FreeBSD/Linux Kernel Cross Reference
sys/kern/sched_ule.c
1 /*-
2 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice unmodified, this list of conditions, and the following
10 * disclaimer.
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.
14 *
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.
25 */
26
27 /*
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.
31 *
32 * etymology:
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.
36 */
37
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_kdtrace.h"
43 #include "opt_sched.h"
44
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kdb.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/sdt.h>
57 #include <sys/smp.h>
58 #include <sys/sx.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
62 #include <sys/umtx.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
65 #include <sys/sbuf.h>
66 #ifdef KTRACE
67 #include <sys/uio.h>
68 #include <sys/ktrace.h>
69 #endif
70
71 #ifdef HWPMC_HOOKS
72 #include <sys/pmckern.h>
73 #endif
74
75 #ifdef KDTRACE_HOOKS
76 #include <sys/dtrace_bsd.h>
77 int dtrace_vtime_active;
78 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
79 #endif
80
81 #include <machine/cpu.h>
82 #include <machine/smp.h>
83
84 #if defined(__sparc64__)
85 #error "This architecture is not currently compatible with ULE"
86 #endif
87
88 #define KTR_ULE 0
89
90 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
91 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
92 #define TDQ_LOADNAME_LEN (PCPU_NAME_LEN + sizeof(" load"))
93
94 /*
95 * Thread scheduler specific section. All fields are protected
96 * by the thread lock.
97 */
98 struct td_sched {
99 struct runq *ts_runq; /* Run-queue we're queued on. */
100 short ts_flags; /* TSF_* flags. */
101 u_char ts_cpu; /* CPU that we have affinity for. */
102 int ts_rltick; /* Real last tick, for affinity. */
103 int ts_slice; /* Ticks of slice remaining. */
104 u_int ts_slptime; /* Number of ticks we vol. slept */
105 u_int ts_runtime; /* Number of ticks we were running */
106 int ts_ltick; /* Last tick that we were running on */
107 int ts_incrtick; /* Last tick that we incremented on */
108 int ts_ftick; /* First tick that we were running on */
109 int ts_ticks; /* Tick count */
110 #ifdef KTR
111 char ts_name[TS_NAME_LEN];
112 #endif
113 };
114 /* flags kept in ts_flags */
115 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
116 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
117
118 static struct td_sched td_sched0;
119
120 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
121 #define THREAD_CAN_SCHED(td, cpu) \
122 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
123
124 /*
125 * Priority ranges used for interactive and non-interactive timeshare
126 * threads. Interactive threads use realtime priorities.
127 */
128 #define PRI_MIN_INTERACT PRI_MIN_REALTIME
129 #define PRI_MAX_INTERACT PRI_MAX_REALTIME
130 #define PRI_MIN_BATCH PRI_MIN_TIMESHARE
131 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
132
133 /*
134 * Cpu percentage computation macros and defines.
135 *
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.
142 */
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))
149
150 /*
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.
156 *
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.
161 */
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)
171
172 /*
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.
177 *
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.
183 */
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)
189
190 /* Flags kept in td_flags. */
191 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
192
193 /*
194 * tickincr: Converts a stathz tick into a hz domain scaled by
195 * the shift factor. Without the shift the error rate
196 * due to rounding would be unacceptably high.
197 * realstathz: stathz is sometimes 0 and run off of hz.
198 * sched_slice: Runtime of each thread before rescheduling.
199 * preempt_thresh: Priority threshold for preemption and remote IPIs.
200 */
201 static int sched_interact = SCHED_INTERACT_THRESH;
202 static int realstathz = 127;
203 static int tickincr = 8 << SCHED_TICK_SHIFT;
204 static int sched_slice = 12;
205 #ifdef PREEMPTION
206 #ifdef FULL_PREEMPTION
207 static int preempt_thresh = PRI_MAX_IDLE;
208 #else
209 static int preempt_thresh = PRI_MIN_KERN;
210 #endif
211 #else
212 static int preempt_thresh = 0;
213 #endif
214 static int static_boost = PRI_MIN_BATCH;
215 static int sched_idlespins = 10000;
216 static int sched_idlespinthresh = 4;
217
218 /*
219 * tdq - per processor runqs and statistics. All fields are protected by the
220 * tdq_lock. The load and lowpri may be accessed without to avoid excess
221 * locking in sched_pickcpu();
222 */
223 struct tdq {
224 /* Ordered to improve efficiency of cpu_search() and switch(). */
225 struct mtx tdq_lock; /* run queue lock. */
226 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
227 volatile int tdq_load; /* Aggregate load. */
228 int tdq_sysload; /* For loadavg, !ITHD load. */
229 int tdq_transferable; /* Transferable thread count. */
230 short tdq_switchcnt; /* Switches this tick. */
231 short tdq_oldswitchcnt; /* Switches last tick. */
232 u_char tdq_lowpri; /* Lowest priority thread. */
233 u_char tdq_ipipending; /* IPI pending. */
234 u_char tdq_idx; /* Current insert index. */
235 u_char tdq_ridx; /* Current removal index. */
236 struct runq tdq_realtime; /* real-time run queue. */
237 struct runq tdq_timeshare; /* timeshare run queue. */
238 struct runq tdq_idle; /* Queue of IDLE threads. */
239 char tdq_name[TDQ_NAME_LEN];
240 #ifdef KTR
241 char tdq_loadname[TDQ_LOADNAME_LEN];
242 #endif
243 } __aligned(64);
244
245 /* Idle thread states and config. */
246 #define TDQ_RUNNING 1
247 #define TDQ_IDLE 2
248
249 #ifdef SMP
250 struct cpu_group *cpu_top; /* CPU topology */
251
252 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
253 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
254
255 /*
256 * Run-time tunables.
257 */
258 static int rebalance = 1;
259 static int balance_interval = 128; /* Default set in sched_initticks(). */
260 static int affinity;
261 static int steal_idle = 1;
262 static int steal_thresh = 2;
263
264 /*
265 * One thread queue per processor.
266 */
267 static struct tdq tdq_cpu[MAXCPU];
268 static struct tdq *balance_tdq;
269 static int balance_ticks;
270 static DPCPU_DEFINE(uint32_t, randomval);
271
272 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
273 #define TDQ_CPU(x) (&tdq_cpu[(x)])
274 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
275 #else /* !SMP */
276 static struct tdq tdq_cpu;
277
278 #define TDQ_ID(x) (0)
279 #define TDQ_SELF() (&tdq_cpu)
280 #define TDQ_CPU(x) (&tdq_cpu)
281 #endif
282
283 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
284 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
285 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
286 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
287 #define TDQ_LOCKPTR(t) (&(t)->tdq_lock)
288
289 static void sched_priority(struct thread *);
290 static void sched_thread_priority(struct thread *, u_char);
291 static int sched_interact_score(struct thread *);
292 static void sched_interact_update(struct thread *);
293 static void sched_interact_fork(struct thread *);
294 static void sched_pctcpu_update(struct td_sched *);
295
296 /* Operations on per processor queues */
297 static struct thread *tdq_choose(struct tdq *);
298 static void tdq_setup(struct tdq *);
299 static void tdq_load_add(struct tdq *, struct thread *);
300 static void tdq_load_rem(struct tdq *, struct thread *);
301 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
302 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
303 static inline int sched_shouldpreempt(int, int, int);
304 void tdq_print(int cpu);
305 static void runq_print(struct runq *rq);
306 static void tdq_add(struct tdq *, struct thread *, int);
307 #ifdef SMP
308 static int tdq_move(struct tdq *, struct tdq *);
309 static int tdq_idled(struct tdq *);
310 static void tdq_notify(struct tdq *, struct thread *);
311 static struct thread *tdq_steal(struct tdq *, int);
312 static struct thread *runq_steal(struct runq *, int);
313 static int sched_pickcpu(struct thread *, int);
314 static void sched_balance(void);
315 static int sched_balance_pair(struct tdq *, struct tdq *);
316 static inline struct tdq *sched_setcpu(struct thread *, int, int);
317 static inline void thread_unblock_switch(struct thread *, struct mtx *);
318 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
319 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
320 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
321 struct cpu_group *cg, int indent);
322 #endif
323
324 static void sched_setup(void *dummy);
325 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
326
327 static void sched_initticks(void *dummy);
328 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
329 NULL);
330
331 SDT_PROVIDER_DEFINE(sched);
332
333 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
334 "struct proc *", "uint8_t");
335 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
336 "struct proc *", "void *");
337 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
338 "struct proc *", "void *", "int");
339 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
340 "struct proc *", "uint8_t", "struct thread *");
341 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
342 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
343 "struct proc *");
344 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
345 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
346 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
347 "struct proc *");
348
349 /*
350 * Print the threads waiting on a run-queue.
351 */
352 static void
353 runq_print(struct runq *rq)
354 {
355 struct rqhead *rqh;
356 struct thread *td;
357 int pri;
358 int j;
359 int i;
360
361 for (i = 0; i < RQB_LEN; i++) {
362 printf("\t\trunq bits %d 0x%zx\n",
363 i, rq->rq_status.rqb_bits[i]);
364 for (j = 0; j < RQB_BPW; j++)
365 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
366 pri = j + (i << RQB_L2BPW);
367 rqh = &rq->rq_queues[pri];
368 TAILQ_FOREACH(td, rqh, td_runq) {
369 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
370 td, td->td_name, td->td_priority,
371 td->td_rqindex, pri);
372 }
373 }
374 }
375 }
376
377 /*
378 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
379 */
380 void
381 tdq_print(int cpu)
382 {
383 struct tdq *tdq;
384
385 tdq = TDQ_CPU(cpu);
386
387 printf("tdq %d:\n", TDQ_ID(tdq));
388 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
389 printf("\tLock name: %s\n", tdq->tdq_name);
390 printf("\tload: %d\n", tdq->tdq_load);
391 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
392 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
393 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
394 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
395 printf("\tload transferable: %d\n", tdq->tdq_transferable);
396 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
397 printf("\trealtime runq:\n");
398 runq_print(&tdq->tdq_realtime);
399 printf("\ttimeshare runq:\n");
400 runq_print(&tdq->tdq_timeshare);
401 printf("\tidle runq:\n");
402 runq_print(&tdq->tdq_idle);
403 }
404
405 static inline int
406 sched_shouldpreempt(int pri, int cpri, int remote)
407 {
408 /*
409 * If the new priority is not better than the current priority there is
410 * nothing to do.
411 */
412 if (pri >= cpri)
413 return (0);
414 /*
415 * Always preempt idle.
416 */
417 if (cpri >= PRI_MIN_IDLE)
418 return (1);
419 /*
420 * If preemption is disabled don't preempt others.
421 */
422 if (preempt_thresh == 0)
423 return (0);
424 /*
425 * Preempt if we exceed the threshold.
426 */
427 if (pri <= preempt_thresh)
428 return (1);
429 /*
430 * If we're interactive or better and there is non-interactive
431 * or worse running preempt only remote processors.
432 */
433 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
434 return (1);
435 return (0);
436 }
437
438 /*
439 * Add a thread to the actual run-queue. Keeps transferable counts up to
440 * date with what is actually on the run-queue. Selects the correct
441 * queue position for timeshare threads.
442 */
443 static __inline void
444 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
445 {
446 struct td_sched *ts;
447 u_char pri;
448
449 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
450 THREAD_LOCK_ASSERT(td, MA_OWNED);
451
452 pri = td->td_priority;
453 ts = td->td_sched;
454 TD_SET_RUNQ(td);
455 if (THREAD_CAN_MIGRATE(td)) {
456 tdq->tdq_transferable++;
457 ts->ts_flags |= TSF_XFERABLE;
458 }
459 if (pri < PRI_MIN_BATCH) {
460 ts->ts_runq = &tdq->tdq_realtime;
461 } else if (pri <= PRI_MAX_BATCH) {
462 ts->ts_runq = &tdq->tdq_timeshare;
463 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
464 ("Invalid priority %d on timeshare runq", pri));
465 /*
466 * This queue contains only priorities between MIN and MAX
467 * realtime. Use the whole queue to represent these values.
468 */
469 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
470 pri = RQ_NQS * (pri - PRI_MIN_BATCH) /
471 (PRI_MAX_BATCH - PRI_MIN_BATCH + 1);
472 pri = (pri + tdq->tdq_idx) % RQ_NQS;
473 /*
474 * This effectively shortens the queue by one so we
475 * can have a one slot difference between idx and
476 * ridx while we wait for threads to drain.
477 */
478 if (tdq->tdq_ridx != tdq->tdq_idx &&
479 pri == tdq->tdq_ridx)
480 pri = (unsigned char)(pri - 1) % RQ_NQS;
481 } else
482 pri = tdq->tdq_ridx;
483 runq_add_pri(ts->ts_runq, td, pri, flags);
484 return;
485 } else
486 ts->ts_runq = &tdq->tdq_idle;
487 runq_add(ts->ts_runq, td, flags);
488 }
489
490 /*
491 * Remove a thread from a run-queue. This typically happens when a thread
492 * is selected to run. Running threads are not on the queue and the
493 * transferable count does not reflect them.
494 */
495 static __inline void
496 tdq_runq_rem(struct tdq *tdq, struct thread *td)
497 {
498 struct td_sched *ts;
499
500 ts = td->td_sched;
501 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
502 KASSERT(ts->ts_runq != NULL,
503 ("tdq_runq_remove: thread %p null ts_runq", td));
504 if (ts->ts_flags & TSF_XFERABLE) {
505 tdq->tdq_transferable--;
506 ts->ts_flags &= ~TSF_XFERABLE;
507 }
508 if (ts->ts_runq == &tdq->tdq_timeshare) {
509 if (tdq->tdq_idx != tdq->tdq_ridx)
510 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
511 else
512 runq_remove_idx(ts->ts_runq, td, NULL);
513 } else
514 runq_remove(ts->ts_runq, td);
515 }
516
517 /*
518 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
519 * for this thread to the referenced thread queue.
520 */
521 static void
522 tdq_load_add(struct tdq *tdq, struct thread *td)
523 {
524
525 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
526 THREAD_LOCK_ASSERT(td, MA_OWNED);
527
528 tdq->tdq_load++;
529 if ((td->td_flags & TDF_NOLOAD) == 0)
530 tdq->tdq_sysload++;
531 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
532 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
533 }
534
535 /*
536 * Remove the load from a thread that is transitioning to a sleep state or
537 * exiting.
538 */
539 static void
540 tdq_load_rem(struct tdq *tdq, struct thread *td)
541 {
542
543 THREAD_LOCK_ASSERT(td, MA_OWNED);
544 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
545 KASSERT(tdq->tdq_load != 0,
546 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
547
548 tdq->tdq_load--;
549 if ((td->td_flags & TDF_NOLOAD) == 0)
550 tdq->tdq_sysload--;
551 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
552 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
553 }
554
555 /*
556 * Set lowpri to its exact value by searching the run-queue and
557 * evaluating curthread. curthread may be passed as an optimization.
558 */
559 static void
560 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
561 {
562 struct thread *td;
563
564 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
565 if (ctd == NULL)
566 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
567 td = tdq_choose(tdq);
568 if (td == NULL || td->td_priority > ctd->td_priority)
569 tdq->tdq_lowpri = ctd->td_priority;
570 else
571 tdq->tdq_lowpri = td->td_priority;
572 }
573
574 #ifdef SMP
575 struct cpu_search {
576 cpuset_t cs_mask;
577 u_int cs_prefer;
578 int cs_pri; /* Min priority for low. */
579 int cs_limit; /* Max load for low, min load for high. */
580 int cs_cpu;
581 int cs_load;
582 };
583
584 #define CPU_SEARCH_LOWEST 0x1
585 #define CPU_SEARCH_HIGHEST 0x2
586 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
587
588 #define CPUSET_FOREACH(cpu, mask) \
589 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \
590 if ((mask) & 1 << (cpu))
591
592 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
593 struct cpu_search *high, const int match);
594 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
595 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
596 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
597 struct cpu_search *high);
598
599 /*
600 * Search the tree of cpu_groups for the lowest or highest loaded cpu
601 * according to the match argument. This routine actually compares the
602 * load on all paths through the tree and finds the least loaded cpu on
603 * the least loaded path, which may differ from the least loaded cpu in
604 * the system. This balances work among caches and busses.
605 *
606 * This inline is instantiated in three forms below using constants for the
607 * match argument. It is reduced to the minimum set for each case. It is
608 * also recursive to the depth of the tree.
609 */
610 static __inline int
611 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
612 struct cpu_search *high, const int match)
613 {
614 struct cpu_search lgroup;
615 struct cpu_search hgroup;
616 cpumask_t cpumask;
617 struct cpu_group *child;
618 struct tdq *tdq;
619 int cpu, i, hload, lload, load, total, rnd, *rndptr;
620
621 total = 0;
622 cpumask = cg->cg_mask;
623 if (match & CPU_SEARCH_LOWEST) {
624 lload = INT_MAX;
625 lgroup = *low;
626 }
627 if (match & CPU_SEARCH_HIGHEST) {
628 hload = INT_MIN;
629 hgroup = *high;
630 }
631
632 /* Iterate through the child CPU groups and then remaining CPUs. */
633 for (i = cg->cg_children, cpu = mp_maxid; i >= 0; ) {
634 if (i == 0) {
635 while (cpu >= 0 && ((1 << cpu) & cpumask) == 0)
636 cpu--;
637 if (cpu < 0)
638 break;
639 child = NULL;
640 } else
641 child = &cg->cg_child[i - 1];
642
643 if (match & CPU_SEARCH_LOWEST)
644 lgroup.cs_cpu = -1;
645 if (match & CPU_SEARCH_HIGHEST)
646 hgroup.cs_cpu = -1;
647 if (child) { /* Handle child CPU group. */
648 cpumask &= ~child->cg_mask;
649 switch (match) {
650 case CPU_SEARCH_LOWEST:
651 load = cpu_search_lowest(child, &lgroup);
652 break;
653 case CPU_SEARCH_HIGHEST:
654 load = cpu_search_highest(child, &hgroup);
655 break;
656 case CPU_SEARCH_BOTH:
657 load = cpu_search_both(child, &lgroup, &hgroup);
658 break;
659 }
660 } else { /* Handle child CPU. */
661 tdq = TDQ_CPU(cpu);
662 load = tdq->tdq_load * 256;
663 rndptr = DPCPU_PTR(randomval);
664 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
665 if (match & CPU_SEARCH_LOWEST) {
666 if (cpu == low->cs_prefer)
667 load -= 64;
668 /* If that CPU is allowed and get data. */
669 if (tdq->tdq_lowpri > lgroup.cs_pri &&
670 tdq->tdq_load <= lgroup.cs_limit &&
671 CPU_ISSET(cpu, &lgroup.cs_mask)) {
672 lgroup.cs_cpu = cpu;
673 lgroup.cs_load = load - rnd;
674 }
675 }
676 if (match & CPU_SEARCH_HIGHEST)
677 if (tdq->tdq_load >= hgroup.cs_limit &&
678 tdq->tdq_transferable &&
679 CPU_ISSET(cpu, &hgroup.cs_mask)) {
680 hgroup.cs_cpu = cpu;
681 hgroup.cs_load = load - rnd;
682 }
683 }
684 total += load;
685
686 /* We have info about child item. Compare it. */
687 if (match & CPU_SEARCH_LOWEST) {
688 if (lgroup.cs_cpu >= 0 &&
689 (load < lload ||
690 (load == lload && lgroup.cs_load < low->cs_load))) {
691 lload = load;
692 low->cs_cpu = lgroup.cs_cpu;
693 low->cs_load = lgroup.cs_load;
694 }
695 }
696 if (match & CPU_SEARCH_HIGHEST)
697 if (hgroup.cs_cpu >= 0 &&
698 (load > hload ||
699 (load == hload && hgroup.cs_load > high->cs_load))) {
700 hload = load;
701 high->cs_cpu = hgroup.cs_cpu;
702 high->cs_load = hgroup.cs_load;
703 }
704 if (child) {
705 i--;
706 if (i == 0 && cpumask == 0)
707 break;
708 } else
709 cpu--;
710 }
711 return (total);
712 }
713
714 /*
715 * cpu_search instantiations must pass constants to maintain the inline
716 * optimization.
717 */
718 int
719 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
720 {
721 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
722 }
723
724 int
725 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
726 {
727 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
728 }
729
730 int
731 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
732 struct cpu_search *high)
733 {
734 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
735 }
736
737 /*
738 * Find the cpu with the least load via the least loaded path that has a
739 * lowpri greater than pri pri. A pri of -1 indicates any priority is
740 * acceptable.
741 */
742 static inline int
743 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
744 int prefer)
745 {
746 struct cpu_search low;
747
748 low.cs_cpu = -1;
749 low.cs_prefer = prefer;
750 low.cs_mask = mask;
751 low.cs_pri = pri;
752 low.cs_limit = maxload;
753 cpu_search_lowest(cg, &low);
754 return low.cs_cpu;
755 }
756
757 /*
758 * Find the cpu with the highest load via the highest loaded path.
759 */
760 static inline int
761 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
762 {
763 struct cpu_search high;
764
765 high.cs_cpu = -1;
766 high.cs_mask = mask;
767 high.cs_limit = minload;
768 cpu_search_highest(cg, &high);
769 return high.cs_cpu;
770 }
771
772 /*
773 * Simultaneously find the highest and lowest loaded cpu reachable via
774 * cg.
775 */
776 static inline void
777 sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
778 {
779 struct cpu_search high;
780 struct cpu_search low;
781
782 low.cs_cpu = -1;
783 low.cs_prefer = -1;
784 low.cs_pri = -1;
785 low.cs_limit = INT_MAX;
786 low.cs_mask = mask;
787 high.cs_cpu = -1;
788 high.cs_limit = -1;
789 high.cs_mask = mask;
790 cpu_search_both(cg, &low, &high);
791 *lowcpu = low.cs_cpu;
792 *highcpu = high.cs_cpu;
793 return;
794 }
795
796 static void
797 sched_balance_group(struct cpu_group *cg)
798 {
799 cpuset_t hmask, lmask;
800 int high, low, anylow;
801
802 CPU_FILL(&hmask);
803 for (;;) {
804 high = sched_highest(cg, hmask, 1);
805 /* Stop if there is no more CPU with transferrable threads. */
806 if (high == -1)
807 break;
808 CPU_CLR(high, &hmask);
809 CPU_COPY(&hmask, &lmask);
810 /* Stop if there is no more CPU left for low. */
811 if (CPU_EMPTY(&lmask))
812 break;
813 anylow = 1;
814 nextlow:
815 low = sched_lowest(cg, lmask, -1,
816 TDQ_CPU(high)->tdq_load - 1, high);
817 /* Stop if we looked well and found no less loaded CPU. */
818 if (anylow && low == -1)
819 break;
820 /* Go to next high if we found no less loaded CPU. */
821 if (low == -1)
822 continue;
823 /* Transfer thread from high to low. */
824 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
825 /* CPU that got thread can no longer be a donor. */
826 CPU_CLR(low, &hmask);
827 } else {
828 /*
829 * If failed, then there is no threads on high
830 * that can run on this low. Drop low from low
831 * mask and look for different one.
832 */
833 CPU_CLR(low, &lmask);
834 anylow = 0;
835 goto nextlow;
836 }
837 }
838 }
839
840 static void
841 sched_balance(void)
842 {
843 struct tdq *tdq;
844
845 /*
846 * Select a random time between .5 * balance_interval and
847 * 1.5 * balance_interval.
848 */
849 balance_ticks = max(balance_interval / 2, 1);
850 balance_ticks += random() % balance_interval;
851 if (smp_started == 0 || rebalance == 0)
852 return;
853 tdq = TDQ_SELF();
854 TDQ_UNLOCK(tdq);
855 sched_balance_group(cpu_top);
856 TDQ_LOCK(tdq);
857 }
858
859 /*
860 * Lock two thread queues using their address to maintain lock order.
861 */
862 static void
863 tdq_lock_pair(struct tdq *one, struct tdq *two)
864 {
865 if (one < two) {
866 TDQ_LOCK(one);
867 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
868 } else {
869 TDQ_LOCK(two);
870 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
871 }
872 }
873
874 /*
875 * Unlock two thread queues. Order is not important here.
876 */
877 static void
878 tdq_unlock_pair(struct tdq *one, struct tdq *two)
879 {
880 TDQ_UNLOCK(one);
881 TDQ_UNLOCK(two);
882 }
883
884 /*
885 * Transfer load between two imbalanced thread queues.
886 */
887 static int
888 sched_balance_pair(struct tdq *high, struct tdq *low)
889 {
890 int moved;
891
892 tdq_lock_pair(high, low);
893 moved = 0;
894 /*
895 * Determine what the imbalance is and then adjust that to how many
896 * threads we actually have to give up (transferable).
897 */
898 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
899 (moved = tdq_move(high, low)) > 0) {
900 /*
901 * IPI the target cpu to force it to reschedule with the new
902 * workload.
903 */
904 ipi_cpu(TDQ_ID(low), IPI_PREEMPT);
905 }
906 tdq_unlock_pair(high, low);
907 return (moved);
908 }
909
910 /*
911 * Move a thread from one thread queue to another.
912 */
913 static int
914 tdq_move(struct tdq *from, struct tdq *to)
915 {
916 struct td_sched *ts;
917 struct thread *td;
918 struct tdq *tdq;
919 int cpu;
920
921 TDQ_LOCK_ASSERT(from, MA_OWNED);
922 TDQ_LOCK_ASSERT(to, MA_OWNED);
923
924 tdq = from;
925 cpu = TDQ_ID(to);
926 td = tdq_steal(tdq, cpu);
927 if (td == NULL)
928 return (0);
929 ts = td->td_sched;
930 /*
931 * Although the run queue is locked the thread may be blocked. Lock
932 * it to clear this and acquire the run-queue lock.
933 */
934 thread_lock(td);
935 /* Drop recursive lock on from acquired via thread_lock(). */
936 TDQ_UNLOCK(from);
937 sched_rem(td);
938 ts->ts_cpu = cpu;
939 td->td_lock = TDQ_LOCKPTR(to);
940 tdq_add(to, td, SRQ_YIELDING);
941 return (1);
942 }
943
944 /*
945 * This tdq has idled. Try to steal a thread from another cpu and switch
946 * to it.
947 */
948 static int
949 tdq_idled(struct tdq *tdq)
950 {
951 struct cpu_group *cg;
952 struct tdq *steal;
953 cpuset_t mask;
954 int thresh;
955 int cpu;
956
957 if (smp_started == 0 || steal_idle == 0)
958 return (1);
959 CPU_FILL(&mask);
960 CPU_CLR(PCPU_GET(cpuid), &mask);
961 /* We don't want to be preempted while we're iterating. */
962 spinlock_enter();
963 for (cg = tdq->tdq_cg; cg != NULL; ) {
964 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
965 thresh = steal_thresh;
966 else
967 thresh = 1;
968 cpu = sched_highest(cg, mask, thresh);
969 if (cpu == -1) {
970 cg = cg->cg_parent;
971 continue;
972 }
973 steal = TDQ_CPU(cpu);
974 CPU_CLR(cpu, &mask);
975 tdq_lock_pair(tdq, steal);
976 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
977 tdq_unlock_pair(tdq, steal);
978 continue;
979 }
980 /*
981 * If a thread was added while interrupts were disabled don't
982 * steal one here. If we fail to acquire one due to affinity
983 * restrictions loop again with this cpu removed from the
984 * set.
985 */
986 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
987 tdq_unlock_pair(tdq, steal);
988 continue;
989 }
990 spinlock_exit();
991 TDQ_UNLOCK(steal);
992 mi_switch(SW_VOL | SWT_IDLE, NULL);
993 thread_unlock(curthread);
994
995 return (0);
996 }
997 spinlock_exit();
998 return (1);
999 }
1000
1001 /*
1002 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1003 */
1004 static void
1005 tdq_notify(struct tdq *tdq, struct thread *td)
1006 {
1007 struct thread *ctd;
1008 int pri;
1009 int cpu;
1010
1011 if (tdq->tdq_ipipending)
1012 return;
1013 cpu = td->td_sched->ts_cpu;
1014 pri = td->td_priority;
1015 ctd = pcpu_find(cpu)->pc_curthread;
1016 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1017 return;
1018 if (TD_IS_IDLETHREAD(ctd)) {
1019 /*
1020 * If the MD code has an idle wakeup routine try that before
1021 * falling back to IPI.
1022 */
1023 if (cpu_idle_wakeup(cpu))
1024 return;
1025 }
1026 tdq->tdq_ipipending = 1;
1027 ipi_cpu(cpu, IPI_PREEMPT);
1028 }
1029
1030 /*
1031 * Steals load from a timeshare queue. Honors the rotating queue head
1032 * index.
1033 */
1034 static struct thread *
1035 runq_steal_from(struct runq *rq, int cpu, u_char start)
1036 {
1037 struct rqbits *rqb;
1038 struct rqhead *rqh;
1039 struct thread *td, *first;
1040 int bit;
1041 int pri;
1042 int i;
1043
1044 rqb = &rq->rq_status;
1045 bit = start & (RQB_BPW -1);
1046 pri = 0;
1047 first = NULL;
1048 again:
1049 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1050 if (rqb->rqb_bits[i] == 0)
1051 continue;
1052 if (bit != 0) {
1053 for (pri = bit; pri < RQB_BPW; pri++)
1054 if (rqb->rqb_bits[i] & (1ul << pri))
1055 break;
1056 if (pri >= RQB_BPW)
1057 continue;
1058 } else
1059 pri = RQB_FFS(rqb->rqb_bits[i]);
1060 pri += (i << RQB_L2BPW);
1061 rqh = &rq->rq_queues[pri];
1062 TAILQ_FOREACH(td, rqh, td_runq) {
1063 if (first && THREAD_CAN_MIGRATE(td) &&
1064 THREAD_CAN_SCHED(td, cpu))
1065 return (td);
1066 first = td;
1067 }
1068 }
1069 if (start != 0) {
1070 start = 0;
1071 goto again;
1072 }
1073
1074 if (first && THREAD_CAN_MIGRATE(first) &&
1075 THREAD_CAN_SCHED(first, cpu))
1076 return (first);
1077 return (NULL);
1078 }
1079
1080 /*
1081 * Steals load from a standard linear queue.
1082 */
1083 static struct thread *
1084 runq_steal(struct runq *rq, int cpu)
1085 {
1086 struct rqhead *rqh;
1087 struct rqbits *rqb;
1088 struct thread *td;
1089 int word;
1090 int bit;
1091
1092 rqb = &rq->rq_status;
1093 for (word = 0; word < RQB_LEN; word++) {
1094 if (rqb->rqb_bits[word] == 0)
1095 continue;
1096 for (bit = 0; bit < RQB_BPW; bit++) {
1097 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1098 continue;
1099 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1100 TAILQ_FOREACH(td, rqh, td_runq)
1101 if (THREAD_CAN_MIGRATE(td) &&
1102 THREAD_CAN_SCHED(td, cpu))
1103 return (td);
1104 }
1105 }
1106 return (NULL);
1107 }
1108
1109 /*
1110 * Attempt to steal a thread in priority order from a thread queue.
1111 */
1112 static struct thread *
1113 tdq_steal(struct tdq *tdq, int cpu)
1114 {
1115 struct thread *td;
1116
1117 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1118 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1119 return (td);
1120 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1121 cpu, tdq->tdq_ridx)) != NULL)
1122 return (td);
1123 return (runq_steal(&tdq->tdq_idle, cpu));
1124 }
1125
1126 /*
1127 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1128 * current lock and returns with the assigned queue locked.
1129 */
1130 static inline struct tdq *
1131 sched_setcpu(struct thread *td, int cpu, int flags)
1132 {
1133
1134 struct tdq *tdq;
1135
1136 THREAD_LOCK_ASSERT(td, MA_OWNED);
1137 tdq = TDQ_CPU(cpu);
1138 td->td_sched->ts_cpu = cpu;
1139 /*
1140 * If the lock matches just return the queue.
1141 */
1142 if (td->td_lock == TDQ_LOCKPTR(tdq))
1143 return (tdq);
1144 #ifdef notyet
1145 /*
1146 * If the thread isn't running its lockptr is a
1147 * turnstile or a sleepqueue. We can just lock_set without
1148 * blocking.
1149 */
1150 if (TD_CAN_RUN(td)) {
1151 TDQ_LOCK(tdq);
1152 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1153 return (tdq);
1154 }
1155 #endif
1156 /*
1157 * The hard case, migration, we need to block the thread first to
1158 * prevent order reversals with other cpus locks.
1159 */
1160 spinlock_enter();
1161 thread_lock_block(td);
1162 TDQ_LOCK(tdq);
1163 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1164 spinlock_exit();
1165 return (tdq);
1166 }
1167
1168 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1169 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1170 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1171 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1172 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1173 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1174
1175 static int
1176 sched_pickcpu(struct thread *td, int flags)
1177 {
1178 struct cpu_group *cg, *ccg;
1179 struct td_sched *ts;
1180 struct tdq *tdq;
1181 cpuset_t mask;
1182 int cpu, pri, self;
1183
1184 self = PCPU_GET(cpuid);
1185 ts = td->td_sched;
1186 if (smp_started == 0)
1187 return (self);
1188 /*
1189 * Don't migrate a running thread from sched_switch().
1190 */
1191 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1192 return (ts->ts_cpu);
1193 /*
1194 * Prefer to run interrupt threads on the processors that generate
1195 * the interrupt.
1196 */
1197 pri = td->td_priority;
1198 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1199 curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1200 SCHED_STAT_INC(pickcpu_intrbind);
1201 ts->ts_cpu = self;
1202 if (TDQ_CPU(self)->tdq_lowpri > pri) {
1203 SCHED_STAT_INC(pickcpu_affinity);
1204 return (ts->ts_cpu);
1205 }
1206 }
1207 /*
1208 * If the thread can run on the last cpu and the affinity has not
1209 * expired or it is idle run it there.
1210 */
1211 tdq = TDQ_CPU(ts->ts_cpu);
1212 cg = tdq->tdq_cg;
1213 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1214 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1215 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1216 if (cg->cg_flags & CG_FLAG_THREAD) {
1217 CPUSET_FOREACH(cpu, cg->cg_mask) {
1218 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1219 break;
1220 }
1221 } else
1222 cpu = INT_MAX;
1223 if (cpu > mp_maxid) {
1224 SCHED_STAT_INC(pickcpu_idle_affinity);
1225 return (ts->ts_cpu);
1226 }
1227 }
1228 /*
1229 * Search for the last level cache CPU group in the tree.
1230 * Skip caches with expired affinity time and SMT groups.
1231 * Affinity to higher level caches will be handled less aggressively.
1232 */
1233 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1234 if (cg->cg_flags & CG_FLAG_THREAD)
1235 continue;
1236 if (!SCHED_AFFINITY(ts, cg->cg_level))
1237 continue;
1238 ccg = cg;
1239 }
1240 if (ccg != NULL)
1241 cg = ccg;
1242 cpu = -1;
1243 /* Search the group for the less loaded idle CPU we can run now. */
1244 mask = td->td_cpuset->cs_mask;
1245 if (cg != NULL && cg != cpu_top &&
1246 cg->cg_mask != cpu_top->cg_mask)
1247 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1248 INT_MAX, ts->ts_cpu);
1249 /* Search globally for the less loaded CPU we can run now. */
1250 if (cpu == -1)
1251 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1252 /* Search globally for the less loaded CPU. */
1253 if (cpu == -1)
1254 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1255 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1256 /*
1257 * Compare the lowest loaded cpu to current cpu.
1258 */
1259 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1260 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1261 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1262 SCHED_STAT_INC(pickcpu_local);
1263 cpu = self;
1264 } else
1265 SCHED_STAT_INC(pickcpu_lowest);
1266 if (cpu != ts->ts_cpu)
1267 SCHED_STAT_INC(pickcpu_migration);
1268 return (cpu);
1269 }
1270 #endif
1271
1272 /*
1273 * Pick the highest priority task we have and return it.
1274 */
1275 static struct thread *
1276 tdq_choose(struct tdq *tdq)
1277 {
1278 struct thread *td;
1279
1280 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1281 td = runq_choose(&tdq->tdq_realtime);
1282 if (td != NULL)
1283 return (td);
1284 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1285 if (td != NULL) {
1286 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1287 ("tdq_choose: Invalid priority on timeshare queue %d",
1288 td->td_priority));
1289 return (td);
1290 }
1291 td = runq_choose(&tdq->tdq_idle);
1292 if (td != NULL) {
1293 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1294 ("tdq_choose: Invalid priority on idle queue %d",
1295 td->td_priority));
1296 return (td);
1297 }
1298
1299 return (NULL);
1300 }
1301
1302 /*
1303 * Initialize a thread queue.
1304 */
1305 static void
1306 tdq_setup(struct tdq *tdq)
1307 {
1308
1309 if (bootverbose)
1310 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1311 runq_init(&tdq->tdq_realtime);
1312 runq_init(&tdq->tdq_timeshare);
1313 runq_init(&tdq->tdq_idle);
1314 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1315 "sched lock %d", (int)TDQ_ID(tdq));
1316 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1317 MTX_SPIN | MTX_RECURSE);
1318 #ifdef KTR
1319 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1320 "CPU %d load", (int)TDQ_ID(tdq));
1321 #endif
1322 }
1323
1324 #ifdef SMP
1325 static void
1326 sched_setup_smp(void)
1327 {
1328 struct tdq *tdq;
1329 int i;
1330
1331 cpu_top = smp_topo();
1332 CPU_FOREACH(i) {
1333 tdq = TDQ_CPU(i);
1334 tdq_setup(tdq);
1335 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1336 if (tdq->tdq_cg == NULL)
1337 panic("Can't find cpu group for %d\n", i);
1338 }
1339 balance_tdq = TDQ_SELF();
1340 sched_balance();
1341 }
1342 #endif
1343
1344 /*
1345 * Setup the thread queues and initialize the topology based on MD
1346 * information.
1347 */
1348 static void
1349 sched_setup(void *dummy)
1350 {
1351 struct tdq *tdq;
1352
1353 tdq = TDQ_SELF();
1354 #ifdef SMP
1355 sched_setup_smp();
1356 #else
1357 tdq_setup(tdq);
1358 #endif
1359
1360 /* Add thread0's load since it's running. */
1361 TDQ_LOCK(tdq);
1362 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1363 tdq_load_add(tdq, &thread0);
1364 tdq->tdq_lowpri = thread0.td_priority;
1365 TDQ_UNLOCK(tdq);
1366 }
1367
1368 /*
1369 * This routine determines time constants after stathz and hz are setup.
1370 */
1371 /* ARGSUSED */
1372 static void
1373 sched_initticks(void *dummy)
1374 {
1375 int incr;
1376
1377 realstathz = stathz ? stathz : hz;
1378 sched_slice = realstathz / 10; /* ~100ms */
1379 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1380 realstathz);
1381
1382 /*
1383 * tickincr is shifted out by 10 to avoid rounding errors due to
1384 * hz not being evenly divisible by stathz on all platforms.
1385 */
1386 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1387 /*
1388 * This does not work for values of stathz that are more than
1389 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1390 */
1391 if (incr == 0)
1392 incr = 1;
1393 tickincr = incr;
1394 #ifdef SMP
1395 /*
1396 * Set the default balance interval now that we know
1397 * what realstathz is.
1398 */
1399 balance_interval = realstathz;
1400 /*
1401 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1402 * This prevents excess thrashing on large machines and excess idle
1403 * on smaller machines.
1404 */
1405 steal_thresh = min(fls(mp_ncpus) - 1, 3);
1406 affinity = SCHED_AFFINITY_DEFAULT;
1407 #endif
1408 }
1409
1410
1411 /*
1412 * This is the core of the interactivity algorithm. Determines a score based
1413 * on past behavior. It is the ratio of sleep time to run time scaled to
1414 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1415 * differs from the cpu usage because it does not account for time spent
1416 * waiting on a run-queue. Would be prettier if we had floating point.
1417 */
1418 static int
1419 sched_interact_score(struct thread *td)
1420 {
1421 struct td_sched *ts;
1422 int div;
1423
1424 ts = td->td_sched;
1425 /*
1426 * The score is only needed if this is likely to be an interactive
1427 * task. Don't go through the expense of computing it if there's
1428 * no chance.
1429 */
1430 if (sched_interact <= SCHED_INTERACT_HALF &&
1431 ts->ts_runtime >= ts->ts_slptime)
1432 return (SCHED_INTERACT_HALF);
1433
1434 if (ts->ts_runtime > ts->ts_slptime) {
1435 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1436 return (SCHED_INTERACT_HALF +
1437 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1438 }
1439 if (ts->ts_slptime > ts->ts_runtime) {
1440 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1441 return (ts->ts_runtime / div);
1442 }
1443 /* runtime == slptime */
1444 if (ts->ts_runtime)
1445 return (SCHED_INTERACT_HALF);
1446
1447 /*
1448 * This can happen if slptime and runtime are 0.
1449 */
1450 return (0);
1451
1452 }
1453
1454 /*
1455 * Scale the scheduling priority according to the "interactivity" of this
1456 * process.
1457 */
1458 static void
1459 sched_priority(struct thread *td)
1460 {
1461 int score;
1462 int pri;
1463
1464 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1465 return;
1466 /*
1467 * If the score is interactive we place the thread in the realtime
1468 * queue with a priority that is less than kernel and interrupt
1469 * priorities. These threads are not subject to nice restrictions.
1470 *
1471 * Scores greater than this are placed on the normal timeshare queue
1472 * where the priority is partially decided by the most recent cpu
1473 * utilization and the rest is decided by nice value.
1474 *
1475 * The nice value of the process has a linear effect on the calculated
1476 * score. Negative nice values make it easier for a thread to be
1477 * considered interactive.
1478 */
1479 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1480 if (score < sched_interact) {
1481 pri = PRI_MIN_INTERACT;
1482 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1483 sched_interact) * score;
1484 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1485 ("sched_priority: invalid interactive priority %d score %d",
1486 pri, score));
1487 } else {
1488 pri = SCHED_PRI_MIN;
1489 if (td->td_sched->ts_ticks)
1490 pri += min(SCHED_PRI_TICKS(td->td_sched),
1491 SCHED_PRI_RANGE - 1);
1492 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1493 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1494 ("sched_priority: invalid priority %d: nice %d, "
1495 "ticks %d ftick %d ltick %d tick pri %d",
1496 pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1497 td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1498 SCHED_PRI_TICKS(td->td_sched)));
1499 }
1500 sched_user_prio(td, pri);
1501
1502 return;
1503 }
1504
1505 /*
1506 * This routine enforces a maximum limit on the amount of scheduling history
1507 * kept. It is called after either the slptime or runtime is adjusted. This
1508 * function is ugly due to integer math.
1509 */
1510 static void
1511 sched_interact_update(struct thread *td)
1512 {
1513 struct td_sched *ts;
1514 u_int sum;
1515
1516 ts = td->td_sched;
1517 sum = ts->ts_runtime + ts->ts_slptime;
1518 if (sum < SCHED_SLP_RUN_MAX)
1519 return;
1520 /*
1521 * This only happens from two places:
1522 * 1) We have added an unusual amount of run time from fork_exit.
1523 * 2) We have added an unusual amount of sleep time from sched_sleep().
1524 */
1525 if (sum > SCHED_SLP_RUN_MAX * 2) {
1526 if (ts->ts_runtime > ts->ts_slptime) {
1527 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1528 ts->ts_slptime = 1;
1529 } else {
1530 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1531 ts->ts_runtime = 1;
1532 }
1533 return;
1534 }
1535 /*
1536 * If we have exceeded by more than 1/5th then the algorithm below
1537 * will not bring us back into range. Dividing by two here forces
1538 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1539 */
1540 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1541 ts->ts_runtime /= 2;
1542 ts->ts_slptime /= 2;
1543 return;
1544 }
1545 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1546 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1547 }
1548
1549 /*
1550 * Scale back the interactivity history when a child thread is created. The
1551 * history is inherited from the parent but the thread may behave totally
1552 * differently. For example, a shell spawning a compiler process. We want
1553 * to learn that the compiler is behaving badly very quickly.
1554 */
1555 static void
1556 sched_interact_fork(struct thread *td)
1557 {
1558 int ratio;
1559 int sum;
1560
1561 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1562 if (sum > SCHED_SLP_RUN_FORK) {
1563 ratio = sum / SCHED_SLP_RUN_FORK;
1564 td->td_sched->ts_runtime /= ratio;
1565 td->td_sched->ts_slptime /= ratio;
1566 }
1567 }
1568
1569 /*
1570 * Called from proc0_init() to setup the scheduler fields.
1571 */
1572 void
1573 schedinit(void)
1574 {
1575
1576 /*
1577 * Set up the scheduler specific parts of proc0.
1578 */
1579 proc0.p_sched = NULL; /* XXX */
1580 thread0.td_sched = &td_sched0;
1581 td_sched0.ts_ltick = ticks;
1582 td_sched0.ts_ftick = ticks;
1583 td_sched0.ts_slice = sched_slice;
1584 }
1585
1586 /*
1587 * This is only somewhat accurate since given many processes of the same
1588 * priority they will switch when their slices run out, which will be
1589 * at most sched_slice stathz ticks.
1590 */
1591 int
1592 sched_rr_interval(void)
1593 {
1594
1595 /* Convert sched_slice from stathz to hz. */
1596 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1597 }
1598
1599 /*
1600 * Update the percent cpu tracking information when it is requested or
1601 * the total history exceeds the maximum. We keep a sliding history of
1602 * tick counts that slowly decays. This is less precise than the 4BSD
1603 * mechanism since it happens with less regular and frequent events.
1604 */
1605 static void
1606 sched_pctcpu_update(struct td_sched *ts)
1607 {
1608
1609 if (ts->ts_ticks == 0)
1610 return;
1611 if (ticks - (hz / 10) < ts->ts_ltick &&
1612 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1613 return;
1614 /*
1615 * Adjust counters and watermark for pctcpu calc.
1616 */
1617 if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1618 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1619 SCHED_TICK_TARG;
1620 else
1621 ts->ts_ticks = 0;
1622 ts->ts_ltick = ticks;
1623 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1624 }
1625
1626 /*
1627 * Adjust the priority of a thread. Move it to the appropriate run-queue
1628 * if necessary. This is the back-end for several priority related
1629 * functions.
1630 */
1631 static void
1632 sched_thread_priority(struct thread *td, u_char prio)
1633 {
1634 struct td_sched *ts;
1635 struct tdq *tdq;
1636 int oldpri;
1637
1638 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1639 "prio:%d", td->td_priority, "new prio:%d", prio,
1640 KTR_ATTR_LINKED, sched_tdname(curthread));
1641 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
1642 if (td != curthread && prio < td->td_priority) {
1643 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1644 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1645 prio, KTR_ATTR_LINKED, sched_tdname(td));
1646 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
1647 curthread);
1648 }
1649 ts = td->td_sched;
1650 THREAD_LOCK_ASSERT(td, MA_OWNED);
1651 if (td->td_priority == prio)
1652 return;
1653 /*
1654 * If the priority has been elevated due to priority
1655 * propagation, we may have to move ourselves to a new
1656 * queue. This could be optimized to not re-add in some
1657 * cases.
1658 */
1659 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1660 sched_rem(td);
1661 td->td_priority = prio;
1662 sched_add(td, SRQ_BORROWING);
1663 return;
1664 }
1665 /*
1666 * If the thread is currently running we may have to adjust the lowpri
1667 * information so other cpus are aware of our current priority.
1668 */
1669 if (TD_IS_RUNNING(td)) {
1670 tdq = TDQ_CPU(ts->ts_cpu);
1671 oldpri = td->td_priority;
1672 td->td_priority = prio;
1673 if (prio < tdq->tdq_lowpri)
1674 tdq->tdq_lowpri = prio;
1675 else if (tdq->tdq_lowpri == oldpri)
1676 tdq_setlowpri(tdq, td);
1677 return;
1678 }
1679 td->td_priority = prio;
1680 }
1681
1682 /*
1683 * Update a thread's priority when it is lent another thread's
1684 * priority.
1685 */
1686 void
1687 sched_lend_prio(struct thread *td, u_char prio)
1688 {
1689
1690 td->td_flags |= TDF_BORROWING;
1691 sched_thread_priority(td, prio);
1692 }
1693
1694 /*
1695 * Restore a thread's priority when priority propagation is
1696 * over. The prio argument is the minimum priority the thread
1697 * needs to have to satisfy other possible priority lending
1698 * requests. If the thread's regular priority is less
1699 * important than prio, the thread will keep a priority boost
1700 * of prio.
1701 */
1702 void
1703 sched_unlend_prio(struct thread *td, u_char prio)
1704 {
1705 u_char base_pri;
1706
1707 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1708 td->td_base_pri <= PRI_MAX_TIMESHARE)
1709 base_pri = td->td_user_pri;
1710 else
1711 base_pri = td->td_base_pri;
1712 if (prio >= base_pri) {
1713 td->td_flags &= ~TDF_BORROWING;
1714 sched_thread_priority(td, base_pri);
1715 } else
1716 sched_lend_prio(td, prio);
1717 }
1718
1719 /*
1720 * Standard entry for setting the priority to an absolute value.
1721 */
1722 void
1723 sched_prio(struct thread *td, u_char prio)
1724 {
1725 u_char oldprio;
1726
1727 /* First, update the base priority. */
1728 td->td_base_pri = prio;
1729
1730 /*
1731 * If the thread is borrowing another thread's priority, don't
1732 * ever lower the priority.
1733 */
1734 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1735 return;
1736
1737 /* Change the real priority. */
1738 oldprio = td->td_priority;
1739 sched_thread_priority(td, prio);
1740
1741 /*
1742 * If the thread is on a turnstile, then let the turnstile update
1743 * its state.
1744 */
1745 if (TD_ON_LOCK(td) && oldprio != prio)
1746 turnstile_adjust(td, oldprio);
1747 }
1748
1749 /*
1750 * Set the base user priority, does not effect current running priority.
1751 */
1752 void
1753 sched_user_prio(struct thread *td, u_char prio)
1754 {
1755 u_char oldprio;
1756
1757 td->td_base_user_pri = prio;
1758 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1759 return;
1760 oldprio = td->td_user_pri;
1761 td->td_user_pri = prio;
1762 }
1763
1764 void
1765 sched_lend_user_prio(struct thread *td, u_char prio)
1766 {
1767 u_char oldprio;
1768
1769 THREAD_LOCK_ASSERT(td, MA_OWNED);
1770 td->td_flags |= TDF_UBORROWING;
1771 oldprio = td->td_user_pri;
1772 td->td_user_pri = prio;
1773 }
1774
1775 void
1776 sched_unlend_user_prio(struct thread *td, u_char prio)
1777 {
1778 u_char base_pri;
1779
1780 THREAD_LOCK_ASSERT(td, MA_OWNED);
1781 base_pri = td->td_base_user_pri;
1782 if (prio >= base_pri) {
1783 td->td_flags &= ~TDF_UBORROWING;
1784 sched_user_prio(td, base_pri);
1785 } else {
1786 sched_lend_user_prio(td, prio);
1787 }
1788 }
1789
1790 /*
1791 * Handle migration from sched_switch(). This happens only for
1792 * cpu binding.
1793 */
1794 static struct mtx *
1795 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1796 {
1797 struct tdq *tdn;
1798
1799 tdn = TDQ_CPU(td->td_sched->ts_cpu);
1800 #ifdef SMP
1801 tdq_load_rem(tdq, td);
1802 /*
1803 * Do the lock dance required to avoid LOR. We grab an extra
1804 * spinlock nesting to prevent preemption while we're
1805 * not holding either run-queue lock.
1806 */
1807 spinlock_enter();
1808 thread_lock_block(td); /* This releases the lock on tdq. */
1809
1810 /*
1811 * Acquire both run-queue locks before placing the thread on the new
1812 * run-queue to avoid deadlocks created by placing a thread with a
1813 * blocked lock on the run-queue of a remote processor. The deadlock
1814 * occurs when a third processor attempts to lock the two queues in
1815 * question while the target processor is spinning with its own
1816 * run-queue lock held while waiting for the blocked lock to clear.
1817 */
1818 tdq_lock_pair(tdn, tdq);
1819 tdq_add(tdn, td, flags);
1820 tdq_notify(tdn, td);
1821 TDQ_UNLOCK(tdn);
1822 spinlock_exit();
1823 #endif
1824 return (TDQ_LOCKPTR(tdn));
1825 }
1826
1827 /*
1828 * Variadic version of thread_lock_unblock() that does not assume td_lock
1829 * is blocked.
1830 */
1831 static inline void
1832 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1833 {
1834 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1835 (uintptr_t)mtx);
1836 }
1837
1838 /*
1839 * Switch threads. This function has to handle threads coming in while
1840 * blocked for some reason, running, or idle. It also must deal with
1841 * migrating a thread from one queue to another as running threads may
1842 * be assigned elsewhere via binding.
1843 */
1844 void
1845 sched_switch(struct thread *td, struct thread *newtd, int flags)
1846 {
1847 struct tdq *tdq;
1848 struct td_sched *ts;
1849 struct mtx *mtx;
1850 int srqflag;
1851 int cpuid, preempted;
1852
1853 THREAD_LOCK_ASSERT(td, MA_OWNED);
1854 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1855
1856 cpuid = PCPU_GET(cpuid);
1857 tdq = TDQ_CPU(cpuid);
1858 ts = td->td_sched;
1859 mtx = td->td_lock;
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 tdq->tdq_switchcnt++;
1867 /*
1868 * The lock pointer in an idle thread should never change. Reset it
1869 * to CAN_RUN as well.
1870 */
1871 if (TD_IS_IDLETHREAD(td)) {
1872 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1873 TD_SET_CAN_RUN(td);
1874 } else if (TD_IS_RUNNING(td)) {
1875 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1876 srqflag = preempted ?
1877 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1878 SRQ_OURSELF|SRQ_YIELDING;
1879 #ifdef SMP
1880 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1881 ts->ts_cpu = sched_pickcpu(td, 0);
1882 #endif
1883 if (ts->ts_cpu == cpuid)
1884 tdq_runq_add(tdq, td, srqflag);
1885 else {
1886 KASSERT(THREAD_CAN_MIGRATE(td) ||
1887 (ts->ts_flags & TSF_BOUND) != 0,
1888 ("Thread %p shouldn't migrate", td));
1889 mtx = sched_switch_migrate(tdq, td, srqflag);
1890 }
1891 } else {
1892 /* This thread must be going to sleep. */
1893 TDQ_LOCK(tdq);
1894 mtx = thread_lock_block(td);
1895 tdq_load_rem(tdq, td);
1896 }
1897 /*
1898 * We enter here with the thread blocked and assigned to the
1899 * appropriate cpu run-queue or sleep-queue and with the current
1900 * thread-queue locked.
1901 */
1902 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1903 newtd = choosethread();
1904 /*
1905 * Call the MD code to switch contexts if necessary.
1906 */
1907 if (td != newtd) {
1908 #ifdef HWPMC_HOOKS
1909 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1910 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1911 #endif
1912 SDT_PROBE2(sched, , , off_cpu, newtd, newtd->td_proc);
1913 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1914 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1915
1916 #ifdef KDTRACE_HOOKS
1917 /*
1918 * If DTrace has set the active vtime enum to anything
1919 * other than INACTIVE (0), then it should have set the
1920 * function to call.
1921 */
1922 if (dtrace_vtime_active)
1923 (*dtrace_vtime_switch_func)(newtd);
1924 #endif
1925
1926 cpu_switch(td, newtd, mtx);
1927 /*
1928 * We may return from cpu_switch on a different cpu. However,
1929 * we always return with td_lock pointing to the current cpu's
1930 * run queue lock.
1931 */
1932 cpuid = PCPU_GET(cpuid);
1933 tdq = TDQ_CPU(cpuid);
1934 lock_profile_obtain_lock_success(
1935 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1936
1937 SDT_PROBE0(sched, , , on_cpu);
1938 #ifdef HWPMC_HOOKS
1939 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1940 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1941 #endif
1942 } else {
1943 thread_unblock_switch(td, mtx);
1944 SDT_PROBE0(sched, , , remain_cpu);
1945 }
1946 /*
1947 * Assert that all went well and return.
1948 */
1949 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1950 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1951 td->td_oncpu = cpuid;
1952 }
1953
1954 /*
1955 * Adjust thread priorities as a result of a nice request.
1956 */
1957 void
1958 sched_nice(struct proc *p, int nice)
1959 {
1960 struct thread *td;
1961
1962 PROC_LOCK_ASSERT(p, MA_OWNED);
1963
1964 p->p_nice = nice;
1965 FOREACH_THREAD_IN_PROC(p, td) {
1966 thread_lock(td);
1967 sched_priority(td);
1968 sched_prio(td, td->td_base_user_pri);
1969 thread_unlock(td);
1970 }
1971 }
1972
1973 /*
1974 * Record the sleep time for the interactivity scorer.
1975 */
1976 void
1977 sched_sleep(struct thread *td, int prio)
1978 {
1979
1980 THREAD_LOCK_ASSERT(td, MA_OWNED);
1981
1982 td->td_slptick = ticks;
1983 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1984 td->td_flags |= TDF_CANSWAP;
1985 if (static_boost == 1 && prio)
1986 sched_prio(td, prio);
1987 else if (static_boost && td->td_priority > static_boost)
1988 sched_prio(td, static_boost);
1989 }
1990
1991 /*
1992 * Schedule a thread to resume execution and record how long it voluntarily
1993 * slept. We also update the pctcpu, interactivity, and priority.
1994 */
1995 void
1996 sched_wakeup(struct thread *td)
1997 {
1998 struct td_sched *ts;
1999 int slptick;
2000
2001 THREAD_LOCK_ASSERT(td, MA_OWNED);
2002 ts = td->td_sched;
2003 td->td_flags &= ~TDF_CANSWAP;
2004 /*
2005 * If we slept for more than a tick update our interactivity and
2006 * priority.
2007 */
2008 slptick = td->td_slptick;
2009 td->td_slptick = 0;
2010 if (slptick && slptick != ticks) {
2011 u_int hzticks;
2012
2013 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
2014 ts->ts_slptime += hzticks;
2015 sched_interact_update(td);
2016 sched_pctcpu_update(ts);
2017 }
2018 /* Reset the slice value after we sleep. */
2019 ts->ts_slice = sched_slice;
2020 sched_add(td, SRQ_BORING);
2021 }
2022
2023 /*
2024 * Penalize the parent for creating a new child and initialize the child's
2025 * priority.
2026 */
2027 void
2028 sched_fork(struct thread *td, struct thread *child)
2029 {
2030 THREAD_LOCK_ASSERT(td, MA_OWNED);
2031 sched_fork_thread(td, child);
2032 /*
2033 * Penalize the parent and child for forking.
2034 */
2035 sched_interact_fork(child);
2036 sched_priority(child);
2037 td->td_sched->ts_runtime += tickincr;
2038 sched_interact_update(td);
2039 sched_priority(td);
2040 }
2041
2042 /*
2043 * Fork a new thread, may be within the same process.
2044 */
2045 void
2046 sched_fork_thread(struct thread *td, struct thread *child)
2047 {
2048 struct td_sched *ts;
2049 struct td_sched *ts2;
2050
2051 THREAD_LOCK_ASSERT(td, MA_OWNED);
2052 /*
2053 * Initialize child.
2054 */
2055 ts = td->td_sched;
2056 ts2 = child->td_sched;
2057 child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
2058 child->td_cpuset = cpuset_ref(td->td_cpuset);
2059 ts2->ts_cpu = ts->ts_cpu;
2060 ts2->ts_flags = 0;
2061 /*
2062 * Grab our parents cpu estimation information.
2063 */
2064 ts2->ts_ticks = ts->ts_ticks;
2065 ts2->ts_ltick = ts->ts_ltick;
2066 ts2->ts_incrtick = ts->ts_incrtick;
2067 ts2->ts_ftick = ts->ts_ftick;
2068 /*
2069 * Do not inherit any borrowed priority from the parent.
2070 */
2071 child->td_priority = child->td_base_pri;
2072 /*
2073 * And update interactivity score.
2074 */
2075 ts2->ts_slptime = ts->ts_slptime;
2076 ts2->ts_runtime = ts->ts_runtime;
2077 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
2078 #ifdef KTR
2079 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2080 #endif
2081 }
2082
2083 /*
2084 * Adjust the priority class of a thread.
2085 */
2086 void
2087 sched_class(struct thread *td, int class)
2088 {
2089
2090 THREAD_LOCK_ASSERT(td, MA_OWNED);
2091 if (td->td_pri_class == class)
2092 return;
2093 td->td_pri_class = class;
2094 }
2095
2096 /*
2097 * Return some of the child's priority and interactivity to the parent.
2098 */
2099 void
2100 sched_exit(struct proc *p, struct thread *child)
2101 {
2102 struct thread *td;
2103
2104 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2105 "prio:%d", child->td_priority);
2106 PROC_LOCK_ASSERT(p, MA_OWNED);
2107 td = FIRST_THREAD_IN_PROC(p);
2108 sched_exit_thread(td, child);
2109 }
2110
2111 /*
2112 * Penalize another thread for the time spent on this one. This helps to
2113 * worsen the priority and interactivity of processes which schedule batch
2114 * jobs such as make. This has little effect on the make process itself but
2115 * causes new processes spawned by it to receive worse scores immediately.
2116 */
2117 void
2118 sched_exit_thread(struct thread *td, struct thread *child)
2119 {
2120
2121 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2122 "prio:%d", child->td_priority);
2123 /*
2124 * Give the child's runtime to the parent without returning the
2125 * sleep time as a penalty to the parent. This causes shells that
2126 * launch expensive things to mark their children as expensive.
2127 */
2128 thread_lock(td);
2129 td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2130 sched_interact_update(td);
2131 sched_priority(td);
2132 thread_unlock(td);
2133 }
2134
2135 void
2136 sched_preempt(struct thread *td)
2137 {
2138 struct tdq *tdq;
2139
2140 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2141
2142 thread_lock(td);
2143 tdq = TDQ_SELF();
2144 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2145 tdq->tdq_ipipending = 0;
2146 if (td->td_priority > tdq->tdq_lowpri) {
2147 int flags;
2148
2149 flags = SW_INVOL | SW_PREEMPT;
2150 if (td->td_critnest > 1)
2151 td->td_owepreempt = 1;
2152 else if (TD_IS_IDLETHREAD(td))
2153 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2154 else
2155 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2156 }
2157 thread_unlock(td);
2158 }
2159
2160 /*
2161 * Fix priorities on return to user-space. Priorities may be elevated due
2162 * to static priorities in msleep() or similar.
2163 */
2164 void
2165 sched_userret(struct thread *td)
2166 {
2167 /*
2168 * XXX we cheat slightly on the locking here to avoid locking in
2169 * the usual case. Setting td_priority here is essentially an
2170 * incomplete workaround for not setting it properly elsewhere.
2171 * Now that some interrupt handlers are threads, not setting it
2172 * properly elsewhere can clobber it in the window between setting
2173 * it here and returning to user mode, so don't waste time setting
2174 * it perfectly here.
2175 */
2176 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2177 ("thread with borrowed priority returning to userland"));
2178 if (td->td_priority != td->td_user_pri) {
2179 thread_lock(td);
2180 td->td_priority = td->td_user_pri;
2181 td->td_base_pri = td->td_user_pri;
2182 tdq_setlowpri(TDQ_SELF(), td);
2183 thread_unlock(td);
2184 }
2185 }
2186
2187 /*
2188 * Handle a stathz tick. This is really only relevant for timeshare
2189 * threads.
2190 */
2191 void
2192 sched_clock(struct thread *td)
2193 {
2194 struct tdq *tdq;
2195 struct td_sched *ts;
2196
2197 THREAD_LOCK_ASSERT(td, MA_OWNED);
2198 tdq = TDQ_SELF();
2199 #ifdef SMP
2200 /*
2201 * We run the long term load balancer infrequently on the first cpu.
2202 */
2203 if (balance_tdq == tdq) {
2204 if (balance_ticks && --balance_ticks == 0)
2205 sched_balance();
2206 }
2207 #endif
2208 /*
2209 * Save the old switch count so we have a record of the last ticks
2210 * activity. Initialize the new switch count based on our load.
2211 * If there is some activity seed it to reflect that.
2212 */
2213 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2214 tdq->tdq_switchcnt = tdq->tdq_load;
2215 /*
2216 * Advance the insert index once for each tick to ensure that all
2217 * threads get a chance to run.
2218 */
2219 if (tdq->tdq_idx == tdq->tdq_ridx) {
2220 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2221 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2222 tdq->tdq_ridx = tdq->tdq_idx;
2223 }
2224 ts = td->td_sched;
2225 if (td->td_pri_class & PRI_FIFO_BIT)
2226 return;
2227 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2228 /*
2229 * We used a tick; charge it to the thread so
2230 * that we can compute our interactivity.
2231 */
2232 td->td_sched->ts_runtime += tickincr;
2233 sched_interact_update(td);
2234 sched_priority(td);
2235 }
2236
2237 /*
2238 * Force a context switch if the current thread has used up a full
2239 * time slice (default is 100ms).
2240 */
2241 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
2242 ts->ts_slice = sched_slice;
2243 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2244 }
2245 }
2246
2247 /*
2248 * Called once per hz tick. Used for cpu utilization information. This
2249 * is easier than trying to scale based on stathz.
2250 */
2251 void
2252 sched_tick(void)
2253 {
2254 struct td_sched *ts;
2255
2256 ts = curthread->td_sched;
2257 /*
2258 * Ticks is updated asynchronously on a single cpu. Check here to
2259 * avoid incrementing ts_ticks multiple times in a single tick.
2260 */
2261 if (ts->ts_incrtick == ticks)
2262 return;
2263 /* Adjust ticks for pctcpu */
2264 ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2265 ts->ts_ltick = ticks;
2266 ts->ts_incrtick = ticks;
2267 /*
2268 * Update if we've exceeded our desired tick threshold by over one
2269 * second.
2270 */
2271 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2272 sched_pctcpu_update(ts);
2273 }
2274
2275 /*
2276 * Return whether the current CPU has runnable tasks. Used for in-kernel
2277 * cooperative idle threads.
2278 */
2279 int
2280 sched_runnable(void)
2281 {
2282 struct tdq *tdq;
2283 int load;
2284
2285 load = 1;
2286
2287 tdq = TDQ_SELF();
2288 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2289 if (tdq->tdq_load > 0)
2290 goto out;
2291 } else
2292 if (tdq->tdq_load - 1 > 0)
2293 goto out;
2294 load = 0;
2295 out:
2296 return (load);
2297 }
2298
2299 /*
2300 * Choose the highest priority thread to run. The thread is removed from
2301 * the run-queue while running however the load remains. For SMP we set
2302 * the tdq in the global idle bitmask if it idles here.
2303 */
2304 struct thread *
2305 sched_choose(void)
2306 {
2307 struct thread *td;
2308 struct tdq *tdq;
2309
2310 tdq = TDQ_SELF();
2311 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2312 td = tdq_choose(tdq);
2313 if (td) {
2314 td->td_sched->ts_ltick = ticks;
2315 tdq_runq_rem(tdq, td);
2316 tdq->tdq_lowpri = td->td_priority;
2317 return (td);
2318 }
2319 tdq->tdq_lowpri = PRI_MAX_IDLE;
2320 return (PCPU_GET(idlethread));
2321 }
2322
2323 /*
2324 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2325 * we always request it once we exit a critical section.
2326 */
2327 static inline void
2328 sched_setpreempt(struct thread *td)
2329 {
2330 struct thread *ctd;
2331 int cpri;
2332 int pri;
2333
2334 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2335
2336 ctd = curthread;
2337 pri = td->td_priority;
2338 cpri = ctd->td_priority;
2339 if (pri < cpri)
2340 ctd->td_flags |= TDF_NEEDRESCHED;
2341 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2342 return;
2343 if (!sched_shouldpreempt(pri, cpri, 0))
2344 return;
2345 ctd->td_owepreempt = 1;
2346 }
2347
2348 /*
2349 * Add a thread to a thread queue. Select the appropriate runq and add the
2350 * thread to it. This is the internal function called when the tdq is
2351 * predetermined.
2352 */
2353 void
2354 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2355 {
2356
2357 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2358 KASSERT((td->td_inhibitors == 0),
2359 ("sched_add: trying to run inhibited thread"));
2360 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2361 ("sched_add: bad thread state"));
2362 KASSERT(td->td_flags & TDF_INMEM,
2363 ("sched_add: thread swapped out"));
2364
2365 if (td->td_priority < tdq->tdq_lowpri)
2366 tdq->tdq_lowpri = td->td_priority;
2367 tdq_runq_add(tdq, td, flags);
2368 tdq_load_add(tdq, td);
2369 }
2370
2371 /*
2372 * Select the target thread queue and add a thread to it. Request
2373 * preemption or IPI a remote processor if required.
2374 */
2375 void
2376 sched_add(struct thread *td, int flags)
2377 {
2378 struct tdq *tdq;
2379 #ifdef SMP
2380 int cpu;
2381 #endif
2382
2383 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2384 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2385 sched_tdname(curthread));
2386 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2387 KTR_ATTR_LINKED, sched_tdname(td));
2388 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2389 flags & SRQ_PREEMPTED);
2390 THREAD_LOCK_ASSERT(td, MA_OWNED);
2391 /*
2392 * Recalculate the priority before we select the target cpu or
2393 * run-queue.
2394 */
2395 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2396 sched_priority(td);
2397 #ifdef SMP
2398 /*
2399 * Pick the destination cpu and if it isn't ours transfer to the
2400 * target cpu.
2401 */
2402 cpu = sched_pickcpu(td, flags);
2403 tdq = sched_setcpu(td, cpu, flags);
2404 tdq_add(tdq, td, flags);
2405 if (cpu != PCPU_GET(cpuid)) {
2406 tdq_notify(tdq, td);
2407 return;
2408 }
2409 #else
2410 tdq = TDQ_SELF();
2411 TDQ_LOCK(tdq);
2412 /*
2413 * Now that the thread is moving to the run-queue, set the lock
2414 * to the scheduler's lock.
2415 */
2416 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2417 tdq_add(tdq, td, flags);
2418 #endif
2419 if (!(flags & SRQ_YIELDING))
2420 sched_setpreempt(td);
2421 }
2422
2423 /*
2424 * Remove a thread from a run-queue without running it. This is used
2425 * when we're stealing a thread from a remote queue. Otherwise all threads
2426 * exit by calling sched_exit_thread() and sched_throw() themselves.
2427 */
2428 void
2429 sched_rem(struct thread *td)
2430 {
2431 struct tdq *tdq;
2432
2433 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2434 "prio:%d", td->td_priority);
2435 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2436 tdq = TDQ_CPU(td->td_sched->ts_cpu);
2437 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2438 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2439 KASSERT(TD_ON_RUNQ(td),
2440 ("sched_rem: thread not on run queue"));
2441 tdq_runq_rem(tdq, td);
2442 tdq_load_rem(tdq, td);
2443 TD_SET_CAN_RUN(td);
2444 if (td->td_priority == tdq->tdq_lowpri)
2445 tdq_setlowpri(tdq, NULL);
2446 }
2447
2448 /*
2449 * Fetch cpu utilization information. Updates on demand.
2450 */
2451 fixpt_t
2452 sched_pctcpu(struct thread *td)
2453 {
2454 fixpt_t pctcpu;
2455 struct td_sched *ts;
2456
2457 pctcpu = 0;
2458 ts = td->td_sched;
2459 if (ts == NULL)
2460 return (0);
2461
2462 THREAD_LOCK_ASSERT(td, MA_OWNED);
2463 if (ts->ts_ticks) {
2464 int rtick;
2465
2466 sched_pctcpu_update(ts);
2467 /* How many rtick per second ? */
2468 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2469 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2470 }
2471
2472 return (pctcpu);
2473 }
2474
2475 /*
2476 * Enforce affinity settings for a thread. Called after adjustments to
2477 * cpumask.
2478 */
2479 void
2480 sched_affinity(struct thread *td)
2481 {
2482 #ifdef SMP
2483 struct td_sched *ts;
2484
2485 THREAD_LOCK_ASSERT(td, MA_OWNED);
2486 ts = td->td_sched;
2487 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2488 return;
2489 if (TD_ON_RUNQ(td)) {
2490 sched_rem(td);
2491 sched_add(td, SRQ_BORING);
2492 return;
2493 }
2494 if (!TD_IS_RUNNING(td))
2495 return;
2496 /*
2497 * Force a switch before returning to userspace. If the
2498 * target thread is not running locally send an ipi to force
2499 * the issue.
2500 */
2501 td->td_flags |= TDF_NEEDRESCHED;
2502 if (td != curthread)
2503 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2504 #endif
2505 }
2506
2507 /*
2508 * Bind a thread to a target cpu.
2509 */
2510 void
2511 sched_bind(struct thread *td, int cpu)
2512 {
2513 struct td_sched *ts;
2514
2515 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2516 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2517 ts = td->td_sched;
2518 if (ts->ts_flags & TSF_BOUND)
2519 sched_unbind(td);
2520 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2521 ts->ts_flags |= TSF_BOUND;
2522 sched_pin();
2523 if (PCPU_GET(cpuid) == cpu)
2524 return;
2525 ts->ts_cpu = cpu;
2526 /* When we return from mi_switch we'll be on the correct cpu. */
2527 mi_switch(SW_VOL, NULL);
2528 }
2529
2530 /*
2531 * Release a bound thread.
2532 */
2533 void
2534 sched_unbind(struct thread *td)
2535 {
2536 struct td_sched *ts;
2537
2538 THREAD_LOCK_ASSERT(td, MA_OWNED);
2539 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2540 ts = td->td_sched;
2541 if ((ts->ts_flags & TSF_BOUND) == 0)
2542 return;
2543 ts->ts_flags &= ~TSF_BOUND;
2544 sched_unpin();
2545 }
2546
2547 int
2548 sched_is_bound(struct thread *td)
2549 {
2550 THREAD_LOCK_ASSERT(td, MA_OWNED);
2551 return (td->td_sched->ts_flags & TSF_BOUND);
2552 }
2553
2554 /*
2555 * Basic yield call.
2556 */
2557 void
2558 sched_relinquish(struct thread *td)
2559 {
2560 thread_lock(td);
2561 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2562 thread_unlock(td);
2563 }
2564
2565 /*
2566 * Return the total system load.
2567 */
2568 int
2569 sched_load(void)
2570 {
2571 #ifdef SMP
2572 int total;
2573 int i;
2574
2575 total = 0;
2576 CPU_FOREACH(i)
2577 total += TDQ_CPU(i)->tdq_sysload;
2578 return (total);
2579 #else
2580 return (TDQ_SELF()->tdq_sysload);
2581 #endif
2582 }
2583
2584 int
2585 sched_sizeof_proc(void)
2586 {
2587 return (sizeof(struct proc));
2588 }
2589
2590 int
2591 sched_sizeof_thread(void)
2592 {
2593 return (sizeof(struct thread) + sizeof(struct td_sched));
2594 }
2595
2596 #ifdef SMP
2597 #define TDQ_IDLESPIN(tdq) \
2598 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2599 #else
2600 #define TDQ_IDLESPIN(tdq) 1
2601 #endif
2602
2603 /*
2604 * The actual idle process.
2605 */
2606 void
2607 sched_idletd(void *dummy)
2608 {
2609 struct thread *td;
2610 struct tdq *tdq;
2611 int switchcnt;
2612 int i;
2613
2614 mtx_assert(&Giant, MA_NOTOWNED);
2615 td = curthread;
2616 tdq = TDQ_SELF();
2617 for (;;) {
2618 #ifdef SMP
2619 if (tdq_idled(tdq) == 0)
2620 continue;
2621 #endif
2622 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
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 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2638 if (tdq->tdq_load == 0)
2639 cpu_idle(switchcnt > 1);
2640 if (tdq->tdq_load) {
2641 thread_lock(td);
2642 mi_switch(SW_VOL | SWT_IDLE, NULL);
2643 thread_unlock(td);
2644 }
2645 }
2646 }
2647
2648 /*
2649 * A CPU is entering for the first time or a thread is exiting.
2650 */
2651 void
2652 sched_throw(struct thread *td)
2653 {
2654 struct thread *newtd;
2655 struct tdq *tdq;
2656
2657 tdq = TDQ_SELF();
2658 if (td == NULL) {
2659 /* Correct spinlock nesting and acquire the correct lock. */
2660 TDQ_LOCK(tdq);
2661 spinlock_exit();
2662 PCPU_SET(switchtime, cpu_ticks());
2663 PCPU_SET(switchticks, ticks);
2664 } else {
2665 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2666 tdq_load_rem(tdq, td);
2667 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2668 }
2669 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2670 newtd = choosethread();
2671 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2672 cpu_throw(td, newtd); /* doesn't return */
2673 }
2674
2675 /*
2676 * This is called from fork_exit(). Just acquire the correct locks and
2677 * let fork do the rest of the work.
2678 */
2679 void
2680 sched_fork_exit(struct thread *td)
2681 {
2682 struct td_sched *ts;
2683 struct tdq *tdq;
2684 int cpuid;
2685
2686 /*
2687 * Finish setting up thread glue so that it begins execution in a
2688 * non-nested critical section with the scheduler lock held.
2689 */
2690 cpuid = PCPU_GET(cpuid);
2691 tdq = TDQ_CPU(cpuid);
2692 ts = td->td_sched;
2693 if (TD_IS_IDLETHREAD(td))
2694 td->td_lock = TDQ_LOCKPTR(tdq);
2695 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2696 td->td_oncpu = cpuid;
2697 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2698 lock_profile_obtain_lock_success(
2699 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2700 }
2701
2702 /*
2703 * Create on first use to catch odd startup conditons.
2704 */
2705 char *
2706 sched_tdname(struct thread *td)
2707 {
2708 #ifdef KTR
2709 struct td_sched *ts;
2710
2711 ts = td->td_sched;
2712 if (ts->ts_name[0] == '\0')
2713 snprintf(ts->ts_name, sizeof(ts->ts_name),
2714 "%s tid %d", td->td_name, td->td_tid);
2715 return (ts->ts_name);
2716 #else
2717 return (td->td_name);
2718 #endif
2719 }
2720
2721 #ifdef KTR
2722 void
2723 sched_clear_tdname(struct thread *td)
2724 {
2725 struct td_sched *ts;
2726
2727 ts = td->td_sched;
2728 ts->ts_name[0] = '\0';
2729 }
2730 #endif
2731
2732 #ifdef SMP
2733
2734 /*
2735 * Build the CPU topology dump string. Is recursively called to collect
2736 * the topology tree.
2737 */
2738 static int
2739 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2740 int indent)
2741 {
2742 int i, first;
2743
2744 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2745 "", 1 + indent / 2, cg->cg_level);
2746 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2747 cg->cg_count, cg->cg_mask);
2748 first = TRUE;
2749 for (i = 0; i < MAXCPU; i++) {
2750 if ((cg->cg_mask & (1 << i)) != 0) {
2751 if (!first)
2752 sbuf_printf(sb, ", ");
2753 else
2754 first = FALSE;
2755 sbuf_printf(sb, "%d", i);
2756 }
2757 }
2758 sbuf_printf(sb, "</cpu>\n");
2759
2760 if (cg->cg_flags != 0) {
2761 sbuf_printf(sb, "%*s <flags>", indent, "");
2762 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2763 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2764 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2765 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2766 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2767 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2768 sbuf_printf(sb, "</flags>\n");
2769 }
2770
2771 if (cg->cg_children > 0) {
2772 sbuf_printf(sb, "%*s <children>\n", indent, "");
2773 for (i = 0; i < cg->cg_children; i++)
2774 sysctl_kern_sched_topology_spec_internal(sb,
2775 &cg->cg_child[i], indent+2);
2776 sbuf_printf(sb, "%*s </children>\n", indent, "");
2777 }
2778 sbuf_printf(sb, "%*s</group>\n", indent, "");
2779 return (0);
2780 }
2781
2782 /*
2783 * Sysctl handler for retrieving topology dump. It's a wrapper for
2784 * the recursive sysctl_kern_smp_topology_spec_internal().
2785 */
2786 static int
2787 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2788 {
2789 struct sbuf *topo;
2790 int err;
2791
2792 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2793
2794 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2795 if (topo == NULL)
2796 return (ENOMEM);
2797
2798 sbuf_printf(topo, "<groups>\n");
2799 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2800 sbuf_printf(topo, "</groups>\n");
2801
2802 if (err == 0) {
2803 sbuf_finish(topo);
2804 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2805 }
2806 sbuf_delete(topo);
2807 return (err);
2808 }
2809 #endif
2810
2811 static int
2812 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2813 {
2814 int error, new_val, period;
2815
2816 period = 1000000 / realstathz;
2817 new_val = period * sched_slice;
2818 error = sysctl_handle_int(oidp, &new_val, 0, req);
2819 if (error != 0 || req->newptr == NULL)
2820 return (error);
2821 if (new_val <= 0)
2822 return (EINVAL);
2823 sched_slice = imax(1, (new_val + period / 2) / period);
2824 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2825 realstathz);
2826 return (0);
2827 }
2828
2829 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2830 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2831 "Scheduler name");
2832 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2833 NULL, 0, sysctl_kern_quantum, "I",
2834 "Quantum for timeshare threads in microseconds");
2835 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2836 "Quantum for timeshare threads in stathz ticks");
2837 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2838 "Interactivity score threshold");
2839 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2840 &preempt_thresh, 0,
2841 "Maximal (lowest) priority for preemption");
2842 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2843 "Assign static kernel priorities to sleeping threads");
2844 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2845 "Number of times idle thread will spin waiting for new work");
2846 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2847 &sched_idlespinthresh, 0,
2848 "Threshold before we will permit idle thread spinning");
2849 #ifdef SMP
2850 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2851 "Number of hz ticks to keep thread affinity for");
2852 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2853 "Enables the long-term load balancer");
2854 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2855 &balance_interval, 0,
2856 "Average period in stathz ticks to run the long-term balancer");
2857 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2858 "Attempts to steal work from other cores before idling");
2859 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2860 "Minimum load on remote CPU before we'll steal");
2861 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2862 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2863 "XML dump of detected CPU topology");
2864 #endif
2865
2866 /* ps compat. All cpu percentages from ULE are weighted. */
2867 static int ccpu = 0;
2868 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
Cache object: 9455665199fcf981824a07bed9c42a1f
|