start & (RQB_BPW -1);
1093 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1094 if (rqb->rqb_bits[i] == 0)
1097 bit = RQB_FFS(rqb->rqb_bits[i]);
1098 for (; bit < RQB_BPW; bit++) {
1099 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1101 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1102 TAILQ_FOREACH(td, rqh, td_runq) {
1103 if (first && THREAD_CAN_MIGRATE(td) &&
1104 THREAD_CAN_SCHED(td, cpu))
1115 if (first && THREAD_CAN_MIGRATE(first) &&
1116 THREAD_CAN_SCHED(first, cpu))
1122 * Steals load from a standard linear queue.
1124 static struct thread *
1125 runq_steal(struct runq *rq, int cpu)
1133 rqb = &rq->rq_status;
1134 for (word = 0; word < RQB_LEN; word++) {
1135 if (rqb->rqb_bits[word] == 0)
1137 for (bit = 0; bit < RQB_BPW; bit++) {
1138 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1140 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1141 TAILQ_FOREACH(td, rqh, td_runq)
1142 if (THREAD_CAN_MIGRATE(td) &&
1143 THREAD_CAN_SCHED(td, cpu))
1151 * Attempt to steal a thread in priority order from a thread queue.
1153 static struct thread *
1154 tdq_steal(struct tdq *tdq, int cpu)
1158 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1159 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1161 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1162 cpu, tdq->tdq_ridx)) != NULL)
1164 return (runq_steal(&tdq->tdq_idle, cpu));
1168 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1169 * current lock and returns with the assigned queue locked.
1171 static inline struct tdq *
1172 sched_setcpu(struct thread *td, int cpu, int flags)
1177 THREAD_LOCK_ASSERT(td, MA_OWNED);
1179 td_get_sched(td)->ts_cpu = cpu;
1181 * If the lock matches just return the queue.
1183 if (td->td_lock == TDQ_LOCKPTR(tdq))
1187 * If the thread isn't running its lockptr is a
1188 * turnstile or a sleepqueue. We can just lock_set without
1191 if (TD_CAN_RUN(td)) {
1193 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1198 * The hard case, migration, we need to block the thread first to
1199 * prevent order reversals with other cpus locks.
1202 thread_lock_block(td);
1204 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1209 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1210 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1211 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1212 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1213 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1214 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1217 sched_pickcpu(struct thread *td, int flags)
1219 struct cpu_group *cg, *ccg;
1220 struct td_sched *ts;
1225 self = PCPU_GET(cpuid);
1226 ts = td_get_sched(td);
1227 if (smp_started == 0)
1230 * Don't migrate a running thread from sched_switch().
1232 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1233 return (ts->ts_cpu);
1235 * Prefer to run interrupt threads on the processors that generate
1238 pri = td->td_priority;
1239 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1240 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1241 SCHED_STAT_INC(pickcpu_intrbind);
1243 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1244 SCHED_STAT_INC(pickcpu_affinity);
1245 return (ts->ts_cpu);
1249 * If the thread can run on the last cpu and the affinity has not
1250 * expired or it is idle run it there.
1252 tdq = TDQ_CPU(ts->ts_cpu);
1254 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1255 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1256 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1257 if (cg->cg_flags & CG_FLAG_THREAD) {
1258 CPUSET_FOREACH(cpu, cg->cg_mask) {
1259 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1264 if (cpu > mp_maxid) {
1265 SCHED_STAT_INC(pickcpu_idle_affinity);
1266 return (ts->ts_cpu);
1270 * Search for the last level cache CPU group in the tree.
1271 * Skip caches with expired affinity time and SMT groups.
1272 * Affinity to higher level caches will be handled less aggressively.
1274 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1275 if (cg->cg_flags & CG_FLAG_THREAD)
1277 if (!SCHED_AFFINITY(ts, cg->cg_level))
1284 /* Search the group for the less loaded idle CPU we can run now. */
1285 mask = td->td_cpuset->cs_mask;
1286 if (cg != NULL && cg != cpu_top &&
1287 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1288 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1289 INT_MAX, ts->ts_cpu);
1290 /* Search globally for the less loaded CPU we can run now. */
1292 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1293 /* Search globally for the less loaded CPU. */
1295 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1296 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1298 * Compare the lowest loaded cpu to current cpu.
1300 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1301 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1302 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1303 SCHED_STAT_INC(pickcpu_local);
1306 SCHED_STAT_INC(pickcpu_lowest);
1307 if (cpu != ts->ts_cpu)
1308 SCHED_STAT_INC(pickcpu_migration);
1314 * Pick the highest priority task we have and return it.
1316 static struct thread *
1317 tdq_choose(struct tdq *tdq)
1321 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1322 td = runq_choose(&tdq->tdq_realtime);
1325 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1327 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1328 ("tdq_choose: Invalid priority on timeshare queue %d",
1332 td = runq_choose(&tdq->tdq_idle);
1334 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1335 ("tdq_choose: Invalid priority on idle queue %d",
1344 * Initialize a thread queue.
1347 tdq_setup(struct tdq *tdq)
1351 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1352 runq_init(&tdq->tdq_realtime);
1353 runq_init(&tdq->tdq_timeshare);
1354 runq_init(&tdq->tdq_idle);
1355 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1356 "sched lock %d", (int)TDQ_ID(tdq));
1357 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1358 MTX_SPIN | MTX_RECURSE);
1360 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1361 "CPU %d load", (int)TDQ_ID(tdq));
1367 sched_setup_smp(void)
1372 cpu_top = smp_topo();
1376 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1377 if (tdq->tdq_cg == NULL)
1378 panic("Can't find cpu group for %d\n", i);
1380 balance_tdq = TDQ_SELF();
1386 * Setup the thread queues and initialize the topology based on MD
1390 sched_setup(void *dummy)
1401 /* Add thread0's load since it's running. */
1403 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1404 tdq_load_add(tdq, &thread0);
1405 tdq->tdq_lowpri = thread0.td_priority;
1410 * This routine determines time constants after stathz and hz are setup.
1414 sched_initticks(void *dummy)
1418 realstathz = stathz ? stathz : hz;
1419 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1420 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1421 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1425 * tickincr is shifted out by 10 to avoid rounding errors due to
1426 * hz not being evenly divisible by stathz on all platforms.
1428 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1430 * This does not work for values of stathz that are more than
1431 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1438 * Set the default balance interval now that we know
1439 * what realstathz is.
1441 balance_interval = realstathz;
1442 affinity = SCHED_AFFINITY_DEFAULT;
1444 if (sched_idlespinthresh < 0)
1445 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1450 * This is the core of the interactivity algorithm. Determines a score based
1451 * on past behavior. It is the ratio of sleep time to run time scaled to
1452 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1453 * differs from the cpu usage because it does not account for time spent
1454 * waiting on a run-queue. Would be prettier if we had floating point.
1456 * When a thread's sleep time is greater than its run time the
1460 * interactivity score = ---------------------
1461 * sleep time / run time
1464 * When a thread's run time is greater than its sleep time the
1468 * interactivity score = --------------------- + scaling factor
1469 * run time / sleep time
1472 sched_interact_score(struct thread *td)
1474 struct td_sched *ts;
1477 ts = td_get_sched(td);
1479 * The score is only needed if this is likely to be an interactive
1480 * task. Don't go through the expense of computing it if there's
1483 if (sched_interact <= SCHED_INTERACT_HALF &&
1484 ts->ts_runtime >= ts->ts_slptime)
1485 return (SCHED_INTERACT_HALF);
1487 if (ts->ts_runtime > ts->ts_slptime) {
1488 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1489 return (SCHED_INTERACT_HALF +
1490 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1492 if (ts->ts_slptime > ts->ts_runtime) {
1493 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1494 return (ts->ts_runtime / div);
1496 /* runtime == slptime */
1498 return (SCHED_INTERACT_HALF);
1501 * This can happen if slptime and runtime are 0.
1508 * Scale the scheduling priority according to the "interactivity" of this
1512 sched_priority(struct thread *td)
1517 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1520 * If the score is interactive we place the thread in the realtime
1521 * queue with a priority that is less than kernel and interrupt
1522 * priorities. These threads are not subject to nice restrictions.
1524 * Scores greater than this are placed on the normal timeshare queue
1525 * where the priority is partially decided by the most recent cpu
1526 * utilization and the rest is decided by nice value.
1528 * The nice value of the process has a linear effect on the calculated
1529 * score. Negative nice values make it easier for a thread to be
1530 * considered interactive.
1532 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1533 if (score < sched_interact) {
1534 pri = PRI_MIN_INTERACT;
1535 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1536 sched_interact) * score;
1537 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1538 ("sched_priority: invalid interactive priority %d score %d",
1541 pri = SCHED_PRI_MIN;
1542 if (td_get_sched(td)->ts_ticks)
1543 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1544 SCHED_PRI_RANGE - 1);
1545 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1546 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1547 ("sched_priority: invalid priority %d: nice %d, "
1548 "ticks %d ftick %d ltick %d tick pri %d",
1549 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1550 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1551 SCHED_PRI_TICKS(td_get_sched(td))));
1553 sched_user_prio(td, pri);
1559 * This routine enforces a maximum limit on the amount of scheduling history
1560 * kept. It is called after either the slptime or runtime is adjusted. This
1561 * function is ugly due to integer math.
1564 sched_interact_update(struct thread *td)
1566 struct td_sched *ts;
1569 ts = td_get_sched(td);
1570 sum = ts->ts_runtime + ts->ts_slptime;
1571 if (sum < SCHED_SLP_RUN_MAX)
1574 * This only happens from two places:
1575 * 1) We have added an unusual amount of run time from fork_exit.
1576 * 2) We have added an unusual amount of sleep time from sched_sleep().
1578 if (sum > SCHED_SLP_RUN_MAX * 2) {
1579 if (ts->ts_runtime > ts->ts_slptime) {
1580 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1583 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1589 * If we have exceeded by more than 1/5th then the algorithm below
1590 * will not bring us back into range. Dividing by two here forces
1591 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1593 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1594 ts->ts_runtime /= 2;
1595 ts->ts_slptime /= 2;
1598 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1599 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1603 * Scale back the interactivity history when a child thread is created. The
1604 * history is inherited from the parent but the thread may behave totally
1605 * differently. For example, a shell spawning a compiler process. We want
1606 * to learn that the compiler is behaving badly very quickly.
1609 sched_interact_fork(struct thread *td)
1611 struct td_sched *ts;
1615 ts = td_get_sched(td);
1616 sum = ts->ts_runtime + ts->ts_slptime;
1617 if (sum > SCHED_SLP_RUN_FORK) {
1618 ratio = sum / SCHED_SLP_RUN_FORK;
1619 ts->ts_runtime /= ratio;
1620 ts->ts_slptime /= ratio;
1625 * Called from proc0_init() to setup the scheduler fields.
1630 struct td_sched *ts0;
1633 * Set up the scheduler specific parts of thread0.
1635 ts0 = td_get_sched(&thread0);
1636 ts0->ts_ltick = ticks;
1637 ts0->ts_ftick = ticks;
1642 * This is only somewhat accurate since given many processes of the same
1643 * priority they will switch when their slices run out, which will be
1644 * at most sched_slice stathz ticks.
1647 sched_rr_interval(void)
1650 /* Convert sched_slice from stathz to hz. */
1651 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1655 * Update the percent cpu tracking information when it is requested or
1656 * the total history exceeds the maximum. We keep a sliding history of
1657 * tick counts that slowly decays. This is less precise than the 4BSD
1658 * mechanism since it happens with less regular and frequent events.
1661 sched_pctcpu_update(struct td_sched *ts, int run)
1665 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1667 ts->ts_ftick = t - SCHED_TICK_TARG;
1668 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1669 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1670 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1671 ts->ts_ftick = t - SCHED_TICK_TARG;
1674 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1679 * Adjust the priority of a thread. Move it to the appropriate run-queue
1680 * if necessary. This is the back-end for several priority related
1684 sched_thread_priority(struct thread *td, u_char prio)
1686 struct td_sched *ts;
1690 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1691 "prio:%d", td->td_priority, "new prio:%d", prio,
1692 KTR_ATTR_LINKED, sched_tdname(curthread));
1693 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1694 if (td != curthread && prio < td->td_priority) {
1695 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1696 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1697 prio, KTR_ATTR_LINKED, sched_tdname(td));
1698 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1701 ts = td_get_sched(td);
1702 THREAD_LOCK_ASSERT(td, MA_OWNED);
1703 if (td->td_priority == prio)
1706 * If the priority has been elevated due to priority
1707 * propagation, we may have to move ourselves to a new
1708 * queue. This could be optimized to not re-add in some
1711 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1713 td->td_priority = prio;
1714 sched_add(td, SRQ_BORROWING);
1718 * If the thread is currently running we may have to adjust the lowpri
1719 * information so other cpus are aware of our current priority.
1721 if (TD_IS_RUNNING(td)) {
1722 tdq = TDQ_CPU(ts->ts_cpu);
1723 oldpri = td->td_priority;
1724 td->td_priority = prio;
1725 if (prio < tdq->tdq_lowpri)
1726 tdq->tdq_lowpri = prio;
1727 else if (tdq->tdq_lowpri == oldpri)
1728 tdq_setlowpri(tdq, td);
1731 td->td_priority = prio;
1735 * Update a thread's priority when it is lent another thread's
1739 sched_lend_prio(struct thread *td, u_char prio)
1742 td->td_flags |= TDF_BORROWING;
1743 sched_thread_priority(td, prio);
1747 * Restore a thread's priority when priority propagation is
1748 * over. The prio argument is the minimum priority the thread
1749 * needs to have to satisfy other possible priority lending
1750 * requests. If the thread's regular priority is less
1751 * important than prio, the thread will keep a priority boost
1755 sched_unlend_prio(struct thread *td, u_char prio)
1759 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1760 td->td_base_pri <= PRI_MAX_TIMESHARE)
1761 base_pri = td->td_user_pri;
1763 base_pri = td->td_base_pri;
1764 if (prio >= base_pri) {
1765 td->td_flags &= ~TDF_BORROWING;
1766 sched_thread_priority(td, base_pri);
1768 sched_lend_prio(td, prio);
1772 * Standard entry for setting the priority to an absolute value.
1775 sched_prio(struct thread *td, u_char prio)
1779 /* First, update the base priority. */
1780 td->td_base_pri = prio;
1783 * If the thread is borrowing another thread's priority, don't
1784 * ever lower the priority.
1786 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1789 /* Change the real priority. */
1790 oldprio = td->td_priority;
1791 sched_thread_priority(td, prio);
1794 * If the thread is on a turnstile, then let the turnstile update
1797 if (TD_ON_LOCK(td) && oldprio != prio)
1798 turnstile_adjust(td, oldprio);
1802 * Set the base user priority, does not effect current running priority.
1805 sched_user_prio(struct thread *td, u_char prio)
1808 td->td_base_user_pri = prio;
1809 if (td->td_lend_user_pri <= prio)
1811 td->td_user_pri = prio;
1815 sched_lend_user_prio(struct thread *td, u_char prio)
1818 THREAD_LOCK_ASSERT(td, MA_OWNED);
1819 td->td_lend_user_pri = prio;
1820 td->td_user_pri = min(prio, td->td_base_user_pri);
1821 if (td->td_priority > td->td_user_pri)
1822 sched_prio(td, td->td_user_pri);
1823 else if (td->td_priority != td->td_user_pri)
1824 td->td_flags |= TDF_NEEDRESCHED;
1828 * Handle migration from sched_switch(). This happens only for
1832 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1836 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1838 tdq_load_rem(tdq, td);
1840 * Do the lock dance required to avoid LOR. We grab an extra
1841 * spinlock nesting to prevent preemption while we're
1842 * not holding either run-queue lock.
1845 thread_lock_block(td); /* This releases the lock on tdq. */
1848 * Acquire both run-queue locks before placing the thread on the new
1849 * run-queue to avoid deadlocks created by placing a thread with a
1850 * blocked lock on the run-queue of a remote processor. The deadlock
1851 * occurs when a third processor attempts to lock the two queues in
1852 * question while the target processor is spinning with its own
1853 * run-queue lock held while waiting for the blocked lock to clear.
1855 tdq_lock_pair(tdn, tdq);
1856 tdq_add(tdn, td, flags);
1857 tdq_notify(tdn, td);
1861 return (TDQ_LOCKPTR(tdn));
1865 * Variadic version of thread_lock_unblock() that does not assume td_lock
1869 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1871 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1876 * Switch threads. This function has to handle threads coming in while
1877 * blocked for some reason, running, or idle. It also must deal with
1878 * migrating a thread from one queue to another as running threads may
1879 * be assigned elsewhere via binding.
1882 sched_switch(struct thread *td, struct thread *newtd, int flags)
1885 struct td_sched *ts;
1888 int cpuid, preempted;
1890 THREAD_LOCK_ASSERT(td, MA_OWNED);
1891 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1893 cpuid = PCPU_GET(cpuid);
1894 tdq = TDQ_CPU(cpuid);
1895 ts = td_get_sched(td);
1897 sched_pctcpu_update(ts, 1);
1898 ts->ts_rltick = ticks;
1899 td->td_lastcpu = td->td_oncpu;
1900 td->td_oncpu = NOCPU;
1901 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
1902 (flags & SW_PREEMPT) != 0;
1903 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1904 td->td_owepreempt = 0;
1905 if (!TD_IS_IDLETHREAD(td))
1906 tdq->tdq_switchcnt++;
1908 * The lock pointer in an idle thread should never change. Reset it
1909 * to CAN_RUN as well.
1911 if (TD_IS_IDLETHREAD(td)) {
1912 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1914 } else if (TD_IS_RUNNING(td)) {
1915 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1916 srqflag = preempted ?
1917 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1918 SRQ_OURSELF|SRQ_YIELDING;
1920 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1921 ts->ts_cpu = sched_pickcpu(td, 0);
1923 if (ts->ts_cpu == cpuid)
1924 tdq_runq_add(tdq, td, srqflag);
1926 KASSERT(THREAD_CAN_MIGRATE(td) ||
1927 (ts->ts_flags & TSF_BOUND) != 0,
1928 ("Thread %p shouldn't migrate", td));
1929 mtx = sched_switch_migrate(tdq, td, srqflag);
1932 /* This thread must be going to sleep. */
1934 mtx = thread_lock_block(td);
1935 tdq_load_rem(tdq, td);
1938 * We enter here with the thread blocked and assigned to the
1939 * appropriate cpu run-queue or sleep-queue and with the current
1940 * thread-queue locked.
1942 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1943 newtd = choosethread();
1945 * Call the MD code to switch contexts if necessary.
1949 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1950 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1952 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1953 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1954 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1955 sched_pctcpu_update(td_get_sched(newtd), 0);
1957 #ifdef KDTRACE_HOOKS
1959 * If DTrace has set the active vtime enum to anything
1960 * other than INACTIVE (0), then it should have set the
1963 if (dtrace_vtime_active)
1964 (*dtrace_vtime_switch_func)(newtd);
1967 cpu_switch(td, newtd, mtx);
1969 * We may return from cpu_switch on a different cpu. However,
1970 * we always return with td_lock pointing to the current cpu's
1973 cpuid = PCPU_GET(cpuid);
1974 tdq = TDQ_CPU(cpuid);
1975 lock_profile_obtain_lock_success(
1976 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1978 SDT_PROBE0(sched, , , on__cpu);
1980 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1981 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1984 thread_unblock_switch(td, mtx);
1985 SDT_PROBE0(sched, , , remain__cpu);
1988 * Assert that all went well and return.
1990 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1991 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1992 td->td_oncpu = cpuid;
1996 * Adjust thread priorities as a result of a nice request.
1999 sched_nice(struct proc *p, int nice)
2003 PROC_LOCK_ASSERT(p, MA_OWNED);
2006 FOREACH_THREAD_IN_PROC(p, td) {
2009 sched_prio(td, td->td_base_user_pri);
2015 * Record the sleep time for the interactivity scorer.
2018 sched_sleep(struct thread *td, int prio)
2021 THREAD_LOCK_ASSERT(td, MA_OWNED);
2023 td->td_slptick = ticks;
2024 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2025 td->td_flags |= TDF_CANSWAP;
2026 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2028 if (static_boost == 1 && prio)
2029 sched_prio(td, prio);
2030 else if (static_boost && td->td_priority > static_boost)
2031 sched_prio(td, static_boost);
2035 * Schedule a thread to resume execution and record how long it voluntarily
2036 * slept. We also update the pctcpu, interactivity, and priority.
2039 sched_wakeup(struct thread *td)
2041 struct td_sched *ts;
2044 THREAD_LOCK_ASSERT(td, MA_OWNED);
2045 ts = td_get_sched(td);
2046 td->td_flags &= ~TDF_CANSWAP;
2048 * If we slept for more than a tick update our interactivity and
2051 slptick = td->td_slptick;
2053 if (slptick && slptick != ticks) {
2054 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2055 sched_interact_update(td);
2056 sched_pctcpu_update(ts, 0);
2059 * Reset the slice value since we slept and advanced the round-robin.
2062 sched_add(td, SRQ_BORING);
2066 * Penalize the parent for creating a new child and initialize the child's
2070 sched_fork(struct thread *td, struct thread *child)
2072 THREAD_LOCK_ASSERT(td, MA_OWNED);
2073 sched_pctcpu_update(td_get_sched(td), 1);
2074 sched_fork_thread(td, child);
2076 * Penalize the parent and child for forking.
2078 sched_interact_fork(child);
2079 sched_priority(child);
2080 td_get_sched(td)->ts_runtime += tickincr;
2081 sched_interact_update(td);
2086 * Fork a new thread, may be within the same process.
2089 sched_fork_thread(struct thread *td, struct thread *child)
2091 struct td_sched *ts;
2092 struct td_sched *ts2;
2096 THREAD_LOCK_ASSERT(td, MA_OWNED);
2100 ts = td_get_sched(td);
2101 ts2 = td_get_sched(child);
2102 child->td_oncpu = NOCPU;
2103 child->td_lastcpu = NOCPU;
2104 child->td_lock = TDQ_LOCKPTR(tdq);
2105 child->td_cpuset = cpuset_ref(td->td_cpuset);
2106 ts2->ts_cpu = ts->ts_cpu;
2109 * Grab our parents cpu estimation information.
2111 ts2->ts_ticks = ts->ts_ticks;
2112 ts2->ts_ltick = ts->ts_ltick;
2113 ts2->ts_ftick = ts->ts_ftick;
2115 * Do not inherit any borrowed priority from the parent.
2117 child->td_priority = child->td_base_pri;
2119 * And update interactivity score.
2121 ts2->ts_slptime = ts->ts_slptime;
2122 ts2->ts_runtime = ts->ts_runtime;
2123 /* Attempt to quickly learn interactivity. */
2124 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2126 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2131 * Adjust the priority class of a thread.
2134 sched_class(struct thread *td, int class)
2137 THREAD_LOCK_ASSERT(td, MA_OWNED);
2138 if (td->td_pri_class == class)
2140 td->td_pri_class = class;
2144 * Return some of the child's priority and interactivity to the parent.
2147 sched_exit(struct proc *p, struct thread *child)
2151 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2152 "prio:%d", child->td_priority);
2153 PROC_LOCK_ASSERT(p, MA_OWNED);
2154 td = FIRST_THREAD_IN_PROC(p);
2155 sched_exit_thread(td, child);
2159 * Penalize another thread for the time spent on this one. This helps to
2160 * worsen the priority and interactivity of processes which schedule batch
2161 * jobs such as make. This has little effect on the make process itself but
2162 * causes new processes spawned by it to receive worse scores immediately.
2165 sched_exit_thread(struct thread *td, struct thread *child)
2168 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2169 "prio:%d", child->td_priority);
2171 * Give the child's runtime to the parent without returning the
2172 * sleep time as a penalty to the parent. This causes shells that
2173 * launch expensive things to mark their children as expensive.
2176 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2177 sched_interact_update(td);
2183 sched_preempt(struct thread *td)
2187 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2191 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2192 tdq->tdq_ipipending = 0;
2193 if (td->td_priority > tdq->tdq_lowpri) {
2196 flags = SW_INVOL | SW_PREEMPT;
2197 if (td->td_critnest > 1)
2198 td->td_owepreempt = 1;
2199 else if (TD_IS_IDLETHREAD(td))
2200 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2202 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2208 * Fix priorities on return to user-space. Priorities may be elevated due
2209 * to static priorities in msleep() or similar.
2212 sched_userret(struct thread *td)
2215 * XXX we cheat slightly on the locking here to avoid locking in
2216 * the usual case. Setting td_priority here is essentially an
2217 * incomplete workaround for not setting it properly elsewhere.
2218 * Now that some interrupt handlers are threads, not setting it
2219 * properly elsewhere can clobber it in the window between setting
2220 * it here and returning to user mode, so don't waste time setting
2221 * it perfectly here.
2223 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2224 ("thread with borrowed priority returning to userland"));
2225 if (td->td_priority != td->td_user_pri) {
2227 td->td_priority = td->td_user_pri;
2228 td->td_base_pri = td->td_user_pri;
2229 tdq_setlowpri(TDQ_SELF(), td);
2235 * Handle a stathz tick. This is really only relevant for timeshare
2239 sched_clock(struct thread *td)
2242 struct td_sched *ts;
2244 THREAD_LOCK_ASSERT(td, MA_OWNED);
2248 * We run the long term load balancer infrequently on the first cpu.
2250 if (balance_tdq == tdq) {
2251 if (balance_ticks && --balance_ticks == 0)
2256 * Save the old switch count so we have a record of the last ticks
2257 * activity. Initialize the new switch count based on our load.
2258 * If there is some activity seed it to reflect that.
2260 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2261 tdq->tdq_switchcnt = tdq->tdq_load;
2263 * Advance the insert index once for each tick to ensure that all
2264 * threads get a chance to run.
2266 if (tdq->tdq_idx == tdq->tdq_ridx) {
2267 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2268 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2269 tdq->tdq_ridx = tdq->tdq_idx;
2271 ts = td_get_sched(td);
2272 sched_pctcpu_update(ts, 1);
2273 if (td->td_pri_class & PRI_FIFO_BIT)
2275 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2277 * We used a tick; charge it to the thread so
2278 * that we can compute our interactivity.
2280 td_get_sched(td)->ts_runtime += tickincr;
2281 sched_interact_update(td);
2286 * Force a context switch if the current thread has used up a full
2287 * time slice (default is 100ms).
2289 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2291 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2296 sched_estcpu(struct thread *td __unused)
2303 * Return whether the current CPU has runnable tasks. Used for in-kernel
2304 * cooperative idle threads.
2307 sched_runnable(void)
2315 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2316 if (tdq->tdq_load > 0)
2319 if (tdq->tdq_load - 1 > 0)
2327 * Choose the highest priority thread to run. The thread is removed from
2328 * the run-queue while running however the load remains. For SMP we set
2329 * the tdq in the global idle bitmask if it idles here.
2338 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2339 td = tdq_choose(tdq);
2341 tdq_runq_rem(tdq, td);
2342 tdq->tdq_lowpri = td->td_priority;
2345 tdq->tdq_lowpri = PRI_MAX_IDLE;
2346 return (PCPU_GET(idlethread));
2350 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2351 * we always request it once we exit a critical section.
2354 sched_setpreempt(struct thread *td)
2360 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2363 pri = td->td_priority;
2364 cpri = ctd->td_priority;
2366 ctd->td_flags |= TDF_NEEDRESCHED;
2367 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2369 if (!sched_shouldpreempt(pri, cpri, 0))
2371 ctd->td_owepreempt = 1;
2375 * Add a thread to a thread queue. Select the appropriate runq and add the
2376 * thread to it. This is the internal function called when the tdq is
2380 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2383 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2384 KASSERT((td->td_inhibitors == 0),
2385 ("sched_add: trying to run inhibited thread"));
2386 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2387 ("sched_add: bad thread state"));
2388 KASSERT(td->td_flags & TDF_INMEM,
2389 ("sched_add: thread swapped out"));
2391 if (td->td_priority < tdq->tdq_lowpri)
2392 tdq->tdq_lowpri = td->td_priority;
2393 tdq_runq_add(tdq, td, flags);
2394 tdq_load_add(tdq, td);
2398 * Select the target thread queue and add a thread to it. Request
2399 * preemption or IPI a remote processor if required.
2402 sched_add(struct thread *td, int flags)
2409 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2410 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2411 sched_tdname(curthread));
2412 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2413 KTR_ATTR_LINKED, sched_tdname(td));
2414 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2415 flags & SRQ_PREEMPTED);
2416 THREAD_LOCK_ASSERT(td, MA_OWNED);
2418 * Recalculate the priority before we select the target cpu or
2421 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2425 * Pick the destination cpu and if it isn't ours transfer to the
2428 cpu = sched_pickcpu(td, flags);
2429 tdq = sched_setcpu(td, cpu, flags);
2430 tdq_add(tdq, td, flags);
2431 if (cpu != PCPU_GET(cpuid)) {
2432 tdq_notify(tdq, td);
2439 * Now that the thread is moving to the run-queue, set the lock
2440 * to the scheduler's lock.
2442 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2443 tdq_add(tdq, td, flags);
2445 if (!(flags & SRQ_YIELDING))
2446 sched_setpreempt(td);
2450 * Remove a thread from a run-queue without running it. This is used
2451 * when we're stealing a thread from a remote queue. Otherwise all threads
2452 * exit by calling sched_exit_thread() and sched_throw() themselves.
2455 sched_rem(struct thread *td)
2459 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2460 "prio:%d", td->td_priority);
2461 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2462 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2463 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2464 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2465 KASSERT(TD_ON_RUNQ(td),
2466 ("sched_rem: thread not on run queue"));
2467 tdq_runq_rem(tdq, td);
2468 tdq_load_rem(tdq, td);
2470 if (td->td_priority == tdq->tdq_lowpri)
2471 tdq_setlowpri(tdq, NULL);
2475 * Fetch cpu utilization information. Updates on demand.
2478 sched_pctcpu(struct thread *td)
2481 struct td_sched *ts;
2484 ts = td_get_sched(td);
2486 THREAD_LOCK_ASSERT(td, MA_OWNED);
2487 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2491 /* How many rtick per second ? */
2492 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2493 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2500 * Enforce affinity settings for a thread. Called after adjustments to
2504 sched_affinity(struct thread *td)
2507 struct td_sched *ts;
2509 THREAD_LOCK_ASSERT(td, MA_OWNED);
2510 ts = td_get_sched(td);
2511 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2513 if (TD_ON_RUNQ(td)) {
2515 sched_add(td, SRQ_BORING);
2518 if (!TD_IS_RUNNING(td))
2521 * Force a switch before returning to userspace. If the
2522 * target thread is not running locally send an ipi to force
2525 td->td_flags |= TDF_NEEDRESCHED;
2526 if (td != curthread)
2527 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2532 * Bind a thread to a target cpu.
2535 sched_bind(struct thread *td, int cpu)
2537 struct td_sched *ts;
2539 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2540 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2541 ts = td_get_sched(td);
2542 if (ts->ts_flags & TSF_BOUND)
2544 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2545 ts->ts_flags |= TSF_BOUND;
2547 if (PCPU_GET(cpuid) == cpu)
2550 /* When we return from mi_switch we'll be on the correct cpu. */
2551 mi_switch(SW_VOL, NULL);
2555 * Release a bound thread.
2558 sched_unbind(struct thread *td)
2560 struct td_sched *ts;
2562 THREAD_LOCK_ASSERT(td, MA_OWNED);
2563 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2564 ts = td_get_sched(td);
2565 if ((ts->ts_flags & TSF_BOUND) == 0)
2567 ts->ts_flags &= ~TSF_BOUND;
2572 sched_is_bound(struct thread *td)
2574 THREAD_LOCK_ASSERT(td, MA_OWNED);
2575 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2582 sched_relinquish(struct thread *td)
2585 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2590 * Return the total system load.
2601 total += TDQ_CPU(i)->tdq_sysload;
2604 return (TDQ_SELF()->tdq_sysload);
2609 sched_sizeof_proc(void)
2611 return (sizeof(struct proc));
2615 sched_sizeof_thread(void)
2617 return (sizeof(struct thread) + sizeof(struct td_sched));
2621 #define TDQ_IDLESPIN(tdq) \
2622 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2624 #define TDQ_IDLESPIN(tdq) 1
2628 * The actual idle process.
2631 sched_idletd(void *dummy)
2635 int oldswitchcnt, switchcnt;
2638 mtx_assert(&Giant, MA_NOTOWNED);
2641 THREAD_NO_SLEEPING();
2644 if (tdq->tdq_load) {
2646 mi_switch(SW_VOL | SWT_IDLE, NULL);
2649 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2651 if (switchcnt != oldswitchcnt) {
2652 oldswitchcnt = switchcnt;
2653 if (tdq_idled(tdq) == 0)
2656 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2658 oldswitchcnt = switchcnt;
2661 * If we're switching very frequently, spin while checking
2662 * for load rather than entering a low power state that
2663 * may require an IPI. However, don't do any busy
2664 * loops while on SMT machines as this simply steals
2665 * cycles from cores doing useful work.
2667 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2668 for (i = 0; i < sched_idlespins; i++) {
2675 /* If there was context switch during spin, restart it. */
2676 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2677 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2680 /* Run main MD idle handler. */
2681 tdq->tdq_cpu_idle = 1;
2683 * Make sure that tdq_cpu_idle update is globally visible
2684 * before cpu_idle() read tdq_load. The order is important
2685 * to avoid race with tdq_notify.
2687 atomic_thread_fence_seq_cst();
2688 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2689 tdq->tdq_cpu_idle = 0;
2692 * Account thread-less hardware interrupts and
2693 * other wakeup reasons equal to context switches.
2695 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2696 if (switchcnt != oldswitchcnt)
2698 tdq->tdq_switchcnt++;
2704 * A CPU is entering for the first time or a thread is exiting.
2707 sched_throw(struct thread *td)
2709 struct thread *newtd;
2714 /* Correct spinlock nesting and acquire the correct lock. */
2717 PCPU_SET(switchtime, cpu_ticks());
2718 PCPU_SET(switchticks, ticks);
2720 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2721 tdq_load_rem(tdq, td);
2722 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2723 td->td_lastcpu = td->td_oncpu;
2724 td->td_oncpu = NOCPU;
2726 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2727 newtd = choosethread();
2728 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2729 cpu_throw(td, newtd); /* doesn't return */
2733 * This is called from fork_exit(). Just acquire the correct locks and
2734 * let fork do the rest of the work.
2737 sched_fork_exit(struct thread *td)
2743 * Finish setting up thread glue so that it begins execution in a
2744 * non-nested critical section with the scheduler lock held.
2746 cpuid = PCPU_GET(cpuid);
2747 tdq = TDQ_CPU(cpuid);
2748 if (TD_IS_IDLETHREAD(td))
2749 td->td_lock = TDQ_LOCKPTR(tdq);
2750 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2751 td->td_oncpu = cpuid;
2752 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2753 lock_profile_obtain_lock_success(
2754 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2758 * Create on first use to catch odd startup conditons.
2761 sched_tdname(struct thread *td)
2764 struct td_sched *ts;
2766 ts = td_get_sched(td);
2767 if (ts->ts_name[0] == '\0')
2768 snprintf(ts->ts_name, sizeof(ts->ts_name),
2769 "%s tid %d", td->td_name, td->td_tid);
2770 return (ts->ts_name);
2772 return (td->td_name);
2778 sched_clear_tdname(struct thread *td)
2780 struct td_sched *ts;
2782 ts = td_get_sched(td);
2783 ts->ts_name[0] = '\0';
2790 * Build the CPU topology dump string. Is recursively called to collect
2791 * the topology tree.
2794 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2797 char cpusetbuf[CPUSETBUFSIZ];
2800 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2801 "", 1 + indent / 2, cg->cg_level);
2802 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2803 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2805 for (i = 0; i < MAXCPU; i++) {
2806 if (CPU_ISSET(i, &cg->cg_mask)) {
2808 sbuf_printf(sb, ", ");
2811 sbuf_printf(sb, "%d", i);
2814 sbuf_printf(sb, "</cpu>\n");
2816 if (cg->cg_flags != 0) {
2817 sbuf_printf(sb, "%*s <flags>", indent, "");
2818 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2819 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2820 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2821 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2822 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2823 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2824 sbuf_printf(sb, "</flags>\n");
2827 if (cg->cg_children > 0) {
2828 sbuf_printf(sb, "%*s <children>\n", indent, "");
2829 for (i = 0; i < cg->cg_children; i++)
2830 sysctl_kern_sched_topology_spec_internal(sb,
2831 &cg->cg_child[i], indent+2);
2832 sbuf_printf(sb, "%*s </children>\n", indent, "");
2834 sbuf_printf(sb, "%*s</group>\n", indent, "");
2839 * Sysctl handler for retrieving topology dump. It's a wrapper for
2840 * the recursive sysctl_kern_smp_topology_spec_internal().
2843 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2848 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2850 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
2854 sbuf_printf(topo, "<groups>\n");
2855 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2856 sbuf_printf(topo, "</groups>\n");
2859 err = sbuf_finish(topo);
2868 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2870 int error, new_val, period;
2872 period = 1000000 / realstathz;
2873 new_val = period * sched_slice;
2874 error = sysctl_handle_int(oidp, &new_val, 0, req);
2875 if (error != 0 || req->newptr == NULL)
2879 sched_slice = imax(1, (new_val + period / 2) / period);
2880 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2881 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2886 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2887 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2889 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2890 NULL, 0, sysctl_kern_quantum, "I",
2891 "Quantum for timeshare threads in microseconds");
2892 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2893 "Quantum for timeshare threads in stathz ticks");
2894 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2895 "Interactivity score threshold");
2896 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2898 "Maximal (lowest) priority for preemption");
2899 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2900 "Assign static kernel priorities to sleeping threads");
2901 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2902 "Number of times idle thread will spin waiting for new work");
2903 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2904 &sched_idlespinthresh, 0,
2905 "Threshold before we will permit idle thread spinning");
2907 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2908 "Number of hz ticks to keep thread affinity for");
2909 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2910 "Enables the long-term load balancer");
2911 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2912 &balance_interval, 0,
2913 "Average period in stathz ticks to run the long-term balancer");
2914 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2915 "Attempts to steal work from other cores before idling");
2916 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2917 "Minimum load on remote CPU before we'll steal");
2918 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2919 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2920 "XML dump of detected CPU topology");
2923 /* ps compat. All cpu percentages from ULE are weighted. */
2924 static int ccpu = 0;
2925 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");