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_kdtrace.h"
43 #include "opt_sched.h"
45 #include <sys/param.h>
46 #include <sys/systm.h>
48 #include <sys/kernel.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>
80 #if defined(__powerpc__) && defined(BOOKE_E500)
81 #error "This architecture is not currently compatible with ULE"
86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
87 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
88 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
91 * Thread scheduler specific section. All fields are protected
95 struct runq *ts_runq; /* Run-queue we're queued on. */
96 short ts_flags; /* TSF_* flags. */
97 u_char ts_cpu; /* CPU that we have affinity for. */
98 int ts_rltick; /* Real last tick, for affinity. */
99 int ts_slice; /* Ticks of slice remaining. */
100 u_int ts_slptime; /* Number of ticks we vol. slept */
101 u_int ts_runtime; /* Number of ticks we were running */
102 int ts_ltick; /* Last tick that we were running on */
103 int ts_ftick; /* First tick that we were running on */
104 int ts_ticks; /* Tick count */
106 char ts_name[TS_NAME_LEN];
109 /* flags kept in ts_flags */
110 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
111 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
113 static struct td_sched td_sched0;
115 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
116 #define THREAD_CAN_SCHED(td, cpu) \
117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
120 * Priority ranges used for interactive and non-interactive timeshare
121 * threads. The timeshare priorities are split up into four ranges.
122 * The first range handles interactive threads. The last three ranges
123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124 * ranges supporting nice values.
126 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
130 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
131 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
136 * Cpu percentage computation macros and defines.
138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
145 #define SCHED_TICK_SECS 10
146 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
147 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
148 #define SCHED_TICK_SHIFT 10
149 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
153 * These macros determine priorities for non-interactive threads. They are
154 * assigned a priority based on their recent cpu utilization as expressed
155 * by the ratio of ticks to the tick total. NHALF priorities at the start
156 * and end of the MIN to MAX timeshare range are only reachable with negative
157 * or positive nice respectively.
159 * PRI_RANGE: Priority range for utilization dependent priorities.
160 * PRI_NRESV: Number of nice values.
161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
162 * PRI_NICE: Determines the part of the priority inherited from nice.
164 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
165 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
166 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define SCHED_PRI_TICKS(ts) \
170 (SCHED_TICK_HZ((ts)) / \
171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define SCHED_PRI_NICE(nice) (nice)
175 * These determine the interactivity of a process. Interactivity differs from
176 * cpu utilization in that it expresses the voluntary time slept vs time ran
177 * while cpu utilization includes all time not running. This more accurately
178 * models the intent of the thread.
180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
181 * before throttling back.
182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
183 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
184 * INTERACT_THRESH: Threshold for placement on the current runq.
186 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
187 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
188 #define SCHED_INTERACT_MAX (100)
189 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
190 #define SCHED_INTERACT_THRESH (30)
193 * These parameters determine the slice behavior for batch work.
195 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
196 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
198 /* Flags kept in td_flags. */
199 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
202 * tickincr: Converts a stathz tick into a hz domain scaled by
203 * the shift factor. Without the shift the error rate
204 * due to rounding would be unacceptably high.
205 * realstathz: stathz is sometimes 0 and run off of hz.
206 * sched_slice: Runtime of each thread before rescheduling.
207 * preempt_thresh: Priority threshold for preemption and remote IPIs.
209 static int sched_interact = SCHED_INTERACT_THRESH;
210 static int tickincr = 8 << SCHED_TICK_SHIFT;
211 static int realstathz = 127; /* reset during boot. */
212 static int sched_slice = 10; /* reset during boot. */
213 static int sched_slice_min = 1; /* reset during boot. */
215 #ifdef FULL_PREEMPTION
216 static int preempt_thresh = PRI_MAX_IDLE;
218 static int preempt_thresh = PRI_MIN_KERN;
221 static int preempt_thresh = 0;
223 static int static_boost = PRI_MIN_BATCH;
224 static int sched_idlespins = 10000;
225 static int sched_idlespinthresh = -1;
228 * tdq - per processor runqs and statistics. All fields are protected by the
229 * tdq_lock. The load and lowpri may be accessed without to avoid excess
230 * locking in sched_pickcpu();
234 * Ordered to improve efficiency of cpu_search() and switch().
235 * tdq_lock is padded to avoid false sharing with tdq_load and
238 struct mtx_padalign tdq_lock; /* run queue lock. */
239 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
240 volatile int tdq_load; /* Aggregate load. */
241 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
242 int tdq_sysload; /* For loadavg, !ITHD load. */
243 int tdq_transferable; /* Transferable thread count. */
244 short tdq_switchcnt; /* Switches this tick. */
245 short tdq_oldswitchcnt; /* Switches last tick. */
246 u_char tdq_lowpri; /* Lowest priority thread. */
247 u_char tdq_ipipending; /* IPI pending. */
248 u_char tdq_idx; /* Current insert index. */
249 u_char tdq_ridx; /* Current removal index. */
250 struct runq tdq_realtime; /* real-time run queue. */
251 struct runq tdq_timeshare; /* timeshare run queue. */
252 struct runq tdq_idle; /* Queue of IDLE threads. */
253 char tdq_name[TDQ_NAME_LEN];
255 char tdq_loadname[TDQ_LOADNAME_LEN];
259 /* Idle thread states and config. */
260 #define TDQ_RUNNING 1
264 struct cpu_group *cpu_top; /* CPU topology */
266 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
267 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
272 static int rebalance = 1;
273 static int balance_interval = 128; /* Default set in sched_initticks(). */
275 static int steal_idle = 1;
276 static int steal_thresh = 2;
279 * One thread queue per processor.
281 static struct tdq tdq_cpu[MAXCPU];
282 static struct tdq *balance_tdq;
283 static int balance_ticks;
284 static DPCPU_DEFINE(uint32_t, randomval);
286 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
287 #define TDQ_CPU(x) (&tdq_cpu[(x)])
288 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
290 static struct tdq tdq_cpu;
292 #define TDQ_ID(x) (0)
293 #define TDQ_SELF() (&tdq_cpu)
294 #define TDQ_CPU(x) (&tdq_cpu)
297 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
298 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
299 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
300 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
301 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
303 static void sched_priority(struct thread *);
304 static void sched_thread_priority(struct thread *, u_char);
305 static int sched_interact_score(struct thread *);
306 static void sched_interact_update(struct thread *);
307 static void sched_interact_fork(struct thread *);
308 static void sched_pctcpu_update(struct td_sched *, int);
310 /* Operations on per processor queues */
311 static struct thread *tdq_choose(struct tdq *);
312 static void tdq_setup(struct tdq *);
313 static void tdq_load_add(struct tdq *, struct thread *);
314 static void tdq_load_rem(struct tdq *, struct thread *);
315 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
316 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
317 static inline int sched_shouldpreempt(int, int, int);
318 void tdq_print(int cpu);
319 static void runq_print(struct runq *rq);
320 static void tdq_add(struct tdq *, struct thread *, int);
322 static int tdq_move(struct tdq *, struct tdq *);
323 static int tdq_idled(struct tdq *);
324 static void tdq_notify(struct tdq *, struct thread *);
325 static struct thread *tdq_steal(struct tdq *, int);
326 static struct thread *runq_steal(struct runq *, int);
327 static int sched_pickcpu(struct thread *, int);
328 static void sched_balance(void);
329 static int sched_balance_pair(struct tdq *, struct tdq *);
330 static inline struct tdq *sched_setcpu(struct thread *, int, int);
331 static inline void thread_unblock_switch(struct thread *, struct mtx *);
332 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
333 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
334 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
335 struct cpu_group *cg, int indent);
338 static void sched_setup(void *dummy);
339 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
341 static void sched_initticks(void *dummy);
342 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
345 SDT_PROVIDER_DEFINE(sched);
347 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
348 "struct proc *", "uint8_t");
349 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
350 "struct proc *", "void *");
351 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
352 "struct proc *", "void *", "int");
353 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
354 "struct proc *", "uint8_t", "struct thread *");
355 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
356 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
358 SDT_PROBE_DEFINE(sched, , , on__cpu);
359 SDT_PROBE_DEFINE(sched, , , remain__cpu);
360 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
364 * Print the threads waiting on a run-queue.
367 runq_print(struct runq *rq)
375 for (i = 0; i < RQB_LEN; i++) {
376 printf("\t\trunq bits %d 0x%zx\n",
377 i, rq->rq_status.rqb_bits[i]);
378 for (j = 0; j < RQB_BPW; j++)
379 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
380 pri = j + (i << RQB_L2BPW);
381 rqh = &rq->rq_queues[pri];
382 TAILQ_FOREACH(td, rqh, td_runq) {
383 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
384 td, td->td_name, td->td_priority,
385 td->td_rqindex, pri);
392 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
401 printf("tdq %d:\n", TDQ_ID(tdq));
402 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
403 printf("\tLock name: %s\n", tdq->tdq_name);
404 printf("\tload: %d\n", tdq->tdq_load);
405 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
406 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
407 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
408 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
409 printf("\tload transferable: %d\n", tdq->tdq_transferable);
410 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
411 printf("\trealtime runq:\n");
412 runq_print(&tdq->tdq_realtime);
413 printf("\ttimeshare runq:\n");
414 runq_print(&tdq->tdq_timeshare);
415 printf("\tidle runq:\n");
416 runq_print(&tdq->tdq_idle);
420 sched_shouldpreempt(int pri, int cpri, int remote)
423 * If the new priority is not better than the current priority there is
429 * Always preempt idle.
431 if (cpri >= PRI_MIN_IDLE)
434 * If preemption is disabled don't preempt others.
436 if (preempt_thresh == 0)
439 * Preempt if we exceed the threshold.
441 if (pri <= preempt_thresh)
444 * If we're interactive or better and there is non-interactive
445 * or worse running preempt only remote processors.
447 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
453 * Add a thread to the actual run-queue. Keeps transferable counts up to
454 * date with what is actually on the run-queue. Selects the correct
455 * queue position for timeshare threads.
458 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
463 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
464 THREAD_LOCK_ASSERT(td, MA_OWNED);
466 pri = td->td_priority;
469 if (THREAD_CAN_MIGRATE(td)) {
470 tdq->tdq_transferable++;
471 ts->ts_flags |= TSF_XFERABLE;
473 if (pri < PRI_MIN_BATCH) {
474 ts->ts_runq = &tdq->tdq_realtime;
475 } else if (pri <= PRI_MAX_BATCH) {
476 ts->ts_runq = &tdq->tdq_timeshare;
477 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
478 ("Invalid priority %d on timeshare runq", pri));
480 * This queue contains only priorities between MIN and MAX
481 * realtime. Use the whole queue to represent these values.
483 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
484 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
485 pri = (pri + tdq->tdq_idx) % RQ_NQS;
487 * This effectively shortens the queue by one so we
488 * can have a one slot difference between idx and
489 * ridx while we wait for threads to drain.
491 if (tdq->tdq_ridx != tdq->tdq_idx &&
492 pri == tdq->tdq_ridx)
493 pri = (unsigned char)(pri - 1) % RQ_NQS;
496 runq_add_pri(ts->ts_runq, td, pri, flags);
499 ts->ts_runq = &tdq->tdq_idle;
500 runq_add(ts->ts_runq, td, flags);
504 * Remove a thread from a run-queue. This typically happens when a thread
505 * is selected to run. Running threads are not on the queue and the
506 * transferable count does not reflect them.
509 tdq_runq_rem(struct tdq *tdq, struct thread *td)
514 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
515 KASSERT(ts->ts_runq != NULL,
516 ("tdq_runq_remove: thread %p null ts_runq", td));
517 if (ts->ts_flags & TSF_XFERABLE) {
518 tdq->tdq_transferable--;
519 ts->ts_flags &= ~TSF_XFERABLE;
521 if (ts->ts_runq == &tdq->tdq_timeshare) {
522 if (tdq->tdq_idx != tdq->tdq_ridx)
523 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
525 runq_remove_idx(ts->ts_runq, td, NULL);
527 runq_remove(ts->ts_runq, td);
531 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
532 * for this thread to the referenced thread queue.
535 tdq_load_add(struct tdq *tdq, struct thread *td)
538 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
539 THREAD_LOCK_ASSERT(td, MA_OWNED);
542 if ((td->td_flags & TDF_NOLOAD) == 0)
544 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
545 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
549 * Remove the load from a thread that is transitioning to a sleep state or
553 tdq_load_rem(struct tdq *tdq, struct thread *td)
556 THREAD_LOCK_ASSERT(td, MA_OWNED);
557 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
558 KASSERT(tdq->tdq_load != 0,
559 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
562 if ((td->td_flags & TDF_NOLOAD) == 0)
564 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
565 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
569 * Bound timeshare latency by decreasing slice size as load increases. We
570 * consider the maximum latency as the sum of the threads waiting to run
571 * aside from curthread and target no more than sched_slice latency but
572 * no less than sched_slice_min runtime.
575 tdq_slice(struct tdq *tdq)
580 * It is safe to use sys_load here because this is called from
581 * contexts where timeshare threads are running and so there
582 * cannot be higher priority load in the system.
584 load = tdq->tdq_sysload - 1;
585 if (load >= SCHED_SLICE_MIN_DIVISOR)
586 return (sched_slice_min);
588 return (sched_slice);
589 return (sched_slice / load);
593 * Set lowpri to its exact value by searching the run-queue and
594 * evaluating curthread. curthread may be passed as an optimization.
597 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
601 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
603 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
604 td = tdq_choose(tdq);
605 if (td == NULL || td->td_priority > ctd->td_priority)
606 tdq->tdq_lowpri = ctd->td_priority;
608 tdq->tdq_lowpri = td->td_priority;
615 int cs_pri; /* Min priority for low. */
616 int cs_limit; /* Max load for low, min load for high. */
621 #define CPU_SEARCH_LOWEST 0x1
622 #define CPU_SEARCH_HIGHEST 0x2
623 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
625 #define CPUSET_FOREACH(cpu, mask) \
626 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
627 if (CPU_ISSET(cpu, &mask))
629 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
630 struct cpu_search *high, const int match);
631 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
632 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
633 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
634 struct cpu_search *high);
637 * Search the tree of cpu_groups for the lowest or highest loaded cpu
638 * according to the match argument. This routine actually compares the
639 * load on all paths through the tree and finds the least loaded cpu on
640 * the least loaded path, which may differ from the least loaded cpu in
641 * the system. This balances work among caches and busses.
643 * This inline is instantiated in three forms below using constants for the
644 * match argument. It is reduced to the minimum set for each case. It is
645 * also recursive to the depth of the tree.
648 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
649 struct cpu_search *high, const int match)
651 struct cpu_search lgroup;
652 struct cpu_search hgroup;
654 struct cpu_group *child;
656 int cpu, i, hload, lload, load, total, rnd, *rndptr;
659 cpumask = cg->cg_mask;
660 if (match & CPU_SEARCH_LOWEST) {
664 if (match & CPU_SEARCH_HIGHEST) {
669 /* Iterate through the child CPU groups and then remaining CPUs. */
670 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
672 #ifdef HAVE_INLINE_FFSL
673 cpu = CPU_FFS(&cpumask) - 1;
675 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
682 child = &cg->cg_child[i - 1];
684 if (match & CPU_SEARCH_LOWEST)
686 if (match & CPU_SEARCH_HIGHEST)
688 if (child) { /* Handle child CPU group. */
689 CPU_NAND(&cpumask, &child->cg_mask);
691 case CPU_SEARCH_LOWEST:
692 load = cpu_search_lowest(child, &lgroup);
694 case CPU_SEARCH_HIGHEST:
695 load = cpu_search_highest(child, &hgroup);
697 case CPU_SEARCH_BOTH:
698 load = cpu_search_both(child, &lgroup, &hgroup);
701 } else { /* Handle child CPU. */
702 CPU_CLR(cpu, &cpumask);
704 load = tdq->tdq_load * 256;
705 rndptr = DPCPU_PTR(randomval);
706 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
707 if (match & CPU_SEARCH_LOWEST) {
708 if (cpu == low->cs_prefer)
710 /* If that CPU is allowed and get data. */
711 if (tdq->tdq_lowpri > lgroup.cs_pri &&
712 tdq->tdq_load <= lgroup.cs_limit &&
713 CPU_ISSET(cpu, &lgroup.cs_mask)) {
715 lgroup.cs_load = load - rnd;
718 if (match & CPU_SEARCH_HIGHEST)
719 if (tdq->tdq_load >= hgroup.cs_limit &&
720 tdq->tdq_transferable &&
721 CPU_ISSET(cpu, &hgroup.cs_mask)) {
723 hgroup.cs_load = load - rnd;
728 /* We have info about child item. Compare it. */
729 if (match & CPU_SEARCH_LOWEST) {
730 if (lgroup.cs_cpu >= 0 &&
732 (load == lload && lgroup.cs_load < low->cs_load))) {
734 low->cs_cpu = lgroup.cs_cpu;
735 low->cs_load = lgroup.cs_load;
738 if (match & CPU_SEARCH_HIGHEST)
739 if (hgroup.cs_cpu >= 0 &&
741 (load == hload && hgroup.cs_load > high->cs_load))) {
743 high->cs_cpu = hgroup.cs_cpu;
744 high->cs_load = hgroup.cs_load;
748 if (i == 0 && CPU_EMPTY(&cpumask))
751 #ifndef HAVE_INLINE_FFSL
760 * cpu_search instantiations must pass constants to maintain the inline
764 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
766 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
770 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
772 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
776 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
777 struct cpu_search *high)
779 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
783 * Find the cpu with the least load via the least loaded path that has a
784 * lowpri greater than pri pri. A pri of -1 indicates any priority is
788 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
791 struct cpu_search low;
794 low.cs_prefer = prefer;
797 low.cs_limit = maxload;
798 cpu_search_lowest(cg, &low);
803 * Find the cpu with the highest load via the highest loaded path.
806 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
808 struct cpu_search high;
812 high.cs_limit = minload;
813 cpu_search_highest(cg, &high);
818 sched_balance_group(struct cpu_group *cg)
820 cpuset_t hmask, lmask;
821 int high, low, anylow;
825 high = sched_highest(cg, hmask, 1);
826 /* Stop if there is no more CPU with transferrable threads. */
829 CPU_CLR(high, &hmask);
830 CPU_COPY(&hmask, &lmask);
831 /* Stop if there is no more CPU left for low. */
832 if (CPU_EMPTY(&lmask))
836 low = sched_lowest(cg, lmask, -1,
837 TDQ_CPU(high)->tdq_load - 1, high);
838 /* Stop if we looked well and found no less loaded CPU. */
839 if (anylow && low == -1)
841 /* Go to next high if we found no less loaded CPU. */
844 /* Transfer thread from high to low. */
845 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
846 /* CPU that got thread can no longer be a donor. */
847 CPU_CLR(low, &hmask);
850 * If failed, then there is no threads on high
851 * that can run on this low. Drop low from low
852 * mask and look for different one.
854 CPU_CLR(low, &lmask);
867 * Select a random time between .5 * balance_interval and
868 * 1.5 * balance_interval.
870 balance_ticks = max(balance_interval / 2, 1);
871 balance_ticks += random() % balance_interval;
872 if (smp_started == 0 || rebalance == 0)
876 sched_balance_group(cpu_top);
881 * Lock two thread queues using their address to maintain lock order.
884 tdq_lock_pair(struct tdq *one, struct tdq *two)
888 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
891 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
896 * Unlock two thread queues. Order is not important here.
899 tdq_unlock_pair(struct tdq *one, struct tdq *two)
906 * Transfer load between two imbalanced thread queues.
909 sched_balance_pair(struct tdq *high, struct tdq *low)
914 tdq_lock_pair(high, low);
917 * Determine what the imbalance is and then adjust that to how many
918 * threads we actually have to give up (transferable).
920 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
921 (moved = tdq_move(high, low)) > 0) {
923 * In case the target isn't the current cpu IPI it to force a
924 * reschedule with the new workload.
927 if (cpu != PCPU_GET(cpuid))
928 ipi_cpu(cpu, IPI_PREEMPT);
930 tdq_unlock_pair(high, low);
935 * Move a thread from one thread queue to another.
938 tdq_move(struct tdq *from, struct tdq *to)
945 TDQ_LOCK_ASSERT(from, MA_OWNED);
946 TDQ_LOCK_ASSERT(to, MA_OWNED);
950 td = tdq_steal(tdq, cpu);
955 * Although the run queue is locked the thread may be blocked. Lock
956 * it to clear this and acquire the run-queue lock.
959 /* Drop recursive lock on from acquired via thread_lock(). */
963 td->td_lock = TDQ_LOCKPTR(to);
964 tdq_add(to, td, SRQ_YIELDING);
969 * This tdq has idled. Try to steal a thread from another cpu and switch
973 tdq_idled(struct tdq *tdq)
975 struct cpu_group *cg;
981 if (smp_started == 0 || steal_idle == 0)
984 CPU_CLR(PCPU_GET(cpuid), &mask);
985 /* We don't want to be preempted while we're iterating. */
987 for (cg = tdq->tdq_cg; cg != NULL; ) {
988 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
989 thresh = steal_thresh;
992 cpu = sched_highest(cg, mask, thresh);
997 steal = TDQ_CPU(cpu);
999 tdq_lock_pair(tdq, steal);
1000 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1001 tdq_unlock_pair(tdq, steal);
1005 * If a thread was added while interrupts were disabled don't
1006 * steal one here. If we fail to acquire one due to affinity
1007 * restrictions loop again with this cpu removed from the
1010 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1011 tdq_unlock_pair(tdq, steal);
1016 mi_switch(SW_VOL | SWT_IDLE, NULL);
1017 thread_unlock(curthread);
1026 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1029 tdq_notify(struct tdq *tdq, struct thread *td)
1035 if (tdq->tdq_ipipending)
1037 cpu = td->td_sched->ts_cpu;
1038 pri = td->td_priority;
1039 ctd = pcpu_find(cpu)->pc_curthread;
1040 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1042 if (TD_IS_IDLETHREAD(ctd)) {
1044 * If the MD code has an idle wakeup routine try that before
1045 * falling back to IPI.
1047 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1050 tdq->tdq_ipipending = 1;
1051 ipi_cpu(cpu, IPI_PREEMPT);
1055 * Steals load from a timeshare queue. Honors the rotating queue head
1058 static struct thread *
1059 runq_steal_from(struct runq *rq, int cpu, u_char start)
1063 struct thread *td, *first;
1068 rqb = &rq->rq_status;
1069 bit = start & (RQB_BPW -1);
1073 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1074 if (rqb->rqb_bits[i] == 0)
1077 for (pri = bit; pri < RQB_BPW; pri++)
1078 if (rqb->rqb_bits[i] & (1ul << pri))
1083 pri = RQB_FFS(rqb->rqb_bits[i]);
1084 pri += (i << RQB_L2BPW);
1085 rqh = &rq->rq_queues[pri];
1086 TAILQ_FOREACH(td, rqh, td_runq) {
1087 if (first && THREAD_CAN_MIGRATE(td) &&
1088 THREAD_CAN_SCHED(td, cpu))
1098 if (first && THREAD_CAN_MIGRATE(first) &&
1099 THREAD_CAN_SCHED(first, cpu))
1105 * Steals load from a standard linear queue.
1107 static struct thread *
1108 runq_steal(struct runq *rq, int cpu)
1116 rqb = &rq->rq_status;
1117 for (word = 0; word < RQB_LEN; word++) {
1118 if (rqb->rqb_bits[word] == 0)
1120 for (bit = 0; bit < RQB_BPW; bit++) {
1121 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1123 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1124 TAILQ_FOREACH(td, rqh, td_runq)
1125 if (THREAD_CAN_MIGRATE(td) &&
1126 THREAD_CAN_SCHED(td, cpu))
1134 * Attempt to steal a thread in priority order from a thread queue.
1136 static struct thread *
1137 tdq_steal(struct tdq *tdq, int cpu)
1141 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1142 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1144 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1145 cpu, tdq->tdq_ridx)) != NULL)
1147 return (runq_steal(&tdq->tdq_idle, cpu));
1151 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1152 * current lock and returns with the assigned queue locked.
1154 static inline struct tdq *
1155 sched_setcpu(struct thread *td, int cpu, int flags)
1160 THREAD_LOCK_ASSERT(td, MA_OWNED);
1162 td->td_sched->ts_cpu = cpu;
1164 * If the lock matches just return the queue.
1166 if (td->td_lock == TDQ_LOCKPTR(tdq))
1170 * If the thread isn't running its lockptr is a
1171 * turnstile or a sleepqueue. We can just lock_set without
1174 if (TD_CAN_RUN(td)) {
1176 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1181 * The hard case, migration, we need to block the thread first to
1182 * prevent order reversals with other cpus locks.
1185 thread_lock_block(td);
1187 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1192 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1193 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1194 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1195 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1196 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1197 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1200 sched_pickcpu(struct thread *td, int flags)
1202 struct cpu_group *cg, *ccg;
1203 struct td_sched *ts;
1208 self = PCPU_GET(cpuid);
1210 if (smp_started == 0)
1213 * Don't migrate a running thread from sched_switch().
1215 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1216 return (ts->ts_cpu);
1218 * Prefer to run interrupt threads on the processors that generate
1221 pri = td->td_priority;
1222 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1223 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1224 SCHED_STAT_INC(pickcpu_intrbind);
1226 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1227 SCHED_STAT_INC(pickcpu_affinity);
1228 return (ts->ts_cpu);
1232 * If the thread can run on the last cpu and the affinity has not
1233 * expired or it is idle run it there.
1235 tdq = TDQ_CPU(ts->ts_cpu);
1237 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1238 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1239 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1240 if (cg->cg_flags & CG_FLAG_THREAD) {
1241 CPUSET_FOREACH(cpu, cg->cg_mask) {
1242 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1247 if (cpu > mp_maxid) {
1248 SCHED_STAT_INC(pickcpu_idle_affinity);
1249 return (ts->ts_cpu);
1253 * Search for the last level cache CPU group in the tree.
1254 * Skip caches with expired affinity time and SMT groups.
1255 * Affinity to higher level caches will be handled less aggressively.
1257 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1258 if (cg->cg_flags & CG_FLAG_THREAD)
1260 if (!SCHED_AFFINITY(ts, cg->cg_level))
1267 /* Search the group for the less loaded idle CPU we can run now. */
1268 mask = td->td_cpuset->cs_mask;
1269 if (cg != NULL && cg != cpu_top &&
1270 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1271 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1272 INT_MAX, ts->ts_cpu);
1273 /* Search globally for the less loaded CPU we can run now. */
1275 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1276 /* Search globally for the less loaded CPU. */
1278 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1279 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1281 * Compare the lowest loaded cpu to current cpu.
1283 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1284 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1285 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1286 SCHED_STAT_INC(pickcpu_local);
1289 SCHED_STAT_INC(pickcpu_lowest);
1290 if (cpu != ts->ts_cpu)
1291 SCHED_STAT_INC(pickcpu_migration);
1297 * Pick the highest priority task we have and return it.
1299 static struct thread *
1300 tdq_choose(struct tdq *tdq)
1304 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1305 td = runq_choose(&tdq->tdq_realtime);
1308 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1310 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1311 ("tdq_choose: Invalid priority on timeshare queue %d",
1315 td = runq_choose(&tdq->tdq_idle);
1317 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1318 ("tdq_choose: Invalid priority on idle queue %d",
1327 * Initialize a thread queue.
1330 tdq_setup(struct tdq *tdq)
1334 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1335 runq_init(&tdq->tdq_realtime);
1336 runq_init(&tdq->tdq_timeshare);
1337 runq_init(&tdq->tdq_idle);
1338 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1339 "sched lock %d", (int)TDQ_ID(tdq));
1340 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1341 MTX_SPIN | MTX_RECURSE);
1343 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1344 "CPU %d load", (int)TDQ_ID(tdq));
1350 sched_setup_smp(void)
1355 cpu_top = smp_topo();
1359 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1360 if (tdq->tdq_cg == NULL)
1361 panic("Can't find cpu group for %d\n", i);
1363 balance_tdq = TDQ_SELF();
1369 * Setup the thread queues and initialize the topology based on MD
1373 sched_setup(void *dummy)
1384 /* Add thread0's load since it's running. */
1386 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1387 tdq_load_add(tdq, &thread0);
1388 tdq->tdq_lowpri = thread0.td_priority;
1393 * This routine determines time constants after stathz and hz are setup.
1397 sched_initticks(void *dummy)
1401 realstathz = stathz ? stathz : hz;
1402 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1403 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1404 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1408 * tickincr is shifted out by 10 to avoid rounding errors due to
1409 * hz not being evenly divisible by stathz on all platforms.
1411 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1413 * This does not work for values of stathz that are more than
1414 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1421 * Set the default balance interval now that we know
1422 * what realstathz is.
1424 balance_interval = realstathz;
1425 affinity = SCHED_AFFINITY_DEFAULT;
1427 if (sched_idlespinthresh < 0)
1428 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1433 * This is the core of the interactivity algorithm. Determines a score based
1434 * on past behavior. It is the ratio of sleep time to run time scaled to
1435 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1436 * differs from the cpu usage because it does not account for time spent
1437 * waiting on a run-queue. Would be prettier if we had floating point.
1440 sched_interact_score(struct thread *td)
1442 struct td_sched *ts;
1447 * The score is only needed if this is likely to be an interactive
1448 * task. Don't go through the expense of computing it if there's
1451 if (sched_interact <= SCHED_INTERACT_HALF &&
1452 ts->ts_runtime >= ts->ts_slptime)
1453 return (SCHED_INTERACT_HALF);
1455 if (ts->ts_runtime > ts->ts_slptime) {
1456 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1457 return (SCHED_INTERACT_HALF +
1458 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1460 if (ts->ts_slptime > ts->ts_runtime) {
1461 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1462 return (ts->ts_runtime / div);
1464 /* runtime == slptime */
1466 return (SCHED_INTERACT_HALF);
1469 * This can happen if slptime and runtime are 0.
1476 * Scale the scheduling priority according to the "interactivity" of this
1480 sched_priority(struct thread *td)
1485 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1488 * If the score is interactive we place the thread in the realtime
1489 * queue with a priority that is less than kernel and interrupt
1490 * priorities. These threads are not subject to nice restrictions.
1492 * Scores greater than this are placed on the normal timeshare queue
1493 * where the priority is partially decided by the most recent cpu
1494 * utilization and the rest is decided by nice value.
1496 * The nice value of the process has a linear effect on the calculated
1497 * score. Negative nice values make it easier for a thread to be
1498 * considered interactive.
1500 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1501 if (score < sched_interact) {
1502 pri = PRI_MIN_INTERACT;
1503 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1504 sched_interact) * score;
1505 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1506 ("sched_priority: invalid interactive priority %d score %d",
1509 pri = SCHED_PRI_MIN;
1510 if (td->td_sched->ts_ticks)
1511 pri += min(SCHED_PRI_TICKS(td->td_sched),
1512 SCHED_PRI_RANGE - 1);
1513 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1514 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1515 ("sched_priority: invalid priority %d: nice %d, "
1516 "ticks %d ftick %d ltick %d tick pri %d",
1517 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1518 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1519 SCHED_PRI_TICKS(td->td_sched)));
1521 sched_user_prio(td, pri);
1527 * This routine enforces a maximum limit on the amount of scheduling history
1528 * kept. It is called after either the slptime or runtime is adjusted. This
1529 * function is ugly due to integer math.
1532 sched_interact_update(struct thread *td)
1534 struct td_sched *ts;
1538 sum = ts->ts_runtime + ts->ts_slptime;
1539 if (sum < SCHED_SLP_RUN_MAX)
1542 * This only happens from two places:
1543 * 1) We have added an unusual amount of run time from fork_exit.
1544 * 2) We have added an unusual amount of sleep time from sched_sleep().
1546 if (sum > SCHED_SLP_RUN_MAX * 2) {
1547 if (ts->ts_runtime > ts->ts_slptime) {
1548 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1551 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1557 * If we have exceeded by more than 1/5th then the algorithm below
1558 * will not bring us back into range. Dividing by two here forces
1559 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1561 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1562 ts->ts_runtime /= 2;
1563 ts->ts_slptime /= 2;
1566 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1567 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1571 * Scale back the interactivity history when a child thread is created. The
1572 * history is inherited from the parent but the thread may behave totally
1573 * differently. For example, a shell spawning a compiler process. We want
1574 * to learn that the compiler is behaving badly very quickly.
1577 sched_interact_fork(struct thread *td)
1582 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1583 if (sum > SCHED_SLP_RUN_FORK) {
1584 ratio = sum / SCHED_SLP_RUN_FORK;
1585 td->td_sched->ts_runtime /= ratio;
1586 td->td_sched->ts_slptime /= ratio;
1591 * Called from proc0_init() to setup the scheduler fields.
1598 * Set up the scheduler specific parts of proc0.
1600 proc0.p_sched = NULL; /* XXX */
1601 thread0.td_sched = &td_sched0;
1602 td_sched0.ts_ltick = ticks;
1603 td_sched0.ts_ftick = ticks;
1604 td_sched0.ts_slice = 0;
1608 * This is only somewhat accurate since given many processes of the same
1609 * priority they will switch when their slices run out, which will be
1610 * at most sched_slice stathz ticks.
1613 sched_rr_interval(void)
1616 /* Convert sched_slice from stathz to hz. */
1617 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1621 * Update the percent cpu tracking information when it is requested or
1622 * the total history exceeds the maximum. We keep a sliding history of
1623 * tick counts that slowly decays. This is less precise than the 4BSD
1624 * mechanism since it happens with less regular and frequent events.
1627 sched_pctcpu_update(struct td_sched *ts, int run)
1631 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1633 ts->ts_ftick = t - SCHED_TICK_TARG;
1634 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1635 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1636 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1637 ts->ts_ftick = t - SCHED_TICK_TARG;
1640 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1645 * Adjust the priority of a thread. Move it to the appropriate run-queue
1646 * if necessary. This is the back-end for several priority related
1650 sched_thread_priority(struct thread *td, u_char prio)
1652 struct td_sched *ts;
1656 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1657 "prio:%d", td->td_priority, "new prio:%d", prio,
1658 KTR_ATTR_LINKED, sched_tdname(curthread));
1659 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1660 if (td != curthread && prio < td->td_priority) {
1661 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1662 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1663 prio, KTR_ATTR_LINKED, sched_tdname(td));
1664 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1668 THREAD_LOCK_ASSERT(td, MA_OWNED);
1669 if (td->td_priority == prio)
1672 * If the priority has been elevated due to priority
1673 * propagation, we may have to move ourselves to a new
1674 * queue. This could be optimized to not re-add in some
1677 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1679 td->td_priority = prio;
1680 sched_add(td, SRQ_BORROWING);
1684 * If the thread is currently running we may have to adjust the lowpri
1685 * information so other cpus are aware of our current priority.
1687 if (TD_IS_RUNNING(td)) {
1688 tdq = TDQ_CPU(ts->ts_cpu);
1689 oldpri = td->td_priority;
1690 td->td_priority = prio;
1691 if (prio < tdq->tdq_lowpri)
1692 tdq->tdq_lowpri = prio;
1693 else if (tdq->tdq_lowpri == oldpri)
1694 tdq_setlowpri(tdq, td);
1697 td->td_priority = prio;
1701 * Update a thread's priority when it is lent another thread's
1705 sched_lend_prio(struct thread *td, u_char prio)
1708 td->td_flags |= TDF_BORROWING;
1709 sched_thread_priority(td, prio);
1713 * Restore a thread's priority when priority propagation is
1714 * over. The prio argument is the minimum priority the thread
1715 * needs to have to satisfy other possible priority lending
1716 * requests. If the thread's regular priority is less
1717 * important than prio, the thread will keep a priority boost
1721 sched_unlend_prio(struct thread *td, u_char prio)
1725 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1726 td->td_base_pri <= PRI_MAX_TIMESHARE)
1727 base_pri = td->td_user_pri;
1729 base_pri = td->td_base_pri;
1730 if (prio >= base_pri) {
1731 td->td_flags &= ~TDF_BORROWING;
1732 sched_thread_priority(td, base_pri);
1734 sched_lend_prio(td, prio);
1738 * Standard entry for setting the priority to an absolute value.
1741 sched_prio(struct thread *td, u_char prio)
1745 /* First, update the base priority. */
1746 td->td_base_pri = prio;
1749 * If the thread is borrowing another thread's priority, don't
1750 * ever lower the priority.
1752 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1755 /* Change the real priority. */
1756 oldprio = td->td_priority;
1757 sched_thread_priority(td, prio);
1760 * If the thread is on a turnstile, then let the turnstile update
1763 if (TD_ON_LOCK(td) && oldprio != prio)
1764 turnstile_adjust(td, oldprio);
1768 * Set the base user priority, does not effect current running priority.
1771 sched_user_prio(struct thread *td, u_char prio)
1774 td->td_base_user_pri = prio;
1775 if (td->td_lend_user_pri <= prio)
1777 td->td_user_pri = prio;
1781 sched_lend_user_prio(struct thread *td, u_char prio)
1784 THREAD_LOCK_ASSERT(td, MA_OWNED);
1785 td->td_lend_user_pri = prio;
1786 td->td_user_pri = min(prio, td->td_base_user_pri);
1787 if (td->td_priority > td->td_user_pri)
1788 sched_prio(td, td->td_user_pri);
1789 else if (td->td_priority != td->td_user_pri)
1790 td->td_flags |= TDF_NEEDRESCHED;
1794 * Handle migration from sched_switch(). This happens only for
1798 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1802 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1804 tdq_load_rem(tdq, td);
1806 * Do the lock dance required to avoid LOR. We grab an extra
1807 * spinlock nesting to prevent preemption while we're
1808 * not holding either run-queue lock.
1811 thread_lock_block(td); /* This releases the lock on tdq. */
1814 * Acquire both run-queue locks before placing the thread on the new
1815 * run-queue to avoid deadlocks created by placing a thread with a
1816 * blocked lock on the run-queue of a remote processor. The deadlock
1817 * occurs when a third processor attempts to lock the two queues in
1818 * question while the target processor is spinning with its own
1819 * run-queue lock held while waiting for the blocked lock to clear.
1821 tdq_lock_pair(tdn, tdq);
1822 tdq_add(tdn, td, flags);
1823 tdq_notify(tdn, td);
1827 return (TDQ_LOCKPTR(tdn));
1831 * Variadic version of thread_lock_unblock() that does not assume td_lock
1835 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1837 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1842 * Switch threads. This function has to handle threads coming in while
1843 * blocked for some reason, running, or idle. It also must deal with
1844 * migrating a thread from one queue to another as running threads may
1845 * be assigned elsewhere via binding.
1848 sched_switch(struct thread *td, struct thread *newtd, int flags)
1851 struct td_sched *ts;
1854 int cpuid, preempted;
1856 THREAD_LOCK_ASSERT(td, MA_OWNED);
1857 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1859 cpuid = PCPU_GET(cpuid);
1860 tdq = TDQ_CPU(cpuid);
1863 sched_pctcpu_update(ts, 1);
1864 ts->ts_rltick = ticks;
1865 td->td_lastcpu = td->td_oncpu;
1866 td->td_oncpu = NOCPU;
1867 preempted = !(td->td_flags & TDF_SLICEEND);
1868 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1869 td->td_owepreempt = 0;
1870 if (!TD_IS_IDLETHREAD(td))
1871 tdq->tdq_switchcnt++;
1873 * The lock pointer in an idle thread should never change. Reset it
1874 * to CAN_RUN as well.
1876 if (TD_IS_IDLETHREAD(td)) {
1877 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1879 } else if (TD_IS_RUNNING(td)) {
1880 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1881 srqflag = preempted ?
1882 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1883 SRQ_OURSELF|SRQ_YIELDING;
1885 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1886 ts->ts_cpu = sched_pickcpu(td, 0);
1888 if (ts->ts_cpu == cpuid)
1889 tdq_runq_add(tdq, td, srqflag);
1891 KASSERT(THREAD_CAN_MIGRATE(td) ||
1892 (ts->ts_flags & TSF_BOUND) != 0,
1893 ("Thread %p shouldn't migrate", td));
1894 mtx = sched_switch_migrate(tdq, td, srqflag);
1897 /* This thread must be going to sleep. */
1899 mtx = thread_lock_block(td);
1900 tdq_load_rem(tdq, td);
1903 * We enter here with the thread blocked and assigned to the
1904 * appropriate cpu run-queue or sleep-queue and with the current
1905 * thread-queue locked.
1907 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1908 newtd = choosethread();
1910 * Call the MD code to switch contexts if necessary.
1914 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1915 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1917 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1918 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1919 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1920 sched_pctcpu_update(newtd->td_sched, 0);
1922 #ifdef KDTRACE_HOOKS
1924 * If DTrace has set the active vtime enum to anything
1925 * other than INACTIVE (0), then it should have set the
1928 if (dtrace_vtime_active)
1929 (*dtrace_vtime_switch_func)(newtd);
1932 cpu_switch(td, newtd, mtx);
1934 * We may return from cpu_switch on a different cpu. However,
1935 * we always return with td_lock pointing to the current cpu's
1938 cpuid = PCPU_GET(cpuid);
1939 tdq = TDQ_CPU(cpuid);
1940 lock_profile_obtain_lock_success(
1941 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1943 SDT_PROBE0(sched, , , on__cpu);
1945 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1946 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1949 thread_unblock_switch(td, mtx);
1950 SDT_PROBE0(sched, , , remain__cpu);
1953 * Assert that all went well and return.
1955 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1956 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1957 td->td_oncpu = cpuid;
1961 * Adjust thread priorities as a result of a nice request.
1964 sched_nice(struct proc *p, int nice)
1968 PROC_LOCK_ASSERT(p, MA_OWNED);
1971 FOREACH_THREAD_IN_PROC(p, td) {
1974 sched_prio(td, td->td_base_user_pri);
1980 * Record the sleep time for the interactivity scorer.
1983 sched_sleep(struct thread *td, int prio)
1986 THREAD_LOCK_ASSERT(td, MA_OWNED);
1988 td->td_slptick = ticks;
1989 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1990 td->td_flags |= TDF_CANSWAP;
1991 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1993 if (static_boost == 1 && prio)
1994 sched_prio(td, prio);
1995 else if (static_boost && td->td_priority > static_boost)
1996 sched_prio(td, static_boost);
2000 * Schedule a thread to resume execution and record how long it voluntarily
2001 * slept. We also update the pctcpu, interactivity, and priority.
2004 sched_wakeup(struct thread *td)
2006 struct td_sched *ts;
2009 THREAD_LOCK_ASSERT(td, MA_OWNED);
2011 td->td_flags &= ~TDF_CANSWAP;
2013 * If we slept for more than a tick update our interactivity and
2016 slptick = td->td_slptick;
2018 if (slptick && slptick != ticks) {
2019 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2020 sched_interact_update(td);
2021 sched_pctcpu_update(ts, 0);
2024 * Reset the slice value since we slept and advanced the round-robin.
2027 sched_add(td, SRQ_BORING);
2031 * Penalize the parent for creating a new child and initialize the child's
2035 sched_fork(struct thread *td, struct thread *child)
2037 THREAD_LOCK_ASSERT(td, MA_OWNED);
2038 sched_pctcpu_update(td->td_sched, 1);
2039 sched_fork_thread(td, child);
2041 * Penalize the parent and child for forking.
2043 sched_interact_fork(child);
2044 sched_priority(child);
2045 td->td_sched->ts_runtime += tickincr;
2046 sched_interact_update(td);
2051 * Fork a new thread, may be within the same process.
2054 sched_fork_thread(struct thread *td, struct thread *child)
2056 struct td_sched *ts;
2057 struct td_sched *ts2;
2061 THREAD_LOCK_ASSERT(td, MA_OWNED);
2066 ts2 = child->td_sched;
2067 child->td_lock = TDQ_LOCKPTR(tdq);
2068 child->td_cpuset = cpuset_ref(td->td_cpuset);
2069 ts2->ts_cpu = ts->ts_cpu;
2072 * Grab our parents cpu estimation information.
2074 ts2->ts_ticks = ts->ts_ticks;
2075 ts2->ts_ltick = ts->ts_ltick;
2076 ts2->ts_ftick = ts->ts_ftick;
2078 * Do not inherit any borrowed priority from the parent.
2080 child->td_priority = child->td_base_pri;
2082 * And update interactivity score.
2084 ts2->ts_slptime = ts->ts_slptime;
2085 ts2->ts_runtime = ts->ts_runtime;
2086 /* Attempt to quickly learn interactivity. */
2087 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2089 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2094 * Adjust the priority class of a thread.
2097 sched_class(struct thread *td, int class)
2100 THREAD_LOCK_ASSERT(td, MA_OWNED);
2101 if (td->td_pri_class == class)
2103 td->td_pri_class = class;
2107 * Return some of the child's priority and interactivity to the parent.
2110 sched_exit(struct proc *p, struct thread *child)
2114 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2115 "prio:%d", child->td_priority);
2116 PROC_LOCK_ASSERT(p, MA_OWNED);
2117 td = FIRST_THREAD_IN_PROC(p);
2118 sched_exit_thread(td, child);
2122 * Penalize another thread for the time spent on this one. This helps to
2123 * worsen the priority and interactivity of processes which schedule batch
2124 * jobs such as make. This has little effect on the make process itself but
2125 * causes new processes spawned by it to receive worse scores immediately.
2128 sched_exit_thread(struct thread *td, struct thread *child)
2131 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2132 "prio:%d", child->td_priority);
2134 * Give the child's runtime to the parent without returning the
2135 * sleep time as a penalty to the parent. This causes shells that
2136 * launch expensive things to mark their children as expensive.
2139 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2140 sched_interact_update(td);
2146 sched_preempt(struct thread *td)
2150 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2154 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2155 tdq->tdq_ipipending = 0;
2156 if (td->td_priority > tdq->tdq_lowpri) {
2159 flags = SW_INVOL | SW_PREEMPT;
2160 if (td->td_critnest > 1)
2161 td->td_owepreempt = 1;
2162 else if (TD_IS_IDLETHREAD(td))
2163 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2165 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2171 * Fix priorities on return to user-space. Priorities may be elevated due
2172 * to static priorities in msleep() or similar.
2175 sched_userret(struct thread *td)
2178 * XXX we cheat slightly on the locking here to avoid locking in
2179 * the usual case. Setting td_priority here is essentially an
2180 * incomplete workaround for not setting it properly elsewhere.
2181 * Now that some interrupt handlers are threads, not setting it
2182 * properly elsewhere can clobber it in the window between setting
2183 * it here and returning to user mode, so don't waste time setting
2184 * it perfectly here.
2186 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2187 ("thread with borrowed priority returning to userland"));
2188 if (td->td_priority != td->td_user_pri) {
2190 td->td_priority = td->td_user_pri;
2191 td->td_base_pri = td->td_user_pri;
2192 tdq_setlowpri(TDQ_SELF(), td);
2198 * Handle a stathz tick. This is really only relevant for timeshare
2202 sched_clock(struct thread *td)
2205 struct td_sched *ts;
2207 THREAD_LOCK_ASSERT(td, MA_OWNED);
2211 * We run the long term load balancer infrequently on the first cpu.
2213 if (balance_tdq == tdq) {
2214 if (balance_ticks && --balance_ticks == 0)
2219 * Save the old switch count so we have a record of the last ticks
2220 * activity. Initialize the new switch count based on our load.
2221 * If there is some activity seed it to reflect that.
2223 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2224 tdq->tdq_switchcnt = tdq->tdq_load;
2226 * Advance the insert index once for each tick to ensure that all
2227 * threads get a chance to run.
2229 if (tdq->tdq_idx == tdq->tdq_ridx) {
2230 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2231 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2232 tdq->tdq_ridx = tdq->tdq_idx;
2235 sched_pctcpu_update(ts, 1);
2236 if (td->td_pri_class & PRI_FIFO_BIT)
2238 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2240 * We used a tick; charge it to the thread so
2241 * that we can compute our interactivity.
2243 td->td_sched->ts_runtime += tickincr;
2244 sched_interact_update(td);
2249 * Force a context switch if the current thread has used up a full
2250 * time slice (default is 100ms).
2252 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2254 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2259 * Called once per hz tick.
2268 * Return whether the current CPU has runnable tasks. Used for in-kernel
2269 * cooperative idle threads.
2272 sched_runnable(void)
2280 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2281 if (tdq->tdq_load > 0)
2284 if (tdq->tdq_load - 1 > 0)
2292 * Choose the highest priority thread to run. The thread is removed from
2293 * the run-queue while running however the load remains. For SMP we set
2294 * the tdq in the global idle bitmask if it idles here.
2303 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2304 td = tdq_choose(tdq);
2306 tdq_runq_rem(tdq, td);
2307 tdq->tdq_lowpri = td->td_priority;
2310 tdq->tdq_lowpri = PRI_MAX_IDLE;
2311 return (PCPU_GET(idlethread));
2315 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2316 * we always request it once we exit a critical section.
2319 sched_setpreempt(struct thread *td)
2325 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2328 pri = td->td_priority;
2329 cpri = ctd->td_priority;
2331 ctd->td_flags |= TDF_NEEDRESCHED;
2332 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2334 if (!sched_shouldpreempt(pri, cpri, 0))
2336 ctd->td_owepreempt = 1;
2340 * Add a thread to a thread queue. Select the appropriate runq and add the
2341 * thread to it. This is the internal function called when the tdq is
2345 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2348 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2349 KASSERT((td->td_inhibitors == 0),
2350 ("sched_add: trying to run inhibited thread"));
2351 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2352 ("sched_add: bad thread state"));
2353 KASSERT(td->td_flags & TDF_INMEM,
2354 ("sched_add: thread swapped out"));
2356 if (td->td_priority < tdq->tdq_lowpri)
2357 tdq->tdq_lowpri = td->td_priority;
2358 tdq_runq_add(tdq, td, flags);
2359 tdq_load_add(tdq, td);
2363 * Select the target thread queue and add a thread to it. Request
2364 * preemption or IPI a remote processor if required.
2367 sched_add(struct thread *td, int flags)
2374 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2375 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2376 sched_tdname(curthread));
2377 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2378 KTR_ATTR_LINKED, sched_tdname(td));
2379 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2380 flags & SRQ_PREEMPTED);
2381 THREAD_LOCK_ASSERT(td, MA_OWNED);
2383 * Recalculate the priority before we select the target cpu or
2386 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2390 * Pick the destination cpu and if it isn't ours transfer to the
2393 cpu = sched_pickcpu(td, flags);
2394 tdq = sched_setcpu(td, cpu, flags);
2395 tdq_add(tdq, td, flags);
2396 if (cpu != PCPU_GET(cpuid)) {
2397 tdq_notify(tdq, td);
2404 * Now that the thread is moving to the run-queue, set the lock
2405 * to the scheduler's lock.
2407 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2408 tdq_add(tdq, td, flags);
2410 if (!(flags & SRQ_YIELDING))
2411 sched_setpreempt(td);
2415 * Remove a thread from a run-queue without running it. This is used
2416 * when we're stealing a thread from a remote queue. Otherwise all threads
2417 * exit by calling sched_exit_thread() and sched_throw() themselves.
2420 sched_rem(struct thread *td)
2424 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2425 "prio:%d", td->td_priority);
2426 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2427 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2428 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2429 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2430 KASSERT(TD_ON_RUNQ(td),
2431 ("sched_rem: thread not on run queue"));
2432 tdq_runq_rem(tdq, td);
2433 tdq_load_rem(tdq, td);
2435 if (td->td_priority == tdq->tdq_lowpri)
2436 tdq_setlowpri(tdq, NULL);
2440 * Fetch cpu utilization information. Updates on demand.
2443 sched_pctcpu(struct thread *td)
2446 struct td_sched *ts;
2453 THREAD_LOCK_ASSERT(td, MA_OWNED);
2454 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2458 /* How many rtick per second ? */
2459 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2460 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2467 * Enforce affinity settings for a thread. Called after adjustments to
2471 sched_affinity(struct thread *td)
2474 struct td_sched *ts;
2476 THREAD_LOCK_ASSERT(td, MA_OWNED);
2478 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2480 if (TD_ON_RUNQ(td)) {
2482 sched_add(td, SRQ_BORING);
2485 if (!TD_IS_RUNNING(td))
2488 * Force a switch before returning to userspace. If the
2489 * target thread is not running locally send an ipi to force
2492 td->td_flags |= TDF_NEEDRESCHED;
2493 if (td != curthread)
2494 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2499 * Bind a thread to a target cpu.
2502 sched_bind(struct thread *td, int cpu)
2504 struct td_sched *ts;
2506 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2507 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2509 if (ts->ts_flags & TSF_BOUND)
2511 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2512 ts->ts_flags |= TSF_BOUND;
2514 if (PCPU_GET(cpuid) == cpu)
2517 /* When we return from mi_switch we'll be on the correct cpu. */
2518 mi_switch(SW_VOL, NULL);
2522 * Release a bound thread.
2525 sched_unbind(struct thread *td)
2527 struct td_sched *ts;
2529 THREAD_LOCK_ASSERT(td, MA_OWNED);
2530 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2532 if ((ts->ts_flags & TSF_BOUND) == 0)
2534 ts->ts_flags &= ~TSF_BOUND;
2539 sched_is_bound(struct thread *td)
2541 THREAD_LOCK_ASSERT(td, MA_OWNED);
2542 return (td->td_sched->ts_flags & TSF_BOUND);
2549 sched_relinquish(struct thread *td)
2552 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2557 * Return the total system load.
2568 total += TDQ_CPU(i)->tdq_sysload;
2571 return (TDQ_SELF()->tdq_sysload);
2576 sched_sizeof_proc(void)
2578 return (sizeof(struct proc));
2582 sched_sizeof_thread(void)
2584 return (sizeof(struct thread) + sizeof(struct td_sched));
2588 #define TDQ_IDLESPIN(tdq) \
2589 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2591 #define TDQ_IDLESPIN(tdq) 1
2595 * The actual idle process.
2598 sched_idletd(void *dummy)
2602 int oldswitchcnt, switchcnt;
2605 mtx_assert(&Giant, MA_NOTOWNED);
2608 THREAD_NO_SLEEPING();
2611 if (tdq->tdq_load) {
2613 mi_switch(SW_VOL | SWT_IDLE, NULL);
2616 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2618 if (switchcnt != oldswitchcnt) {
2619 oldswitchcnt = switchcnt;
2620 if (tdq_idled(tdq) == 0)
2623 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2625 oldswitchcnt = switchcnt;
2628 * If we're switching very frequently, spin while checking
2629 * for load rather than entering a low power state that
2630 * may require an IPI. However, don't do any busy
2631 * loops while on SMT machines as this simply steals
2632 * cycles from cores doing useful work.
2634 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2635 for (i = 0; i < sched_idlespins; i++) {
2642 /* If there was context switch during spin, restart it. */
2643 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2644 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2647 /* Run main MD idle handler. */
2648 tdq->tdq_cpu_idle = 1;
2649 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2650 tdq->tdq_cpu_idle = 0;
2653 * Account thread-less hardware interrupts and
2654 * other wakeup reasons equal to context switches.
2656 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2657 if (switchcnt != oldswitchcnt)
2659 tdq->tdq_switchcnt++;
2665 * A CPU is entering for the first time or a thread is exiting.
2668 sched_throw(struct thread *td)
2670 struct thread *newtd;
2675 /* Correct spinlock nesting and acquire the correct lock. */
2678 PCPU_SET(switchtime, cpu_ticks());
2679 PCPU_SET(switchticks, ticks);
2681 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2682 tdq_load_rem(tdq, td);
2683 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2685 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2686 newtd = choosethread();
2687 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2688 cpu_throw(td, newtd); /* doesn't return */
2692 * This is called from fork_exit(). Just acquire the correct locks and
2693 * let fork do the rest of the work.
2696 sched_fork_exit(struct thread *td)
2698 struct td_sched *ts;
2703 * Finish setting up thread glue so that it begins execution in a
2704 * non-nested critical section with the scheduler lock held.
2706 cpuid = PCPU_GET(cpuid);
2707 tdq = TDQ_CPU(cpuid);
2709 if (TD_IS_IDLETHREAD(td))
2710 td->td_lock = TDQ_LOCKPTR(tdq);
2711 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2712 td->td_oncpu = cpuid;
2713 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2714 lock_profile_obtain_lock_success(
2715 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2719 * Create on first use to catch odd startup conditons.
2722 sched_tdname(struct thread *td)
2725 struct td_sched *ts;
2728 if (ts->ts_name[0] == '\0')
2729 snprintf(ts->ts_name, sizeof(ts->ts_name),
2730 "%s tid %d", td->td_name, td->td_tid);
2731 return (ts->ts_name);
2733 return (td->td_name);
2739 sched_clear_tdname(struct thread *td)
2741 struct td_sched *ts;
2744 ts->ts_name[0] = '\0';
2751 * Build the CPU topology dump string. Is recursively called to collect
2752 * the topology tree.
2755 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2758 char cpusetbuf[CPUSETBUFSIZ];
2761 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2762 "", 1 + indent / 2, cg->cg_level);
2763 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2764 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2766 for (i = 0; i < MAXCPU; i++) {
2767 if (CPU_ISSET(i, &cg->cg_mask)) {
2769 sbuf_printf(sb, ", ");
2772 sbuf_printf(sb, "%d", i);
2775 sbuf_printf(sb, "</cpu>\n");
2777 if (cg->cg_flags != 0) {
2778 sbuf_printf(sb, "%*s <flags>", indent, "");
2779 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2780 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2781 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2782 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2783 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2784 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2785 sbuf_printf(sb, "</flags>\n");
2788 if (cg->cg_children > 0) {
2789 sbuf_printf(sb, "%*s <children>\n", indent, "");
2790 for (i = 0; i < cg->cg_children; i++)
2791 sysctl_kern_sched_topology_spec_internal(sb,
2792 &cg->cg_child[i], indent+2);
2793 sbuf_printf(sb, "%*s </children>\n", indent, "");
2795 sbuf_printf(sb, "%*s</group>\n", indent, "");
2800 * Sysctl handler for retrieving topology dump. It's a wrapper for
2801 * the recursive sysctl_kern_smp_topology_spec_internal().
2804 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2809 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2811 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2815 sbuf_printf(topo, "<groups>\n");
2816 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2817 sbuf_printf(topo, "</groups>\n");
2821 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2830 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2832 int error, new_val, period;
2834 period = 1000000 / realstathz;
2835 new_val = period * sched_slice;
2836 error = sysctl_handle_int(oidp, &new_val, 0, req);
2837 if (error != 0 || req->newptr == NULL)
2841 sched_slice = imax(1, (new_val + period / 2) / period);
2842 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2843 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2848 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2849 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2851 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2852 NULL, 0, sysctl_kern_quantum, "I",
2853 "Quantum for timeshare threads in microseconds");
2854 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2855 "Quantum for timeshare threads in stathz ticks");
2856 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2857 "Interactivity score threshold");
2858 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2860 "Maximal (lowest) priority for preemption");
2861 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2862 "Assign static kernel priorities to sleeping threads");
2863 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2864 "Number of times idle thread will spin waiting for new work");
2865 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2866 &sched_idlespinthresh, 0,
2867 "Threshold before we will permit idle thread spinning");
2869 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2870 "Number of hz ticks to keep thread affinity for");
2871 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2872 "Enables the long-term load balancer");
2873 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2874 &balance_interval, 0,
2875 "Average period in stathz ticks to run the long-term balancer");
2876 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2877 "Attempts to steal work from other cores before idling");
2878 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2879 "Minimum load on remote CPU before we'll steal");
2880 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2881 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2882 "XML dump of detected CPU topology");
2885 /* ps compat. All cpu percentages from ULE are weighted. */
2886 static int ccpu = 0;
2887 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");