80#define KTR_ULE 0
81
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"))
85
86/*
87 * Thread scheduler specific section. All fields are protected
88 * by the thread lock.
89 */
90struct td_sched {
91 struct runq *ts_runq; /* Run-queue we're queued on. */
92 short ts_flags; /* TSF_* flags. */
93 u_char 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 */
101#ifdef KTR
102 char ts_name[TS_NAME_LEN];
103#endif
104};
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. */
108
109static struct td_sched td_sched0;
110
111#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
112#define THREAD_CAN_SCHED(td, cpu) \
113 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114
115/*
116 * Priority ranges used for interactive and non-interactive timeshare
117 * threads. The timeshare priorities are split up into four ranges.
118 * The first range handles interactive threads. The last three ranges
119 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
120 * ranges supporting nice values.
121 */
122#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
123#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
124#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
125
126#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
127#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
128#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
129#define PRI_MAX_BATCH PRI_MAX_TIMESHARE
130
131/*
132 * Cpu percentage computation macros and defines.
133 *
134 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
135 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
136 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
137 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
138 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
139 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
140 */
141#define SCHED_TICK_SECS 10
142#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
143#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
144#define SCHED_TICK_SHIFT 10
145#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
146#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
147
148/*
149 * These macros determine priorities for non-interactive threads. They are
150 * assigned a priority based on their recent cpu utilization as expressed
151 * by the ratio of ticks to the tick total. NHALF priorities at the start
152 * and end of the MIN to MAX timeshare range are only reachable with negative
153 * or positive nice respectively.
154 *
155 * PRI_RANGE: Priority range for utilization dependent priorities.
156 * PRI_NRESV: Number of nice values.
157 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
158 * PRI_NICE: Determines the part of the priority inherited from nice.
159 */
160#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
161#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
162#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
163#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
164#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
165#define SCHED_PRI_TICKS(ts) \
166 (SCHED_TICK_HZ((ts)) / \
167 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
168#define SCHED_PRI_NICE(nice) (nice)
169
170/*
171 * These determine the interactivity of a process. Interactivity differs from
172 * cpu utilization in that it expresses the voluntary time slept vs time ran
173 * while cpu utilization includes all time not running. This more accurately
174 * models the intent of the thread.
175 *
176 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
177 * before throttling back.
178 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
179 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
180 * INTERACT_THRESH: Threshold for placement on the current runq.
181 */
182#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
183#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
184#define SCHED_INTERACT_MAX (100)
185#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
186#define SCHED_INTERACT_THRESH (30)
187
188/*
189 * These parameters determine the slice behavior for batch work.
190 */
191#define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
192#define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
193
194/* Flags kept in td_flags. */
195#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
196
197/*
198 * tickincr: Converts a stathz tick into a hz domain scaled by
199 * the shift factor. Without the shift the error rate
200 * due to rounding would be unacceptably high.
201 * realstathz: stathz is sometimes 0 and run off of hz.
202 * sched_slice: Runtime of each thread before rescheduling.
203 * preempt_thresh: Priority threshold for preemption and remote IPIs.
204 */
205static int sched_interact = SCHED_INTERACT_THRESH;
206static int tickincr = 8 << SCHED_TICK_SHIFT;
207static int realstathz = 127; /* reset during boot. */
208static int sched_slice = 10; /* reset during boot. */
209static int sched_slice_min = 1; /* reset during boot. */
210#ifdef PREEMPTION
211#ifdef FULL_PREEMPTION
212static int preempt_thresh = PRI_MAX_IDLE;
213#else
214static int preempt_thresh = PRI_MIN_KERN;
215#endif
216#else
217static int preempt_thresh = 0;
218#endif
219static int static_boost = PRI_MIN_BATCH;
220static int sched_idlespins = 10000;
221static int sched_idlespinthresh = -1;
222
223/*
224 * tdq - per processor runqs and statistics. All fields are protected by the
225 * tdq_lock. The load and lowpri may be accessed without to avoid excess
226 * locking in sched_pickcpu();
227 */
228struct tdq {
229 /*
230 * Ordered to improve efficiency of cpu_search() and switch().
231 * tdq_lock is padded to avoid false sharing with tdq_load and
232 * tdq_cpu_idle.
233 */
234 struct mtx_padalign tdq_lock; /* run queue lock. */
235 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
236 volatile int tdq_load; /* Aggregate load. */
237 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
238 int tdq_sysload; /* For loadavg, !ITHD load. */
239 int tdq_transferable; /* Transferable thread count. */
240 short tdq_switchcnt; /* Switches this tick. */
241 short tdq_oldswitchcnt; /* Switches last tick. */
242 u_char tdq_lowpri; /* Lowest priority thread. */
243 u_char tdq_ipipending; /* IPI pending. */
244 u_char tdq_idx; /* Current insert index. */
245 u_char tdq_ridx; /* Current removal index. */
246 struct runq tdq_realtime; /* real-time run queue. */
247 struct runq tdq_timeshare; /* timeshare run queue. */
248 struct runq tdq_idle; /* Queue of IDLE threads. */
249 char tdq_name[TDQ_NAME_LEN];
250#ifdef KTR
251 char tdq_loadname[TDQ_LOADNAME_LEN];
252#endif
253} __aligned(64);
254
255/* Idle thread states and config. */
256#define TDQ_RUNNING 1
257#define TDQ_IDLE 2
258
259#ifdef SMP
260struct cpu_group *cpu_top; /* CPU topology */
261
262#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
263#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
264
265/*
266 * Run-time tunables.
267 */
268static int rebalance = 1;
269static int balance_interval = 128; /* Default set in sched_initticks(). */
270static int affinity;
271static int steal_idle = 1;
272static int steal_thresh = 2;
273
274/*
275 * One thread queue per processor.
276 */
277static struct tdq tdq_cpu[MAXCPU];
278static struct tdq *balance_tdq;
279static int balance_ticks;
280static DPCPU_DEFINE(uint32_t, randomval);
281
282#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
283#define TDQ_CPU(x) (&tdq_cpu[(x)])
284#define TDQ_ID(x) ((int)((x) - tdq_cpu))
285#else /* !SMP */
286static struct tdq tdq_cpu;
287
288#define TDQ_ID(x) (0)
289#define TDQ_SELF() (&tdq_cpu)
290#define TDQ_CPU(x) (&tdq_cpu)
291#endif
292
293#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
294#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
295#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
296#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
297#define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
298
299static void sched_priority(struct thread *);
300static void sched_thread_priority(struct thread *, u_char);
301static int sched_interact_score(struct thread *);
302static void sched_interact_update(struct thread *);
303static void sched_interact_fork(struct thread *);
304static void sched_pctcpu_update(struct td_sched *, int);
305
306/* Operations on per processor queues */
307static struct thread *tdq_choose(struct tdq *);
308static void tdq_setup(struct tdq *);
309static void tdq_load_add(struct tdq *, struct thread *);
310static void tdq_load_rem(struct tdq *, struct thread *);
311static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
312static __inline void tdq_runq_rem(struct tdq *, struct thread *);
313static inline int sched_shouldpreempt(int, int, int);
314void tdq_print(int cpu);
315static void runq_print(struct runq *rq);
316static void tdq_add(struct tdq *, struct thread *, int);
317#ifdef SMP
318static int tdq_move(struct tdq *, struct tdq *);
319static int tdq_idled(struct tdq *);
320static void tdq_notify(struct tdq *, struct thread *);
321static struct thread *tdq_steal(struct tdq *, int);
322static struct thread *runq_steal(struct runq *, int);
323static int sched_pickcpu(struct thread *, int);
324static void sched_balance(void);
325static int sched_balance_pair(struct tdq *, struct tdq *);
326static inline struct tdq *sched_setcpu(struct thread *, int, int);
327static inline void thread_unblock_switch(struct thread *, struct mtx *);
328static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
329static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
330static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
331 struct cpu_group *cg, int indent);
332#endif
333
334static void sched_setup(void *dummy);
335SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
336
337static void sched_initticks(void *dummy);
338SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
339 NULL);
340
341SDT_PROVIDER_DEFINE(sched);
342
343SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
344 "struct proc *", "uint8_t");
345SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
346 "struct proc *", "void *");
347SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
348 "struct proc *", "void *", "int");
349SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
350 "struct proc *", "uint8_t", "struct thread *");
351SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
352SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
353 "struct proc *");
354SDT_PROBE_DEFINE(sched, , , on__cpu);
355SDT_PROBE_DEFINE(sched, , , remain__cpu);
356SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
357 "struct proc *");
358
359/*
360 * Print the threads waiting on a run-queue.
361 */
362static void
363runq_print(struct runq *rq)
364{
365 struct rqhead *rqh;
366 struct thread *td;
367 int pri;
368 int j;
369 int i;
370
371 for (i = 0; i < RQB_LEN; i++) {
372 printf("\t\trunq bits %d 0x%zx\n",
373 i, rq->rq_status.rqb_bits[i]);
374 for (j = 0; j < RQB_BPW; j++)
375 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
376 pri = j + (i << RQB_L2BPW);
377 rqh = &rq->rq_queues[pri];
378 TAILQ_FOREACH(td, rqh, td_runq) {
379 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
380 td, td->td_name, td->td_priority,
381 td->td_rqindex, pri);
382 }
383 }
384 }
385}
386
387/*
388 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
389 */
390void
391tdq_print(int cpu)
392{
393 struct tdq *tdq;
394
395 tdq = TDQ_CPU(cpu);
396
397 printf("tdq %d:\n", TDQ_ID(tdq));
398 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
399 printf("\tLock name: %s\n", tdq->tdq_name);
400 printf("\tload: %d\n", tdq->tdq_load);
401 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
402 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
403 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
404 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
405 printf("\tload transferable: %d\n", tdq->tdq_transferable);
406 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
407 printf("\trealtime runq:\n");
408 runq_print(&tdq->tdq_realtime);
409 printf("\ttimeshare runq:\n");
410 runq_print(&tdq->tdq_timeshare);
411 printf("\tidle runq:\n");
412 runq_print(&tdq->tdq_idle);
413}
414
415static inline int
416sched_shouldpreempt(int pri, int cpri, int remote)
417{
418 /*
419 * If the new priority is not better than the current priority there is
420 * nothing to do.
421 */
422 if (pri >= cpri)
423 return (0);
424 /*
425 * Always preempt idle.
426 */
427 if (cpri >= PRI_MIN_IDLE)
428 return (1);
429 /*
430 * If preemption is disabled don't preempt others.
431 */
432 if (preempt_thresh == 0)
433 return (0);
434 /*
435 * Preempt if we exceed the threshold.
436 */
437 if (pri <= preempt_thresh)
438 return (1);
439 /*
440 * If we're interactive or better and there is non-interactive
441 * or worse running preempt only remote processors.
442 */
443 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
444 return (1);
445 return (0);
446}
447
448/*
449 * Add a thread to the actual run-queue. Keeps transferable counts up to
450 * date with what is actually on the run-queue. Selects the correct
451 * queue position for timeshare threads.
452 */
453static __inline void
454tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
455{
456 struct td_sched *ts;
457 u_char pri;
458
459 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
460 THREAD_LOCK_ASSERT(td, MA_OWNED);
461
462 pri = td->td_priority;
463 ts = td->td_sched;
464 TD_SET_RUNQ(td);
465 if (THREAD_CAN_MIGRATE(td)) {
466 tdq->tdq_transferable++;
467 ts->ts_flags |= TSF_XFERABLE;
468 }
469 if (pri < PRI_MIN_BATCH) {
470 ts->ts_runq = &tdq->tdq_realtime;
471 } else if (pri <= PRI_MAX_BATCH) {
472 ts->ts_runq = &tdq->tdq_timeshare;
473 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
474 ("Invalid priority %d on timeshare runq", pri));
475 /*
476 * This queue contains only priorities between MIN and MAX
477 * realtime. Use the whole queue to represent these values.
478 */
479 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
480 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
481 pri = (pri + tdq->tdq_idx) % RQ_NQS;
482 /*
483 * This effectively shortens the queue by one so we
484 * can have a one slot difference between idx and
485 * ridx while we wait for threads to drain.
486 */
487 if (tdq->tdq_ridx != tdq->tdq_idx &&
488 pri == tdq->tdq_ridx)
489 pri = (unsigned char)(pri - 1) % RQ_NQS;
490 } else
491 pri = tdq->tdq_ridx;
492 runq_add_pri(ts->ts_runq, td, pri, flags);
493 return;
494 } else
495 ts->ts_runq = &tdq->tdq_idle;
496 runq_add(ts->ts_runq, td, flags);
497}
498
499/*
500 * Remove a thread from a run-queue. This typically happens when a thread
501 * is selected to run. Running threads are not on the queue and the
502 * transferable count does not reflect them.
503 */
504static __inline void
505tdq_runq_rem(struct tdq *tdq, struct thread *td)
506{
507 struct td_sched *ts;
508
509 ts = td->td_sched;
510 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
511 KASSERT(ts->ts_runq != NULL,
512 ("tdq_runq_remove: thread %p null ts_runq", td));
513 if (ts->ts_flags & TSF_XFERABLE) {
514 tdq->tdq_transferable--;
515 ts->ts_flags &= ~TSF_XFERABLE;
516 }
517 if (ts->ts_runq == &tdq->tdq_timeshare) {
518 if (tdq->tdq_idx != tdq->tdq_ridx)
519 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
520 else
521 runq_remove_idx(ts->ts_runq, td, NULL);
522 } else
523 runq_remove(ts->ts_runq, td);
524}
525
526/*
527 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
528 * for this thread to the referenced thread queue.
529 */
530static void
531tdq_load_add(struct tdq *tdq, struct thread *td)
532{
533
534 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
535 THREAD_LOCK_ASSERT(td, MA_OWNED);
536
537 tdq->tdq_load++;
538 if ((td->td_flags & TDF_NOLOAD) == 0)
539 tdq->tdq_sysload++;
540 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
541 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
542}
543
544/*
545 * Remove the load from a thread that is transitioning to a sleep state or
546 * exiting.
547 */
548static void
549tdq_load_rem(struct tdq *tdq, struct thread *td)
550{
551
552 THREAD_LOCK_ASSERT(td, MA_OWNED);
553 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
554 KASSERT(tdq->tdq_load != 0,
555 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
556
557 tdq->tdq_load--;
558 if ((td->td_flags & TDF_NOLOAD) == 0)
559 tdq->tdq_sysload--;
560 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
561 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
562}
563
564/*
565 * Bound timeshare latency by decreasing slice size as load increases. We
566 * consider the maximum latency as the sum of the threads waiting to run
567 * aside from curthread and target no more than sched_slice latency but
568 * no less than sched_slice_min runtime.
569 */
570static inline int
571tdq_slice(struct tdq *tdq)
572{
573 int load;
574
575 /*
576 * It is safe to use sys_load here because this is called from
577 * contexts where timeshare threads are running and so there
578 * cannot be higher priority load in the system.
579 */
580 load = tdq->tdq_sysload - 1;
581 if (load >= SCHED_SLICE_MIN_DIVISOR)
582 return (sched_slice_min);
583 if (load <= 1)
584 return (sched_slice);
585 return (sched_slice / load);
586}
587
588/*
589 * Set lowpri to its exact value by searching the run-queue and
590 * evaluating curthread. curthread may be passed as an optimization.
591 */
592static void
593tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
594{
595 struct thread *td;
596
597 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
598 if (ctd == NULL)
599 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
600 td = tdq_choose(tdq);
601 if (td == NULL || td->td_priority > ctd->td_priority)
602 tdq->tdq_lowpri = ctd->td_priority;
603 else
604 tdq->tdq_lowpri = td->td_priority;
605}
606
607#ifdef SMP
608struct cpu_search {
609 cpuset_t cs_mask;
610 u_int cs_prefer;
611 int cs_pri; /* Min priority for low. */
612 int cs_limit; /* Max load for low, min load for high. */
613 int cs_cpu;
614 int cs_load;
615};
616
617#define CPU_SEARCH_LOWEST 0x1
618#define CPU_SEARCH_HIGHEST 0x2
619#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
620
621#define CPUSET_FOREACH(cpu, mask) \
622 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
623 if (CPU_ISSET(cpu, &mask))
624
625static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
626 struct cpu_search *high, const int match);
627int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
628int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
629int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
630 struct cpu_search *high);
631
632/*
633 * Search the tree of cpu_groups for the lowest or highest loaded cpu
634 * according to the match argument. This routine actually compares the
635 * load on all paths through the tree and finds the least loaded cpu on
636 * the least loaded path, which may differ from the least loaded cpu in
637 * the system. This balances work among caches and busses.
638 *
639 * This inline is instantiated in three forms below using constants for the
640 * match argument. It is reduced to the minimum set for each case. It is
641 * also recursive to the depth of the tree.
642 */
643static __inline int
644cpu_search(const struct cpu_group *cg, struct cpu_search *low,
645 struct cpu_search *high, const int match)
646{
647 struct cpu_search lgroup;
648 struct cpu_search hgroup;
649 cpuset_t cpumask;
650 struct cpu_group *child;
651 struct tdq *tdq;
652 int cpu, i, hload, lload, load, total, rnd, *rndptr;
653
654 total = 0;
655 cpumask = cg->cg_mask;
656 if (match & CPU_SEARCH_LOWEST) {
657 lload = INT_MAX;
658 lgroup = *low;
659 }
660 if (match & CPU_SEARCH_HIGHEST) {
661 hload = INT_MIN;
662 hgroup = *high;
663 }
664
665 /* Iterate through the child CPU groups and then remaining CPUs. */
666 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
667 if (i == 0) {
668#ifdef HAVE_INLINE_FFSL
669 cpu = CPU_FFS(&cpumask) - 1;
670#else
671 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
672 cpu--;
673#endif
674 if (cpu < 0)
675 break;
676 child = NULL;
677 } else
678 child = &cg->cg_child[i - 1];
679
680 if (match & CPU_SEARCH_LOWEST)
681 lgroup.cs_cpu = -1;
682 if (match & CPU_SEARCH_HIGHEST)
683 hgroup.cs_cpu = -1;
684 if (child) { /* Handle child CPU group. */
685 CPU_NAND(&cpumask, &child->cg_mask);
686 switch (match) {
687 case CPU_SEARCH_LOWEST:
688 load = cpu_search_lowest(child, &lgroup);
689 break;
690 case CPU_SEARCH_HIGHEST:
691 load = cpu_search_highest(child, &hgroup);
692 break;
693 case CPU_SEARCH_BOTH:
694 load = cpu_search_both(child, &lgroup, &hgroup);
695 break;
696 }
697 } else { /* Handle child CPU. */
698 CPU_CLR(cpu, &cpumask);
699 tdq = TDQ_CPU(cpu);
700 load = tdq->tdq_load * 256;
701 rndptr = DPCPU_PTR(randomval);
702 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
703 if (match & CPU_SEARCH_LOWEST) {
704 if (cpu == low->cs_prefer)
705 load -= 64;
706 /* If that CPU is allowed and get data. */
707 if (tdq->tdq_lowpri > lgroup.cs_pri &&
708 tdq->tdq_load <= lgroup.cs_limit &&
709 CPU_ISSET(cpu, &lgroup.cs_mask)) {
710 lgroup.cs_cpu = cpu;
711 lgroup.cs_load = load - rnd;
712 }
713 }
714 if (match & CPU_SEARCH_HIGHEST)
715 if (tdq->tdq_load >= hgroup.cs_limit &&
716 tdq->tdq_transferable &&
717 CPU_ISSET(cpu, &hgroup.cs_mask)) {
718 hgroup.cs_cpu = cpu;
719 hgroup.cs_load = load - rnd;
720 }
721 }
722 total += load;
723
724 /* We have info about child item. Compare it. */
725 if (match & CPU_SEARCH_LOWEST) {
726 if (lgroup.cs_cpu >= 0 &&
727 (load < lload ||
728 (load == lload && lgroup.cs_load < low->cs_load))) {
729 lload = load;
730 low->cs_cpu = lgroup.cs_cpu;
731 low->cs_load = lgroup.cs_load;
732 }
733 }
734 if (match & CPU_SEARCH_HIGHEST)
735 if (hgroup.cs_cpu >= 0 &&
736 (load > hload ||
737 (load == hload && hgroup.cs_load > high->cs_load))) {
738 hload = load;
739 high->cs_cpu = hgroup.cs_cpu;
740 high->cs_load = hgroup.cs_load;
741 }
742 if (child) {
743 i--;
744 if (i == 0 && CPU_EMPTY(&cpumask))
745 break;
746 }
747#ifndef HAVE_INLINE_FFSL
748 else
749 cpu--;
750#endif
751 }
752 return (total);
753}
754
755/*
756 * cpu_search instantiations must pass constants to maintain the inline
757 * optimization.
758 */
759int
760cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
761{
762 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
763}
764
765int
766cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
767{
768 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
769}
770
771int
772cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
773 struct cpu_search *high)
774{
775 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
776}
777
778/*
779 * Find the cpu with the least load via the least loaded path that has a
780 * lowpri greater than pri pri. A pri of -1 indicates any priority is
781 * acceptable.
782 */
783static inline int
784sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
785 int prefer)
786{
787 struct cpu_search low;
788
789 low.cs_cpu = -1;
790 low.cs_prefer = prefer;
791 low.cs_mask = mask;
792 low.cs_pri = pri;
793 low.cs_limit = maxload;
794 cpu_search_lowest(cg, &low);
795 return low.cs_cpu;
796}
797
798/*
799 * Find the cpu with the highest load via the highest loaded path.
800 */
801static inline int
802sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
803{
804 struct cpu_search high;
805
806 high.cs_cpu = -1;
807 high.cs_mask = mask;
808 high.cs_limit = minload;
809 cpu_search_highest(cg, &high);
810 return high.cs_cpu;
811}
812
813static void
814sched_balance_group(struct cpu_group *cg)
815{
816 cpuset_t hmask, lmask;
817 int high, low, anylow;
818
819 CPU_FILL(&hmask);
820 for (;;) {
821 high = sched_highest(cg, hmask, 1);
822 /* Stop if there is no more CPU with transferrable threads. */
823 if (high == -1)
824 break;
825 CPU_CLR(high, &hmask);
826 CPU_COPY(&hmask, &lmask);
827 /* Stop if there is no more CPU left for low. */
828 if (CPU_EMPTY(&lmask))
829 break;
830 anylow = 1;
831nextlow:
832 low = sched_lowest(cg, lmask, -1,
833 TDQ_CPU(high)->tdq_load - 1, high);
834 /* Stop if we looked well and found no less loaded CPU. */
835 if (anylow && low == -1)
836 break;
837 /* Go to next high if we found no less loaded CPU. */
838 if (low == -1)
839 continue;
840 /* Transfer thread from high to low. */
841 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
842 /* CPU that got thread can no longer be a donor. */
843 CPU_CLR(low, &hmask);
844 } else {
845 /*
846 * If failed, then there is no threads on high
847 * that can run on this low. Drop low from low
848 * mask and look for different one.
849 */
850 CPU_CLR(low, &lmask);
851 anylow = 0;
852 goto nextlow;
853 }
854 }
855}
856
857static void
858sched_balance(void)
859{
860 struct tdq *tdq;
861
862 /*
863 * Select a random time between .5 * balance_interval and
864 * 1.5 * balance_interval.
865 */
866 balance_ticks = max(balance_interval / 2, 1);
867 balance_ticks += random() % balance_interval;
868 if (smp_started == 0 || rebalance == 0)
869 return;
870 tdq = TDQ_SELF();
871 TDQ_UNLOCK(tdq);
872 sched_balance_group(cpu_top);
873 TDQ_LOCK(tdq);
874}
875
876/*
877 * Lock two thread queues using their address to maintain lock order.
878 */
879static void
880tdq_lock_pair(struct tdq *one, struct tdq *two)
881{
882 if (one < two) {
883 TDQ_LOCK(one);
884 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
885 } else {
886 TDQ_LOCK(two);
887 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
888 }
889}
890
891/*
892 * Unlock two thread queues. Order is not important here.
893 */
894static void
895tdq_unlock_pair(struct tdq *one, struct tdq *two)
896{
897 TDQ_UNLOCK(one);
898 TDQ_UNLOCK(two);
899}
900
901/*
902 * Transfer load between two imbalanced thread queues.
903 */
904static int
905sched_balance_pair(struct tdq *high, struct tdq *low)
906{
907 int moved;
908 int cpu;
909
910 tdq_lock_pair(high, low);
911 moved = 0;
912 /*
913 * Determine what the imbalance is and then adjust that to how many
914 * threads we actually have to give up (transferable).
915 */
916 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
917 (moved = tdq_move(high, low)) > 0) {
918 /*
919 * In case the target isn't the current cpu IPI it to force a
920 * reschedule with the new workload.
921 */
922 cpu = TDQ_ID(low);
923 if (cpu != PCPU_GET(cpuid))
924 ipi_cpu(cpu, IPI_PREEMPT);
925 }
926 tdq_unlock_pair(high, low);
927 return (moved);
928}
929
930/*
931 * Move a thread from one thread queue to another.
932 */
933static int
934tdq_move(struct tdq *from, struct tdq *to)
935{
936 struct td_sched *ts;
937 struct thread *td;
938 struct tdq *tdq;
939 int cpu;
940
941 TDQ_LOCK_ASSERT(from, MA_OWNED);
942 TDQ_LOCK_ASSERT(to, MA_OWNED);
943
944 tdq = from;
945 cpu = TDQ_ID(to);
946 td = tdq_steal(tdq, cpu);
947 if (td == NULL)
948 return (0);
949 ts = td->td_sched;
950 /*
951 * Although the run queue is locked the thread may be blocked. Lock
952 * it to clear this and acquire the run-queue lock.
953 */
954 thread_lock(td);
955 /* Drop recursive lock on from acquired via thread_lock(). */
956 TDQ_UNLOCK(from);
957 sched_rem(td);
958 ts->ts_cpu = cpu;
959 td->td_lock = TDQ_LOCKPTR(to);
960 tdq_add(to, td, SRQ_YIELDING);
961 return (1);
962}
963
964/*
965 * This tdq has idled. Try to steal a thread from another cpu and switch
966 * to it.
967 */
968static int
969tdq_idled(struct tdq *tdq)
970{
971 struct cpu_group *cg;
972 struct tdq *steal;
973 cpuset_t mask;
974 int thresh;
975 int cpu;
976
977 if (smp_started == 0 || steal_idle == 0)
978 return (1);
979 CPU_FILL(&mask);
980 CPU_CLR(PCPU_GET(cpuid), &mask);
981 /* We don't want to be preempted while we're iterating. */
982 spinlock_enter();
983 for (cg = tdq->tdq_cg; cg != NULL; ) {
984 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
985 thresh = steal_thresh;
986 else
987 thresh = 1;
988 cpu = sched_highest(cg, mask, thresh);
989 if (cpu == -1) {
990 cg = cg->cg_parent;
991 continue;
992 }
993 steal = TDQ_CPU(cpu);
994 CPU_CLR(cpu, &mask);
995 tdq_lock_pair(tdq, steal);
996 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
997 tdq_unlock_pair(tdq, steal);
998 continue;
999 }
1000 /*
1001 * If a thread was added while interrupts were disabled don't
1002 * steal one here. If we fail to acquire one due to affinity
1003 * restrictions loop again with this cpu removed from the
1004 * set.
1005 */
1006 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1007 tdq_unlock_pair(tdq, steal);
1008 continue;
1009 }
1010 spinlock_exit();
1011 TDQ_UNLOCK(steal);
1012 mi_switch(SW_VOL | SWT_IDLE, NULL);
1013 thread_unlock(curthread);
1014
1015 return (0);
1016 }
1017 spinlock_exit();
1018 return (1);
1019}
1020
1021/*
1022 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1023 */
1024static void
1025tdq_notify(struct tdq *tdq, struct thread *td)
1026{
1027 struct thread *ctd;
1028 int pri;
1029 int cpu;
1030
1031 if (tdq->tdq_ipipending)
1032 return;
1033 cpu = td->td_sched->ts_cpu;
1034 pri = td->td_priority;
1035 ctd = pcpu_find(cpu)->pc_curthread;
1036 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1037 return;
1038 if (TD_IS_IDLETHREAD(ctd)) {
1039 /*
1040 * If the MD code has an idle wakeup routine try that before
1041 * falling back to IPI.
1042 */
1043 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1044 return;
1045 }
1046 tdq->tdq_ipipending = 1;
1047 ipi_cpu(cpu, IPI_PREEMPT);
1048}
1049
1050/*
1051 * Steals load from a timeshare queue. Honors the rotating queue head
1052 * index.
1053 */
1054static struct thread *
1055runq_steal_from(struct runq *rq, int cpu, u_char start)
1056{
1057 struct rqbits *rqb;
1058 struct rqhead *rqh;
1059 struct thread *td, *first;
1060 int bit;
1061 int pri;
1062 int i;
1063
1064 rqb = &rq->rq_status;
1065 bit = start & (RQB_BPW -1);
1066 pri = 0;
1067 first = NULL;
1068again:
1069 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1070 if (rqb->rqb_bits[i] == 0)
1071 continue;
1072 if (bit != 0) {
1073 for (pri = bit; pri < RQB_BPW; pri++)
1074 if (rqb->rqb_bits[i] & (1ul << pri))
1075 break;
1076 if (pri >= RQB_BPW)
1077 continue;
1078 } else
1079 pri = RQB_FFS(rqb->rqb_bits[i]);
1080 pri += (i << RQB_L2BPW);
1081 rqh = &rq->rq_queues[pri];
1082 TAILQ_FOREACH(td, rqh, td_runq) {
1083 if (first && THREAD_CAN_MIGRATE(td) &&
1084 THREAD_CAN_SCHED(td, cpu))
1085 return (td);
1086 first = td;
1087 }
1088 }
1089 if (start != 0) {
1090 start = 0;
1091 goto again;
1092 }
1093
1094 if (first && THREAD_CAN_MIGRATE(first) &&
1095 THREAD_CAN_SCHED(first, cpu))
1096 return (first);
1097 return (NULL);
1098}
1099
1100/*
1101 * Steals load from a standard linear queue.
1102 */
1103static struct thread *
1104runq_steal(struct runq *rq, int cpu)
1105{
1106 struct rqhead *rqh;
1107 struct rqbits *rqb;
1108 struct thread *td;
1109 int word;
1110 int bit;
1111
1112 rqb = &rq->rq_status;
1113 for (word = 0; word < RQB_LEN; word++) {
1114 if (rqb->rqb_bits[word] == 0)
1115 continue;
1116 for (bit = 0; bit < RQB_BPW; bit++) {
1117 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1118 continue;
1119 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1120 TAILQ_FOREACH(td, rqh, td_runq)
1121 if (THREAD_CAN_MIGRATE(td) &&
1122 THREAD_CAN_SCHED(td, cpu))
1123 return (td);
1124 }
1125 }
1126 return (NULL);
1127}
1128
1129/*
1130 * Attempt to steal a thread in priority order from a thread queue.
1131 */
1132static struct thread *
1133tdq_steal(struct tdq *tdq, int cpu)
1134{
1135 struct thread *td;
1136
1137 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1138 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1139 return (td);
1140 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1141 cpu, tdq->tdq_ridx)) != NULL)
1142 return (td);
1143 return (runq_steal(&tdq->tdq_idle, cpu));
1144}
1145
1146/*
1147 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1148 * current lock and returns with the assigned queue locked.
1149 */
1150static inline struct tdq *
1151sched_setcpu(struct thread *td, int cpu, int flags)
1152{
1153
1154 struct tdq *tdq;
1155
1156 THREAD_LOCK_ASSERT(td, MA_OWNED);
1157 tdq = TDQ_CPU(cpu);
1158 td->td_sched->ts_cpu = cpu;
1159 /*
1160 * If the lock matches just return the queue.
1161 */
1162 if (td->td_lock == TDQ_LOCKPTR(tdq))
1163 return (tdq);
1164#ifdef notyet
1165 /*
1166 * If the thread isn't running its lockptr is a
1167 * turnstile or a sleepqueue. We can just lock_set without
1168 * blocking.
1169 */
1170 if (TD_CAN_RUN(td)) {
1171 TDQ_LOCK(tdq);
1172 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1173 return (tdq);
1174 }
1175#endif
1176 /*
1177 * The hard case, migration, we need to block the thread first to
1178 * prevent order reversals with other cpus locks.
1179 */
1180 spinlock_enter();
1181 thread_lock_block(td);
1182 TDQ_LOCK(tdq);
1183 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1184 spinlock_exit();
1185 return (tdq);
1186}
1187
1188SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1189SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1190SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1191SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1192SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1193SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1194
1195static int
1196sched_pickcpu(struct thread *td, int flags)
1197{
1198 struct cpu_group *cg, *ccg;
1199 struct td_sched *ts;
1200 struct tdq *tdq;
1201 cpuset_t mask;
1202 int cpu, pri, self;
1203
1204 self = PCPU_GET(cpuid);
1205 ts = td->td_sched;
1206 if (smp_started == 0)
1207 return (self);
1208 /*
1209 * Don't migrate a running thread from sched_switch().
1210 */
1211 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1212 return (ts->ts_cpu);
1213 /*
1214 * Prefer to run interrupt threads on the processors that generate
1215 * the interrupt.
1216 */
1217 pri = td->td_priority;
1218 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1219 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1220 SCHED_STAT_INC(pickcpu_intrbind);
1221 ts->ts_cpu = self;
1222 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1223 SCHED_STAT_INC(pickcpu_affinity);
1224 return (ts->ts_cpu);
1225 }
1226 }
1227 /*
1228 * If the thread can run on the last cpu and the affinity has not
1229 * expired or it is idle run it there.
1230 */
1231 tdq = TDQ_CPU(ts->ts_cpu);
1232 cg = tdq->tdq_cg;
1233 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1234 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1235 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1236 if (cg->cg_flags & CG_FLAG_THREAD) {
1237 CPUSET_FOREACH(cpu, cg->cg_mask) {
1238 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1239 break;
1240 }
1241 } else
1242 cpu = INT_MAX;
1243 if (cpu > mp_maxid) {
1244 SCHED_STAT_INC(pickcpu_idle_affinity);
1245 return (ts->ts_cpu);
1246 }
1247 }
1248 /*
1249 * Search for the last level cache CPU group in the tree.
1250 * Skip caches with expired affinity time and SMT groups.
1251 * Affinity to higher level caches will be handled less aggressively.
1252 */
1253 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1254 if (cg->cg_flags & CG_FLAG_THREAD)
1255 continue;
1256 if (!SCHED_AFFINITY(ts, cg->cg_level))
1257 continue;
1258 ccg = cg;
1259 }
1260 if (ccg != NULL)
1261 cg = ccg;
1262 cpu = -1;
1263 /* Search the group for the less loaded idle CPU we can run now. */
1264 mask = td->td_cpuset->cs_mask;
1265 if (cg != NULL && cg != cpu_top &&
1266 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1267 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1268 INT_MAX, ts->ts_cpu);
1269 /* Search globally for the less loaded CPU we can run now. */
1270 if (cpu == -1)
1271 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1272 /* Search globally for the less loaded CPU. */
1273 if (cpu == -1)
1274 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1275 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1276 /*
1277 * Compare the lowest loaded cpu to current cpu.
1278 */
1279 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1280 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1281 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1282 SCHED_STAT_INC(pickcpu_local);
1283 cpu = self;
1284 } else
1285 SCHED_STAT_INC(pickcpu_lowest);
1286 if (cpu != ts->ts_cpu)
1287 SCHED_STAT_INC(pickcpu_migration);
1288 return (cpu);
1289}
1290#endif
1291
1292/*
1293 * Pick the highest priority task we have and return it.
1294 */
1295static struct thread *
1296tdq_choose(struct tdq *tdq)
1297{
1298 struct thread *td;
1299
1300 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1301 td = runq_choose(&tdq->tdq_realtime);
1302 if (td != NULL)
1303 return (td);
1304 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1305 if (td != NULL) {
1306 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1307 ("tdq_choose: Invalid priority on timeshare queue %d",
1308 td->td_priority));
1309 return (td);
1310 }
1311 td = runq_choose(&tdq->tdq_idle);
1312 if (td != NULL) {
1313 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1314 ("tdq_choose: Invalid priority on idle queue %d",
1315 td->td_priority));
1316 return (td);
1317 }
1318
1319 return (NULL);
1320}
1321
1322/*
1323 * Initialize a thread queue.
1324 */
1325static void
1326tdq_setup(struct tdq *tdq)
1327{
1328
1329 if (bootverbose)
1330 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1331 runq_init(&tdq->tdq_realtime);
1332 runq_init(&tdq->tdq_timeshare);
1333 runq_init(&tdq->tdq_idle);
1334 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1335 "sched lock %d", (int)TDQ_ID(tdq));
1336 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1337 MTX_SPIN | MTX_RECURSE);
1338#ifdef KTR
1339 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1340 "CPU %d load", (int)TDQ_ID(tdq));
1341#endif
1342}
1343
1344#ifdef SMP
1345static void
1346sched_setup_smp(void)
1347{
1348 struct tdq *tdq;
1349 int i;
1350
1351 cpu_top = smp_topo();
1352 CPU_FOREACH(i) {
1353 tdq = TDQ_CPU(i);
1354 tdq_setup(tdq);
1355 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1356 if (tdq->tdq_cg == NULL)
1357 panic("Can't find cpu group for %d\n", i);
1358 }
1359 balance_tdq = TDQ_SELF();
1360 sched_balance();
1361}
1362#endif
1363
1364/*
1365 * Setup the thread queues and initialize the topology based on MD
1366 * information.
1367 */
1368static void
1369sched_setup(void *dummy)
1370{
1371 struct tdq *tdq;
1372
1373 tdq = TDQ_SELF();
1374#ifdef SMP
1375 sched_setup_smp();
1376#else
1377 tdq_setup(tdq);
1378#endif
1379
1380 /* Add thread0's load since it's running. */
1381 TDQ_LOCK(tdq);
1382 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1383 tdq_load_add(tdq, &thread0);
1384 tdq->tdq_lowpri = thread0.td_priority;
1385 TDQ_UNLOCK(tdq);
1386}
1387
1388/*
1389 * This routine determines time constants after stathz and hz are setup.
1390 */
1391/* ARGSUSED */
1392static void
1393sched_initticks(void *dummy)
1394{
1395 int incr;
1396
1397 realstathz = stathz ? stathz : hz;
1398 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1399 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1400 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1401 realstathz);
1402
1403 /*
1404 * tickincr is shifted out by 10 to avoid rounding errors due to
1405 * hz not being evenly divisible by stathz on all platforms.
1406 */
1407 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1408 /*
1409 * This does not work for values of stathz that are more than
1410 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1411 */
1412 if (incr == 0)
1413 incr = 1;
1414 tickincr = incr;
1415#ifdef SMP
1416 /*
1417 * Set the default balance interval now that we know
1418 * what realstathz is.
1419 */
1420 balance_interval = realstathz;
1421 affinity = SCHED_AFFINITY_DEFAULT;
1422#endif
1423 if (sched_idlespinthresh < 0)
1424 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1425}
1426
1427
1428/*
1429 * This is the core of the interactivity algorithm. Determines a score based
1430 * on past behavior. It is the ratio of sleep time to run time scaled to
1431 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1432 * differs from the cpu usage because it does not account for time spent
1433 * waiting on a run-queue. Would be prettier if we had floating point.
1434 */
1435static int
1436sched_interact_score(struct thread *td)
1437{
1438 struct td_sched *ts;
1439 int div;
1440
1441 ts = td->td_sched;
1442 /*
1443 * The score is only needed if this is likely to be an interactive
1444 * task. Don't go through the expense of computing it if there's
1445 * no chance.
1446 */
1447 if (sched_interact <= SCHED_INTERACT_HALF &&
1448 ts->ts_runtime >= ts->ts_slptime)
1449 return (SCHED_INTERACT_HALF);
1450
1451 if (ts->ts_runtime > ts->ts_slptime) {
1452 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1453 return (SCHED_INTERACT_HALF +
1454 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1455 }
1456 if (ts->ts_slptime > ts->ts_runtime) {
1457 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1458 return (ts->ts_runtime / div);
1459 }
1460 /* runtime == slptime */
1461 if (ts->ts_runtime)
1462 return (SCHED_INTERACT_HALF);
1463
1464 /*
1465 * This can happen if slptime and runtime are 0.
1466 */
1467 return (0);
1468
1469}
1470
1471/*
1472 * Scale the scheduling priority according to the "interactivity" of this
1473 * process.
1474 */
1475static void
1476sched_priority(struct thread *td)
1477{
1478 int score;
1479 int pri;
1480
1481 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1482 return;
1483 /*
1484 * If the score is interactive we place the thread in the realtime
1485 * queue with a priority that is less than kernel and interrupt
1486 * priorities. These threads are not subject to nice restrictions.
1487 *
1488 * Scores greater than this are placed on the normal timeshare queue
1489 * where the priority is partially decided by the most recent cpu
1490 * utilization and the rest is decided by nice value.
1491 *
1492 * The nice value of the process has a linear effect on the calculated
1493 * score. Negative nice values make it easier for a thread to be
1494 * considered interactive.
1495 */
1496 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1497 if (score < sched_interact) {
1498 pri = PRI_MIN_INTERACT;
1499 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1500 sched_interact) * score;
1501 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1502 ("sched_priority: invalid interactive priority %d score %d",
1503 pri, score));
1504 } else {
1505 pri = SCHED_PRI_MIN;
1506 if (td->td_sched->ts_ticks)
1507 pri += min(SCHED_PRI_TICKS(td->td_sched),
1508 SCHED_PRI_RANGE - 1);
1509 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1510 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1511 ("sched_priority: invalid priority %d: nice %d, "
1512 "ticks %d ftick %d ltick %d tick pri %d",
1513 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1514 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1515 SCHED_PRI_TICKS(td->td_sched)));
1516 }
1517 sched_user_prio(td, pri);
1518
1519 return;
1520}
1521
1522/*
1523 * This routine enforces a maximum limit on the amount of scheduling history
1524 * kept. It is called after either the slptime or runtime is adjusted. This
1525 * function is ugly due to integer math.
1526 */
1527static void
1528sched_interact_update(struct thread *td)
1529{
1530 struct td_sched *ts;
1531 u_int sum;
1532
1533 ts = td->td_sched;
1534 sum = ts->ts_runtime + ts->ts_slptime;
1535 if (sum < SCHED_SLP_RUN_MAX)
1536 return;
1537 /*
1538 * This only happens from two places:
1539 * 1) We have added an unusual amount of run time from fork_exit.
1540 * 2) We have added an unusual amount of sleep time from sched_sleep().
1541 */
1542 if (sum > SCHED_SLP_RUN_MAX * 2) {
1543 if (ts->ts_runtime > ts->ts_slptime) {
1544 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1545 ts->ts_slptime = 1;
1546 } else {
1547 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1548 ts->ts_runtime = 1;
1549 }
1550 return;
1551 }
1552 /*
1553 * If we have exceeded by more than 1/5th then the algorithm below
1554 * will not bring us back into range. Dividing by two here forces
1555 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1556 */
1557 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1558 ts->ts_runtime /= 2;
1559 ts->ts_slptime /= 2;
1560 return;
1561 }
1562 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1563 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1564}
1565
1566/*
1567 * Scale back the interactivity history when a child thread is created. The
1568 * history is inherited from the parent but the thread may behave totally
1569 * differently. For example, a shell spawning a compiler process. We want
1570 * to learn that the compiler is behaving badly very quickly.
1571 */
1572static void
1573sched_interact_fork(struct thread *td)
1574{
1575 int ratio;
1576 int sum;
1577
1578 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1579 if (sum > SCHED_SLP_RUN_FORK) {
1580 ratio = sum / SCHED_SLP_RUN_FORK;
1581 td->td_sched->ts_runtime /= ratio;
1582 td->td_sched->ts_slptime /= ratio;
1583 }
1584}
1585
1586/*
1587 * Called from proc0_init() to setup the scheduler fields.
1588 */
1589void
1590schedinit(void)
1591{
1592
1593 /*
1594 * Set up the scheduler specific parts of proc0.
1595 */
1596 proc0.p_sched = NULL; /* XXX */
1597 thread0.td_sched = &td_sched0;
1598 td_sched0.ts_ltick = ticks;
1599 td_sched0.ts_ftick = ticks;
1600 td_sched0.ts_slice = 0;
1601}
1602
1603/*
1604 * This is only somewhat accurate since given many processes of the same
1605 * priority they will switch when their slices run out, which will be
1606 * at most sched_slice stathz ticks.
1607 */
1608int
1609sched_rr_interval(void)
1610{
1611
1612 /* Convert sched_slice from stathz to hz. */
1613 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1614}
1615
1616/*
1617 * Update the percent cpu tracking information when it is requested or
1618 * the total history exceeds the maximum. We keep a sliding history of
1619 * tick counts that slowly decays. This is less precise than the 4BSD
1620 * mechanism since it happens with less regular and frequent events.
1621 */
1622static void
1623sched_pctcpu_update(struct td_sched *ts, int run)
1624{
1625 int t = ticks;
1626
1627 if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1628 ts->ts_ticks = 0;
1629 ts->ts_ftick = t - SCHED_TICK_TARG;
1630 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1631 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1632 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1633 ts->ts_ftick = t - SCHED_TICK_TARG;
1634 }
1635 if (run)
1636 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1637 ts->ts_ltick = t;
1638}
1639
1640/*
1641 * Adjust the priority of a thread. Move it to the appropriate run-queue
1642 * if necessary. This is the back-end for several priority related
1643 * functions.
1644 */
1645static void
1646sched_thread_priority(struct thread *td, u_char prio)
1647{
1648 struct td_sched *ts;
1649 struct tdq *tdq;
1650 int oldpri;
1651
1652 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1653 "prio:%d", td->td_priority, "new prio:%d", prio,
1654 KTR_ATTR_LINKED, sched_tdname(curthread));
1655 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1656 if (td != curthread && prio < td->td_priority) {
1657 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1658 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1659 prio, KTR_ATTR_LINKED, sched_tdname(td));
1660 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1661 curthread);
1662 }
1663 ts = td->td_sched;
1664 THREAD_LOCK_ASSERT(td, MA_OWNED);
1665 if (td->td_priority == prio)
1666 return;
1667 /*
1668 * If the priority has been elevated due to priority
1669 * propagation, we may have to move ourselves to a new
1670 * queue. This could be optimized to not re-add in some
1671 * cases.
1672 */
1673 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1674 sched_rem(td);
1675 td->td_priority = prio;
1676 sched_add(td, SRQ_BORROWING);
1677 return;
1678 }
1679 /*
1680 * If the thread is currently running we may have to adjust the lowpri
1681 * information so other cpus are aware of our current priority.
1682 */
1683 if (TD_IS_RUNNING(td)) {
1684 tdq = TDQ_CPU(ts->ts_cpu);
1685 oldpri = td->td_priority;
1686 td->td_priority = prio;
1687 if (prio < tdq->tdq_lowpri)
1688 tdq->tdq_lowpri = prio;
1689 else if (tdq->tdq_lowpri == oldpri)
1690 tdq_setlowpri(tdq, td);
1691 return;
1692 }
1693 td->td_priority = prio;
1694}
1695
1696/*
1697 * Update a thread's priority when it is lent another thread's
1698 * priority.
1699 */
1700void
1701sched_lend_prio(struct thread *td, u_char prio)
1702{
1703
1704 td->td_flags |= TDF_BORROWING;
1705 sched_thread_priority(td, prio);
1706}
1707
1708/*
1709 * Restore a thread's priority when priority propagation is
1710 * over. The prio argument is the minimum priority the thread
1711 * needs to have to satisfy other possible priority lending
1712 * requests. If the thread's regular priority is less
1713 * important than prio, the thread will keep a priority boost
1714 * of prio.
1715 */
1716void
1717sched_unlend_prio(struct thread *td, u_char prio)
1718{
1719 u_char base_pri;
1720
1721 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1722 td->td_base_pri <= PRI_MAX_TIMESHARE)
1723 base_pri = td->td_user_pri;
1724 else
1725 base_pri = td->td_base_pri;
1726 if (prio >= base_pri) {
1727 td->td_flags &= ~TDF_BORROWING;
1728 sched_thread_priority(td, base_pri);
1729 } else
1730 sched_lend_prio(td, prio);
1731}
1732
1733/*
1734 * Standard entry for setting the priority to an absolute value.
1735 */
1736void
1737sched_prio(struct thread *td, u_char prio)
1738{
1739 u_char oldprio;
1740
1741 /* First, update the base priority. */
1742 td->td_base_pri = prio;
1743
1744 /*
1745 * If the thread is borrowing another thread's priority, don't
1746 * ever lower the priority.
1747 */
1748 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1749 return;
1750
1751 /* Change the real priority. */
1752 oldprio = td->td_priority;
1753 sched_thread_priority(td, prio);
1754
1755 /*
1756 * If the thread is on a turnstile, then let the turnstile update
1757 * its state.
1758 */
1759 if (TD_ON_LOCK(td) && oldprio != prio)
1760 turnstile_adjust(td, oldprio);
1761}
1762
1763/*
1764 * Set the base user priority, does not effect current running priority.
1765 */
1766void
1767sched_user_prio(struct thread *td, u_char prio)
1768{
1769
1770 td->td_base_user_pri = prio;
1771 if (td->td_lend_user_pri <= prio)
1772 return;
1773 td->td_user_pri = prio;
1774}
1775
1776void
1777sched_lend_user_prio(struct thread *td, u_char prio)
1778{
1779
1780 THREAD_LOCK_ASSERT(td, MA_OWNED);
1781 td->td_lend_user_pri = prio;
1782 td->td_user_pri = min(prio, td->td_base_user_pri);
1783 if (td->td_priority > td->td_user_pri)
1784 sched_prio(td, td->td_user_pri);
1785 else if (td->td_priority != td->td_user_pri)
1786 td->td_flags |= TDF_NEEDRESCHED;
1787}
1788
1789/*
1790 * Handle migration from sched_switch(). This happens only for
1791 * cpu binding.
1792 */
1793static struct mtx *
1794sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1795{
1796 struct tdq *tdn;
1797
1798 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1799#ifdef SMP
1800 tdq_load_rem(tdq, td);
1801 /*
1802 * Do the lock dance required to avoid LOR. We grab an extra
1803 * spinlock nesting to prevent preemption while we're
1804 * not holding either run-queue lock.
1805 */
1806 spinlock_enter();
1807 thread_lock_block(td); /* This releases the lock on tdq. */
1808
1809 /*
1810 * Acquire both run-queue locks before placing the thread on the new
1811 * run-queue to avoid deadlocks created by placing a thread with a
1812 * blocked lock on the run-queue of a remote processor. The deadlock
1813 * occurs when a third processor attempts to lock the two queues in
1814 * question while the target processor is spinning with its own
1815 * run-queue lock held while waiting for the blocked lock to clear.
1816 */
1817 tdq_lock_pair(tdn, tdq);
1818 tdq_add(tdn, td, flags);
1819 tdq_notify(tdn, td);
1820 TDQ_UNLOCK(tdn);
1821 spinlock_exit();
1822#endif
1823 return (TDQ_LOCKPTR(tdn));
1824}
1825
1826/*
1827 * Variadic version of thread_lock_unblock() that does not assume td_lock
1828 * is blocked.
1829 */
1830static inline void
1831thread_unblock_switch(struct thread *td, struct mtx *mtx)
1832{
1833 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1834 (uintptr_t)mtx);
1835}
1836
1837/*
1838 * Switch threads. This function has to handle threads coming in while
1839 * blocked for some reason, running, or idle. It also must deal with
1840 * migrating a thread from one queue to another as running threads may
1841 * be assigned elsewhere via binding.
1842 */
1843void
1844sched_switch(struct thread *td, struct thread *newtd, int flags)
1845{
1846 struct tdq *tdq;
1847 struct td_sched *ts;
1848 struct mtx *mtx;
1849 int srqflag;
1850 int cpuid, preempted;
1851
1852 THREAD_LOCK_ASSERT(td, MA_OWNED);
1853 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1854
1855 cpuid = PCPU_GET(cpuid);
1856 tdq = TDQ_CPU(cpuid);
1857 ts = td->td_sched;
1858 mtx = td->td_lock;
1859 sched_pctcpu_update(ts, 1);
1860 ts->ts_rltick = ticks;
1861 td->td_lastcpu = td->td_oncpu;
1862 td->td_oncpu = NOCPU;
1863 preempted = !(td->td_flags & TDF_SLICEEND);
1864 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1865 td->td_owepreempt = 0;
1866 if (!TD_IS_IDLETHREAD(td))
1867 tdq->tdq_switchcnt++;
1868 /*
1869 * The lock pointer in an idle thread should never change. Reset it
1870 * to CAN_RUN as well.
1871 */
1872 if (TD_IS_IDLETHREAD(td)) {
1873 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1874 TD_SET_CAN_RUN(td);
1875 } else if (TD_IS_RUNNING(td)) {
1876 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1877 srqflag = preempted ?
1878 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1879 SRQ_OURSELF|SRQ_YIELDING;
1880#ifdef SMP
1881 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1882 ts->ts_cpu = sched_pickcpu(td, 0);
1883#endif
1884 if (ts->ts_cpu == cpuid)
1885 tdq_runq_add(tdq, td, srqflag);
1886 else {
1887 KASSERT(THREAD_CAN_MIGRATE(td) ||
1888 (ts->ts_flags & TSF_BOUND) != 0,
1889 ("Thread %p shouldn't migrate", td));
1890 mtx = sched_switch_migrate(tdq, td, srqflag);
1891 }
1892 } else {
1893 /* This thread must be going to sleep. */
1894 TDQ_LOCK(tdq);
1895 mtx = thread_lock_block(td);
1896 tdq_load_rem(tdq, td);
1897 }
1898 /*
1899 * We enter here with the thread blocked and assigned to the
1900 * appropriate cpu run-queue or sleep-queue and with the current
1901 * thread-queue locked.
1902 */
1903 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1904 newtd = choosethread();
1905 /*
1906 * Call the MD code to switch contexts if necessary.
1907 */
1908 if (td != newtd) {
1909#ifdef HWPMC_HOOKS
1910 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1911 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1912#endif
1913 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1914 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1915 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1916 sched_pctcpu_update(newtd->td_sched, 0);
1917
1918#ifdef KDTRACE_HOOKS
1919 /*
1920 * If DTrace has set the active vtime enum to anything
1921 * other than INACTIVE (0), then it should have set the
1922 * function to call.
1923 */
1924 if (dtrace_vtime_active)
1925 (*dtrace_vtime_switch_func)(newtd);
1926#endif
1927
1928 cpu_switch(td, newtd, mtx);
1929 /*
1930 * We may return from cpu_switch on a different cpu. However,
1931 * we always return with td_lock pointing to the current cpu's
1932 * run queue lock.
1933 */
1934 cpuid = PCPU_GET(cpuid);
1935 tdq = TDQ_CPU(cpuid);
1936 lock_profile_obtain_lock_success(
1937 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1938
1939 SDT_PROBE0(sched, , , on__cpu);
1940#ifdef HWPMC_HOOKS
1941 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1942 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1943#endif
1944 } else {
1945 thread_unblock_switch(td, mtx);
1946 SDT_PROBE0(sched, , , remain__cpu);
1947 }
1948 /*
1949 * Assert that all went well and return.
1950 */
1951 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1952 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1953 td->td_oncpu = cpuid;
1954}
1955
1956/*
1957 * Adjust thread priorities as a result of a nice request.
1958 */
1959void
1960sched_nice(struct proc *p, int nice)
1961{
1962 struct thread *td;
1963
1964 PROC_LOCK_ASSERT(p, MA_OWNED);
1965
1966 p->p_nice = nice;
1967 FOREACH_THREAD_IN_PROC(p, td) {
1968 thread_lock(td);
1969 sched_priority(td);
1970 sched_prio(td, td->td_base_user_pri);
1971 thread_unlock(td);
1972 }
1973}
1974
1975/*
1976 * Record the sleep time for the interactivity scorer.
1977 */
1978void
1979sched_sleep(struct thread *td, int prio)
1980{
1981
1982 THREAD_LOCK_ASSERT(td, MA_OWNED);
1983
1984 td->td_slptick = ticks;
1985 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1986 td->td_flags |= TDF_CANSWAP;
1987 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1988 return;
1989 if (static_boost == 1 && prio)
1990 sched_prio(td, prio);
1991 else if (static_boost && td->td_priority > static_boost)
1992 sched_prio(td, static_boost);
1993}
1994
1995/*
1996 * Schedule a thread to resume execution and record how long it voluntarily
1997 * slept. We also update the pctcpu, interactivity, and priority.
1998 */
1999void
2000sched_wakeup(struct thread *td)
2001{
2002 struct td_sched *ts;
2003 int slptick;
2004
2005 THREAD_LOCK_ASSERT(td, MA_OWNED);
2006 ts = td->td_sched;
2007 td->td_flags &= ~TDF_CANSWAP;
2008 /*
2009 * If we slept for more than a tick update our interactivity and
2010 * priority.
2011 */
2012 slptick = td->td_slptick;
2013 td->td_slptick = 0;
2014 if (slptick && slptick != ticks) {
2015 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2016 sched_interact_update(td);
2017 sched_pctcpu_update(ts, 0);
2018 }
2019 /*
2020 * Reset the slice value since we slept and advanced the round-robin.
2021 */
2022 ts->ts_slice = 0;
2023 sched_add(td, SRQ_BORING);
2024}
2025
2026/*
2027 * Penalize the parent for creating a new child and initialize the child's
2028 * priority.
2029 */
2030void
2031sched_fork(struct thread *td, struct thread *child)
2032{
2033 THREAD_LOCK_ASSERT(td, MA_OWNED);
2034 sched_pctcpu_update(td->td_sched, 1);
2035 sched_fork_thread(td, child);
2036 /*
2037 * Penalize the parent and child for forking.
2038 */
2039 sched_interact_fork(child);
2040 sched_priority(child);
2041 td->td_sched->ts_runtime += tickincr;
2042 sched_interact_update(td);
2043 sched_priority(td);
2044}
2045
2046/*
2047 * Fork a new thread, may be within the same process.
2048 */
2049void
2050sched_fork_thread(struct thread *td, struct thread *child)
2051{
2052 struct td_sched *ts;
2053 struct td_sched *ts2;
2054 struct tdq *tdq;
2055
2056 tdq = TDQ_SELF();
2057 THREAD_LOCK_ASSERT(td, MA_OWNED);
2058 /*
2059 * Initialize child.
2060 */
2061 ts = td->td_sched;
2062 ts2 = child->td_sched;
2063 child->td_lock = TDQ_LOCKPTR(tdq);
2064 child->td_cpuset = cpuset_ref(td->td_cpuset);
2065 ts2->ts_cpu = ts->ts_cpu;
2066 ts2->ts_flags = 0;
2067 /*
2068 * Grab our parents cpu estimation information.
2069 */
2070 ts2->ts_ticks = ts->ts_ticks;
2071 ts2->ts_ltick = ts->ts_ltick;
2072 ts2->ts_ftick = ts->ts_ftick;
2073 /*
2074 * Do not inherit any borrowed priority from the parent.
2075 */
2076 child->td_priority = child->td_base_pri;
2077 /*
2078 * And update interactivity score.
2079 */
2080 ts2->ts_slptime = ts->ts_slptime;
2081 ts2->ts_runtime = ts->ts_runtime;
2082 /* Attempt to quickly learn interactivity. */
2083 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2084#ifdef KTR
2085 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2086#endif
2087}
2088
2089/*
2090 * Adjust the priority class of a thread.
2091 */
2092void
2093sched_class(struct thread *td, int class)
2094{
2095
2096 THREAD_LOCK_ASSERT(td, MA_OWNED);
2097 if (td->td_pri_class == class)
2098 return;
2099 td->td_pri_class = class;
2100}
2101
2102/*
2103 * Return some of the child's priority and interactivity to the parent.
2104 */
2105void
2106sched_exit(struct proc *p, struct thread *child)
2107{
2108 struct thread *td;
2109
2110 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2111 "prio:%d", child->td_priority);
2112 PROC_LOCK_ASSERT(p, MA_OWNED);
2113 td = FIRST_THREAD_IN_PROC(p);
2114 sched_exit_thread(td, child);
2115}
2116
2117/*
2118 * Penalize another thread for the time spent on this one. This helps to
2119 * worsen the priority and interactivity of processes which schedule batch
2120 * jobs such as make. This has little effect on the make process itself but
2121 * causes new processes spawned by it to receive worse scores immediately.
2122 */
2123void
2124sched_exit_thread(struct thread *td, struct thread *child)
2125{
2126
2127 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2128 "prio:%d", child->td_priority);
2129 /*
2130 * Give the child's runtime to the parent without returning the
2131 * sleep time as a penalty to the parent. This causes shells that
2132 * launch expensive things to mark their children as expensive.
2133 */
2134 thread_lock(td);
2135 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2136 sched_interact_update(td);
2137 sched_priority(td);
2138 thread_unlock(td);
2139}
2140
2141void
2142sched_preempt(struct thread *td)
2143{
2144 struct tdq *tdq;
2145
2146 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2147
2148 thread_lock(td);
2149 tdq = TDQ_SELF();
2150 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2151 tdq->tdq_ipipending = 0;
2152 if (td->td_priority > tdq->tdq_lowpri) {
2153 int flags;
2154
2155 flags = SW_INVOL | SW_PREEMPT;
2156 if (td->td_critnest > 1)
2157 td->td_owepreempt = 1;
2158 else if (TD_IS_IDLETHREAD(td))
2159 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2160 else
2161 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2162 }
2163 thread_unlock(td);
2164}
2165
2166/*
2167 * Fix priorities on return to user-space. Priorities may be elevated due
2168 * to static priorities in msleep() or similar.
2169 */
2170void
2171sched_userret(struct thread *td)
2172{
2173 /*
2174 * XXX we cheat slightly on the locking here to avoid locking in
2175 * the usual case. Setting td_priority here is essentially an
2176 * incomplete workaround for not setting it properly elsewhere.
2177 * Now that some interrupt handlers are threads, not setting it
2178 * properly elsewhere can clobber it in the window between setting
2179 * it here and returning to user mode, so don't waste time setting
2180 * it perfectly here.
2181 */
2182 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2183 ("thread with borrowed priority returning to userland"));
2184 if (td->td_priority != td->td_user_pri) {
2185 thread_lock(td);
2186 td->td_priority = td->td_user_pri;
2187 td->td_base_pri = td->td_user_pri;
2188 tdq_setlowpri(TDQ_SELF(), td);
2189 thread_unlock(td);
2190 }
2191}
2192
2193/*
2194 * Handle a stathz tick. This is really only relevant for timeshare
2195 * threads.
2196 */
2197void
2198sched_clock(struct thread *td)
2199{
2200 struct tdq *tdq;
2201 struct td_sched *ts;
2202
2203 THREAD_LOCK_ASSERT(td, MA_OWNED);
2204 tdq = TDQ_SELF();
2205#ifdef SMP
2206 /*
2207 * We run the long term load balancer infrequently on the first cpu.
2208 */
2209 if (balance_tdq == tdq) {
2210 if (balance_ticks && --balance_ticks == 0)
2211 sched_balance();
2212 }
2213#endif
2214 /*
2215 * Save the old switch count so we have a record of the last ticks
2216 * activity. Initialize the new switch count based on our load.
2217 * If there is some activity seed it to reflect that.
2218 */
2219 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2220 tdq->tdq_switchcnt = tdq->tdq_load;
2221 /*
2222 * Advance the insert index once for each tick to ensure that all
2223 * threads get a chance to run.
2224 */
2225 if (tdq->tdq_idx == tdq->tdq_ridx) {
2226 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2227 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2228 tdq->tdq_ridx = tdq->tdq_idx;
2229 }
2230 ts = td->td_sched;
2231 sched_pctcpu_update(ts, 1);
2232 if (td->td_pri_class & PRI_FIFO_BIT)
2233 return;
2234 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2235 /*
2236 * We used a tick; charge it to the thread so
2237 * that we can compute our interactivity.
2238 */
2239 td->td_sched->ts_runtime += tickincr;
2240 sched_interact_update(td);
2241 sched_priority(td);
2242 }
2243
2244 /*
2245 * Force a context switch if the current thread has used up a full
2246 * time slice (default is 100ms).
2247 */
2248 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2249 ts->ts_slice = 0;
2250 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2251 }
2252}
2253
2254/*
2255 * Called once per hz tick.
2256 */
2257void
2258sched_tick(int cnt)
2259{
2260
2261}
2262
2263/*
2264 * Return whether the current CPU has runnable tasks. Used for in-kernel
2265 * cooperative idle threads.
2266 */
2267int
2268sched_runnable(void)
2269{
2270 struct tdq *tdq;
2271 int load;
2272
2273 load = 1;
2274
2275 tdq = TDQ_SELF();
2276 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2277 if (tdq->tdq_load > 0)
2278 goto out;
2279 } else
2280 if (tdq->tdq_load - 1 > 0)
2281 goto out;
2282 load = 0;
2283out:
2284 return (load);
2285}
2286
2287/*
2288 * Choose the highest priority thread to run. The thread is removed from
2289 * the run-queue while running however the load remains. For SMP we set
2290 * the tdq in the global idle bitmask if it idles here.
2291 */
2292struct thread *
2293sched_choose(void)
2294{
2295 struct thread *td;
2296 struct tdq *tdq;
2297
2298 tdq = TDQ_SELF();
2299 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2300 td = tdq_choose(tdq);
2301 if (td) {
2302 tdq_runq_rem(tdq, td);
2303 tdq->tdq_lowpri = td->td_priority;
2304 return (td);
2305 }
2306 tdq->tdq_lowpri = PRI_MAX_IDLE;
2307 return (PCPU_GET(idlethread));
2308}
2309
2310/*
2311 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2312 * we always request it once we exit a critical section.
2313 */
2314static inline void
2315sched_setpreempt(struct thread *td)
2316{
2317 struct thread *ctd;
2318 int cpri;
2319 int pri;
2320
2321 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2322
2323 ctd = curthread;
2324 pri = td->td_priority;
2325 cpri = ctd->td_priority;
2326 if (pri < cpri)
2327 ctd->td_flags |= TDF_NEEDRESCHED;
2328 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2329 return;
2330 if (!sched_shouldpreempt(pri, cpri, 0))
2331 return;
2332 ctd->td_owepreempt = 1;
2333}
2334
2335/*
2336 * Add a thread to a thread queue. Select the appropriate runq and add the
2337 * thread to it. This is the internal function called when the tdq is
2338 * predetermined.
2339 */
2340void
2341tdq_add(struct tdq *tdq, struct thread *td, int flags)
2342{
2343
2344 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2345 KASSERT((td->td_inhibitors == 0),
2346 ("sched_add: trying to run inhibited thread"));
2347 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2348 ("sched_add: bad thread state"));
2349 KASSERT(td->td_flags & TDF_INMEM,
2350 ("sched_add: thread swapped out"));
2351
2352 if (td->td_priority < tdq->tdq_lowpri)
2353 tdq->tdq_lowpri = td->td_priority;
2354 tdq_runq_add(tdq, td, flags);
2355 tdq_load_add(tdq, td);
2356}
2357
2358/*
2359 * Select the target thread queue and add a thread to it. Request
2360 * preemption or IPI a remote processor if required.
2361 */
2362void
2363sched_add(struct thread *td, int flags)
2364{
2365 struct tdq *tdq;
2366#ifdef SMP
2367 int cpu;
2368#endif
2369
2370 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2371 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2372 sched_tdname(curthread));
2373 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2374 KTR_ATTR_LINKED, sched_tdname(td));
2375 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2376 flags & SRQ_PREEMPTED);
2377 THREAD_LOCK_ASSERT(td, MA_OWNED);
2378 /*
2379 * Recalculate the priority before we select the target cpu or
2380 * run-queue.
2381 */
2382 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2383 sched_priority(td);
2384#ifdef SMP
2385 /*
2386 * Pick the destination cpu and if it isn't ours transfer to the
2387 * target cpu.
2388 */
2389 cpu = sched_pickcpu(td, flags);
2390 tdq = sched_setcpu(td, cpu, flags);
2391 tdq_add(tdq, td, flags);
2392 if (cpu != PCPU_GET(cpuid)) {
2393 tdq_notify(tdq, td);
2394 return;
2395 }
2396#else
2397 tdq = TDQ_SELF();
2398 TDQ_LOCK(tdq);
2399 /*
2400 * Now that the thread is moving to the run-queue, set the lock
2401 * to the scheduler's lock.
2402 */
2403 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2404 tdq_add(tdq, td, flags);
2405#endif
2406 if (!(flags & SRQ_YIELDING))
2407 sched_setpreempt(td);
2408}
2409
2410/*
2411 * Remove a thread from a run-queue without running it. This is used
2412 * when we're stealing a thread from a remote queue. Otherwise all threads
2413 * exit by calling sched_exit_thread() and sched_throw() themselves.
2414 */
2415void
2416sched_rem(struct thread *td)
2417{
2418 struct tdq *tdq;
2419
2420 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2421 "prio:%d", td->td_priority);
2422 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2423 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2424 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2425 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2426 KASSERT(TD_ON_RUNQ(td),
2427 ("sched_rem: thread not on run queue"));
2428 tdq_runq_rem(tdq, td);
2429 tdq_load_rem(tdq, td);
2430 TD_SET_CAN_RUN(td);
2431 if (td->td_priority == tdq->tdq_lowpri)
2432 tdq_setlowpri(tdq, NULL);
2433}
2434
2435/*
2436 * Fetch cpu utilization information. Updates on demand.
2437 */
2438fixpt_t
2439sched_pctcpu(struct thread *td)
2440{
2441 fixpt_t pctcpu;
2442 struct td_sched *ts;
2443
2444 pctcpu = 0;
2445 ts = td->td_sched;
2446 if (ts == NULL)
2447 return (0);
2448
2449 THREAD_LOCK_ASSERT(td, MA_OWNED);
2450 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2451 if (ts->ts_ticks) {
2452 int rtick;
2453
2454 /* How many rtick per second ? */
2455 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2456 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2457 }
2458
2459 return (pctcpu);
2460}
2461
2462/*
2463 * Enforce affinity settings for a thread. Called after adjustments to
2464 * cpumask.
2465 */
2466void
2467sched_affinity(struct thread *td)
2468{
2469#ifdef SMP
2470 struct td_sched *ts;
2471
2472 THREAD_LOCK_ASSERT(td, MA_OWNED);
2473 ts = td->td_sched;
2474 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2475 return;
2476 if (TD_ON_RUNQ(td)) {
2477 sched_rem(td);
2478 sched_add(td, SRQ_BORING);
2479 return;
2480 }
2481 if (!TD_IS_RUNNING(td))
2482 return;
2483 /*
2484 * Force a switch before returning to userspace. If the
2485 * target thread is not running locally send an ipi to force
2486 * the issue.
2487 */
2488 td->td_flags |= TDF_NEEDRESCHED;
2489 if (td != curthread)
2490 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2491#endif
2492}
2493
2494/*
2495 * Bind a thread to a target cpu.
2496 */
2497void
2498sched_bind(struct thread *td, int cpu)
2499{
2500 struct td_sched *ts;
2501
2502 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2503 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2504 ts = td->td_sched;
2505 if (ts->ts_flags & TSF_BOUND)
2506 sched_unbind(td);
2507 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2508 ts->ts_flags |= TSF_BOUND;
2509 sched_pin();
2510 if (PCPU_GET(cpuid) == cpu)
2511 return;
2512 ts->ts_cpu = cpu;
2513 /* When we return from mi_switch we'll be on the correct cpu. */
2514 mi_switch(SW_VOL, NULL);
2515}
2516
2517/*
2518 * Release a bound thread.
2519 */
2520void
2521sched_unbind(struct thread *td)
2522{
2523 struct td_sched *ts;
2524
2525 THREAD_LOCK_ASSERT(td, MA_OWNED);
2526 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2527 ts = td->td_sched;
2528 if ((ts->ts_flags & TSF_BOUND) == 0)
2529 return;
2530 ts->ts_flags &= ~TSF_BOUND;
2531 sched_unpin();
2532}
2533
2534int
2535sched_is_bound(struct thread *td)
2536{
2537 THREAD_LOCK_ASSERT(td, MA_OWNED);
2538 return (td->td_sched->ts_flags & TSF_BOUND);
2539}
2540
2541/*
2542 * Basic yield call.
2543 */
2544void
2545sched_relinquish(struct thread *td)
2546{
2547 thread_lock(td);
2548 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2549 thread_unlock(td);
2550}
2551
2552/*
2553 * Return the total system load.
2554 */
2555int
2556sched_load(void)
2557{
2558#ifdef SMP
2559 int total;
2560 int i;
2561
2562 total = 0;
2563 CPU_FOREACH(i)
2564 total += TDQ_CPU(i)->tdq_sysload;
2565 return (total);
2566#else
2567 return (TDQ_SELF()->tdq_sysload);
2568#endif
2569}
2570
2571int
2572sched_sizeof_proc(void)
2573{
2574 return (sizeof(struct proc));
2575}
2576
2577int
2578sched_sizeof_thread(void)
2579{
2580 return (sizeof(struct thread) + sizeof(struct td_sched));
2581}
2582
2583#ifdef SMP
2584#define TDQ_IDLESPIN(tdq) \
2585 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2586#else
2587#define TDQ_IDLESPIN(tdq) 1
2588#endif
2589
2590/*
2591 * The actual idle process.
2592 */
2593void
2594sched_idletd(void *dummy)
2595{
2596 struct thread *td;
2597 struct tdq *tdq;
2598 int oldswitchcnt, switchcnt;
2599 int i;
2600
2601 mtx_assert(&Giant, MA_NOTOWNED);
2602 td = curthread;
2603 tdq = TDQ_SELF();
2604 THREAD_NO_SLEEPING();
2605 oldswitchcnt = -1;
2606 for (;;) {
2607 if (tdq->tdq_load) {
2608 thread_lock(td);
2609 mi_switch(SW_VOL | SWT_IDLE, NULL);
2610 thread_unlock(td);
2611 }
2612 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2613#ifdef SMP
2614 if (switchcnt != oldswitchcnt) {
2615 oldswitchcnt = switchcnt;
2616 if (tdq_idled(tdq) == 0)
2617 continue;
2618 }
2619 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2620#else
2621 oldswitchcnt = switchcnt;
2622#endif
2623 /*
2624 * If we're switching very frequently, spin while checking
2625 * for load rather than entering a low power state that
2626 * may require an IPI. However, don't do any busy
2627 * loops while on SMT machines as this simply steals
2628 * cycles from cores doing useful work.
2629 */
2630 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2631 for (i = 0; i < sched_idlespins; i++) {
2632 if (tdq->tdq_load)
2633 break;
2634 cpu_spinwait();
2635 }
2636 }
2637
2638 /* If there was context switch during spin, restart it. */
2639 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2640 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2641 continue;
2642
2643 /* Run main MD idle handler. */
2644 tdq->tdq_cpu_idle = 1;
2645 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2646 tdq->tdq_cpu_idle = 0;
2647
2648 /*
2649 * Account thread-less hardware interrupts and
2650 * other wakeup reasons equal to context switches.
2651 */
2652 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2653 if (switchcnt != oldswitchcnt)
2654 continue;
2655 tdq->tdq_switchcnt++;
2656 oldswitchcnt++;
2657 }
2658}
2659
2660/*
2661 * A CPU is entering for the first time or a thread is exiting.
2662 */
2663void
2664sched_throw(struct thread *td)
2665{
2666 struct thread *newtd;
2667 struct tdq *tdq;
2668
2669 tdq = TDQ_SELF();
2670 if (td == NULL) {
2671 /* Correct spinlock nesting and acquire the correct lock. */
2672 TDQ_LOCK(tdq);
2673 spinlock_exit();
2674 PCPU_SET(switchtime, cpu_ticks());
2675 PCPU_SET(switchticks, ticks);
2676 } else {
2677 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2678 tdq_load_rem(tdq, td);
2679 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2680 }
2681 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2682 newtd = choosethread();
2683 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2684 cpu_throw(td, newtd); /* doesn't return */
2685}
2686
2687/*
2688 * This is called from fork_exit(). Just acquire the correct locks and
2689 * let fork do the rest of the work.
2690 */
2691void
2692sched_fork_exit(struct thread *td)
2693{
2694 struct td_sched *ts;
2695 struct tdq *tdq;
2696 int cpuid;
2697
2698 /*
2699 * Finish setting up thread glue so that it begins execution in a
2700 * non-nested critical section with the scheduler lock held.
2701 */
2702 cpuid = PCPU_GET(cpuid);
2703 tdq = TDQ_CPU(cpuid);
2704 ts = td->td_sched;
2705 if (TD_IS_IDLETHREAD(td))
2706 td->td_lock = TDQ_LOCKPTR(tdq);
2707 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2708 td->td_oncpu = cpuid;
2709 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2710 lock_profile_obtain_lock_success(
2711 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2712}
2713
2714/*
2715 * Create on first use to catch odd startup conditons.
2716 */
2717char *
2718sched_tdname(struct thread *td)
2719{
2720#ifdef KTR
2721 struct td_sched *ts;
2722
2723 ts = td->td_sched;
2724 if (ts->ts_name[0] == '\0')
2725 snprintf(ts->ts_name, sizeof(ts->ts_name),
2726 "%s tid %d", td->td_name, td->td_tid);
2727 return (ts->ts_name);
2728#else
2729 return (td->td_name);
2730#endif
2731}
2732
2733#ifdef KTR
2734void
2735sched_clear_tdname(struct thread *td)
2736{
2737 struct td_sched *ts;
2738
2739 ts = td->td_sched;
2740 ts->ts_name[0] = '\0';
2741}
2742#endif
2743
2744#ifdef SMP
2745
2746/*
2747 * Build the CPU topology dump string. Is recursively called to collect
2748 * the topology tree.
2749 */
2750static int
2751sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2752 int indent)
2753{
2754 char cpusetbuf[CPUSETBUFSIZ];
2755 int i, first;
2756
2757 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2758 "", 1 + indent / 2, cg->cg_level);
2759 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2760 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2761 first = TRUE;
2762 for (i = 0; i < MAXCPU; i++) {
2763 if (CPU_ISSET(i, &cg->cg_mask)) {
2764 if (!first)
2765 sbuf_printf(sb, ", ");
2766 else
2767 first = FALSE;
2768 sbuf_printf(sb, "%d", i);
2769 }
2770 }
2771 sbuf_printf(sb, "</cpu>\n");
2772
2773 if (cg->cg_flags != 0) {
2774 sbuf_printf(sb, "%*s <flags>", indent, "");
2775 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2776 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2777 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2778 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2779 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2780 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2781 sbuf_printf(sb, "</flags>\n");
2782 }
2783
2784 if (cg->cg_children > 0) {
2785 sbuf_printf(sb, "%*s <children>\n", indent, "");
2786 for (i = 0; i < cg->cg_children; i++)
2787 sysctl_kern_sched_topology_spec_internal(sb,
2788 &cg->cg_child[i], indent+2);
2789 sbuf_printf(sb, "%*s </children>\n", indent, "");
2790 }
2791 sbuf_printf(sb, "%*s</group>\n", indent, "");
2792 return (0);
2793}
2794
2795/*
2796 * Sysctl handler for retrieving topology dump. It's a wrapper for
2797 * the recursive sysctl_kern_smp_topology_spec_internal().
2798 */
2799static int
2800sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2801{
2802 struct sbuf *topo;
2803 int err;
2804
2805 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2806
2807 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2808 if (topo == NULL)
2809 return (ENOMEM);
2810
2811 sbuf_printf(topo, "<groups>\n");
2812 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2813 sbuf_printf(topo, "</groups>\n");
2814
2815 if (err == 0) {
2816 sbuf_finish(topo);
2817 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2818 }
2819 sbuf_delete(topo);
2820 return (err);
2821}
2822
2823#endif
2824
2825static int
2826sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2827{
2828 int error, new_val, period;
2829
2830 period = 1000000 / realstathz;
2831 new_val = period * sched_slice;
2832 error = sysctl_handle_int(oidp, &new_val, 0, req);
2833 if (error != 0 || req->newptr == NULL)
2834 return (error);
2835 if (new_val <= 0)
2836 return (EINVAL);
2837 sched_slice = imax(1, (new_val + period / 2) / period);
2838 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2839 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2840 realstathz);
2841 return (0);
2842}
2843
2844SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2845SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2846 "Scheduler name");
2847SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2848 NULL, 0, sysctl_kern_quantum, "I",
2849 "Quantum for timeshare threads in microseconds");
2850SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2851 "Quantum for timeshare threads in stathz ticks");
2852SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2853 "Interactivity score threshold");
2854SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2855 &preempt_thresh, 0,
2856 "Maximal (lowest) priority for preemption");
2857SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2858 "Assign static kernel priorities to sleeping threads");
2859SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2860 "Number of times idle thread will spin waiting for new work");
2861SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2862 &sched_idlespinthresh, 0,
2863 "Threshold before we will permit idle thread spinning");
2864#ifdef SMP
2865SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2866 "Number of hz ticks to keep thread affinity for");
2867SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2868 "Enables the long-term load balancer");
2869SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2870 &balance_interval, 0,
2871 "Average period in stathz ticks to run the long-term balancer");
2872SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2873 "Attempts to steal work from other cores before idling");
2874SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2875 "Minimum load on remote CPU before we'll steal");
2876SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2877 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2878 "XML dump of detected CPU topology");
2879#endif
2880
2881/* ps compat. All cpu percentages from ULE are weighted. */
2882static int ccpu = 0;
2883SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
|