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_sched.h"
43
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/limits.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
67 #ifdef HWPMC_HOOKS
68 #include <sys/pmckern.h>
69 #endif
70
71 #ifdef KDTRACE_HOOKS
72 #include <sys/dtrace_bsd.h>
73 int dtrace_vtime_active;
74 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
75 #endif
76
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
79
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 */
90 struct td_sched {
91 struct runq *ts_runq; /* Run-queue we're queued on. */
92 short ts_flags; /* TSF_* flags. */
93 int ts_cpu; /* CPU that we have affinity for. */
94 int ts_rltick; /* Real last tick, for affinity. */
95 int ts_slice; /* Ticks of slice remaining. */
96 u_int ts_slptime; /* Number of ticks we vol. slept */
97 u_int ts_runtime; /* Number of ticks we were running */
98 int ts_ltick; /* Last tick that we were running on */
99 int ts_ftick; /* First tick that we were running on */
100 int ts_ticks; /* Tick count */
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
109 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
110 #define THREAD_CAN_SCHED(td, cpu) \
111 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
112
113 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
114 sizeof(struct thread0_storage),
115 "increase struct thread0_storage.t0st_sched size");
116
117 /*
118 * Priority ranges used for interactive and non-interactive timeshare
119 * threads. The timeshare priorities are split up into four ranges.
120 * The first range handles interactive threads. The last three ranges
121 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
122 * ranges supporting nice values.
123 */
124 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
125 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
126 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
127
128 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
129 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
130 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
131 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
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 /*
191 * These parameters determine the slice behavior for batch work.
192 */
193 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
194 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
195
196 /* Flags kept in td_flags. */
197 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
198
199 /*
200 * tickincr: Converts a stathz tick into a hz domain scaled by
201 * the shift factor. Without the shift the error rate
202 * due to rounding would be unacceptably high.
203 * realstathz: stathz is sometimes 0 and run off of hz.
204 * sched_slice: Runtime of each thread before rescheduling.
205 * preempt_thresh: Priority threshold for preemption and remote IPIs.
206 */
207 static int sched_interact = SCHED_INTERACT_THRESH;
208 static int tickincr = 8 << SCHED_TICK_SHIFT;
209 static int realstathz = 127; /* reset during boot. */
210 static int sched_slice = 10; /* reset during boot. */
211 static int sched_slice_min = 1; /* reset during boot. */
212 #ifdef PREEMPTION
213 #ifdef FULL_PREEMPTION
214 static int preempt_thresh = PRI_MAX_IDLE;
215 #else
216 static int preempt_thresh = PRI_MIN_KERN;
217 #endif
218 #else
219 static int preempt_thresh = 0;
220 #endif
221 static int static_boost = PRI_MIN_BATCH;
222 static int sched_idlespins = 10000;
223 static int sched_idlespinthresh = -1;
224
225 /*
226 * tdq - per processor runqs and statistics. All fields are protected by the
227 * tdq_lock. The load and lowpri may be accessed without to avoid excess
228 * locking in sched_pickcpu();
229 */
230 struct tdq {
231 /*
232 * Ordered to improve efficiency of cpu_search() and switch().
233 * tdq_lock is padded to avoid false sharing with tdq_load and
234 * tdq_cpu_idle.
235 */
236 struct mtx_padalign tdq_lock; /* run queue lock. */
237 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
238 volatile int tdq_load; /* Aggregate load. */
239 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
240 int tdq_sysload; /* For loadavg, !ITHD load. */
241 volatile int tdq_transferable; /* Transferable thread count. */
242 volatile short tdq_switchcnt; /* Switches this tick. */
243 volatile short tdq_oldswitchcnt; /* Switches last tick. */
244 u_char tdq_lowpri; /* Lowest priority thread. */
245 u_char tdq_ipipending; /* IPI pending. */
246 u_char tdq_idx; /* Current insert index. */
247 u_char tdq_ridx; /* Current removal index. */
248 struct runq tdq_realtime; /* real-time run queue. */
249 struct runq tdq_timeshare; /* timeshare run queue. */
250 struct runq tdq_idle; /* Queue of IDLE threads. */
251 char tdq_name[TDQ_NAME_LEN];
252 #ifdef KTR
253 char tdq_loadname[TDQ_LOADNAME_LEN];
254 #endif
255 } __aligned(64);
256
257 /* Idle thread states and config. */
258 #define TDQ_RUNNING 1
259 #define TDQ_IDLE 2
260
261 #ifdef SMP
262 struct cpu_group *cpu_top; /* CPU topology */
263
264 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
265 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
266
267 /*
268 * Run-time tunables.
269 */
270 static int rebalance = 1;
271 static int balance_interval = 128; /* Default set in sched_initticks(). */
272 static int affinity;
273 static int steal_idle = 1;
274 static int steal_thresh = 2;
275 static int always_steal = 0;
276 static int trysteal_limit = 2;
277
278 /*
279 * One thread queue per processor.
280 */
281 static struct tdq tdq_cpu[MAXCPU];
282 static struct tdq *balance_tdq;
283 static int balance_ticks;
284 static DPCPU_DEFINE(uint32_t, randomval);
285
286 #define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
287 #define TDQ_CPU(x) (&tdq_cpu[(x)])
288 #define TDQ_ID(x) ((int)((x) - tdq_cpu))
289 #else /* !SMP */
290 static struct tdq tdq_cpu;
291
292 #define TDQ_ID(x) (0)
293 #define TDQ_SELF() (&tdq_cpu)
294 #define TDQ_CPU(x) (&tdq_cpu)
295 #endif
296
297 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
298 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
299 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
300 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
301 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
302
303 static void sched_priority(struct thread *);
304 static void sched_thread_priority(struct thread *, u_char);
305 static int sched_interact_score(struct thread *);
306 static void sched_interact_update(struct thread *);
307 static void sched_interact_fork(struct thread *);
308 static void sched_pctcpu_update(struct td_sched *, int);
309
310 /* Operations on per processor queues */
311 static struct thread *tdq_choose(struct tdq *);
312 static void tdq_setup(struct tdq *);
313 static void tdq_load_add(struct tdq *, struct thread *);
314 static void tdq_load_rem(struct tdq *, struct thread *);
315 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
316 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
317 static inline int sched_shouldpreempt(int, int, int);
318 void tdq_print(int cpu);
319 static void runq_print(struct runq *rq);
320 static void tdq_add(struct tdq *, struct thread *, int);
321 #ifdef SMP
322 static struct thread *tdq_move(struct tdq *, struct tdq *);
323 static int tdq_idled(struct tdq *);
324 static void tdq_notify(struct tdq *, struct thread *);
325 static struct thread *tdq_steal(struct tdq *, int);
326 static struct thread *runq_steal(struct runq *, int);
327 static int sched_pickcpu(struct thread *, int);
328 static void sched_balance(void);
329 static int sched_balance_pair(struct tdq *, struct tdq *);
330 static inline struct tdq *sched_setcpu(struct thread *, int, int);
331 static inline void thread_unblock_switch(struct thread *, struct mtx *);
332 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
333 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
334 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
335 struct cpu_group *cg, int indent);
336 #endif
337
338 static void sched_setup(void *dummy);
339 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
340
341 static void sched_initticks(void *dummy);
342 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
343 NULL);
344
345 SDT_PROVIDER_DEFINE(sched);
346
347 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
348 "struct proc *", "uint8_t");
349 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
350 "struct proc *", "void *");
351 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
352 "struct proc *", "void *", "int");
353 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
354 "struct proc *", "uint8_t", "struct thread *");
355 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
356 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
357 "struct proc *");
358 SDT_PROBE_DEFINE(sched, , , on__cpu);
359 SDT_PROBE_DEFINE(sched, , , remain__cpu);
360 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
361 "struct proc *");
362
363 /*
364 * Print the threads waiting on a run-queue.
365 */
366 static void
367 runq_print(struct runq *rq)
368 {
369 struct rqhead *rqh;
370 struct thread *td;
371 int pri;
372 int j;
373 int i;
374
375 for (i = 0; i < RQB_LEN; i++) {
376 printf("\t\trunq bits %d 0x%zx\n",
377 i, rq->rq_status.rqb_bits[i]);
378 for (j = 0; j < RQB_BPW; j++)
379 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
380 pri = j + (i << RQB_L2BPW);
381 rqh = &rq->rq_queues[pri];
382 TAILQ_FOREACH(td, rqh, td_runq) {
383 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
384 td, td->td_name, td->td_priority,
385 td->td_rqindex, pri);
386 }
387 }
388 }
389 }
390
391 /*
392 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
393 */
394 void
395 tdq_print(int cpu)
396 {
397 struct tdq *tdq;
398
399 tdq = TDQ_CPU(cpu);
400
401 printf("tdq %d:\n", TDQ_ID(tdq));
402 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
403 printf("\tLock name: %s\n", tdq->tdq_name);
404 printf("\tload: %d\n", tdq->tdq_load);
405 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
406 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
407 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
408 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
409 printf("\tload transferable: %d\n", tdq->tdq_transferable);
410 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
411 printf("\trealtime runq:\n");
412 runq_print(&tdq->tdq_realtime);
413 printf("\ttimeshare runq:\n");
414 runq_print(&tdq->tdq_timeshare);
415 printf("\tidle runq:\n");
416 runq_print(&tdq->tdq_idle);
417 }
418
419 static inline int
420 sched_shouldpreempt(int pri, int cpri, int remote)
421 {
422 /*
423 * If the new priority is not better than the current priority there is
424 * nothing to do.
425 */
426 if (pri >= cpri)
427 return (0);
428 /*
429 * Always preempt idle.
430 */
431 if (cpri >= PRI_MIN_IDLE)
432 return (1);
433 /*
434 * If preemption is disabled don't preempt others.
435 */
436 if (preempt_thresh == 0)
437 return (0);
438 /*
439 * Preempt if we exceed the threshold.
440 */
441 if (pri <= preempt_thresh)
442 return (1);
443 /*
444 * If we're interactive or better and there is non-interactive
445 * or worse running preempt only remote processors.
446 */
447 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
448 return (1);
449 return (0);
450 }
451
452 /*
453 * Add a thread to the actual run-queue. Keeps transferable counts up to
454 * date with what is actually on the run-queue. Selects the correct
455 * queue position for timeshare threads.
456 */
457 static __inline void
458 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
459 {
460 struct td_sched *ts;
461 u_char pri;
462
463 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
464 THREAD_LOCK_ASSERT(td, MA_OWNED);
465
466 pri = td->td_priority;
467 ts = td_get_sched(td);
468 TD_SET_RUNQ(td);
469 if (THREAD_CAN_MIGRATE(td)) {
470 tdq->tdq_transferable++;
471 ts->ts_flags |= TSF_XFERABLE;
472 }
473 if (pri < PRI_MIN_BATCH) {
474 ts->ts_runq = &tdq->tdq_realtime;
475 } else if (pri <= PRI_MAX_BATCH) {
476 ts->ts_runq = &tdq->tdq_timeshare;
477 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
478 ("Invalid priority %d on timeshare runq", pri));
479 /*
480 * This queue contains only priorities between MIN and MAX
481 * realtime. Use the whole queue to represent these values.
482 */
483 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
484 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
485 pri = (pri + tdq->tdq_idx) % RQ_NQS;
486 /*
487 * This effectively shortens the queue by one so we
488 * can have a one slot difference between idx and
489 * ridx while we wait for threads to drain.
490 */
491 if (tdq->tdq_ridx != tdq->tdq_idx &&
492 pri == tdq->tdq_ridx)
493 pri = (unsigned char)(pri - 1) % RQ_NQS;
494 } else
495 pri = tdq->tdq_ridx;
496 runq_add_pri(ts->ts_runq, td, pri, flags);
497 return;
498 } else
499 ts->ts_runq = &tdq->tdq_idle;
500 runq_add(ts->ts_runq, td, flags);
501 }
502
503 /*
504 * Remove a thread from a run-queue. This typically happens when a thread
505 * is selected to run. Running threads are not on the queue and the
506 * transferable count does not reflect them.
507 */
508 static __inline void
509 tdq_runq_rem(struct tdq *tdq, struct thread *td)
510 {
511 struct td_sched *ts;
512
513 ts = td_get_sched(td);
514 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
515 KASSERT(ts->ts_runq != NULL,
516 ("tdq_runq_remove: thread %p null ts_runq", td));
517 if (ts->ts_flags & TSF_XFERABLE) {
518 tdq->tdq_transferable--;
519 ts->ts_flags &= ~TSF_XFERABLE;
520 }
521 if (ts->ts_runq == &tdq->tdq_timeshare) {
522 if (tdq->tdq_idx != tdq->tdq_ridx)
523 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
524 else
525 runq_remove_idx(ts->ts_runq, td, NULL);
526 } else
527 runq_remove(ts->ts_runq, td);
528 }
529
530 /*
531 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
532 * for this thread to the referenced thread queue.
533 */
534 static void
535 tdq_load_add(struct tdq *tdq, struct thread *td)
536 {
537
538 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
539 THREAD_LOCK_ASSERT(td, MA_OWNED);
540
541 tdq->tdq_load++;
542 if ((td->td_flags & TDF_NOLOAD) == 0)
543 tdq->tdq_sysload++;
544 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
545 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
546 }
547
548 /*
549 * Remove the load from a thread that is transitioning to a sleep state or
550 * exiting.
551 */
552 static void
553 tdq_load_rem(struct tdq *tdq, struct thread *td)
554 {
555
556 THREAD_LOCK_ASSERT(td, MA_OWNED);
557 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
558 KASSERT(tdq->tdq_load != 0,
559 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
560
561 tdq->tdq_load--;
562 if ((td->td_flags & TDF_NOLOAD) == 0)
563 tdq->tdq_sysload--;
564 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
565 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
566 }
567
568 /*
569 * Bound timeshare latency by decreasing slice size as load increases. We
570 * consider the maximum latency as the sum of the threads waiting to run
571 * aside from curthread and target no more than sched_slice latency but
572 * no less than sched_slice_min runtime.
573 */
574 static inline int
575 tdq_slice(struct tdq *tdq)
576 {
577 int load;
578
579 /*
580 * It is safe to use sys_load here because this is called from
581 * contexts where timeshare threads are running and so there
582 * cannot be higher priority load in the system.
583 */
584 load = tdq->tdq_sysload - 1;
585 if (load >= SCHED_SLICE_MIN_DIVISOR)
586 return (sched_slice_min);
587 if (load <= 1)
588 return (sched_slice);
589 return (sched_slice / load);
590 }
591
592 /*
593 * Set lowpri to its exact value by searching the run-queue and
594 * evaluating curthread. curthread may be passed as an optimization.
595 */
596 static void
597 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
598 {
599 struct thread *td;
600
601 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
602 if (ctd == NULL)
603 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
604 td = tdq_choose(tdq);
605 if (td == NULL || td->td_priority > ctd->td_priority)
606 tdq->tdq_lowpri = ctd->td_priority;
607 else
608 tdq->tdq_lowpri = td->td_priority;
609 }
610
611 #ifdef SMP
612 /*
613 * We need some randomness. Implement a classic Linear Congruential
614 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
615 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
616 * of the random state (in the low bits of our answer) to keep
617 * the maximum randomness.
618 */
619 static uint32_t
620 sched_random(void)
621 {
622 uint32_t *rndptr;
623
624 rndptr = DPCPU_PTR(randomval);
625 *rndptr = *rndptr * 69069 + 5;
626
627 return (*rndptr >> 16);
628 }
629
630 struct cpu_search {
631 cpuset_t cs_mask;
632 u_int cs_prefer;
633 int cs_pri; /* Min priority for low. */
634 int cs_limit; /* Max load for low, min load for high. */
635 int cs_cpu;
636 int cs_load;
637 };
638
639 #define CPU_SEARCH_LOWEST 0x1
640 #define CPU_SEARCH_HIGHEST 0x2
641 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
642
643 static __always_inline int cpu_search(const struct cpu_group *cg,
644 struct cpu_search *low, struct cpu_search *high, const int match);
645 int __noinline cpu_search_lowest(const struct cpu_group *cg,
646 struct cpu_search *low);
647 int __noinline cpu_search_highest(const struct cpu_group *cg,
648 struct cpu_search *high);
649 int __noinline cpu_search_both(const struct cpu_group *cg,
650 struct cpu_search *low, struct cpu_search *high);
651
652 /*
653 * Search the tree of cpu_groups for the lowest or highest loaded cpu
654 * according to the match argument. This routine actually compares the
655 * load on all paths through the tree and finds the least loaded cpu on
656 * the least loaded path, which may differ from the least loaded cpu in
657 * the system. This balances work among caches and busses.
658 *
659 * This inline is instantiated in three forms below using constants for the
660 * match argument. It is reduced to the minimum set for each case. It is
661 * also recursive to the depth of the tree.
662 */
663 static __always_inline int
664 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
665 struct cpu_search *high, const int match)
666 {
667 struct cpu_search lgroup;
668 struct cpu_search hgroup;
669 cpuset_t cpumask;
670 struct cpu_group *child;
671 struct tdq *tdq;
672 int cpu, i, hload, lload, load, total, rnd;
673
674 total = 0;
675 cpumask = cg->cg_mask;
676 if (match & CPU_SEARCH_LOWEST) {
677 lload = INT_MAX;
678 lgroup = *low;
679 }
680 if (match & CPU_SEARCH_HIGHEST) {
681 hload = INT_MIN;
682 hgroup = *high;
683 }
684
685 /* Iterate through the child CPU groups and then remaining CPUs. */
686 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
687 if (i == 0) {
688 #ifdef HAVE_INLINE_FFSL
689 cpu = CPU_FFS(&cpumask) - 1;
690 #else
691 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
692 cpu--;
693 #endif
694 if (cpu < 0)
695 break;
696 child = NULL;
697 } else
698 child = &cg->cg_child[i - 1];
699
700 if (match & CPU_SEARCH_LOWEST)
701 lgroup.cs_cpu = -1;
702 if (match & CPU_SEARCH_HIGHEST)
703 hgroup.cs_cpu = -1;
704 if (child) { /* Handle child CPU group. */
705 CPU_NAND(&cpumask, &child->cg_mask);
706 switch (match) {
707 case CPU_SEARCH_LOWEST:
708 load = cpu_search_lowest(child, &lgroup);
709 break;
710 case CPU_SEARCH_HIGHEST:
711 load = cpu_search_highest(child, &hgroup);
712 break;
713 case CPU_SEARCH_BOTH:
714 load = cpu_search_both(child, &lgroup, &hgroup);
715 break;
716 }
717 } else { /* Handle child CPU. */
718 CPU_CLR(cpu, &cpumask);
719 tdq = TDQ_CPU(cpu);
720 load = tdq->tdq_load * 256;
721 rnd = sched_random() % 32;
722 if (match & CPU_SEARCH_LOWEST) {
723 if (cpu == low->cs_prefer)
724 load -= 64;
725 /* If that CPU is allowed and get data. */
726 if (tdq->tdq_lowpri > lgroup.cs_pri &&
727 tdq->tdq_load <= lgroup.cs_limit &&
728 CPU_ISSET(cpu, &lgroup.cs_mask)) {
729 lgroup.cs_cpu = cpu;
730 lgroup.cs_load = load - rnd;
731 }
732 }
733 if (match & CPU_SEARCH_HIGHEST)
734 if (tdq->tdq_load >= hgroup.cs_limit &&
735 tdq->tdq_transferable &&
736 CPU_ISSET(cpu, &hgroup.cs_mask)) {
737 hgroup.cs_cpu = cpu;
738 hgroup.cs_load = load - rnd;
739 }
740 }
741 total += load;
742
743 /* We have info about child item. Compare it. */
744 if (match & CPU_SEARCH_LOWEST) {
745 if (lgroup.cs_cpu >= 0 &&
746 (load < lload ||
747 (load == lload && lgroup.cs_load < low->cs_load))) {
748 lload = load;
749 low->cs_cpu = lgroup.cs_cpu;
750 low->cs_load = lgroup.cs_load;
751 }
752 }
753 if (match & CPU_SEARCH_HIGHEST)
754 if (hgroup.cs_cpu >= 0 &&
755 (load > hload ||
756 (load == hload && hgroup.cs_load > high->cs_load))) {
757 hload = load;
758 high->cs_cpu = hgroup.cs_cpu;
759 high->cs_load = hgroup.cs_load;
760 }
761 if (child) {
762 i--;
763 if (i == 0 && CPU_EMPTY(&cpumask))
764 break;
765 }
766 #ifndef HAVE_INLINE_FFSL
767 else
768 cpu--;
769 #endif
770 }
771 return (total);
772 }
773
774 /*
775 * cpu_search instantiations must pass constants to maintain the inline
776 * optimization.
777 */
778 int
779 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
780 {
781 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
782 }
783
784 int
785 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
786 {
787 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
788 }
789
790 int
791 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
792 struct cpu_search *high)
793 {
794 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
795 }
796
797 /*
798 * Find the cpu with the least load via the least loaded path that has a
799 * lowpri greater than pri pri. A pri of -1 indicates any priority is
800 * acceptable.
801 */
802 static inline int
803 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
804 int prefer)
805 {
806 struct cpu_search low;
807
808 low.cs_cpu = -1;
809 low.cs_prefer = prefer;
810 low.cs_mask = mask;
811 low.cs_pri = pri;
812 low.cs_limit = maxload;
813 cpu_search_lowest(cg, &low);
814 return low.cs_cpu;
815 }
816
817 /*
818 * Find the cpu with the highest load via the highest loaded path.
819 */
820 static inline int
821 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
822 {
823 struct cpu_search high;
824
825 high.cs_cpu = -1;
826 high.cs_mask = mask;
827 high.cs_limit = minload;
828 cpu_search_highest(cg, &high);
829 return high.cs_cpu;
830 }
831
832 static void
833 sched_balance_group(struct cpu_group *cg)
834 {
835 cpuset_t hmask, lmask;
836 int high, low, anylow;
837
838 CPU_FILL(&hmask);
839 for (;;) {
840 high = sched_highest(cg, hmask, 2);
841 /* Stop if there is no more CPU with transferrable threads. */
842 if (high == -1)
843 break;
844 CPU_CLR(high, &hmask);
845 CPU_COPY(&hmask, &lmask);
846 /* Stop if there is no more CPU left for low. */
847 if (CPU_EMPTY(&lmask))
848 break;
849 anylow = 1;
850 nextlow:
851 low = sched_lowest(cg, lmask, -1,
852 TDQ_CPU(high)->tdq_load - 1, high);
853 /* Stop if we looked well and found no less loaded CPU. */
854 if (anylow && low == -1)
855 break;
856 /* Go to next high if we found no less loaded CPU. */
857 if (low == -1)
858 continue;
859 /* Transfer thread from high to low. */
860 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
861 /* CPU that got thread can no longer be a donor. */
862 CPU_CLR(low, &hmask);
863 } else {
864 /*
865 * If failed, then there is no threads on high
866 * that can run on this low. Drop low from low
867 * mask and look for different one.
868 */
869 CPU_CLR(low, &lmask);
870 anylow = 0;
871 goto nextlow;
872 }
873 }
874 }
875
876 static void
877 sched_balance(void)
878 {
879 struct tdq *tdq;
880
881 balance_ticks = max(balance_interval / 2, 1) +
882 (sched_random() % balance_interval);
883 tdq = TDQ_SELF();
884 TDQ_UNLOCK(tdq);
885 sched_balance_group(cpu_top);
886 TDQ_LOCK(tdq);
887 }
888
889 /*
890 * Lock two thread queues using their address to maintain lock order.
891 */
892 static void
893 tdq_lock_pair(struct tdq *one, struct tdq *two)
894 {
895 if (one < two) {
896 TDQ_LOCK(one);
897 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
898 } else {
899 TDQ_LOCK(two);
900 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
901 }
902 }
903
904 /*
905 * Unlock two thread queues. Order is not important here.
906 */
907 static void
908 tdq_unlock_pair(struct tdq *one, struct tdq *two)
909 {
910 TDQ_UNLOCK(one);
911 TDQ_UNLOCK(two);
912 }
913
914 /*
915 * Transfer load between two imbalanced thread queues.
916 */
917 static int
918 sched_balance_pair(struct tdq *high, struct tdq *low)
919 {
920 struct thread *td;
921 int cpu;
922
923 tdq_lock_pair(high, low);
924 td = NULL;
925 /*
926 * Transfer a thread from high to low.
927 */
928 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
929 (td = tdq_move(high, low)) != NULL) {
930 /*
931 * In case the target isn't the current cpu notify it of the
932 * new load, possibly sending an IPI to force it to reschedule.
933 */
934 cpu = TDQ_ID(low);
935 if (cpu != PCPU_GET(cpuid))
936 tdq_notify(low, td);
937 }
938 tdq_unlock_pair(high, low);
939 return (td != NULL);
940 }
941
942 /*
943 * Move a thread from one thread queue to another.
944 */
945 static struct thread *
946 tdq_move(struct tdq *from, struct tdq *to)
947 {
948 struct td_sched *ts;
949 struct thread *td;
950 struct tdq *tdq;
951 int cpu;
952
953 TDQ_LOCK_ASSERT(from, MA_OWNED);
954 TDQ_LOCK_ASSERT(to, MA_OWNED);
955
956 tdq = from;
957 cpu = TDQ_ID(to);
958 td = tdq_steal(tdq, cpu);
959 if (td == NULL)
960 return (NULL);
961 ts = td_get_sched(td);
962 /*
963 * Although the run queue is locked the thread may be blocked. Lock
964 * it to clear this and acquire the run-queue lock.
965 */
966 thread_lock(td);
967 /* Drop recursive lock on from acquired via thread_lock(). */
968 TDQ_UNLOCK(from);
969 sched_rem(td);
970 ts->ts_cpu = cpu;
971 td->td_lock = TDQ_LOCKPTR(to);
972 tdq_add(to, td, SRQ_YIELDING);
973 return (td);
974 }
975
976 /*
977 * This tdq has idled. Try to steal a thread from another cpu and switch
978 * to it.
979 */
980 static int
981 tdq_idled(struct tdq *tdq)
982 {
983 struct cpu_group *cg;
984 struct tdq *steal;
985 cpuset_t mask;
986 int cpu, switchcnt;
987
988 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
989 return (1);
990 CPU_FILL(&mask);
991 CPU_CLR(PCPU_GET(cpuid), &mask);
992 restart:
993 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
994 for (cg = tdq->tdq_cg; ; ) {
995 cpu = sched_highest(cg, mask, steal_thresh);
996 /*
997 * We were assigned a thread but not preempted. Returning
998 * 0 here will cause our caller to switch to it.
999 */
1000 if (tdq->tdq_load)
1001 return (0);
1002 if (cpu == -1) {
1003 cg = cg->cg_parent;
1004 if (cg == NULL)
1005 return (1);
1006 continue;
1007 }
1008 steal = TDQ_CPU(cpu);
1009 /*
1010 * The data returned by sched_highest() is stale and
1011 * the chosen CPU no longer has an eligible thread.
1012 *
1013 * Testing this ahead of tdq_lock_pair() only catches
1014 * this situation about 20% of the time on an 8 core
1015 * 16 thread Ryzen 7, but it still helps performance.
1016 */
1017 if (steal->tdq_load < steal_thresh ||
1018 steal->tdq_transferable == 0)
1019 goto restart;
1020 tdq_lock_pair(tdq, steal);
1021 /*
1022 * We were assigned a thread while waiting for the locks.
1023 * Switch to it now instead of stealing a thread.
1024 */
1025 if (tdq->tdq_load)
1026 break;
1027 /*
1028 * The data returned by sched_highest() is stale and
1029 * the chosen CPU no longer has an eligible thread, or
1030 * we were preempted and the CPU loading info may be out
1031 * of date. The latter is rare. In either case restart
1032 * the search.
1033 */
1034 if (steal->tdq_load < steal_thresh ||
1035 steal->tdq_transferable == 0 ||
1036 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1037 tdq_unlock_pair(tdq, steal);
1038 goto restart;
1039 }
1040 /*
1041 * Steal the thread and switch to it.
1042 */
1043 if (tdq_move(steal, tdq) != NULL)
1044 break;
1045 /*
1046 * We failed to acquire a thread even though it looked
1047 * like one was available. This could be due to affinity
1048 * restrictions or for other reasons. Loop again after
1049 * removing this CPU from the set. The restart logic
1050 * above does not restore this CPU to the set due to the
1051 * likelyhood of failing here again.
1052 */
1053 CPU_CLR(cpu, &mask);
1054 tdq_unlock_pair(tdq, steal);
1055 }
1056 TDQ_UNLOCK(steal);
1057 mi_switch(SW_VOL | SWT_IDLE, NULL);
1058 thread_unlock(curthread);
1059 return (0);
1060 }
1061
1062 /*
1063 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1064 */
1065 static void
1066 tdq_notify(struct tdq *tdq, struct thread *td)
1067 {
1068 struct thread *ctd;
1069 int pri;
1070 int cpu;
1071
1072 if (tdq->tdq_ipipending)
1073 return;
1074 cpu = td_get_sched(td)->ts_cpu;
1075 pri = td->td_priority;
1076 ctd = pcpu_find(cpu)->pc_curthread;
1077 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1078 return;
1079
1080 /*
1081 * Make sure that our caller's earlier update to tdq_load is
1082 * globally visible before we read tdq_cpu_idle. Idle thread
1083 * accesses both of them without locks, and the order is important.
1084 */
1085 atomic_thread_fence_seq_cst();
1086
1087 if (TD_IS_IDLETHREAD(ctd)) {
1088 /*
1089 * If the MD code has an idle wakeup routine try that before
1090 * falling back to IPI.
1091 */
1092 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1093 return;
1094 }
1095 tdq->tdq_ipipending = 1;
1096 ipi_cpu(cpu, IPI_PREEMPT);
1097 }
1098
1099 /*
1100 * Steals load from a timeshare queue. Honors the rotating queue head
1101 * index.
1102 */
1103 static struct thread *
1104 runq_steal_from(struct runq *rq, int cpu, u_char start)
1105 {
1106 struct rqbits *rqb;
1107 struct rqhead *rqh;
1108 struct thread *td, *first;
1109 int bit;
1110 int i;
1111
1112 rqb = &rq->rq_status;
1113 bit = start & (RQB_BPW -1);
1114 first = NULL;
1115 again:
1116 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1117 if (rqb->rqb_bits[i] == 0)
1118 continue;
1119 if (bit == 0)
1120 bit = RQB_FFS(rqb->rqb_bits[i]);
1121 for (; bit < RQB_BPW; bit++) {
1122 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1123 continue;
1124 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1125 TAILQ_FOREACH(td, rqh, td_runq) {
1126 if (first && THREAD_CAN_MIGRATE(td) &&
1127 THREAD_CAN_SCHED(td, cpu))
1128 return (td);
1129 first = td;
1130 }
1131 }
1132 }
1133 if (start != 0) {
1134 start = 0;
1135 goto again;
1136 }
1137
1138 if (first && THREAD_CAN_MIGRATE(first) &&
1139 THREAD_CAN_SCHED(first, cpu))
1140 return (first);
1141 return (NULL);
1142 }
1143
1144 /*
1145 * Steals load from a standard linear queue.
1146 */
1147 static struct thread *
1148 runq_steal(struct runq *rq, int cpu)
1149 {
1150 struct rqhead *rqh;
1151 struct rqbits *rqb;
1152 struct thread *td;
1153 int word;
1154 int bit;
1155
1156 rqb = &rq->rq_status;
1157 for (word = 0; word < RQB_LEN; word++) {
1158 if (rqb->rqb_bits[word] == 0)
1159 continue;
1160 for (bit = 0; bit < RQB_BPW; bit++) {
1161 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1162 continue;
1163 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1164 TAILQ_FOREACH(td, rqh, td_runq)
1165 if (THREAD_CAN_MIGRATE(td) &&
1166 THREAD_CAN_SCHED(td, cpu))
1167 return (td);
1168 }
1169 }
1170 return (NULL);
1171 }
1172
1173 /*
1174 * Attempt to steal a thread in priority order from a thread queue.
1175 */
1176 static struct thread *
1177 tdq_steal(struct tdq *tdq, int cpu)
1178 {
1179 struct thread *td;
1180
1181 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1182 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1183 return (td);
1184 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1185 cpu, tdq->tdq_ridx)) != NULL)
1186 return (td);
1187 return (runq_steal(&tdq->tdq_idle, cpu));
1188 }
1189
1190 /*
1191 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1192 * current lock and returns with the assigned queue locked.
1193 */
1194 static inline struct tdq *
1195 sched_setcpu(struct thread *td, int cpu, int flags)
1196 {
1197
1198 struct tdq *tdq;
1199
1200 THREAD_LOCK_ASSERT(td, MA_OWNED);
1201 tdq = TDQ_CPU(cpu);
1202 td_get_sched(td)->ts_cpu = cpu;
1203 /*
1204 * If the lock matches just return the queue.
1205 */
1206 if (td->td_lock == TDQ_LOCKPTR(tdq))
1207 return (tdq);
1208 #ifdef notyet
1209 /*
1210 * If the thread isn't running its lockptr is a
1211 * turnstile or a sleepqueue. We can just lock_set without
1212 * blocking.
1213 */
1214 if (TD_CAN_RUN(td)) {
1215 TDQ_LOCK(tdq);
1216 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1217 return (tdq);
1218 }
1219 #endif
1220 /*
1221 * The hard case, migration, we need to block the thread first to
1222 * prevent order reversals with other cpus locks.
1223 */
1224 spinlock_enter();
1225 thread_lock_block(td);
1226 TDQ_LOCK(tdq);
1227 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1228 spinlock_exit();
1229 return (tdq);
1230 }
1231
1232 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1233 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1234 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1235 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1236 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1237 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1238
1239 static int
1240 sched_pickcpu(struct thread *td, int flags)
1241 {
1242 struct cpu_group *cg, *ccg;
1243 struct td_sched *ts;
1244 struct tdq *tdq;
1245 cpuset_t mask;
1246 int cpu, pri, self, intr;
1247
1248 self = PCPU_GET(cpuid);
1249 ts = td_get_sched(td);
1250 if (smp_started == 0)
1251 return (self);
1252 /*
1253 * Don't migrate a running thread from sched_switch().
1254 */
1255 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1256 return (ts->ts_cpu);
1257 /*
1258 * Prefer to run interrupt threads on the processors that generate
1259 * the interrupt.
1260 */
1261 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1262 curthread->td_intr_nesting_level) {
1263 tdq = TDQ_SELF();
1264 if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1265 SCHED_STAT_INC(pickcpu_idle_affinity);
1266 return (self);
1267 }
1268 ts->ts_cpu = self;
1269 intr = 1;
1270 cg = tdq->tdq_cg;
1271 goto llc;
1272 } else {
1273 intr = 0;
1274 tdq = TDQ_CPU(ts->ts_cpu);
1275 cg = tdq->tdq_cg;
1276 }
1277 /*
1278 * If the thread can run on the last cpu and the affinity has not
1279 * expired or it is idle run it there.
1280 */
1281 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1282 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1283 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1284 if (cg->cg_flags & CG_FLAG_THREAD) {
1285 /* Check all SMT threads for being idle. */
1286 for (cpu = CPU_FFS(&cg->cg_mask) - 1; ; cpu++) {
1287 if (CPU_ISSET(cpu, &cg->cg_mask) &&
1288 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1289 break;
1290 if (cpu >= mp_maxid) {
1291 SCHED_STAT_INC(pickcpu_idle_affinity);
1292 return (ts->ts_cpu);
1293 }
1294 }
1295 } else {
1296 SCHED_STAT_INC(pickcpu_idle_affinity);
1297 return (ts->ts_cpu);
1298 }
1299 }
1300 llc:
1301 /*
1302 * Search for the last level cache CPU group in the tree.
1303 * Skip SMT, identical groups and caches with expired affinity.
1304 * Interrupt threads affinity is explicit and never expires.
1305 */
1306 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1307 if (cg->cg_flags & CG_FLAG_THREAD)
1308 continue;
1309 if (cg->cg_children == 1 || cg->cg_count == 1)
1310 continue;
1311 if (cg->cg_level == CG_SHARE_NONE ||
1312 (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1313 continue;
1314 ccg = cg;
1315 }
1316 /* Found LLC shared by all CPUs, so do a global search. */
1317 if (ccg == cpu_top)
1318 ccg = NULL;
1319 cpu = -1;
1320 mask = td->td_cpuset->cs_mask;
1321 pri = td->td_priority;
1322 /*
1323 * Try hard to keep interrupts within found LLC. Search the LLC for
1324 * the least loaded CPU we can run now. For NUMA systems it should
1325 * be within target domain, and it also reduces scheduling overhead.
1326 */
1327 if (ccg != NULL && intr) {
1328 cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu);
1329 if (cpu >= 0)
1330 SCHED_STAT_INC(pickcpu_intrbind);
1331 } else
1332 /* Search the LLC for the least loaded idle CPU we can run now. */
1333 if (ccg != NULL) {
1334 cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1335 INT_MAX, ts->ts_cpu);
1336 if (cpu >= 0)
1337 SCHED_STAT_INC(pickcpu_affinity);
1338 }
1339 /* Search globally for the least loaded CPU we can run now. */
1340 if (cpu < 0) {
1341 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1342 if (cpu >= 0)
1343 SCHED_STAT_INC(pickcpu_lowest);
1344 }
1345 /* Search globally for the least loaded CPU. */
1346 if (cpu < 0) {
1347 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1348 if (cpu >= 0)
1349 SCHED_STAT_INC(pickcpu_lowest);
1350 }
1351 KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1352 /*
1353 * Compare the lowest loaded cpu to current cpu.
1354 */
1355 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1356 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1357 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1358 SCHED_STAT_INC(pickcpu_local);
1359 cpu = self;
1360 }
1361 if (cpu != ts->ts_cpu)
1362 SCHED_STAT_INC(pickcpu_migration);
1363 return (cpu);
1364 }
1365 #endif
1366
1367 /*
1368 * Pick the highest priority task we have and return it.
1369 */
1370 static struct thread *
1371 tdq_choose(struct tdq *tdq)
1372 {
1373 struct thread *td;
1374
1375 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1376 td = runq_choose(&tdq->tdq_realtime);
1377 if (td != NULL)
1378 return (td);
1379 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1380 if (td != NULL) {
1381 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1382 ("tdq_choose: Invalid priority on timeshare queue %d",
1383 td->td_priority));
1384 return (td);
1385 }
1386 td = runq_choose(&tdq->tdq_idle);
1387 if (td != NULL) {
1388 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1389 ("tdq_choose: Invalid priority on idle queue %d",
1390 td->td_priority));
1391 return (td);
1392 }
1393
1394 return (NULL);
1395 }
1396
1397 /*
1398 * Initialize a thread queue.
1399 */
1400 static void
1401 tdq_setup(struct tdq *tdq)
1402 {
1403
1404 if (bootverbose)
1405 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1406 runq_init(&tdq->tdq_realtime);
1407 runq_init(&tdq->tdq_timeshare);
1408 runq_init(&tdq->tdq_idle);
1409 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1410 "sched lock %d", (int)TDQ_ID(tdq));
1411 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1412 MTX_SPIN | MTX_RECURSE);
1413 #ifdef KTR
1414 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1415 "CPU %d load", (int)TDQ_ID(tdq));
1416 #endif
1417 }
1418
1419 #ifdef SMP
1420 static void
1421 sched_setup_smp(void)
1422 {
1423 struct tdq *tdq;
1424 int i;
1425
1426 cpu_top = smp_topo();
1427 CPU_FOREACH(i) {
1428 tdq = TDQ_CPU(i);
1429 tdq_setup(tdq);
1430 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1431 if (tdq->tdq_cg == NULL)
1432 panic("Can't find cpu group for %d\n", i);
1433 }
1434 balance_tdq = TDQ_SELF();
1435 }
1436 #endif
1437
1438 /*
1439 * Setup the thread queues and initialize the topology based on MD
1440 * information.
1441 */
1442 static void
1443 sched_setup(void *dummy)
1444 {
1445 struct tdq *tdq;
1446
1447 tdq = TDQ_SELF();
1448 #ifdef SMP
1449 sched_setup_smp();
1450 #else
1451 tdq_setup(tdq);
1452 #endif
1453
1454 /* Add thread0's load since it's running. */
1455 TDQ_LOCK(tdq);
1456 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1457 tdq_load_add(tdq, &thread0);
1458 tdq->tdq_lowpri = thread0.td_priority;
1459 TDQ_UNLOCK(tdq);
1460 }
1461
1462 /*
1463 * This routine determines time constants after stathz and hz are setup.
1464 */
1465 /* ARGSUSED */
1466 static void
1467 sched_initticks(void *dummy)
1468 {
1469 int incr;
1470
1471 realstathz = stathz ? stathz : hz;
1472 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1473 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1474 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1475 realstathz);
1476
1477 /*
1478 * tickincr is shifted out by 10 to avoid rounding errors due to
1479 * hz not being evenly divisible by stathz on all platforms.
1480 */
1481 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1482 /*
1483 * This does not work for values of stathz that are more than
1484 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1485 */
1486 if (incr == 0)
1487 incr = 1;
1488 tickincr = incr;
1489 #ifdef SMP
1490 /*
1491 * Set the default balance interval now that we know
1492 * what realstathz is.
1493 */
1494 balance_interval = realstathz;
1495 balance_ticks = balance_interval;
1496 affinity = SCHED_AFFINITY_DEFAULT;
1497 #endif
1498 if (sched_idlespinthresh < 0)
1499 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1500 }
1501
1502
1503 /*
1504 * This is the core of the interactivity algorithm. Determines a score based
1505 * on past behavior. It is the ratio of sleep time to run time scaled to
1506 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1507 * differs from the cpu usage because it does not account for time spent
1508 * waiting on a run-queue. Would be prettier if we had floating point.
1509 *
1510 * When a thread's sleep time is greater than its run time the
1511 * calculation is:
1512 *
1513 * scaling factor
1514 * interactivity score = ---------------------
1515 * sleep time / run time
1516 *
1517 *
1518 * When a thread's run time is greater than its sleep time the
1519 * calculation is:
1520 *
1521 * scaling factor
1522 * interactivity score = --------------------- + scaling factor
1523 * run time / sleep time
1524 */
1525 static int
1526 sched_interact_score(struct thread *td)
1527 {
1528 struct td_sched *ts;
1529 int div;
1530
1531 ts = td_get_sched(td);
1532 /*
1533 * The score is only needed if this is likely to be an interactive
1534 * task. Don't go through the expense of computing it if there's
1535 * no chance.
1536 */
1537 if (sched_interact <= SCHED_INTERACT_HALF &&
1538 ts->ts_runtime >= ts->ts_slptime)
1539 return (SCHED_INTERACT_HALF);
1540
1541 if (ts->ts_runtime > ts->ts_slptime) {
1542 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1543 return (SCHED_INTERACT_HALF +
1544 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1545 }
1546 if (ts->ts_slptime > ts->ts_runtime) {
1547 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1548 return (ts->ts_runtime / div);
1549 }
1550 /* runtime == slptime */
1551 if (ts->ts_runtime)
1552 return (SCHED_INTERACT_HALF);
1553
1554 /*
1555 * This can happen if slptime and runtime are 0.
1556 */
1557 return (0);
1558
1559 }
1560
1561 /*
1562 * Scale the scheduling priority according to the "interactivity" of this
1563 * process.
1564 */
1565 static void
1566 sched_priority(struct thread *td)
1567 {
1568 int score;
1569 int pri;
1570
1571 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1572 return;
1573 /*
1574 * If the score is interactive we place the thread in the realtime
1575 * queue with a priority that is less than kernel and interrupt
1576 * priorities. These threads are not subject to nice restrictions.
1577 *
1578 * Scores greater than this are placed on the normal timeshare queue
1579 * where the priority is partially decided by the most recent cpu
1580 * utilization and the rest is decided by nice value.
1581 *
1582 * The nice value of the process has a linear effect on the calculated
1583 * score. Negative nice values make it easier for a thread to be
1584 * considered interactive.
1585 */
1586 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1587 if (score < sched_interact) {
1588 pri = PRI_MIN_INTERACT;
1589 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1590 sched_interact) * score;
1591 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1592 ("sched_priority: invalid interactive priority %d score %d",
1593 pri, score));
1594 } else {
1595 pri = SCHED_PRI_MIN;
1596 if (td_get_sched(td)->ts_ticks)
1597 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1598 SCHED_PRI_RANGE - 1);
1599 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1600 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1601 ("sched_priority: invalid priority %d: nice %d, "
1602 "ticks %d ftick %d ltick %d tick pri %d",
1603 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1604 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1605 SCHED_PRI_TICKS(td_get_sched(td))));
1606 }
1607 sched_user_prio(td, pri);
1608
1609 return;
1610 }
1611
1612 /*
1613 * This routine enforces a maximum limit on the amount of scheduling history
1614 * kept. It is called after either the slptime or runtime is adjusted. This
1615 * function is ugly due to integer math.
1616 */
1617 static void
1618 sched_interact_update(struct thread *td)
1619 {
1620 struct td_sched *ts;
1621 u_int sum;
1622
1623 ts = td_get_sched(td);
1624 sum = ts->ts_runtime + ts->ts_slptime;
1625 if (sum < SCHED_SLP_RUN_MAX)
1626 return;
1627 /*
1628 * This only happens from two places:
1629 * 1) We have added an unusual amount of run time from fork_exit.
1630 * 2) We have added an unusual amount of sleep time from sched_sleep().
1631 */
1632 if (sum > SCHED_SLP_RUN_MAX * 2) {
1633 if (ts->ts_runtime > ts->ts_slptime) {
1634 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1635 ts->ts_slptime = 1;
1636 } else {
1637 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1638 ts->ts_runtime = 1;
1639 }
1640 return;
1641 }
1642 /*
1643 * If we have exceeded by more than 1/5th then the algorithm below
1644 * will not bring us back into range. Dividing by two here forces
1645 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1646 */
1647 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1648 ts->ts_runtime /= 2;
1649 ts->ts_slptime /= 2;
1650 return;
1651 }
1652 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1653 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1654 }
1655
1656 /*
1657 * Scale back the interactivity history when a child thread is created. The
1658 * history is inherited from the parent but the thread may behave totally
1659 * differently. For example, a shell spawning a compiler process. We want
1660 * to learn that the compiler is behaving badly very quickly.
1661 */
1662 static void
1663 sched_interact_fork(struct thread *td)
1664 {
1665 struct td_sched *ts;
1666 int ratio;
1667 int sum;
1668
1669 ts = td_get_sched(td);
1670 sum = ts->ts_runtime + ts->ts_slptime;
1671 if (sum > SCHED_SLP_RUN_FORK) {
1672 ratio = sum / SCHED_SLP_RUN_FORK;
1673 ts->ts_runtime /= ratio;
1674 ts->ts_slptime /= ratio;
1675 }
1676 }
1677
1678 /*
1679 * Called from proc0_init() to setup the scheduler fields.
1680 */
1681 void
1682 schedinit(void)
1683 {
1684 struct td_sched *ts0;
1685
1686 /*
1687 * Set up the scheduler specific parts of thread0.
1688 */
1689 ts0 = td_get_sched(&thread0);
1690 ts0->ts_ltick = ticks;
1691 ts0->ts_ftick = ticks;
1692 ts0->ts_slice = 0;
1693 }
1694
1695 /*
1696 * This is only somewhat accurate since given many processes of the same
1697 * priority they will switch when their slices run out, which will be
1698 * at most sched_slice stathz ticks.
1699 */
1700 int
1701 sched_rr_interval(void)
1702 {
1703
1704 /* Convert sched_slice from stathz to hz. */
1705 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1706 }
1707
1708 /*
1709 * Update the percent cpu tracking information when it is requested or
1710 * the total history exceeds the maximum. We keep a sliding history of
1711 * tick counts that slowly decays. This is less precise than the 4BSD
1712 * mechanism since it happens with less regular and frequent events.
1713 */
1714 static void
1715 sched_pctcpu_update(struct td_sched *ts, int run)
1716 {
1717 int t = ticks;
1718
1719 /*
1720 * The signed difference may be negative if the thread hasn't run for
1721 * over half of the ticks rollover period.
1722 */
1723 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1724 ts->ts_ticks = 0;
1725 ts->ts_ftick = t - SCHED_TICK_TARG;
1726 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1727 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1728 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1729 ts->ts_ftick = t - SCHED_TICK_TARG;
1730 }
1731 if (run)
1732 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1733 ts->ts_ltick = t;
1734 }
1735
1736 /*
1737 * Adjust the priority of a thread. Move it to the appropriate run-queue
1738 * if necessary. This is the back-end for several priority related
1739 * functions.
1740 */
1741 static void
1742 sched_thread_priority(struct thread *td, u_char prio)
1743 {
1744 struct td_sched *ts;
1745 struct tdq *tdq;
1746 int oldpri;
1747
1748 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1749 "prio:%d", td->td_priority, "new prio:%d", prio,
1750 KTR_ATTR_LINKED, sched_tdname(curthread));
1751 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1752 if (td != curthread && prio < td->td_priority) {
1753 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1754 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1755 prio, KTR_ATTR_LINKED, sched_tdname(td));
1756 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1757 curthread);
1758 }
1759 ts = td_get_sched(td);
1760 THREAD_LOCK_ASSERT(td, MA_OWNED);
1761 if (td->td_priority == prio)
1762 return;
1763 /*
1764 * If the priority has been elevated due to priority
1765 * propagation, we may have to move ourselves to a new
1766 * queue. This could be optimized to not re-add in some
1767 * cases.
1768 */
1769 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1770 sched_rem(td);
1771 td->td_priority = prio;
1772 sched_add(td, SRQ_BORROWING);
1773 return;
1774 }
1775 /*
1776 * If the thread is currently running we may have to adjust the lowpri
1777 * information so other cpus are aware of our current priority.
1778 */
1779 if (TD_IS_RUNNING(td)) {
1780 tdq = TDQ_CPU(ts->ts_cpu);
1781 oldpri = td->td_priority;
1782 td->td_priority = prio;
1783 if (prio < tdq->tdq_lowpri)
1784 tdq->tdq_lowpri = prio;
1785 else if (tdq->tdq_lowpri == oldpri)
1786 tdq_setlowpri(tdq, td);
1787 return;
1788 }
1789 td->td_priority = prio;
1790 }
1791
1792 /*
1793 * Update a thread's priority when it is lent another thread's
1794 * priority.
1795 */
1796 void
1797 sched_lend_prio(struct thread *td, u_char prio)
1798 {
1799
1800 td->td_flags |= TDF_BORROWING;
1801 sched_thread_priority(td, prio);
1802 }
1803
1804 /*
1805 * Restore a thread's priority when priority propagation is
1806 * over. The prio argument is the minimum priority the thread
1807 * needs to have to satisfy other possible priority lending
1808 * requests. If the thread's regular priority is less
1809 * important than prio, the thread will keep a priority boost
1810 * of prio.
1811 */
1812 void
1813 sched_unlend_prio(struct thread *td, u_char prio)
1814 {
1815 u_char base_pri;
1816
1817 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1818 td->td_base_pri <= PRI_MAX_TIMESHARE)
1819 base_pri = td->td_user_pri;
1820 else
1821 base_pri = td->td_base_pri;
1822 if (prio >= base_pri) {
1823 td->td_flags &= ~TDF_BORROWING;
1824 sched_thread_priority(td, base_pri);
1825 } else
1826 sched_lend_prio(td, prio);
1827 }
1828
1829 /*
1830 * Standard entry for setting the priority to an absolute value.
1831 */
1832 void
1833 sched_prio(struct thread *td, u_char prio)
1834 {
1835 u_char oldprio;
1836
1837 /* First, update the base priority. */
1838 td->td_base_pri = prio;
1839
1840 /*
1841 * If the thread is borrowing another thread's priority, don't
1842 * ever lower the priority.
1843 */
1844 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1845 return;
1846
1847 /* Change the real priority. */
1848 oldprio = td->td_priority;
1849 sched_thread_priority(td, prio);
1850
1851 /*
1852 * If the thread is on a turnstile, then let the turnstile update
1853 * its state.
1854 */
1855 if (TD_ON_LOCK(td) && oldprio != prio)
1856 turnstile_adjust(td, oldprio);
1857 }
1858
1859 /*
1860 * Set the base user priority, does not effect current running priority.
1861 */
1862 void
1863 sched_user_prio(struct thread *td, u_char prio)
1864 {
1865
1866 td->td_base_user_pri = prio;
1867 if (td->td_lend_user_pri <= prio)
1868 return;
1869 td->td_user_pri = prio;
1870 }
1871
1872 void
1873 sched_lend_user_prio(struct thread *td, u_char prio)
1874 {
1875
1876 THREAD_LOCK_ASSERT(td, MA_OWNED);
1877 td->td_lend_user_pri = prio;
1878 td->td_user_pri = min(prio, td->td_base_user_pri);
1879 if (td->td_priority > td->td_user_pri)
1880 sched_prio(td, td->td_user_pri);
1881 else if (td->td_priority != td->td_user_pri)
1882 td->td_flags |= TDF_NEEDRESCHED;
1883 }
1884
1885 #ifdef SMP
1886 /*
1887 * This tdq is about to idle. Try to steal a thread from another CPU before
1888 * choosing the idle thread.
1889 */
1890 static void
1891 tdq_trysteal(struct tdq *tdq)
1892 {
1893 struct cpu_group *cg;
1894 struct tdq *steal;
1895 cpuset_t mask;
1896 int cpu, i;
1897
1898 if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1899 return;
1900 CPU_FILL(&mask);
1901 CPU_CLR(PCPU_GET(cpuid), &mask);
1902 /* We don't want to be preempted while we're iterating. */
1903 spinlock_enter();
1904 TDQ_UNLOCK(tdq);
1905 for (i = 1, cg = tdq->tdq_cg; ; ) {
1906 cpu = sched_highest(cg, mask, steal_thresh);
1907 /*
1908 * If a thread was added while interrupts were disabled don't
1909 * steal one here.
1910 */
1911 if (tdq->tdq_load > 0) {
1912 TDQ_LOCK(tdq);
1913 break;
1914 }
1915 if (cpu == -1) {
1916 i++;
1917 cg = cg->cg_parent;
1918 if (cg == NULL || i > trysteal_limit) {
1919 TDQ_LOCK(tdq);
1920 break;
1921 }
1922 continue;
1923 }
1924 steal = TDQ_CPU(cpu);
1925 /*
1926 * The data returned by sched_highest() is stale and
1927 * the chosen CPU no longer has an eligible thread.
1928 */
1929 if (steal->tdq_load < steal_thresh ||
1930 steal->tdq_transferable == 0)
1931 continue;
1932 tdq_lock_pair(tdq, steal);
1933 /*
1934 * If we get to this point, unconditonally exit the loop
1935 * to bound the time spent in the critcal section.
1936 *
1937 * If a thread was added while interrupts were disabled don't
1938 * steal one here.
1939 */
1940 if (tdq->tdq_load > 0) {
1941 TDQ_UNLOCK(steal);
1942 break;
1943 }
1944 /*
1945 * The data returned by sched_highest() is stale and
1946 * the chosen CPU no longer has an eligible thread.
1947 */
1948 if (steal->tdq_load < steal_thresh ||
1949 steal->tdq_transferable == 0) {
1950 TDQ_UNLOCK(steal);
1951 break;
1952 }
1953 /*
1954 * If we fail to acquire one due to affinity restrictions,
1955 * bail out and let the idle thread to a more complete search
1956 * outside of a critical section.
1957 */
1958 if (tdq_move(steal, tdq) == NULL) {
1959 TDQ_UNLOCK(steal);
1960 break;
1961 }
1962 TDQ_UNLOCK(steal);
1963 break;
1964 }
1965 spinlock_exit();
1966 }
1967 #endif
1968
1969 /*
1970 * Handle migration from sched_switch(). This happens only for
1971 * cpu binding.
1972 */
1973 static struct mtx *
1974 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1975 {
1976 struct tdq *tdn;
1977
1978 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1979 #ifdef SMP
1980 tdq_load_rem(tdq, td);
1981 /*
1982 * Do the lock dance required to avoid LOR. We grab an extra
1983 * spinlock nesting to prevent preemption while we're
1984 * not holding either run-queue lock.
1985 */
1986 spinlock_enter();
1987 thread_lock_block(td); /* This releases the lock on tdq. */
1988
1989 /*
1990 * Acquire both run-queue locks before placing the thread on the new
1991 * run-queue to avoid deadlocks created by placing a thread with a
1992 * blocked lock on the run-queue of a remote processor. The deadlock
1993 * occurs when a third processor attempts to lock the two queues in
1994 * question while the target processor is spinning with its own
1995 * run-queue lock held while waiting for the blocked lock to clear.
1996 */
1997 tdq_lock_pair(tdn, tdq);
1998 tdq_add(tdn, td, flags);
1999 tdq_notify(tdn, td);
2000 TDQ_UNLOCK(tdn);
2001 spinlock_exit();
2002 #endif
2003 return (TDQ_LOCKPTR(tdn));
2004 }
2005
2006 /*
2007 * Variadic version of thread_lock_unblock() that does not assume td_lock
2008 * is blocked.
2009 */
2010 static inline void
2011 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2012 {
2013 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2014 (uintptr_t)mtx);
2015 }
2016
2017 /*
2018 * Switch threads. This function has to handle threads coming in while
2019 * blocked for some reason, running, or idle. It also must deal with
2020 * migrating a thread from one queue to another as running threads may
2021 * be assigned elsewhere via binding.
2022 */
2023 void
2024 sched_switch(struct thread *td, struct thread *newtd, int flags)
2025 {
2026 struct tdq *tdq;
2027 struct td_sched *ts;
2028 struct mtx *mtx;
2029 int srqflag;
2030 int cpuid, preempted;
2031
2032 THREAD_LOCK_ASSERT(td, MA_OWNED);
2033 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2034
2035 cpuid = PCPU_GET(cpuid);
2036 tdq = TDQ_CPU(cpuid);
2037 ts = td_get_sched(td);
2038 mtx = td->td_lock;
2039 sched_pctcpu_update(ts, 1);
2040 ts->ts_rltick = ticks;
2041 td->td_lastcpu = td->td_oncpu;
2042 td->td_oncpu = NOCPU;
2043 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2044 (flags & SW_PREEMPT) != 0;
2045 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2046 td->td_owepreempt = 0;
2047 if (!TD_IS_IDLETHREAD(td))
2048 tdq->tdq_switchcnt++;
2049 /*
2050 * The lock pointer in an idle thread should never change. Reset it
2051 * to CAN_RUN as well.
2052 */
2053 if (TD_IS_IDLETHREAD(td)) {
2054 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2055 TD_SET_CAN_RUN(td);
2056 } else if (TD_IS_RUNNING(td)) {
2057 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2058 srqflag = preempted ?
2059 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2060 SRQ_OURSELF|SRQ_YIELDING;
2061 #ifdef SMP
2062 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2063 ts->ts_cpu = sched_pickcpu(td, 0);
2064 #endif
2065 if (ts->ts_cpu == cpuid)
2066 tdq_runq_add(tdq, td, srqflag);
2067 else {
2068 KASSERT(THREAD_CAN_MIGRATE(td) ||
2069 (ts->ts_flags & TSF_BOUND) != 0,
2070 ("Thread %p shouldn't migrate", td));
2071 mtx = sched_switch_migrate(tdq, td, srqflag);
2072 }
2073 } else {
2074 /* This thread must be going to sleep. */
2075 TDQ_LOCK(tdq);
2076 mtx = thread_lock_block(td);
2077 tdq_load_rem(tdq, td);
2078 #ifdef SMP
2079 if (tdq->tdq_load == 0)
2080 tdq_trysteal(tdq);
2081 #endif
2082 }
2083
2084 #if (KTR_COMPILE & KTR_SCHED) != 0
2085 if (TD_IS_IDLETHREAD(td))
2086 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2087 "prio:%d", td->td_priority);
2088 else
2089 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2090 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2091 "lockname:\"%s\"", td->td_lockname);
2092 #endif
2093
2094 /*
2095 * We enter here with the thread blocked and assigned to the
2096 * appropriate cpu run-queue or sleep-queue and with the current
2097 * thread-queue locked.
2098 */
2099 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2100 newtd = choosethread();
2101 /*
2102 * Call the MD code to switch contexts if necessary.
2103 */
2104 if (td != newtd) {
2105 #ifdef HWPMC_HOOKS
2106 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2107 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2108 #endif
2109 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2110 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2111 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2112 sched_pctcpu_update(td_get_sched(newtd), 0);
2113
2114 #ifdef KDTRACE_HOOKS
2115 /*
2116 * If DTrace has set the active vtime enum to anything
2117 * other than INACTIVE (0), then it should have set the
2118 * function to call.
2119 */
2120 if (dtrace_vtime_active)
2121 (*dtrace_vtime_switch_func)(newtd);
2122 #endif
2123
2124 cpu_switch(td, newtd, mtx);
2125 /*
2126 * We may return from cpu_switch on a different cpu. However,
2127 * we always return with td_lock pointing to the current cpu's
2128 * run queue lock.
2129 */
2130 cpuid = PCPU_GET(cpuid);
2131 tdq = TDQ_CPU(cpuid);
2132 lock_profile_obtain_lock_success(
2133 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2134
2135 SDT_PROBE0(sched, , , on__cpu);
2136 #ifdef HWPMC_HOOKS
2137 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2138 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2139 #endif
2140 } else {
2141 thread_unblock_switch(td, mtx);
2142 SDT_PROBE0(sched, , , remain__cpu);
2143 }
2144
2145 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2146 "prio:%d", td->td_priority);
2147
2148 /*
2149 * Assert that all went well and return.
2150 */
2151 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2152 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2153 td->td_oncpu = cpuid;
2154 }
2155
2156 /*
2157 * Adjust thread priorities as a result of a nice request.
2158 */
2159 void
2160 sched_nice(struct proc *p, int nice)
2161 {
2162 struct thread *td;
2163
2164 PROC_LOCK_ASSERT(p, MA_OWNED);
2165
2166 p->p_nice = nice;
2167 FOREACH_THREAD_IN_PROC(p, td) {
2168 thread_lock(td);
2169 sched_priority(td);
2170 sched_prio(td, td->td_base_user_pri);
2171 thread_unlock(td);
2172 }
2173 }
2174
2175 /*
2176 * Record the sleep time for the interactivity scorer.
2177 */
2178 void
2179 sched_sleep(struct thread *td, int prio)
2180 {
2181
2182 THREAD_LOCK_ASSERT(td, MA_OWNED);
2183
2184 td->td_slptick = ticks;
2185 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2186 td->td_flags |= TDF_CANSWAP;
2187 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2188 return;
2189 if (static_boost == 1 && prio)
2190 sched_prio(td, prio);
2191 else if (static_boost && td->td_priority > static_boost)
2192 sched_prio(td, static_boost);
2193 }
2194
2195 /*
2196 * Schedule a thread to resume execution and record how long it voluntarily
2197 * slept. We also update the pctcpu, interactivity, and priority.
2198 */
2199 void
2200 sched_wakeup(struct thread *td)
2201 {
2202 struct td_sched *ts;
2203 int slptick;
2204
2205 THREAD_LOCK_ASSERT(td, MA_OWNED);
2206 ts = td_get_sched(td);
2207 td->td_flags &= ~TDF_CANSWAP;
2208 /*
2209 * If we slept for more than a tick update our interactivity and
2210 * priority.
2211 */
2212 slptick = td->td_slptick;
2213 td->td_slptick = 0;
2214 if (slptick && slptick != ticks) {
2215 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2216 sched_interact_update(td);
2217 sched_pctcpu_update(ts, 0);
2218 }
2219 /*
2220 * Reset the slice value since we slept and advanced the round-robin.
2221 */
2222 ts->ts_slice = 0;
2223 sched_add(td, SRQ_BORING);
2224 }
2225
2226 /*
2227 * Penalize the parent for creating a new child and initialize the child's
2228 * priority.
2229 */
2230 void
2231 sched_fork(struct thread *td, struct thread *child)
2232 {
2233 THREAD_LOCK_ASSERT(td, MA_OWNED);
2234 sched_pctcpu_update(td_get_sched(td), 1);
2235 sched_fork_thread(td, child);
2236 /*
2237 * Penalize the parent and child for forking.
2238 */
2239 sched_interact_fork(child);
2240 sched_priority(child);
2241 td_get_sched(td)->ts_runtime += tickincr;
2242 sched_interact_update(td);
2243 sched_priority(td);
2244 }
2245
2246 /*
2247 * Fork a new thread, may be within the same process.
2248 */
2249 void
2250 sched_fork_thread(struct thread *td, struct thread *child)
2251 {
2252 struct td_sched *ts;
2253 struct td_sched *ts2;
2254 struct tdq *tdq;
2255
2256 tdq = TDQ_SELF();
2257 THREAD_LOCK_ASSERT(td, MA_OWNED);
2258 /*
2259 * Initialize child.
2260 */
2261 ts = td_get_sched(td);
2262 ts2 = td_get_sched(child);
2263 child->td_oncpu = NOCPU;
2264 child->td_lastcpu = NOCPU;
2265 child->td_lock = TDQ_LOCKPTR(tdq);
2266 child->td_cpuset = cpuset_ref(td->td_cpuset);
2267 ts2->ts_cpu = ts->ts_cpu;
2268 ts2->ts_flags = 0;
2269 /*
2270 * Grab our parents cpu estimation information.
2271 */
2272 ts2->ts_ticks = ts->ts_ticks;
2273 ts2->ts_ltick = ts->ts_ltick;
2274 ts2->ts_ftick = ts->ts_ftick;
2275 /*
2276 * Do not inherit any borrowed priority from the parent.
2277 */
2278 child->td_priority = child->td_base_pri;
2279 /*
2280 * And update interactivity score.
2281 */
2282 ts2->ts_slptime = ts->ts_slptime;
2283 ts2->ts_runtime = ts->ts_runtime;
2284 /* Attempt to quickly learn interactivity. */
2285 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2286 #ifdef KTR
2287 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2288 #endif
2289 }
2290
2291 /*
2292 * Adjust the priority class of a thread.
2293 */
2294 void
2295 sched_class(struct thread *td, int class)
2296 {
2297
2298 THREAD_LOCK_ASSERT(td, MA_OWNED);
2299 if (td->td_pri_class == class)
2300 return;
2301 td->td_pri_class = class;
2302 }
2303
2304 /*
2305 * Return some of the child's priority and interactivity to the parent.
2306 */
2307 void
2308 sched_exit(struct proc *p, struct thread *child)
2309 {
2310 struct thread *td;
2311
2312 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2313 "prio:%d", child->td_priority);
2314 PROC_LOCK_ASSERT(p, MA_OWNED);
2315 td = FIRST_THREAD_IN_PROC(p);
2316 sched_exit_thread(td, child);
2317 }
2318
2319 /*
2320 * Penalize another thread for the time spent on this one. This helps to
2321 * worsen the priority and interactivity of processes which schedule batch
2322 * jobs such as make. This has little effect on the make process itself but
2323 * causes new processes spawned by it to receive worse scores immediately.
2324 */
2325 void
2326 sched_exit_thread(struct thread *td, struct thread *child)
2327 {
2328
2329 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2330 "prio:%d", child->td_priority);
2331 /*
2332 * Give the child's runtime to the parent without returning the
2333 * sleep time as a penalty to the parent. This causes shells that
2334 * launch expensive things to mark their children as expensive.
2335 */
2336 thread_lock(td);
2337 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2338 sched_interact_update(td);
2339 sched_priority(td);
2340 thread_unlock(td);
2341 }
2342
2343 void
2344 sched_preempt(struct thread *td)
2345 {
2346 struct tdq *tdq;
2347
2348 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2349
2350 thread_lock(td);
2351 tdq = TDQ_SELF();
2352 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2353 tdq->tdq_ipipending = 0;
2354 if (td->td_priority > tdq->tdq_lowpri) {
2355 int flags;
2356
2357 flags = SW_INVOL | SW_PREEMPT;
2358 if (td->td_critnest > 1)
2359 td->td_owepreempt = 1;
2360 else if (TD_IS_IDLETHREAD(td))
2361 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2362 else
2363 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2364 }
2365 thread_unlock(td);
2366 }
2367
2368 /*
2369 * Fix priorities on return to user-space. Priorities may be elevated due
2370 * to static priorities in msleep() or similar.
2371 */
2372 void
2373 sched_userret(struct thread *td)
2374 {
2375 /*
2376 * XXX we cheat slightly on the locking here to avoid locking in
2377 * the usual case. Setting td_priority here is essentially an
2378 * incomplete workaround for not setting it properly elsewhere.
2379 * Now that some interrupt handlers are threads, not setting it
2380 * properly elsewhere can clobber it in the window between setting
2381 * it here and returning to user mode, so don't waste time setting
2382 * it perfectly here.
2383 */
2384 KASSERT((td->td_flags & TDF_BORROWING) == 0,
2385 ("thread with borrowed priority returning to userland"));
2386 if (td->td_priority != td->td_user_pri) {
2387 thread_lock(td);
2388 td->td_priority = td->td_user_pri;
2389 td->td_base_pri = td->td_user_pri;
2390 tdq_setlowpri(TDQ_SELF(), td);
2391 thread_unlock(td);
2392 }
2393 }
2394
2395 /*
2396 * Handle a stathz tick. This is really only relevant for timeshare
2397 * threads.
2398 */
2399 void
2400 sched_clock(struct thread *td)
2401 {
2402 struct tdq *tdq;
2403 struct td_sched *ts;
2404
2405 THREAD_LOCK_ASSERT(td, MA_OWNED);
2406 tdq = TDQ_SELF();
2407 #ifdef SMP
2408 /*
2409 * We run the long term load balancer infrequently on the first cpu.
2410 */
2411 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2412 if (balance_ticks && --balance_ticks == 0)
2413 sched_balance();
2414 }
2415 #endif
2416 /*
2417 * Save the old switch count so we have a record of the last ticks
2418 * activity. Initialize the new switch count based on our load.
2419 * If there is some activity seed it to reflect that.
2420 */
2421 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2422 tdq->tdq_switchcnt = tdq->tdq_load;
2423 /*
2424 * Advance the insert index once for each tick to ensure that all
2425 * threads get a chance to run.
2426 */
2427 if (tdq->tdq_idx == tdq->tdq_ridx) {
2428 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2429 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2430 tdq->tdq_ridx = tdq->tdq_idx;
2431 }
2432 ts = td_get_sched(td);
2433 sched_pctcpu_update(ts, 1);
2434 if (td->td_pri_class & PRI_FIFO_BIT)
2435 return;
2436 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2437 /*
2438 * We used a tick; charge it to the thread so
2439 * that we can compute our interactivity.
2440 */
2441 td_get_sched(td)->ts_runtime += tickincr;
2442 sched_interact_update(td);
2443 sched_priority(td);
2444 }
2445
2446 /*
2447 * Force a context switch if the current thread has used up a full
2448 * time slice (default is 100ms).
2449 */
2450 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2451 ts->ts_slice = 0;
2452 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2453 }
2454 }
2455
2456 u_int
2457 sched_estcpu(struct thread *td __unused)
2458 {
2459
2460 return (0);
2461 }
2462
2463 /*
2464 * Return whether the current CPU has runnable tasks. Used for in-kernel
2465 * cooperative idle threads.
2466 */
2467 int
2468 sched_runnable(void)
2469 {
2470 struct tdq *tdq;
2471 int load;
2472
2473 load = 1;
2474
2475 tdq = TDQ_SELF();
2476 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2477 if (tdq->tdq_load > 0)
2478 goto out;
2479 } else
2480 if (tdq->tdq_load - 1 > 0)
2481 goto out;
2482 load = 0;
2483 out:
2484 return (load);
2485 }
2486
2487 /*
2488 * Choose the highest priority thread to run. The thread is removed from
2489 * the run-queue while running however the load remains. For SMP we set
2490 * the tdq in the global idle bitmask if it idles here.
2491 */
2492 struct thread *
2493 sched_choose(void)
2494 {
2495 struct thread *td;
2496 struct tdq *tdq;
2497
2498 tdq = TDQ_SELF();
2499 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2500 td = tdq_choose(tdq);
2501 if (td) {
2502 tdq_runq_rem(tdq, td);
2503 tdq->tdq_lowpri = td->td_priority;
2504 return (td);
2505 }
2506 tdq->tdq_lowpri = PRI_MAX_IDLE;
2507 return (PCPU_GET(idlethread));
2508 }
2509
2510 /*
2511 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2512 * we always request it once we exit a critical section.
2513 */
2514 static inline void
2515 sched_setpreempt(struct thread *td)
2516 {
2517 struct thread *ctd;
2518 int cpri;
2519 int pri;
2520
2521 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2522
2523 ctd = curthread;
2524 pri = td->td_priority;
2525 cpri = ctd->td_priority;
2526 if (pri < cpri)
2527 ctd->td_flags |= TDF_NEEDRESCHED;
2528 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2529 return;
2530 if (!sched_shouldpreempt(pri, cpri, 0))
2531 return;
2532 ctd->td_owepreempt = 1;
2533 }
2534
2535 /*
2536 * Add a thread to a thread queue. Select the appropriate runq and add the
2537 * thread to it. This is the internal function called when the tdq is
2538 * predetermined.
2539 */
2540 void
2541 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2542 {
2543
2544 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2545 KASSERT((td->td_inhibitors == 0),
2546 ("sched_add: trying to run inhibited thread"));
2547 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2548 ("sched_add: bad thread state"));
2549 KASSERT(td->td_flags & TDF_INMEM,
2550 ("sched_add: thread swapped out"));
2551
2552 if (td->td_priority < tdq->tdq_lowpri)
2553 tdq->tdq_lowpri = td->td_priority;
2554 tdq_runq_add(tdq, td, flags);
2555 tdq_load_add(tdq, td);
2556 }
2557
2558 /*
2559 * Select the target thread queue and add a thread to it. Request
2560 * preemption or IPI a remote processor if required.
2561 */
2562 void
2563 sched_add(struct thread *td, int flags)
2564 {
2565 struct tdq *tdq;
2566 #ifdef SMP
2567 int cpu;
2568 #endif
2569
2570 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2571 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2572 sched_tdname(curthread));
2573 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2574 KTR_ATTR_LINKED, sched_tdname(td));
2575 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2576 flags & SRQ_PREEMPTED);
2577 THREAD_LOCK_ASSERT(td, MA_OWNED);
2578 /*
2579 * Recalculate the priority before we select the target cpu or
2580 * run-queue.
2581 */
2582 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2583 sched_priority(td);
2584 #ifdef SMP
2585 /*
2586 * Pick the destination cpu and if it isn't ours transfer to the
2587 * target cpu.
2588 */
2589 cpu = sched_pickcpu(td, flags);
2590 tdq = sched_setcpu(td, cpu, flags);
2591 tdq_add(tdq, td, flags);
2592 if (cpu != PCPU_GET(cpuid)) {
2593 tdq_notify(tdq, td);
2594 return;
2595 }
2596 #else
2597 tdq = TDQ_SELF();
2598 TDQ_LOCK(tdq);
2599 /*
2600 * Now that the thread is moving to the run-queue, set the lock
2601 * to the scheduler's lock.
2602 */
2603 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2604 tdq_add(tdq, td, flags);
2605 #endif
2606 if (!(flags & SRQ_YIELDING))
2607 sched_setpreempt(td);
2608 }
2609
2610 /*
2611 * Remove a thread from a run-queue without running it. This is used
2612 * when we're stealing a thread from a remote queue. Otherwise all threads
2613 * exit by calling sched_exit_thread() and sched_throw() themselves.
2614 */
2615 void
2616 sched_rem(struct thread *td)
2617 {
2618 struct tdq *tdq;
2619
2620 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2621 "prio:%d", td->td_priority);
2622 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2623 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2624 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2625 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2626 KASSERT(TD_ON_RUNQ(td),
2627 ("sched_rem: thread not on run queue"));
2628 tdq_runq_rem(tdq, td);
2629 tdq_load_rem(tdq, td);
2630 TD_SET_CAN_RUN(td);
2631 if (td->td_priority == tdq->tdq_lowpri)
2632 tdq_setlowpri(tdq, NULL);
2633 }
2634
2635 /*
2636 * Fetch cpu utilization information. Updates on demand.
2637 */
2638 fixpt_t
2639 sched_pctcpu(struct thread *td)
2640 {
2641 fixpt_t pctcpu;
2642 struct td_sched *ts;
2643
2644 pctcpu = 0;
2645 ts = td_get_sched(td);
2646
2647 THREAD_LOCK_ASSERT(td, MA_OWNED);
2648 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2649 if (ts->ts_ticks) {
2650 int rtick;
2651
2652 /* How many rtick per second ? */
2653 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2654 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2655 }
2656
2657 return (pctcpu);
2658 }
2659
2660 /*
2661 * Enforce affinity settings for a thread. Called after adjustments to
2662 * cpumask.
2663 */
2664 void
2665 sched_affinity(struct thread *td)
2666 {
2667 #ifdef SMP
2668 struct td_sched *ts;
2669
2670 THREAD_LOCK_ASSERT(td, MA_OWNED);
2671 ts = td_get_sched(td);
2672 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2673 return;
2674 if (TD_ON_RUNQ(td)) {
2675 sched_rem(td);
2676 sched_add(td, SRQ_BORING);
2677 return;
2678 }
2679 if (!TD_IS_RUNNING(td))
2680 return;
2681 /*
2682 * Force a switch before returning to userspace. If the
2683 * target thread is not running locally send an ipi to force
2684 * the issue.
2685 */
2686 td->td_flags |= TDF_NEEDRESCHED;
2687 if (td != curthread)
2688 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2689 #endif
2690 }
2691
2692 /*
2693 * Bind a thread to a target cpu.
2694 */
2695 void
2696 sched_bind(struct thread *td, int cpu)
2697 {
2698 struct td_sched *ts;
2699
2700 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2701 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2702 ts = td_get_sched(td);
2703 if (ts->ts_flags & TSF_BOUND)
2704 sched_unbind(td);
2705 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2706 ts->ts_flags |= TSF_BOUND;
2707 sched_pin();
2708 if (PCPU_GET(cpuid) == cpu)
2709 return;
2710 ts->ts_cpu = cpu;
2711 /* When we return from mi_switch we'll be on the correct cpu. */
2712 mi_switch(SW_VOL, NULL);
2713 }
2714
2715 /*
2716 * Release a bound thread.
2717 */
2718 void
2719 sched_unbind(struct thread *td)
2720 {
2721 struct td_sched *ts;
2722
2723 THREAD_LOCK_ASSERT(td, MA_OWNED);
2724 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2725 ts = td_get_sched(td);
2726 if ((ts->ts_flags & TSF_BOUND) == 0)
2727 return;
2728 ts->ts_flags &= ~TSF_BOUND;
2729 sched_unpin();
2730 }
2731
2732 int
2733 sched_is_bound(struct thread *td)
2734 {
2735 THREAD_LOCK_ASSERT(td, MA_OWNED);
2736 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2737 }
2738
2739 /*
2740 * Basic yield call.
2741 */
2742 void
2743 sched_relinquish(struct thread *td)
2744 {
2745 thread_lock(td);
2746 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2747 thread_unlock(td);
2748 }
2749
2750 /*
2751 * Return the total system load.
2752 */
2753 int
2754 sched_load(void)
2755 {
2756 #ifdef SMP
2757 int total;
2758 int i;
2759
2760 total = 0;
2761 CPU_FOREACH(i)
2762 total += TDQ_CPU(i)->tdq_sysload;
2763 return (total);
2764 #else
2765 return (TDQ_SELF()->tdq_sysload);
2766 #endif
2767 }
2768
2769 int
2770 sched_sizeof_proc(void)
2771 {
2772 return (sizeof(struct proc));
2773 }
2774
2775 int
2776 sched_sizeof_thread(void)
2777 {
2778 return (sizeof(struct thread) + sizeof(struct td_sched));
2779 }
2780
2781 #ifdef SMP
2782 #define TDQ_IDLESPIN(tdq) \
2783 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2784 #else
2785 #define TDQ_IDLESPIN(tdq) 1
2786 #endif
2787
2788 /*
2789 * The actual idle process.
2790 */
2791 void
2792 sched_idletd(void *dummy)
2793 {
2794 struct thread *td;
2795 struct tdq *tdq;
2796 int oldswitchcnt, switchcnt;
2797 int i;
2798
2799 mtx_assert(&Giant, MA_NOTOWNED);
2800 td = curthread;
2801 tdq = TDQ_SELF();
2802 THREAD_NO_SLEEPING();
2803 oldswitchcnt = -1;
2804 for (;;) {
2805 if (tdq->tdq_load) {
2806 thread_lock(td);
2807 mi_switch(SW_VOL | SWT_IDLE, NULL);
2808 thread_unlock(td);
2809 }
2810 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2811 #ifdef SMP
2812 if (always_steal || switchcnt != oldswitchcnt) {
2813 oldswitchcnt = switchcnt;
2814 if (tdq_idled(tdq) == 0)
2815 continue;
2816 }
2817 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2818 #else
2819 oldswitchcnt = switchcnt;
2820 #endif
2821 /*
2822 * If we're switching very frequently, spin while checking
2823 * for load rather than entering a low power state that
2824 * may require an IPI. However, don't do any busy
2825 * loops while on SMT machines as this simply steals
2826 * cycles from cores doing useful work.
2827 */
2828 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2829 for (i = 0; i < sched_idlespins; i++) {
2830 if (tdq->tdq_load)
2831 break;
2832 cpu_spinwait();
2833 }
2834 }
2835
2836 /* If there was context switch during spin, restart it. */
2837 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2838 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2839 continue;
2840
2841 /* Run main MD idle handler. */
2842 tdq->tdq_cpu_idle = 1;
2843 /*
2844 * Make sure that tdq_cpu_idle update is globally visible
2845 * before cpu_idle() read tdq_load. The order is important
2846 * to avoid race with tdq_notify.
2847 */
2848 atomic_thread_fence_seq_cst();
2849 /*
2850 * Checking for again after the fence picks up assigned
2851 * threads often enough to make it worthwhile to do so in
2852 * order to avoid calling cpu_idle().
2853 */
2854 if (tdq->tdq_load != 0) {
2855 tdq->tdq_cpu_idle = 0;
2856 continue;
2857 }
2858 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2859 tdq->tdq_cpu_idle = 0;
2860
2861 /*
2862 * Account thread-less hardware interrupts and
2863 * other wakeup reasons equal to context switches.
2864 */
2865 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2866 if (switchcnt != oldswitchcnt)
2867 continue;
2868 tdq->tdq_switchcnt++;
2869 oldswitchcnt++;
2870 }
2871 }
2872
2873 /*
2874 * A CPU is entering for the first time or a thread is exiting.
2875 */
2876 void
2877 sched_throw(struct thread *td)
2878 {
2879 struct thread *newtd;
2880 struct tdq *tdq;
2881
2882 tdq = TDQ_SELF();
2883 if (td == NULL) {
2884 /* Correct spinlock nesting and acquire the correct lock. */
2885 TDQ_LOCK(tdq);
2886 spinlock_exit();
2887 PCPU_SET(switchtime, cpu_ticks());
2888 PCPU_SET(switchticks, ticks);
2889 } else {
2890 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2891 tdq_load_rem(tdq, td);
2892 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2893 td->td_lastcpu = td->td_oncpu;
2894 td->td_oncpu = NOCPU;
2895 }
2896 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2897 newtd = choosethread();
2898 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2899 cpu_throw(td, newtd); /* doesn't return */
2900 }
2901
2902 /*
2903 * This is called from fork_exit(). Just acquire the correct locks and
2904 * let fork do the rest of the work.
2905 */
2906 void
2907 sched_fork_exit(struct thread *td)
2908 {
2909 struct tdq *tdq;
2910 int cpuid;
2911
2912 /*
2913 * Finish setting up thread glue so that it begins execution in a
2914 * non-nested critical section with the scheduler lock held.
2915 */
2916 cpuid = PCPU_GET(cpuid);
2917 tdq = TDQ_CPU(cpuid);
2918 if (TD_IS_IDLETHREAD(td))
2919 td->td_lock = TDQ_LOCKPTR(tdq);
2920 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2921 td->td_oncpu = cpuid;
2922 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2923 lock_profile_obtain_lock_success(
2924 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2925
2926 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2927 "prio:%d", td->td_priority);
2928 SDT_PROBE0(sched, , , on__cpu);
2929 }
2930
2931 /*
2932 * Create on first use to catch odd startup conditons.
2933 */
2934 char *
2935 sched_tdname(struct thread *td)
2936 {
2937 #ifdef KTR
2938 struct td_sched *ts;
2939
2940 ts = td_get_sched(td);
2941 if (ts->ts_name[0] == '\0')
2942 snprintf(ts->ts_name, sizeof(ts->ts_name),
2943 "%s tid %d", td->td_name, td->td_tid);
2944 return (ts->ts_name);
2945 #else
2946 return (td->td_name);
2947 #endif
2948 }
2949
2950 #ifdef KTR
2951 void
2952 sched_clear_tdname(struct thread *td)
2953 {
2954 struct td_sched *ts;
2955
2956 ts = td_get_sched(td);
2957 ts->ts_name[0] = '\0';
2958 }
2959 #endif
2960
2961 #ifdef SMP
2962
2963 /*
2964 * Build the CPU topology dump string. Is recursively called to collect
2965 * the topology tree.
2966 */
2967 static int
2968 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2969 int indent)
2970 {
2971 char cpusetbuf[CPUSETBUFSIZ];
2972 int i, first;
2973
2974 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2975 "", 1 + indent / 2, cg->cg_level);
2976 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2977 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2978 first = TRUE;
2979 for (i = 0; i < MAXCPU; i++) {
2980 if (CPU_ISSET(i, &cg->cg_mask)) {
2981 if (!first)
2982 sbuf_printf(sb, ", ");
2983 else
2984 first = FALSE;
2985 sbuf_printf(sb, "%d", i);
2986 }
2987 }
2988 sbuf_printf(sb, "</cpu>\n");
2989
2990 if (cg->cg_flags != 0) {
2991 sbuf_printf(sb, "%*s <flags>", indent, "");
2992 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2993 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2994 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2995 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2996 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2997 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2998 sbuf_printf(sb, "</flags>\n");
2999 }
3000
3001 if (cg->cg_children > 0) {
3002 sbuf_printf(sb, "%*s <children>\n", indent, "");
3003 for (i = 0; i < cg->cg_children; i++)
3004 sysctl_kern_sched_topology_spec_internal(sb,
3005 &cg->cg_child[i], indent+2);
3006 sbuf_printf(sb, "%*s </children>\n", indent, "");
3007 }
3008 sbuf_printf(sb, "%*s</group>\n", indent, "");
3009 return (0);
3010 }
3011
3012 /*
3013 * Sysctl handler for retrieving topology dump. It's a wrapper for
3014 * the recursive sysctl_kern_smp_topology_spec_internal().
3015 */
3016 static int
3017 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3018 {
3019 struct sbuf *topo;
3020 int err;
3021
3022 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3023
3024 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3025 if (topo == NULL)
3026 return (ENOMEM);
3027
3028 sbuf_printf(topo, "<groups>\n");
3029 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3030 sbuf_printf(topo, "</groups>\n");
3031
3032 if (err == 0) {
3033 err = sbuf_finish(topo);
3034 }
3035 sbuf_delete(topo);
3036 return (err);
3037 }
3038
3039 #endif
3040
3041 static int
3042 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3043 {
3044 int error, new_val, period;
3045
3046 period = 1000000 / realstathz;
3047 new_val = period * sched_slice;
3048 error = sysctl_handle_int(oidp, &new_val, 0, req);
3049 if (error != 0 || req->newptr == NULL)
3050 return (error);
3051 if (new_val <= 0)
3052 return (EINVAL);
3053 sched_slice = imax(1, (new_val + period / 2) / period);
3054 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3055 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3056 realstathz);
3057 return (0);
3058 }
3059
3060 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3061 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3062 "Scheduler name");
3063 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3064 NULL, 0, sysctl_kern_quantum, "I",
3065 "Quantum for timeshare threads in microseconds");
3066 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3067 "Quantum for timeshare threads in stathz ticks");
3068 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3069 "Interactivity score threshold");
3070 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3071 &preempt_thresh, 0,
3072 "Maximal (lowest) priority for preemption");
3073 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3074 "Assign static kernel priorities to sleeping threads");
3075 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3076 "Number of times idle thread will spin waiting for new work");
3077 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3078 &sched_idlespinthresh, 0,
3079 "Threshold before we will permit idle thread spinning");
3080 #ifdef SMP
3081 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3082 "Number of hz ticks to keep thread affinity for");
3083 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3084 "Enables the long-term load balancer");
3085 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3086 &balance_interval, 0,
3087 "Average period in stathz ticks to run the long-term balancer");
3088 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3089 "Attempts to steal work from other cores before idling");
3090 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3091 "Minimum load on remote CPU before we'll steal");
3092 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3093 0, "Topological distance limit for stealing threads in sched_switch()");
3094 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3095 "Always run the stealer from the idle thread");
3096 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3097 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3098 "XML dump of detected CPU topology");
3099 #endif
3100
3101 /* ps compat. All cpu percentages from ULE are weighted. */
3102 static int ccpu = 0;
3103 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
Cache object: be53eef2ac281fee0ecd532172dbf50e
|