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 volatile int tdq_transferable; /* Transferable thread count. */
242 volatile short tdq_switchcnt; /* Switches this tick. */
243 volatile 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;
275 static int always_steal = 0;
276 static int trysteal_limit = 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 struct thread *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;
467 ts = td_get_sched(td);
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)
513 ts = td_get_sched(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;
613 * We need some randomness. Implement a classic Linear Congruential
614 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
615 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
616 * of the random state (in the low bits of our answer) to keep
617 * the maximum randomness.
624 rndptr = DPCPU_PTR(randomval);
625 *rndptr = *rndptr * 69069 + 5;
627 return (*rndptr >> 16);
633 int cs_pri; /* Min priority for low. */
634 int cs_limit; /* Max load for low, min load for high. */
639 #define CPU_SEARCH_LOWEST 0x1
640 #define CPU_SEARCH_HIGHEST 0x2
641 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
643 #define CPUSET_FOREACH(cpu, mask) \
644 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
645 if (CPU_ISSET(cpu, &mask))
647 static __always_inline int cpu_search(const struct cpu_group *cg,
648 struct cpu_search *low, struct cpu_search *high, const int match);
649 int __noinline cpu_search_lowest(const struct cpu_group *cg,
650 struct cpu_search *low);
651 int __noinline cpu_search_highest(const struct cpu_group *cg,
652 struct cpu_search *high);
653 int __noinline cpu_search_both(const struct cpu_group *cg,
654 struct cpu_search *low, struct cpu_search *high);
657 * Search the tree of cpu_groups for the lowest or highest loaded cpu
658 * according to the match argument. This routine actually compares the
659 * load on all paths through the tree and finds the least loaded cpu on
660 * the least loaded path, which may differ from the least loaded cpu in
661 * the system. This balances work among caches and busses.
663 * This inline is instantiated in three forms below using constants for the
664 * match argument. It is reduced to the minimum set for each case. It is
665 * also recursive to the depth of the tree.
667 static __always_inline int
668 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
669 struct cpu_search *high, const int match)
671 struct cpu_search lgroup;
672 struct cpu_search hgroup;
674 struct cpu_group *child;
676 int cpu, i, hload, lload, load, total, rnd;
679 cpumask = cg->cg_mask;
680 if (match & CPU_SEARCH_LOWEST) {
684 if (match & CPU_SEARCH_HIGHEST) {
689 /* Iterate through the child CPU groups and then remaining CPUs. */
690 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
692 #ifdef HAVE_INLINE_FFSL
693 cpu = CPU_FFS(&cpumask) - 1;
695 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
702 child = &cg->cg_child[i - 1];
704 if (match & CPU_SEARCH_LOWEST)
706 if (match & CPU_SEARCH_HIGHEST)
708 if (child) { /* Handle child CPU group. */
709 CPU_NAND(&cpumask, &child->cg_mask);
711 case CPU_SEARCH_LOWEST:
712 load = cpu_search_lowest(child, &lgroup);
714 case CPU_SEARCH_HIGHEST:
715 load = cpu_search_highest(child, &hgroup);
717 case CPU_SEARCH_BOTH:
718 load = cpu_search_both(child, &lgroup, &hgroup);
721 } else { /* Handle child CPU. */
722 CPU_CLR(cpu, &cpumask);
724 load = tdq->tdq_load * 256;
725 rnd = sched_random() % 32;
726 if (match & CPU_SEARCH_LOWEST) {
727 if (cpu == low->cs_prefer)
729 /* If that CPU is allowed and get data. */
730 if (tdq->tdq_lowpri > lgroup.cs_pri &&
731 tdq->tdq_load <= lgroup.cs_limit &&
732 CPU_ISSET(cpu, &lgroup.cs_mask)) {
734 lgroup.cs_load = load - rnd;
737 if (match & CPU_SEARCH_HIGHEST)
738 if (tdq->tdq_load >= hgroup.cs_limit &&
739 tdq->tdq_transferable &&
740 CPU_ISSET(cpu, &hgroup.cs_mask)) {
742 hgroup.cs_load = load - rnd;
747 /* We have info about child item. Compare it. */
748 if (match & CPU_SEARCH_LOWEST) {
749 if (lgroup.cs_cpu >= 0 &&
751 (load == lload && lgroup.cs_load < low->cs_load))) {
753 low->cs_cpu = lgroup.cs_cpu;
754 low->cs_load = lgroup.cs_load;
757 if (match & CPU_SEARCH_HIGHEST)
758 if (hgroup.cs_cpu >= 0 &&
760 (load == hload && hgroup.cs_load > high->cs_load))) {
762 high->cs_cpu = hgroup.cs_cpu;
763 high->cs_load = hgroup.cs_load;
767 if (i == 0 && CPU_EMPTY(&cpumask))
770 #ifndef HAVE_INLINE_FFSL
779 * cpu_search instantiations must pass constants to maintain the inline
783 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
785 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
789 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
791 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
795 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
796 struct cpu_search *high)
798 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
802 * Find the cpu with the least load via the least loaded path that has a
803 * lowpri greater than pri pri. A pri of -1 indicates any priority is
807 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
810 struct cpu_search low;
813 low.cs_prefer = prefer;
816 low.cs_limit = maxload;
817 cpu_search_lowest(cg, &low);
822 * Find the cpu with the highest load via the highest loaded path.
825 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
827 struct cpu_search high;
831 high.cs_limit = minload;
832 cpu_search_highest(cg, &high);
837 sched_balance_group(struct cpu_group *cg)
839 cpuset_t hmask, lmask;
840 int high, low, anylow;
844 high = sched_highest(cg, hmask, 2);
845 /* Stop if there is no more CPU with transferrable threads. */
848 CPU_CLR(high, &hmask);
849 CPU_COPY(&hmask, &lmask);
850 /* Stop if there is no more CPU left for low. */
851 if (CPU_EMPTY(&lmask))
855 low = sched_lowest(cg, lmask, -1,
856 TDQ_CPU(high)->tdq_load - 1, high);
857 /* Stop if we looked well and found no less loaded CPU. */
858 if (anylow && low == -1)
860 /* Go to next high if we found no less loaded CPU. */
863 /* Transfer thread from high to low. */
864 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
865 /* CPU that got thread can no longer be a donor. */
866 CPU_CLR(low, &hmask);
869 * If failed, then there is no threads on high
870 * that can run on this low. Drop low from low
871 * mask and look for different one.
873 CPU_CLR(low, &lmask);
885 balance_ticks = max(balance_interval / 2, 1) +
886 (sched_random() % balance_interval);
889 sched_balance_group(cpu_top);
894 * Lock two thread queues using their address to maintain lock order.
897 tdq_lock_pair(struct tdq *one, struct tdq *two)
901 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
904 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
909 * Unlock two thread queues. Order is not important here.
912 tdq_unlock_pair(struct tdq *one, struct tdq *two)
919 * Transfer load between two imbalanced thread queues.
922 sched_balance_pair(struct tdq *high, struct tdq *low)
927 tdq_lock_pair(high, low);
930 * Transfer a thread from high to low.
932 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
933 (td = tdq_move(high, low)) != NULL) {
935 * In case the target isn't the current cpu notify it of the
936 * new load, possibly sending an IPI to force it to reschedule.
939 if (cpu != PCPU_GET(cpuid))
942 tdq_unlock_pair(high, low);
947 * Move a thread from one thread queue to another.
949 static struct thread *
950 tdq_move(struct tdq *from, struct tdq *to)
957 TDQ_LOCK_ASSERT(from, MA_OWNED);
958 TDQ_LOCK_ASSERT(to, MA_OWNED);
962 td = tdq_steal(tdq, cpu);
965 ts = td_get_sched(td);
967 * Although the run queue is locked the thread may be blocked. Lock
968 * it to clear this and acquire the run-queue lock.
971 /* Drop recursive lock on from acquired via thread_lock(). */
975 td->td_lock = TDQ_LOCKPTR(to);
976 tdq_add(to, td, SRQ_YIELDING);
981 * This tdq has idled. Try to steal a thread from another cpu and switch
985 tdq_idled(struct tdq *tdq)
987 struct cpu_group *cg;
992 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
995 CPU_CLR(PCPU_GET(cpuid), &mask);
997 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
998 for (cg = tdq->tdq_cg; ; ) {
999 cpu = sched_highest(cg, mask, steal_thresh);
1001 * We were assigned a thread but not preempted. Returning
1002 * 0 here will cause our caller to switch to it.
1012 steal = TDQ_CPU(cpu);
1014 * The data returned by sched_highest() is stale and
1015 * the chosen CPU no longer has an eligible thread.
1017 * Testing this ahead of tdq_lock_pair() only catches
1018 * this situation about 20% of the time on an 8 core
1019 * 16 thread Ryzen 7, but it still helps performance.
1021 if (steal->tdq_load < steal_thresh ||
1022 steal->tdq_transferable == 0)
1024 tdq_lock_pair(tdq, steal);
1026 * We were assigned a thread while waiting for the locks.
1027 * Switch to it now instead of stealing a thread.
1032 * The data returned by sched_highest() is stale and
1033 * the chosen CPU no longer has an eligible thread, or
1034 * we were preempted and the CPU loading info may be out
1035 * of date. The latter is rare. In either case restart
1038 if (steal->tdq_load < steal_thresh ||
1039 steal->tdq_transferable == 0 ||
1040 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1041 tdq_unlock_pair(tdq, steal);
1045 * Steal the thread and switch to it.
1047 if (tdq_move(steal, tdq) != NULL)
1050 * We failed to acquire a thread even though it looked
1051 * like one was available. This could be due to affinity
1052 * restrictions or for other reasons. Loop again after
1053 * removing this CPU from the set. The restart logic
1054 * above does not restore this CPU to the set due to the
1055 * likelyhood of failing here again.
1057 CPU_CLR(cpu, &mask);
1058 tdq_unlock_pair(tdq, steal);
1061 mi_switch(SW_VOL | SWT_IDLE, NULL);
1062 thread_unlock(curthread);
1067 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1070 tdq_notify(struct tdq *tdq, struct thread *td)
1076 if (tdq->tdq_ipipending)
1078 cpu = td_get_sched(td)->ts_cpu;
1079 pri = td->td_priority;
1080 ctd = pcpu_find(cpu)->pc_curthread;
1081 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1085 * Make sure that our caller's earlier update to tdq_load is
1086 * globally visible before we read tdq_cpu_idle. Idle thread
1087 * accesses both of them without locks, and the order is important.
1089 atomic_thread_fence_seq_cst();
1091 if (TD_IS_IDLETHREAD(ctd)) {
1093 * If the MD code has an idle wakeup routine try that before
1094 * falling back to IPI.
1096 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1099 tdq->tdq_ipipending = 1;
1100 ipi_cpu(cpu, IPI_PREEMPT);
1104 * Steals load from a timeshare queue. Honors the rotating queue head
1107 static struct thread *
1108 runq_steal_from(struct runq *rq, int cpu, u_char start)
1112 struct thread *td, *first;
1116 rqb = &rq->rq_status;
1117 bit = start & (RQB_BPW -1);
1120 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1121 if (rqb->rqb_bits[i] == 0)
1124 bit = RQB_FFS(rqb->rqb_bits[i]);
1125 for (; bit < RQB_BPW; bit++) {
1126 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1128 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1129 TAILQ_FOREACH(td, rqh, td_runq) {
1130 if (first && THREAD_CAN_MIGRATE(td) &&
1131 THREAD_CAN_SCHED(td, cpu))
1142 if (first && THREAD_CAN_MIGRATE(first) &&
1143 THREAD_CAN_SCHED(first, cpu))
1149 * Steals load from a standard linear queue.
1151 static struct thread *
1152 runq_steal(struct runq *rq, int cpu)
1160 rqb = &rq->rq_status;
1161 for (word = 0; word < RQB_LEN; word++) {
1162 if (rqb->rqb_bits[word] == 0)
1164 for (bit = 0; bit < RQB_BPW; bit++) {
1165 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1167 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1168 TAILQ_FOREACH(td, rqh, td_runq)
1169 if (THREAD_CAN_MIGRATE(td) &&
1170 THREAD_CAN_SCHED(td, cpu))
1178 * Attempt to steal a thread in priority order from a thread queue.
1180 static struct thread *
1181 tdq_steal(struct tdq *tdq, int cpu)
1185 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1186 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1188 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1189 cpu, tdq->tdq_ridx)) != NULL)
1191 return (runq_steal(&tdq->tdq_idle, cpu));
1195 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1196 * current lock and returns with the assigned queue locked.
1198 static inline struct tdq *
1199 sched_setcpu(struct thread *td, int cpu, int flags)
1204 THREAD_LOCK_ASSERT(td, MA_OWNED);
1206 td_get_sched(td)->ts_cpu = cpu;
1208 * If the lock matches just return the queue.
1210 if (td->td_lock == TDQ_LOCKPTR(tdq))
1214 * If the thread isn't running its lockptr is a
1215 * turnstile or a sleepqueue. We can just lock_set without
1218 if (TD_CAN_RUN(td)) {
1220 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1225 * The hard case, migration, we need to block the thread first to
1226 * prevent order reversals with other cpus locks.
1229 thread_lock_block(td);
1231 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1236 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1237 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1238 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1239 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1240 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1241 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1244 sched_pickcpu(struct thread *td, int flags)
1246 struct cpu_group *cg, *ccg;
1247 struct td_sched *ts;
1252 self = PCPU_GET(cpuid);
1253 ts = td_get_sched(td);
1254 if (smp_started == 0)
1257 * Don't migrate a running thread from sched_switch().
1259 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1260 return (ts->ts_cpu);
1262 * Prefer to run interrupt threads on the processors that generate
1265 pri = td->td_priority;
1266 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1267 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1268 SCHED_STAT_INC(pickcpu_intrbind);
1270 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1271 SCHED_STAT_INC(pickcpu_affinity);
1272 return (ts->ts_cpu);
1276 * If the thread can run on the last cpu and the affinity has not
1277 * expired or it is idle run it there.
1279 tdq = TDQ_CPU(ts->ts_cpu);
1281 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1282 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1283 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1284 if (cg->cg_flags & CG_FLAG_THREAD) {
1285 CPUSET_FOREACH(cpu, cg->cg_mask) {
1286 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1291 if (cpu > mp_maxid) {
1292 SCHED_STAT_INC(pickcpu_idle_affinity);
1293 return (ts->ts_cpu);
1297 * Search for the last level cache CPU group in the tree.
1298 * Skip caches with expired affinity time and SMT groups.
1299 * Affinity to higher level caches will be handled less aggressively.
1301 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1302 if (cg->cg_flags & CG_FLAG_THREAD)
1304 if (!SCHED_AFFINITY(ts, cg->cg_level))
1311 /* Search the group for the less loaded idle CPU we can run now. */
1312 mask = td->td_cpuset->cs_mask;
1313 if (cg != NULL && cg != cpu_top &&
1314 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1315 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1316 INT_MAX, ts->ts_cpu);
1317 /* Search globally for the less loaded CPU we can run now. */
1319 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1320 /* Search globally for the less loaded CPU. */
1322 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1323 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1325 * Compare the lowest loaded cpu to current cpu.
1327 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1328 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1329 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1330 SCHED_STAT_INC(pickcpu_local);
1333 SCHED_STAT_INC(pickcpu_lowest);
1334 if (cpu != ts->ts_cpu)
1335 SCHED_STAT_INC(pickcpu_migration);
1341 * Pick the highest priority task we have and return it.
1343 static struct thread *
1344 tdq_choose(struct tdq *tdq)
1348 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1349 td = runq_choose(&tdq->tdq_realtime);
1352 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1354 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1355 ("tdq_choose: Invalid priority on timeshare queue %d",
1359 td = runq_choose(&tdq->tdq_idle);
1361 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1362 ("tdq_choose: Invalid priority on idle queue %d",
1371 * Initialize a thread queue.
1374 tdq_setup(struct tdq *tdq)
1378 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1379 runq_init(&tdq->tdq_realtime);
1380 runq_init(&tdq->tdq_timeshare);
1381 runq_init(&tdq->tdq_idle);
1382 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1383 "sched lock %d", (int)TDQ_ID(tdq));
1384 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1385 MTX_SPIN | MTX_RECURSE);
1387 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1388 "CPU %d load", (int)TDQ_ID(tdq));
1394 sched_setup_smp(void)
1399 cpu_top = smp_topo();
1403 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1404 if (tdq->tdq_cg == NULL)
1405 panic("Can't find cpu group for %d\n", i);
1407 balance_tdq = TDQ_SELF();
1412 * Setup the thread queues and initialize the topology based on MD
1416 sched_setup(void *dummy)
1427 /* Add thread0's load since it's running. */
1429 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1430 tdq_load_add(tdq, &thread0);
1431 tdq->tdq_lowpri = thread0.td_priority;
1436 * This routine determines time constants after stathz and hz are setup.
1440 sched_initticks(void *dummy)
1444 realstathz = stathz ? stathz : hz;
1445 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1446 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1447 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1451 * tickincr is shifted out by 10 to avoid rounding errors due to
1452 * hz not being evenly divisible by stathz on all platforms.
1454 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1456 * This does not work for values of stathz that are more than
1457 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1464 * Set the default balance interval now that we know
1465 * what realstathz is.
1467 balance_interval = realstathz;
1468 balance_ticks = balance_interval;
1469 affinity = SCHED_AFFINITY_DEFAULT;
1471 if (sched_idlespinthresh < 0)
1472 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1477 * This is the core of the interactivity algorithm. Determines a score based
1478 * on past behavior. It is the ratio of sleep time to run time scaled to
1479 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1480 * differs from the cpu usage because it does not account for time spent
1481 * waiting on a run-queue. Would be prettier if we had floating point.
1483 * When a thread's sleep time is greater than its run time the
1487 * interactivity score = ---------------------
1488 * sleep time / run time
1491 * When a thread's run time is greater than its sleep time the
1495 * interactivity score = --------------------- + scaling factor
1496 * run time / sleep time
1499 sched_interact_score(struct thread *td)
1501 struct td_sched *ts;
1504 ts = td_get_sched(td);
1506 * The score is only needed if this is likely to be an interactive
1507 * task. Don't go through the expense of computing it if there's
1510 if (sched_interact <= SCHED_INTERACT_HALF &&
1511 ts->ts_runtime >= ts->ts_slptime)
1512 return (SCHED_INTERACT_HALF);
1514 if (ts->ts_runtime > ts->ts_slptime) {
1515 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1516 return (SCHED_INTERACT_HALF +
1517 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1519 if (ts->ts_slptime > ts->ts_runtime) {
1520 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1521 return (ts->ts_runtime / div);
1523 /* runtime == slptime */
1525 return (SCHED_INTERACT_HALF);
1528 * This can happen if slptime and runtime are 0.
1535 * Scale the scheduling priority according to the "interactivity" of this
1539 sched_priority(struct thread *td)
1544 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1547 * If the score is interactive we place the thread in the realtime
1548 * queue with a priority that is less than kernel and interrupt
1549 * priorities. These threads are not subject to nice restrictions.
1551 * Scores greater than this are placed on the normal timeshare queue
1552 * where the priority is partially decided by the most recent cpu
1553 * utilization and the rest is decided by nice value.
1555 * The nice value of the process has a linear effect on the calculated
1556 * score. Negative nice values make it easier for a thread to be
1557 * considered interactive.
1559 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1560 if (score < sched_interact) {
1561 pri = PRI_MIN_INTERACT;
1562 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1563 sched_interact) * score;
1564 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1565 ("sched_priority: invalid interactive priority %d score %d",
1568 pri = SCHED_PRI_MIN;
1569 if (td_get_sched(td)->ts_ticks)
1570 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1571 SCHED_PRI_RANGE - 1);
1572 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1573 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1574 ("sched_priority: invalid priority %d: nice %d, "
1575 "ticks %d ftick %d ltick %d tick pri %d",
1576 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1577 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1578 SCHED_PRI_TICKS(td_get_sched(td))));
1580 sched_user_prio(td, pri);
1586 * This routine enforces a maximum limit on the amount of scheduling history
1587 * kept. It is called after either the slptime or runtime is adjusted. This
1588 * function is ugly due to integer math.
1591 sched_interact_update(struct thread *td)
1593 struct td_sched *ts;
1596 ts = td_get_sched(td);
1597 sum = ts->ts_runtime + ts->ts_slptime;
1598 if (sum < SCHED_SLP_RUN_MAX)
1601 * This only happens from two places:
1602 * 1) We have added an unusual amount of run time from fork_exit.
1603 * 2) We have added an unusual amount of sleep time from sched_sleep().
1605 if (sum > SCHED_SLP_RUN_MAX * 2) {
1606 if (ts->ts_runtime > ts->ts_slptime) {
1607 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1610 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1616 * If we have exceeded by more than 1/5th then the algorithm below
1617 * will not bring us back into range. Dividing by two here forces
1618 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1620 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1621 ts->ts_runtime /= 2;
1622 ts->ts_slptime /= 2;
1625 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1626 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1630 * Scale back the interactivity history when a child thread is created. The
1631 * history is inherited from the parent but the thread may behave totally
1632 * differently. For example, a shell spawning a compiler process. We want
1633 * to learn that the compiler is behaving badly very quickly.
1636 sched_interact_fork(struct thread *td)
1638 struct td_sched *ts;
1642 ts = td_get_sched(td);
1643 sum = ts->ts_runtime + ts->ts_slptime;
1644 if (sum > SCHED_SLP_RUN_FORK) {
1645 ratio = sum / SCHED_SLP_RUN_FORK;
1646 ts->ts_runtime /= ratio;
1647 ts->ts_slptime /= ratio;
1652 * Called from proc0_init() to setup the scheduler fields.
1657 struct td_sched *ts0;
1660 * Set up the scheduler specific parts of thread0.
1662 ts0 = td_get_sched(&thread0);
1663 ts0->ts_ltick = ticks;
1664 ts0->ts_ftick = ticks;
1669 * This is only somewhat accurate since given many processes of the same
1670 * priority they will switch when their slices run out, which will be
1671 * at most sched_slice stathz ticks.
1674 sched_rr_interval(void)
1677 /* Convert sched_slice from stathz to hz. */
1678 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1682 * Update the percent cpu tracking information when it is requested or
1683 * the total history exceeds the maximum. We keep a sliding history of
1684 * tick counts that slowly decays. This is less precise than the 4BSD
1685 * mechanism since it happens with less regular and frequent events.
1688 sched_pctcpu_update(struct td_sched *ts, int run)
1693 * The signed difference may be negative if the thread hasn't run for
1694 * over half of the ticks rollover period.
1696 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1698 ts->ts_ftick = t - SCHED_TICK_TARG;
1699 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1700 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1701 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1702 ts->ts_ftick = t - SCHED_TICK_TARG;
1705 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1710 * Adjust the priority of a thread. Move it to the appropriate run-queue
1711 * if necessary. This is the back-end for several priority related
1715 sched_thread_priority(struct thread *td, u_char prio)
1717 struct td_sched *ts;
1721 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1722 "prio:%d", td->td_priority, "new prio:%d", prio,
1723 KTR_ATTR_LINKED, sched_tdname(curthread));
1724 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1725 if (td != curthread && prio < td->td_priority) {
1726 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1727 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1728 prio, KTR_ATTR_LINKED, sched_tdname(td));
1729 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1732 ts = td_get_sched(td);
1733 THREAD_LOCK_ASSERT(td, MA_OWNED);
1734 if (td->td_priority == prio)
1737 * If the priority has been elevated due to priority
1738 * propagation, we may have to move ourselves to a new
1739 * queue. This could be optimized to not re-add in some
1742 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1744 td->td_priority = prio;
1745 sched_add(td, SRQ_BORROWING);
1749 * If the thread is currently running we may have to adjust the lowpri
1750 * information so other cpus are aware of our current priority.
1752 if (TD_IS_RUNNING(td)) {
1753 tdq = TDQ_CPU(ts->ts_cpu);
1754 oldpri = td->td_priority;
1755 td->td_priority = prio;
1756 if (prio < tdq->tdq_lowpri)
1757 tdq->tdq_lowpri = prio;
1758 else if (tdq->tdq_lowpri == oldpri)
1759 tdq_setlowpri(tdq, td);
1762 td->td_priority = prio;
1766 * Update a thread's priority when it is lent another thread's
1770 sched_lend_prio(struct thread *td, u_char prio)
1773 td->td_flags |= TDF_BORROWING;
1774 sched_thread_priority(td, prio);
1778 * Restore a thread's priority when priority propagation is
1779 * over. The prio argument is the minimum priority the thread
1780 * needs to have to satisfy other possible priority lending
1781 * requests. If the thread's regular priority is less
1782 * important than prio, the thread will keep a priority boost
1786 sched_unlend_prio(struct thread *td, u_char prio)
1790 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1791 td->td_base_pri <= PRI_MAX_TIMESHARE)
1792 base_pri = td->td_user_pri;
1794 base_pri = td->td_base_pri;
1795 if (prio >= base_pri) {
1796 td->td_flags &= ~TDF_BORROWING;
1797 sched_thread_priority(td, base_pri);
1799 sched_lend_prio(td, prio);
1803 * Standard entry for setting the priority to an absolute value.
1806 sched_prio(struct thread *td, u_char prio)
1810 /* First, update the base priority. */
1811 td->td_base_pri = prio;
1814 * If the thread is borrowing another thread's priority, don't
1815 * ever lower the priority.
1817 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1820 /* Change the real priority. */
1821 oldprio = td->td_priority;
1822 sched_thread_priority(td, prio);
1825 * If the thread is on a turnstile, then let the turnstile update
1828 if (TD_ON_LOCK(td) && oldprio != prio)
1829 turnstile_adjust(td, oldprio);
1833 * Set the base user priority, does not effect current running priority.
1836 sched_user_prio(struct thread *td, u_char prio)
1839 td->td_base_user_pri = prio;
1840 if (td->td_lend_user_pri <= prio)
1842 td->td_user_pri = prio;
1846 sched_lend_user_prio(struct thread *td, u_char prio)
1849 THREAD_LOCK_ASSERT(td, MA_OWNED);
1850 td->td_lend_user_pri = prio;
1851 td->td_user_pri = min(prio, td->td_base_user_pri);
1852 if (td->td_priority > td->td_user_pri)
1853 sched_prio(td, td->td_user_pri);
1854 else if (td->td_priority != td->td_user_pri)
1855 td->td_flags |= TDF_NEEDRESCHED;
1860 * This tdq is about to idle. Try to steal a thread from another CPU before
1861 * choosing the idle thread.
1864 tdq_trysteal(struct tdq *tdq)
1866 struct cpu_group *cg;
1871 if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1874 CPU_CLR(PCPU_GET(cpuid), &mask);
1875 /* We don't want to be preempted while we're iterating. */
1878 for (i = 1, cg = tdq->tdq_cg; ; ) {
1879 cpu = sched_highest(cg, mask, steal_thresh);
1881 * If a thread was added while interrupts were disabled don't
1884 if (tdq->tdq_load > 0) {
1891 if (cg == NULL || i > trysteal_limit) {
1897 steal = TDQ_CPU(cpu);
1899 * The data returned by sched_highest() is stale and
1900 * the chosen CPU no longer has an eligible thread.
1902 if (steal->tdq_load < steal_thresh ||
1903 steal->tdq_transferable == 0)
1905 tdq_lock_pair(tdq, steal);
1907 * If we get to this point, unconditonally exit the loop
1908 * to bound the time spent in the critcal section.
1910 * If a thread was added while interrupts were disabled don't
1913 if (tdq->tdq_load > 0) {
1918 * The data returned by sched_highest() is stale and
1919 * the chosen CPU no longer has an eligible thread.
1921 if (steal->tdq_load < steal_thresh ||
1922 steal->tdq_transferable == 0) {
1927 * If we fail to acquire one due to affinity restrictions,
1928 * bail out and let the idle thread to a more complete search
1929 * outside of a critical section.
1931 if (tdq_move(steal, tdq) == NULL) {
1943 * Handle migration from sched_switch(). This happens only for
1947 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1951 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1953 tdq_load_rem(tdq, td);
1955 * Do the lock dance required to avoid LOR. We grab an extra
1956 * spinlock nesting to prevent preemption while we're
1957 * not holding either run-queue lock.
1960 thread_lock_block(td); /* This releases the lock on tdq. */
1963 * Acquire both run-queue locks before placing the thread on the new
1964 * run-queue to avoid deadlocks created by placing a thread with a
1965 * blocked lock on the run-queue of a remote processor. The deadlock
1966 * occurs when a third processor attempts to lock the two queues in
1967 * question while the target processor is spinning with its own
1968 * run-queue lock held while waiting for the blocked lock to clear.
1970 tdq_lock_pair(tdn, tdq);
1971 tdq_add(tdn, td, flags);
1972 tdq_notify(tdn, td);
1976 return (TDQ_LOCKPTR(tdn));
1980 * Variadic version of thread_lock_unblock() that does not assume td_lock
1984 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1986 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1991 * Switch threads. This function has to handle threads coming in while
1992 * blocked for some reason, running, or idle. It also must deal with
1993 * migrating a thread from one queue to another as running threads may
1994 * be assigned elsewhere via binding.
1997 sched_switch(struct thread *td, struct thread *newtd, int flags)
2000 struct td_sched *ts;
2003 int cpuid, preempted;
2005 THREAD_LOCK_ASSERT(td, MA_OWNED);
2006 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2008 cpuid = PCPU_GET(cpuid);
2009 tdq = TDQ_CPU(cpuid);
2010 ts = td_get_sched(td);
2012 sched_pctcpu_update(ts, 1);
2013 ts->ts_rltick = ticks;
2014 td->td_lastcpu = td->td_oncpu;
2015 td->td_oncpu = NOCPU;
2016 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2017 (flags & SW_PREEMPT) != 0;
2018 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2019 td->td_owepreempt = 0;
2020 if (!TD_IS_IDLETHREAD(td))
2021 tdq->tdq_switchcnt++;
2023 * The lock pointer in an idle thread should never change. Reset it
2024 * to CAN_RUN as well.
2026 if (TD_IS_IDLETHREAD(td)) {
2027 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2029 } else if (TD_IS_RUNNING(td)) {
2030 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2031 srqflag = preempted ?
2032 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2033 SRQ_OURSELF|SRQ_YIELDING;
2035 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2036 ts->ts_cpu = sched_pickcpu(td, 0);
2038 if (ts->ts_cpu == cpuid)
2039 tdq_runq_add(tdq, td, srqflag);
2041 KASSERT(THREAD_CAN_MIGRATE(td) ||
2042 (ts->ts_flags & TSF_BOUND) != 0,
2043 ("Thread %p shouldn't migrate", td));
2044 mtx = sched_switch_migrate(tdq, td, srqflag);
2047 /* This thread must be going to sleep. */
2049 mtx = thread_lock_block(td);
2050 tdq_load_rem(tdq, td);
2052 if (tdq->tdq_load == 0)
2057 #if (KTR_COMPILE & KTR_SCHED) != 0
2058 if (TD_IS_IDLETHREAD(td))
2059 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2060 "prio:%d", td->td_priority);
2062 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2063 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2064 "lockname:\"%s\"", td->td_lockname);
2068 * We enter here with the thread blocked and assigned to the
2069 * appropriate cpu run-queue or sleep-queue and with the current
2070 * thread-queue locked.
2072 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2073 newtd = choosethread();
2075 * Call the MD code to switch contexts if necessary.
2079 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2080 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2082 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2083 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2084 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2085 sched_pctcpu_update(td_get_sched(newtd), 0);
2087 #ifdef KDTRACE_HOOKS
2089 * If DTrace has set the active vtime enum to anything
2090 * other than INACTIVE (0), then it should have set the
2093 if (dtrace_vtime_active)
2094 (*dtrace_vtime_switch_func)(newtd);
2097 cpu_switch(td, newtd, mtx);
2099 * We may return from cpu_switch on a different cpu. However,
2100 * we always return with td_lock pointing to the current cpu's
2103 cpuid = PCPU_GET(cpuid);
2104 tdq = TDQ_CPU(cpuid);
2105 lock_profile_obtain_lock_success(
2106 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2108 SDT_PROBE0(sched, , , on__cpu);
2110 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2111 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2114 thread_unblock_switch(td, mtx);
2115 SDT_PROBE0(sched, , , remain__cpu);
2118 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2119 "prio:%d", td->td_priority);
2122 * Assert that all went well and return.
2124 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2125 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2126 td->td_oncpu = cpuid;
2130 * Adjust thread priorities as a result of a nice request.
2133 sched_nice(struct proc *p, int nice)
2137 PROC_LOCK_ASSERT(p, MA_OWNED);
2140 FOREACH_THREAD_IN_PROC(p, td) {
2143 sched_prio(td, td->td_base_user_pri);
2149 * Record the sleep time for the interactivity scorer.
2152 sched_sleep(struct thread *td, int prio)
2155 THREAD_LOCK_ASSERT(td, MA_OWNED);
2157 td->td_slptick = ticks;
2158 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2159 td->td_flags |= TDF_CANSWAP;
2160 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2162 if (static_boost == 1 && prio)
2163 sched_prio(td, prio);
2164 else if (static_boost && td->td_priority > static_boost)
2165 sched_prio(td, static_boost);
2169 * Schedule a thread to resume execution and record how long it voluntarily
2170 * slept. We also update the pctcpu, interactivity, and priority.
2173 sched_wakeup(struct thread *td)
2175 struct td_sched *ts;
2178 THREAD_LOCK_ASSERT(td, MA_OWNED);
2179 ts = td_get_sched(td);
2180 td->td_flags &= ~TDF_CANSWAP;
2182 * If we slept for more than a tick update our interactivity and
2185 slptick = td->td_slptick;
2187 if (slptick && slptick != ticks) {
2188 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2189 sched_interact_update(td);
2190 sched_pctcpu_update(ts, 0);
2193 * Reset the slice value since we slept and advanced the round-robin.
2196 sched_add(td, SRQ_BORING);
2200 * Penalize the parent for creating a new child and initialize the child's
2204 sched_fork(struct thread *td, struct thread *child)
2206 THREAD_LOCK_ASSERT(td, MA_OWNED);
2207 sched_pctcpu_update(td_get_sched(td), 1);
2208 sched_fork_thread(td, child);
2210 * Penalize the parent and child for forking.
2212 sched_interact_fork(child);
2213 sched_priority(child);
2214 td_get_sched(td)->ts_runtime += tickincr;
2215 sched_interact_update(td);
2220 * Fork a new thread, may be within the same process.
2223 sched_fork_thread(struct thread *td, struct thread *child)
2225 struct td_sched *ts;
2226 struct td_sched *ts2;
2230 THREAD_LOCK_ASSERT(td, MA_OWNED);
2234 ts = td_get_sched(td);
2235 ts2 = td_get_sched(child);
2236 child->td_oncpu = NOCPU;
2237 child->td_lastcpu = NOCPU;
2238 child->td_lock = TDQ_LOCKPTR(tdq);
2239 child->td_cpuset = cpuset_ref(td->td_cpuset);
2240 ts2->ts_cpu = ts->ts_cpu;
2243 * Grab our parents cpu estimation information.
2245 ts2->ts_ticks = ts->ts_ticks;
2246 ts2->ts_ltick = ts->ts_ltick;
2247 ts2->ts_ftick = ts->ts_ftick;
2249 * Do not inherit any borrowed priority from the parent.
2251 child->td_priority = child->td_base_pri;
2253 * And update interactivity score.
2255 ts2->ts_slptime = ts->ts_slptime;
2256 ts2->ts_runtime = ts->ts_runtime;
2257 /* Attempt to quickly learn interactivity. */
2258 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2260 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2265 * Adjust the priority class of a thread.
2268 sched_class(struct thread *td, int class)
2271 THREAD_LOCK_ASSERT(td, MA_OWNED);
2272 if (td->td_pri_class == class)
2274 td->td_pri_class = class;
2278 * Return some of the child's priority and interactivity to the parent.
2281 sched_exit(struct proc *p, struct thread *child)
2285 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2286 "prio:%d", child->td_priority);
2287 PROC_LOCK_ASSERT(p, MA_OWNED);
2288 td = FIRST_THREAD_IN_PROC(p);
2289 sched_exit_thread(td, child);
2293 * Penalize another thread for the time spent on this one. This helps to
2294 * worsen the priority and interactivity of processes which schedule batch
2295 * jobs such as make. This has little effect on the make process itself but
2296 * causes new processes spawned by it to receive worse scores immediately.
2299 sched_exit_thread(struct thread *td, struct thread *child)
2302 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2303 "prio:%d", child->td_priority);
2305 * Give the child's runtime to the parent without returning the
2306 * sleep time as a penalty to the parent. This causes shells that
2307 * launch expensive things to mark their children as expensive.
2310 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2311 sched_interact_update(td);
2317 sched_preempt(struct thread *td)
2321 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2325 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2326 tdq->tdq_ipipending = 0;
2327 if (td->td_priority > tdq->tdq_lowpri) {
2330 flags = SW_INVOL | SW_PREEMPT;
2331 if (td->td_critnest > 1)
2332 td->td_owepreempt = 1;
2333 else if (TD_IS_IDLETHREAD(td))
2334 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2336 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2342 * Fix priorities on return to user-space. Priorities may be elevated due
2343 * to static priorities in msleep() or similar.
2346 sched_userret(struct thread *td)
2349 * XXX we cheat slightly on the locking here to avoid locking in
2350 * the usual case. Setting td_priority here is essentially an
2351 * incomplete workaround for not setting it properly elsewhere.
2352 * Now that some interrupt handlers are threads, not setting it
2353 * properly elsewhere can clobber it in the window between setting
2354 * it here and returning to user mode, so don't waste time setting
2355 * it perfectly here.
2357 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2358 ("thread with borrowed priority returning to userland"));
2359 if (td->td_priority != td->td_user_pri) {
2361 td->td_priority = td->td_user_pri;
2362 td->td_base_pri = td->td_user_pri;
2363 tdq_setlowpri(TDQ_SELF(), td);
2369 * Handle a stathz tick. This is really only relevant for timeshare
2373 sched_clock(struct thread *td)
2376 struct td_sched *ts;
2378 THREAD_LOCK_ASSERT(td, MA_OWNED);
2382 * We run the long term load balancer infrequently on the first cpu.
2384 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2385 if (balance_ticks && --balance_ticks == 0)
2390 * Save the old switch count so we have a record of the last ticks
2391 * activity. Initialize the new switch count based on our load.
2392 * If there is some activity seed it to reflect that.
2394 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2395 tdq->tdq_switchcnt = tdq->tdq_load;
2397 * Advance the insert index once for each tick to ensure that all
2398 * threads get a chance to run.
2400 if (tdq->tdq_idx == tdq->tdq_ridx) {
2401 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2402 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2403 tdq->tdq_ridx = tdq->tdq_idx;
2405 ts = td_get_sched(td);
2406 sched_pctcpu_update(ts, 1);
2407 if (td->td_pri_class & PRI_FIFO_BIT)
2409 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2411 * We used a tick; charge it to the thread so
2412 * that we can compute our interactivity.
2414 td_get_sched(td)->ts_runtime += tickincr;
2415 sched_interact_update(td);
2420 * Force a context switch if the current thread has used up a full
2421 * time slice (default is 100ms).
2423 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2425 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2430 sched_estcpu(struct thread *td __unused)
2437 * Return whether the current CPU has runnable tasks. Used for in-kernel
2438 * cooperative idle threads.
2441 sched_runnable(void)
2449 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2450 if (tdq->tdq_load > 0)
2453 if (tdq->tdq_load - 1 > 0)
2461 * Choose the highest priority thread to run. The thread is removed from
2462 * the run-queue while running however the load remains. For SMP we set
2463 * the tdq in the global idle bitmask if it idles here.
2472 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2473 td = tdq_choose(tdq);
2475 tdq_runq_rem(tdq, td);
2476 tdq->tdq_lowpri = td->td_priority;
2479 tdq->tdq_lowpri = PRI_MAX_IDLE;
2480 return (PCPU_GET(idlethread));
2484 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2485 * we always request it once we exit a critical section.
2488 sched_setpreempt(struct thread *td)
2494 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2497 pri = td->td_priority;
2498 cpri = ctd->td_priority;
2500 ctd->td_flags |= TDF_NEEDRESCHED;
2501 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2503 if (!sched_shouldpreempt(pri, cpri, 0))
2505 ctd->td_owepreempt = 1;
2509 * Add a thread to a thread queue. Select the appropriate runq and add the
2510 * thread to it. This is the internal function called when the tdq is
2514 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2517 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2518 KASSERT((td->td_inhibitors == 0),
2519 ("sched_add: trying to run inhibited thread"));
2520 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2521 ("sched_add: bad thread state"));
2522 KASSERT(td->td_flags & TDF_INMEM,
2523 ("sched_add: thread swapped out"));
2525 if (td->td_priority < tdq->tdq_lowpri)
2526 tdq->tdq_lowpri = td->td_priority;
2527 tdq_runq_add(tdq, td, flags);
2528 tdq_load_add(tdq, td);
2532 * Select the target thread queue and add a thread to it. Request
2533 * preemption or IPI a remote processor if required.
2536 sched_add(struct thread *td, int flags)
2543 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2544 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2545 sched_tdname(curthread));
2546 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2547 KTR_ATTR_LINKED, sched_tdname(td));
2548 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2549 flags & SRQ_PREEMPTED);
2550 THREAD_LOCK_ASSERT(td, MA_OWNED);
2552 * Recalculate the priority before we select the target cpu or
2555 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2559 * Pick the destination cpu and if it isn't ours transfer to the
2562 cpu = sched_pickcpu(td, flags);
2563 tdq = sched_setcpu(td, cpu, flags);
2564 tdq_add(tdq, td, flags);
2565 if (cpu != PCPU_GET(cpuid)) {
2566 tdq_notify(tdq, td);
2573 * Now that the thread is moving to the run-queue, set the lock
2574 * to the scheduler's lock.
2576 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2577 tdq_add(tdq, td, flags);
2579 if (!(flags & SRQ_YIELDING))
2580 sched_setpreempt(td);
2584 * Remove a thread from a run-queue without running it. This is used
2585 * when we're stealing a thread from a remote queue. Otherwise all threads
2586 * exit by calling sched_exit_thread() and sched_throw() themselves.
2589 sched_rem(struct thread *td)
2593 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2594 "prio:%d", td->td_priority);
2595 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2596 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2597 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2598 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2599 KASSERT(TD_ON_RUNQ(td),
2600 ("sched_rem: thread not on run queue"));
2601 tdq_runq_rem(tdq, td);
2602 tdq_load_rem(tdq, td);
2604 if (td->td_priority == tdq->tdq_lowpri)
2605 tdq_setlowpri(tdq, NULL);
2609 * Fetch cpu utilization information. Updates on demand.
2612 sched_pctcpu(struct thread *td)
2615 struct td_sched *ts;
2618 ts = td_get_sched(td);
2620 THREAD_LOCK_ASSERT(td, MA_OWNED);
2621 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2625 /* How many rtick per second ? */
2626 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2627 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2634 * Enforce affinity settings for a thread. Called after adjustments to
2638 sched_affinity(struct thread *td)
2641 struct td_sched *ts;
2643 THREAD_LOCK_ASSERT(td, MA_OWNED);
2644 ts = td_get_sched(td);
2645 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2647 if (TD_ON_RUNQ(td)) {
2649 sched_add(td, SRQ_BORING);
2652 if (!TD_IS_RUNNING(td))
2655 * Force a switch before returning to userspace. If the
2656 * target thread is not running locally send an ipi to force
2659 td->td_flags |= TDF_NEEDRESCHED;
2660 if (td != curthread)
2661 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2666 * Bind a thread to a target cpu.
2669 sched_bind(struct thread *td, int cpu)
2671 struct td_sched *ts;
2673 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2674 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2675 ts = td_get_sched(td);
2676 if (ts->ts_flags & TSF_BOUND)
2678 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2679 ts->ts_flags |= TSF_BOUND;
2681 if (PCPU_GET(cpuid) == cpu)
2684 /* When we return from mi_switch we'll be on the correct cpu. */
2685 mi_switch(SW_VOL, NULL);
2689 * Release a bound thread.
2692 sched_unbind(struct thread *td)
2694 struct td_sched *ts;
2696 THREAD_LOCK_ASSERT(td, MA_OWNED);
2697 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2698 ts = td_get_sched(td);
2699 if ((ts->ts_flags & TSF_BOUND) == 0)
2701 ts->ts_flags &= ~TSF_BOUND;
2706 sched_is_bound(struct thread *td)
2708 THREAD_LOCK_ASSERT(td, MA_OWNED);
2709 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2716 sched_relinquish(struct thread *td)
2719 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2724 * Return the total system load.
2735 total += TDQ_CPU(i)->tdq_sysload;
2738 return (TDQ_SELF()->tdq_sysload);
2743 sched_sizeof_proc(void)
2745 return (sizeof(struct proc));
2749 sched_sizeof_thread(void)
2751 return (sizeof(struct thread) + sizeof(struct td_sched));
2755 #define TDQ_IDLESPIN(tdq) \
2756 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2758 #define TDQ_IDLESPIN(tdq) 1
2762 * The actual idle process.
2765 sched_idletd(void *dummy)
2769 int oldswitchcnt, switchcnt;
2772 mtx_assert(&Giant, MA_NOTOWNED);
2775 THREAD_NO_SLEEPING();
2778 if (tdq->tdq_load) {
2780 mi_switch(SW_VOL | SWT_IDLE, NULL);
2783 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2785 if (always_steal || switchcnt != oldswitchcnt) {
2786 oldswitchcnt = switchcnt;
2787 if (tdq_idled(tdq) == 0)
2790 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2792 oldswitchcnt = switchcnt;
2795 * If we're switching very frequently, spin while checking
2796 * for load rather than entering a low power state that
2797 * may require an IPI. However, don't do any busy
2798 * loops while on SMT machines as this simply steals
2799 * cycles from cores doing useful work.
2801 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2802 for (i = 0; i < sched_idlespins; i++) {
2809 /* If there was context switch during spin, restart it. */
2810 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2811 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2814 /* Run main MD idle handler. */
2815 tdq->tdq_cpu_idle = 1;
2817 * Make sure that tdq_cpu_idle update is globally visible
2818 * before cpu_idle() read tdq_load. The order is important
2819 * to avoid race with tdq_notify.
2821 atomic_thread_fence_seq_cst();
2823 * Checking for again after the fence picks up assigned
2824 * threads often enough to make it worthwhile to do so in
2825 * order to avoid calling cpu_idle().
2827 if (tdq->tdq_load != 0) {
2828 tdq->tdq_cpu_idle = 0;
2831 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2832 tdq->tdq_cpu_idle = 0;
2835 * Account thread-less hardware interrupts and
2836 * other wakeup reasons equal to context switches.
2838 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2839 if (switchcnt != oldswitchcnt)
2841 tdq->tdq_switchcnt++;
2847 * A CPU is entering for the first time or a thread is exiting.
2850 sched_throw(struct thread *td)
2852 struct thread *newtd;
2857 /* Correct spinlock nesting and acquire the correct lock. */
2860 PCPU_SET(switchtime, cpu_ticks());
2861 PCPU_SET(switchticks, ticks);
2863 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2864 tdq_load_rem(tdq, td);
2865 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2866 td->td_lastcpu = td->td_oncpu;
2867 td->td_oncpu = NOCPU;
2869 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2870 newtd = choosethread();
2871 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2872 cpu_throw(td, newtd); /* doesn't return */
2876 * This is called from fork_exit(). Just acquire the correct locks and
2877 * let fork do the rest of the work.
2880 sched_fork_exit(struct thread *td)
2886 * Finish setting up thread glue so that it begins execution in a
2887 * non-nested critical section with the scheduler lock held.
2889 cpuid = PCPU_GET(cpuid);
2890 tdq = TDQ_CPU(cpuid);
2891 if (TD_IS_IDLETHREAD(td))
2892 td->td_lock = TDQ_LOCKPTR(tdq);
2893 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2894 td->td_oncpu = cpuid;
2895 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2896 lock_profile_obtain_lock_success(
2897 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2899 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2900 "prio:%d", td->td_priority);
2901 SDT_PROBE0(sched, , , on__cpu);
2905 * Create on first use to catch odd startup conditons.
2908 sched_tdname(struct thread *td)
2911 struct td_sched *ts;
2913 ts = td_get_sched(td);
2914 if (ts->ts_name[0] == '\0')
2915 snprintf(ts->ts_name, sizeof(ts->ts_name),
2916 "%s tid %d", td->td_name, td->td_tid);
2917 return (ts->ts_name);
2919 return (td->td_name);
2925 sched_clear_tdname(struct thread *td)
2927 struct td_sched *ts;
2929 ts = td_get_sched(td);
2930 ts->ts_name[0] = '\0';
2937 * Build the CPU topology dump string. Is recursively called to collect
2938 * the topology tree.
2941 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2944 char cpusetbuf[CPUSETBUFSIZ];
2947 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2948 "", 1 + indent / 2, cg->cg_level);
2949 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2950 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2952 for (i = 0; i < MAXCPU; i++) {
2953 if (CPU_ISSET(i, &cg->cg_mask)) {
2955 sbuf_printf(sb, ", ");
2958 sbuf_printf(sb, "%d", i);
2961 sbuf_printf(sb, "</cpu>\n");
2963 if (cg->cg_flags != 0) {
2964 sbuf_printf(sb, "%*s <flags>", indent, "");
2965 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2966 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2967 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2968 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2969 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2970 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2971 sbuf_printf(sb, "</flags>\n");
2974 if (cg->cg_children > 0) {
2975 sbuf_printf(sb, "%*s <children>\n", indent, "");
2976 for (i = 0; i < cg->cg_children; i++)
2977 sysctl_kern_sched_topology_spec_internal(sb,
2978 &cg->cg_child[i], indent+2);
2979 sbuf_printf(sb, "%*s </children>\n", indent, "");
2981 sbuf_printf(sb, "%*s</group>\n", indent, "");
2986 * Sysctl handler for retrieving topology dump. It's a wrapper for
2987 * the recursive sysctl_kern_smp_topology_spec_internal().
2990 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2995 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2997 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3001 sbuf_printf(topo, "<groups>\n");
3002 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3003 sbuf_printf(topo, "</groups>\n");
3006 err = sbuf_finish(topo);
3015 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3017 int error, new_val, period;
3019 period = 1000000 / realstathz;
3020 new_val = period * sched_slice;
3021 error = sysctl_handle_int(oidp, &new_val, 0, req);
3022 if (error != 0 || req->newptr == NULL)
3026 sched_slice = imax(1, (new_val + period / 2) / period);
3027 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3028 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3033 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3034 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3036 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3037 NULL, 0, sysctl_kern_quantum, "I",
3038 "Quantum for timeshare threads in microseconds");
3039 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3040 "Quantum for timeshare threads in stathz ticks");
3041 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3042 "Interactivity score threshold");
3043 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3045 "Maximal (lowest) priority for preemption");
3046 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3047 "Assign static kernel priorities to sleeping threads");
3048 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3049 "Number of times idle thread will spin waiting for new work");
3050 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3051 &sched_idlespinthresh, 0,
3052 "Threshold before we will permit idle thread spinning");
3054 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3055 "Number of hz ticks to keep thread affinity for");
3056 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3057 "Enables the long-term load balancer");
3058 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3059 &balance_interval, 0,
3060 "Average period in stathz ticks to run the long-term balancer");
3061 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3062 "Attempts to steal work from other cores before idling");
3063 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3064 "Minimum load on remote CPU before we'll steal");
3065 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3066 0, "Topological distance limit for stealing threads in sched_switch()");
3067 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3068 "Always run the stealer from the idle thread");
3069 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3070 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3071 "XML dump of detected CPU topology");
3074 /* ps compat. All cpu percentages from ULE are weighted. */
3075 static int ccpu = 0;
3076 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");