FreeBSD/Linux Kernel Cross Reference
sys/kern/sched_ule.c
1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29 /*
30 * This file implements the ULE scheduler. ULE supports independent CPU
31 * run queues and fine grain locking. It has superior interactive
32 * performance under load even on uni-processor systems.
33 *
34 * etymology:
35 * ULE is the last three letters in schedule. It owes its name to a
36 * generic user created for a scheduling system by Paul Mikesell at
37 * Isilon Systems and a general lack of creativity on the part of the author.
38 */
39
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
45
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resource.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sdt.h>
59 #include <sys/smp.h>
60 #include <sys/sx.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/turnstile.h>
64 #include <sys/umtx.h>
65 #include <sys/vmmeter.h>
66 #include <sys/cpuset.h>
67 #include <sys/sbuf.h>
68
69 #ifdef HWPMC_HOOKS
70 #include <sys/pmckern.h>
71 #endif
72
73 #ifdef KDTRACE_HOOKS
74 #include <sys/dtrace_bsd.h>
75 int __read_mostly dtrace_vtime_active;
76 dtrace_vtime_switch_func_t dtrace_vtime_switch_func;
77 #endif
78
79 #include <machine/cpu.h>
80 #include <machine/smp.h>
81
82 #define KTR_ULE 0
83
84 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
85 #define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
86 #define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87
88 /*
89 * Thread scheduler specific section. All fields are protected
90 * by the thread lock.
91 */
92 struct td_sched {
93 struct runq *ts_runq; /* Run-queue we're queued on. */
94 short ts_flags; /* TSF_* flags. */
95 int ts_cpu; /* CPU that we have affinity for. */
96 int ts_rltick; /* Real last tick, for affinity. */
97 int ts_slice; /* Ticks of slice remaining. */
98 u_int ts_slptime; /* Number of ticks we vol. slept */
99 u_int ts_runtime; /* Number of ticks we were running */
100 int ts_ltick; /* Last tick that we were running on */
101 int ts_ftick; /* First tick that we were running on */
102 int ts_ticks; /* Tick count */
103 #ifdef KTR
104 char ts_name[TS_NAME_LEN];
105 #endif
106 };
107 /* flags kept in ts_flags */
108 #define TSF_BOUND 0x0001 /* Thread can not migrate. */
109 #define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
110
111 #define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
112 #define THREAD_CAN_SCHED(td, cpu) \
113 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114
115 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
116 sizeof(struct thread0_storage),
117 "increase struct thread0_storage.t0st_sched size");
118
119 /*
120 * Priority ranges used for interactive and non-interactive timeshare
121 * threads. The timeshare priorities are split up into four ranges.
122 * The first range handles interactive threads. The last three ranges
123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124 * ranges supporting nice values.
125 */
126 #define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
129
130 #define PRI_MIN_INTERACT PRI_MIN_TIMESHARE
131 #define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define PRI_MAX_BATCH PRI_MAX_TIMESHARE
134
135 /*
136 * Cpu percentage computation macros and defines.
137 *
138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back.
141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
144 */
145 #define SCHED_TICK_SECS 10
146 #define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
147 #define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
148 #define SCHED_TICK_SHIFT 10
149 #define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151
152 /*
153 * These macros determine priorities for non-interactive threads. They are
154 * assigned a priority based on their recent cpu utilization as expressed
155 * by the ratio of ticks to the tick total. NHALF priorities at the start
156 * and end of the MIN to MAX timeshare range are only reachable with negative
157 * or positive nice respectively.
158 *
159 * PRI_RANGE: Priority range for utilization dependent priorities.
160 * PRI_NRESV: Number of nice values.
161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
162 * PRI_NICE: Determines the part of the priority inherited from nice.
163 */
164 #define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
165 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
166 #define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define SCHED_PRI_TICKS(ts) \
170 (SCHED_TICK_HZ((ts)) / \
171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define SCHED_PRI_NICE(nice) (nice)
173
174 /*
175 * These determine the interactivity of a process. Interactivity differs from
176 * cpu utilization in that it expresses the voluntary time slept vs time ran
177 * while cpu utilization includes all time not running. This more accurately
178 * models the intent of the thread.
179 *
180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
181 * before throttling back.
182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
183 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
184 * INTERACT_THRESH: Threshold for placement on the current runq.
185 */
186 #define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
187 #define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
188 #define SCHED_INTERACT_MAX (100)
189 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
190 #define SCHED_INTERACT_THRESH (30)
191
192 /*
193 * These parameters determine the slice behavior for batch work.
194 */
195 #define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */
196 #define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */
197
198 /* Flags kept in td_flags. */
199 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */
200
201 /*
202 * tickincr: Converts a stathz tick into a hz domain scaled by
203 * the shift factor. Without the shift the error rate
204 * due to rounding would be unacceptably high.
205 * realstathz: stathz is sometimes 0 and run off of hz.
206 * sched_slice: Runtime of each thread before rescheduling.
207 * preempt_thresh: Priority threshold for preemption and remote IPIs.
208 */
209 static u_int __read_mostly sched_interact = SCHED_INTERACT_THRESH;
210 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
211 static int __read_mostly realstathz = 127; /* reset during boot. */
212 static int __read_mostly sched_slice = 10; /* reset during boot. */
213 static int __read_mostly sched_slice_min = 1; /* reset during boot. */
214 #ifdef PREEMPTION
215 #ifdef FULL_PREEMPTION
216 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
217 #else
218 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
219 #endif
220 #else
221 static int __read_mostly preempt_thresh = 0;
222 #endif
223 static int __read_mostly static_boost = PRI_MIN_BATCH;
224 static int __read_mostly sched_idlespins = 10000;
225 static int __read_mostly sched_idlespinthresh = -1;
226
227 /*
228 * tdq - per processor runqs and statistics. All fields are protected by the
229 * tdq_lock. The load and lowpri may be accessed without to avoid excess
230 * locking in sched_pickcpu();
231 */
232 struct tdq {
233 /*
234 * Ordered to improve efficiency of cpu_search() and switch().
235 * tdq_lock is padded to avoid false sharing with tdq_load and
236 * tdq_cpu_idle.
237 */
238 struct mtx_padalign tdq_lock; /* run queue lock. */
239 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */
240 volatile int tdq_load; /* Aggregate load. */
241 volatile int tdq_cpu_idle; /* cpu_idle() is active. */
242 int tdq_sysload; /* For loadavg, !ITHD load. */
243 volatile int tdq_transferable; /* Transferable thread count. */
244 volatile short tdq_switchcnt; /* Switches this tick. */
245 volatile short tdq_oldswitchcnt; /* Switches last tick. */
246 u_char tdq_lowpri; /* Lowest priority thread. */
247 u_char tdq_ipipending; /* IPI pending. */
248 u_char tdq_idx; /* Current insert index. */
249 u_char tdq_ridx; /* Current removal index. */
250 int tdq_id; /* cpuid. */
251 struct runq tdq_realtime; /* real-time run queue. */
252 struct runq tdq_timeshare; /* timeshare run queue. */
253 struct runq tdq_idle; /* Queue of IDLE threads. */
254 char tdq_name[TDQ_NAME_LEN];
255 #ifdef KTR
256 char tdq_loadname[TDQ_LOADNAME_LEN];
257 #endif
258 } __aligned(64);
259
260 /* Idle thread states and config. */
261 #define TDQ_RUNNING 1
262 #define TDQ_IDLE 2
263
264 #ifdef SMP
265 struct cpu_group __read_mostly *cpu_top; /* CPU topology */
266
267 #define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000))
268 #define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity))
269
270 /*
271 * Run-time tunables.
272 */
273 static int rebalance = 1;
274 static int balance_interval = 128; /* Default set in sched_initticks(). */
275 static int __read_mostly affinity;
276 static int __read_mostly steal_idle = 1;
277 static int __read_mostly steal_thresh = 2;
278 static int __read_mostly always_steal = 0;
279 static int __read_mostly trysteal_limit = 2;
280
281 /*
282 * One thread queue per processor.
283 */
284 static struct tdq __read_mostly *balance_tdq;
285 static int balance_ticks;
286 DPCPU_DEFINE_STATIC(struct tdq, tdq);
287 DPCPU_DEFINE_STATIC(uint32_t, randomval);
288
289 #define TDQ_SELF() (DPCPU_PTR(tdq))
290 #define TDQ_CPU(x) (DPCPU_ID_PTR((x), tdq))
291 #define TDQ_ID(x) ((x)->tdq_id)
292 #else /* !SMP */
293 static struct tdq tdq_cpu;
294
295 #define TDQ_ID(x) (0)
296 #define TDQ_SELF() (&tdq_cpu)
297 #define TDQ_CPU(x) (&tdq_cpu)
298 #endif
299
300 #define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type))
301 #define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t)))
302 #define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
303 #define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t)))
304 #define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock))
305
306 static void sched_priority(struct thread *);
307 static void sched_thread_priority(struct thread *, u_char);
308 static int sched_interact_score(struct thread *);
309 static void sched_interact_update(struct thread *);
310 static void sched_interact_fork(struct thread *);
311 static void sched_pctcpu_update(struct td_sched *, int);
312
313 /* Operations on per processor queues */
314 static struct thread *tdq_choose(struct tdq *);
315 static void tdq_setup(struct tdq *, int i);
316 static void tdq_load_add(struct tdq *, struct thread *);
317 static void tdq_load_rem(struct tdq *, struct thread *);
318 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
319 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
320 static inline int sched_shouldpreempt(int, int, int);
321 void tdq_print(int cpu);
322 static void runq_print(struct runq *rq);
323 static void tdq_add(struct tdq *, struct thread *, int);
324 #ifdef SMP
325 static struct thread *tdq_move(struct tdq *, struct tdq *);
326 static int tdq_idled(struct tdq *);
327 static void tdq_notify(struct tdq *, struct thread *);
328 static struct thread *tdq_steal(struct tdq *, int);
329 static struct thread *runq_steal(struct runq *, int);
330 static int sched_pickcpu(struct thread *, int);
331 static void sched_balance(void);
332 static int sched_balance_pair(struct tdq *, struct tdq *);
333 static inline struct tdq *sched_setcpu(struct thread *, int, int);
334 static inline void thread_unblock_switch(struct thread *, struct mtx *);
335 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
336 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
337 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
338 struct cpu_group *cg, int indent);
339 #endif
340
341 static void sched_setup(void *dummy);
342 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
343
344 static void sched_initticks(void *dummy);
345 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
346 NULL);
347
348 SDT_PROVIDER_DEFINE(sched);
349
350 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
351 "struct proc *", "uint8_t");
352 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
353 "struct proc *", "void *");
354 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
355 "struct proc *", "void *", "int");
356 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
357 "struct proc *", "uint8_t", "struct thread *");
358 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
359 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
360 "struct proc *");
361 SDT_PROBE_DEFINE(sched, , , on__cpu);
362 SDT_PROBE_DEFINE(sched, , , remain__cpu);
363 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
364 "struct proc *");
365
366 /*
367 * Print the threads waiting on a run-queue.
368 */
369 static void
370 runq_print(struct runq *rq)
371 {
372 struct rqhead *rqh;
373 struct thread *td;
374 int pri;
375 int j;
376 int i;
377
378 for (i = 0; i < RQB_LEN; i++) {
379 printf("\t\trunq bits %d 0x%zx\n",
380 i, rq->rq_status.rqb_bits[i]);
381 for (j = 0; j < RQB_BPW; j++)
382 if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
383 pri = j + (i << RQB_L2BPW);
384 rqh = &rq->rq_queues[pri];
385 TAILQ_FOREACH(td, rqh, td_runq) {
386 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
387 td, td->td_name, td->td_priority,
388 td->td_rqindex, pri);
389 }
390 }
391 }
392 }
393
394 /*
395 * Print the status of a per-cpu thread queue. Should be a ddb show cmd.
396 */
397 void
398 tdq_print(int cpu)
399 {
400 struct tdq *tdq;
401
402 tdq = TDQ_CPU(cpu);
403
404 printf("tdq %d:\n", TDQ_ID(tdq));
405 printf("\tlock %p\n", TDQ_LOCKPTR(tdq));
406 printf("\tLock name: %s\n", tdq->tdq_name);
407 printf("\tload: %d\n", tdq->tdq_load);
408 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt);
409 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
410 printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
411 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
412 printf("\tload transferable: %d\n", tdq->tdq_transferable);
413 printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
414 printf("\trealtime runq:\n");
415 runq_print(&tdq->tdq_realtime);
416 printf("\ttimeshare runq:\n");
417 runq_print(&tdq->tdq_timeshare);
418 printf("\tidle runq:\n");
419 runq_print(&tdq->tdq_idle);
420 }
421
422 static inline int
423 sched_shouldpreempt(int pri, int cpri, int remote)
424 {
425 /*
426 * If the new priority is not better than the current priority there is
427 * nothing to do.
428 */
429 if (pri >= cpri)
430 return (0);
431 /*
432 * Always preempt idle.
433 */
434 if (cpri >= PRI_MIN_IDLE)
435 return (1);
436 /*
437 * If preemption is disabled don't preempt others.
438 */
439 if (preempt_thresh == 0)
440 return (0);
441 /*
442 * Preempt if we exceed the threshold.
443 */
444 if (pri <= preempt_thresh)
445 return (1);
446 /*
447 * If we're interactive or better and there is non-interactive
448 * or worse running preempt only remote processors.
449 */
450 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
451 return (1);
452 return (0);
453 }
454
455 /*
456 * Add a thread to the actual run-queue. Keeps transferable counts up to
457 * date with what is actually on the run-queue. Selects the correct
458 * queue position for timeshare threads.
459 */
460 static __inline void
461 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
462 {
463 struct td_sched *ts;
464 u_char pri;
465
466 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
467 THREAD_LOCK_ASSERT(td, MA_OWNED);
468
469 pri = td->td_priority;
470 ts = td_get_sched(td);
471 TD_SET_RUNQ(td);
472 if (THREAD_CAN_MIGRATE(td)) {
473 tdq->tdq_transferable++;
474 ts->ts_flags |= TSF_XFERABLE;
475 }
476 if (pri < PRI_MIN_BATCH) {
477 ts->ts_runq = &tdq->tdq_realtime;
478 } else if (pri <= PRI_MAX_BATCH) {
479 ts->ts_runq = &tdq->tdq_timeshare;
480 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
481 ("Invalid priority %d on timeshare runq", pri));
482 /*
483 * This queue contains only priorities between MIN and MAX
484 * realtime. Use the whole queue to represent these values.
485 */
486 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
487 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
488 pri = (pri + tdq->tdq_idx) % RQ_NQS;
489 /*
490 * This effectively shortens the queue by one so we
491 * can have a one slot difference between idx and
492 * ridx while we wait for threads to drain.
493 */
494 if (tdq->tdq_ridx != tdq->tdq_idx &&
495 pri == tdq->tdq_ridx)
496 pri = (unsigned char)(pri - 1) % RQ_NQS;
497 } else
498 pri = tdq->tdq_ridx;
499 runq_add_pri(ts->ts_runq, td, pri, flags);
500 return;
501 } else
502 ts->ts_runq = &tdq->tdq_idle;
503 runq_add(ts->ts_runq, td, flags);
504 }
505
506 /*
507 * Remove a thread from a run-queue. This typically happens when a thread
508 * is selected to run. Running threads are not on the queue and the
509 * transferable count does not reflect them.
510 */
511 static __inline void
512 tdq_runq_rem(struct tdq *tdq, struct thread *td)
513 {
514 struct td_sched *ts;
515
516 ts = td_get_sched(td);
517 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
518 KASSERT(ts->ts_runq != NULL,
519 ("tdq_runq_remove: thread %p null ts_runq", td));
520 if (ts->ts_flags & TSF_XFERABLE) {
521 tdq->tdq_transferable--;
522 ts->ts_flags &= ~TSF_XFERABLE;
523 }
524 if (ts->ts_runq == &tdq->tdq_timeshare) {
525 if (tdq->tdq_idx != tdq->tdq_ridx)
526 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
527 else
528 runq_remove_idx(ts->ts_runq, td, NULL);
529 } else
530 runq_remove(ts->ts_runq, td);
531 }
532
533 /*
534 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load
535 * for this thread to the referenced thread queue.
536 */
537 static void
538 tdq_load_add(struct tdq *tdq, struct thread *td)
539 {
540
541 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
542 THREAD_LOCK_ASSERT(td, MA_OWNED);
543
544 tdq->tdq_load++;
545 if ((td->td_flags & TDF_NOLOAD) == 0)
546 tdq->tdq_sysload++;
547 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
548 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
549 }
550
551 /*
552 * Remove the load from a thread that is transitioning to a sleep state or
553 * exiting.
554 */
555 static void
556 tdq_load_rem(struct tdq *tdq, struct thread *td)
557 {
558
559 THREAD_LOCK_ASSERT(td, MA_OWNED);
560 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
561 KASSERT(tdq->tdq_load != 0,
562 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
563
564 tdq->tdq_load--;
565 if ((td->td_flags & TDF_NOLOAD) == 0)
566 tdq->tdq_sysload--;
567 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
568 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
569 }
570
571 /*
572 * Bound timeshare latency by decreasing slice size as load increases. We
573 * consider the maximum latency as the sum of the threads waiting to run
574 * aside from curthread and target no more than sched_slice latency but
575 * no less than sched_slice_min runtime.
576 */
577 static inline int
578 tdq_slice(struct tdq *tdq)
579 {
580 int load;
581
582 /*
583 * It is safe to use sys_load here because this is called from
584 * contexts where timeshare threads are running and so there
585 * cannot be higher priority load in the system.
586 */
587 load = tdq->tdq_sysload - 1;
588 if (load >= SCHED_SLICE_MIN_DIVISOR)
589 return (sched_slice_min);
590 if (load <= 1)
591 return (sched_slice);
592 return (sched_slice / load);
593 }
594
595 /*
596 * Set lowpri to its exact value by searching the run-queue and
597 * evaluating curthread. curthread may be passed as an optimization.
598 */
599 static void
600 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
601 {
602 struct thread *td;
603
604 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
605 if (ctd == NULL)
606 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
607 td = tdq_choose(tdq);
608 if (td == NULL || td->td_priority > ctd->td_priority)
609 tdq->tdq_lowpri = ctd->td_priority;
610 else
611 tdq->tdq_lowpri = td->td_priority;
612 }
613
614 #ifdef SMP
615 /*
616 * We need some randomness. Implement a classic Linear Congruential
617 * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
618 * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
619 * of the random state (in the low bits of our answer) to keep
620 * the maximum randomness.
621 */
622 static uint32_t
623 sched_random(void)
624 {
625 uint32_t *rndptr;
626
627 rndptr = DPCPU_PTR(randomval);
628 *rndptr = *rndptr * 69069 + 5;
629
630 return (*rndptr >> 16);
631 }
632
633 struct cpu_search {
634 cpuset_t cs_mask;
635 u_int cs_prefer;
636 int cs_pri; /* Min priority for low. */
637 int cs_limit; /* Max load for low, min load for high. */
638 int cs_cpu;
639 int cs_load;
640 };
641
642 #define CPU_SEARCH_LOWEST 0x1
643 #define CPU_SEARCH_HIGHEST 0x2
644 #define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
645
646 static __always_inline int cpu_search(const struct cpu_group *cg,
647 struct cpu_search *low, struct cpu_search *high, const int match);
648 int __noinline cpu_search_lowest(const struct cpu_group *cg,
649 struct cpu_search *low);
650 int __noinline cpu_search_highest(const struct cpu_group *cg,
651 struct cpu_search *high);
652 int __noinline cpu_search_both(const struct cpu_group *cg,
653 struct cpu_search *low, struct cpu_search *high);
654
655 /*
656 * Search the tree of cpu_groups for the lowest or highest loaded cpu
657 * according to the match argument. This routine actually compares the
658 * load on all paths through the tree and finds the least loaded cpu on
659 * the least loaded path, which may differ from the least loaded cpu in
660 * the system. This balances work among caches and buses.
661 *
662 * This inline is instantiated in three forms below using constants for the
663 * match argument. It is reduced to the minimum set for each case. It is
664 * also recursive to the depth of the tree.
665 */
666 static __always_inline int
667 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
668 struct cpu_search *high, const int match)
669 {
670 struct cpu_search lgroup;
671 struct cpu_search hgroup;
672 cpuset_t cpumask;
673 struct cpu_group *child;
674 struct tdq *tdq;
675 int cpu, i, hload, lload, load, total, rnd;
676
677 total = 0;
678 cpumask = cg->cg_mask;
679 if (match & CPU_SEARCH_LOWEST) {
680 lload = INT_MAX;
681 lgroup = *low;
682 }
683 if (match & CPU_SEARCH_HIGHEST) {
684 hload = INT_MIN;
685 hgroup = *high;
686 }
687
688 /* Iterate through the child CPU groups and then remaining CPUs. */
689 for (i = cg->cg_children, cpu = mp_maxid; ; ) {
690 if (i == 0) {
691 #ifdef HAVE_INLINE_FFSL
692 cpu = CPU_FFS(&cpumask) - 1;
693 #else
694 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
695 cpu--;
696 #endif
697 if (cpu < 0)
698 break;
699 child = NULL;
700 } else
701 child = &cg->cg_child[i - 1];
702
703 if (match & CPU_SEARCH_LOWEST)
704 lgroup.cs_cpu = -1;
705 if (match & CPU_SEARCH_HIGHEST)
706 hgroup.cs_cpu = -1;
707 if (child) { /* Handle child CPU group. */
708 CPU_NAND(&cpumask, &child->cg_mask);
709 switch (match) {
710 case CPU_SEARCH_LOWEST:
711 load = cpu_search_lowest(child, &lgroup);
712 break;
713 case CPU_SEARCH_HIGHEST:
714 load = cpu_search_highest(child, &hgroup);
715 break;
716 case CPU_SEARCH_BOTH:
717 load = cpu_search_both(child, &lgroup, &hgroup);
718 break;
719 }
720 } else { /* Handle child CPU. */
721 CPU_CLR(cpu, &cpumask);
722 tdq = TDQ_CPU(cpu);
723 load = tdq->tdq_load * 256;
724 rnd = sched_random() % 32;
725 if (match & CPU_SEARCH_LOWEST) {
726 if (cpu == low->cs_prefer)
727 load -= 64;
728 /* If that CPU is allowed and get data. */
729 if (tdq->tdq_lowpri > lgroup.cs_pri &&
730 tdq->tdq_load <= lgroup.cs_limit &&
731 CPU_ISSET(cpu, &lgroup.cs_mask)) {
732 lgroup.cs_cpu = cpu;
733 lgroup.cs_load = load - rnd;
734 }
735 }
736 if (match & CPU_SEARCH_HIGHEST)
737 if (tdq->tdq_load >= hgroup.cs_limit &&
738 tdq->tdq_transferable &&
739 CPU_ISSET(cpu, &hgroup.cs_mask)) {
740 hgroup.cs_cpu = cpu;
741 hgroup.cs_load = load - rnd;
742 }
743 }
744 total += load;
745
746 /* We have info about child item. Compare it. */
747 if (match & CPU_SEARCH_LOWEST) {
748 if (lgroup.cs_cpu >= 0 &&
749 (load < lload ||
750 (load == lload && lgroup.cs_load < low->cs_load))) {
751 lload = load;
752 low->cs_cpu = lgroup.cs_cpu;
753 low->cs_load = lgroup.cs_load;
754 }
755 }
756 if (match & CPU_SEARCH_HIGHEST)
757 if (hgroup.cs_cpu >= 0 &&
758 (load > hload ||
759 (load == hload && hgroup.cs_load > high->cs_load))) {
760 hload = load;
761 high->cs_cpu = hgroup.cs_cpu;
762 high->cs_load = hgroup.cs_load;
763 }
764 if (child) {
765 i--;
766 if (i == 0 && CPU_EMPTY(&cpumask))
767 break;
768 }
769 #ifndef HAVE_INLINE_FFSL
770 else
771 cpu--;
772 #endif
773 }
774 return (total);
775 }
776
777 /*
778 * cpu_search instantiations must pass constants to maintain the inline
779 * optimization.
780 */
781 int
782 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
783 {
784 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
785 }
786
787 int
788 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
789 {
790 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
791 }
792
793 int
794 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
795 struct cpu_search *high)
796 {
797 return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
798 }
799
800 /*
801 * Find the cpu with the least load via the least loaded path that has a
802 * lowpri greater than pri pri. A pri of -1 indicates any priority is
803 * acceptable.
804 */
805 static inline int
806 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
807 int prefer)
808 {
809 struct cpu_search low;
810
811 low.cs_cpu = -1;
812 low.cs_prefer = prefer;
813 low.cs_mask = mask;
814 low.cs_pri = pri;
815 low.cs_limit = maxload;
816 cpu_search_lowest(cg, &low);
817 return low.cs_cpu;
818 }
819
820 /*
821 * Find the cpu with the highest load via the highest loaded path.
822 */
823 static inline int
824 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
825 {
826 struct cpu_search high;
827
828 high.cs_cpu = -1;
829 high.cs_mask = mask;
830 high.cs_limit = minload;
831 cpu_search_highest(cg, &high);
832 return high.cs_cpu;
833 }
834
835 static void
836 sched_balance_group(struct cpu_group *cg)
837 {
838 struct tdq *tdq;
839 cpuset_t hmask, lmask;
840 int high, low, anylow;
841
842 CPU_FILL(&hmask);
843 for (;;) {
844 high = sched_highest(cg, hmask, 2);
845 /* Stop if there is no more CPU with transferrable threads. */
846 if (high == -1)
847 break;
848 CPU_CLR(high, &hmask);
849 CPU_COPY(&hmask, &lmask);
850 /* Stop if there is no more CPU left for low. */
851 if (CPU_EMPTY(&lmask))
852 break;
853 anylow = 1;
854 tdq = TDQ_CPU(high);
855 nextlow:
856 low = sched_lowest(cg, lmask, -1, tdq->tdq_load - 1, high);
857 /* Stop if we looked well and found no less loaded CPU. */
858 if (anylow && low == -1)
859 break;
860 /* Go to next high if we found no less loaded CPU. */
861 if (low == -1)
862 continue;
863 /* Transfer thread from high to low. */
864 if (sched_balance_pair(tdq, TDQ_CPU(low))) {
865 /* CPU that got thread can no longer be a donor. */
866 CPU_CLR(low, &hmask);
867 } else {
868 /*
869 * If failed, then there is no threads on high
870 * that can run on this low. Drop low from low
871 * mask and look for different one.
872 */
873 CPU_CLR(low, &lmask);
874 anylow = 0;
875 goto nextlow;
876 }
877 }
878 }
879
880 static void
881 sched_balance(void)
882 {
883 struct tdq *tdq;
884
885 balance_ticks = max(balance_interval / 2, 1) +
886 (sched_random() % balance_interval);
887 tdq = TDQ_SELF();
888 TDQ_UNLOCK(tdq);
889 sched_balance_group(cpu_top);
890 TDQ_LOCK(tdq);
891 }
892
893 /*
894 * Lock two thread queues using their address to maintain lock order.
895 */
896 static void
897 tdq_lock_pair(struct tdq *one, struct tdq *two)
898 {
899 if (one < two) {
900 TDQ_LOCK(one);
901 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
902 } else {
903 TDQ_LOCK(two);
904 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
905 }
906 }
907
908 /*
909 * Unlock two thread queues. Order is not important here.
910 */
911 static void
912 tdq_unlock_pair(struct tdq *one, struct tdq *two)
913 {
914 TDQ_UNLOCK(one);
915 TDQ_UNLOCK(two);
916 }
917
918 /*
919 * Transfer load between two imbalanced thread queues.
920 */
921 static int
922 sched_balance_pair(struct tdq *high, struct tdq *low)
923 {
924 struct thread *td;
925 int cpu;
926
927 tdq_lock_pair(high, low);
928 td = NULL;
929 /*
930 * Transfer a thread from high to low.
931 */
932 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
933 (td = tdq_move(high, low)) != NULL) {
934 /*
935 * In case the target isn't the current cpu notify it of the
936 * new load, possibly sending an IPI to force it to reschedule.
937 */
938 cpu = TDQ_ID(low);
939 if (cpu != PCPU_GET(cpuid))
940 tdq_notify(low, td);
941 }
942 tdq_unlock_pair(high, low);
943 return (td != NULL);
944 }
945
946 /*
947 * Move a thread from one thread queue to another.
948 */
949 static struct thread *
950 tdq_move(struct tdq *from, struct tdq *to)
951 {
952 struct td_sched *ts;
953 struct thread *td;
954 struct tdq *tdq;
955 int cpu;
956
957 TDQ_LOCK_ASSERT(from, MA_OWNED);
958 TDQ_LOCK_ASSERT(to, MA_OWNED);
959
960 tdq = from;
961 cpu = TDQ_ID(to);
962 td = tdq_steal(tdq, cpu);
963 if (td == NULL)
964 return (NULL);
965 ts = td_get_sched(td);
966 /*
967 * Although the run queue is locked the thread may be blocked. Lock
968 * it to clear this and acquire the run-queue lock.
969 */
970 thread_lock(td);
971 /* Drop recursive lock on from acquired via thread_lock(). */
972 TDQ_UNLOCK(from);
973 sched_rem(td);
974 ts->ts_cpu = cpu;
975 td->td_lock = TDQ_LOCKPTR(to);
976 tdq_add(to, td, SRQ_YIELDING);
977 return (td);
978 }
979
980 /*
981 * This tdq has idled. Try to steal a thread from another cpu and switch
982 * to it.
983 */
984 static int
985 tdq_idled(struct tdq *tdq)
986 {
987 struct cpu_group *cg;
988 struct tdq *steal;
989 cpuset_t mask;
990 int cpu, switchcnt;
991
992 if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
993 return (1);
994 CPU_FILL(&mask);
995 CPU_CLR(PCPU_GET(cpuid), &mask);
996 restart:
997 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
998 for (cg = tdq->tdq_cg; ; ) {
999 cpu = sched_highest(cg, mask, steal_thresh);
1000 /*
1001 * We were assigned a thread but not preempted. Returning
1002 * 0 here will cause our caller to switch to it.
1003 */
1004 if (tdq->tdq_load)
1005 return (0);
1006 if (cpu == -1) {
1007 cg = cg->cg_parent;
1008 if (cg == NULL)
1009 return (1);
1010 continue;
1011 }
1012 steal = TDQ_CPU(cpu);
1013 /*
1014 * The data returned by sched_highest() is stale and
1015 * the chosen CPU no longer has an eligible thread.
1016 *
1017 * Testing this ahead of tdq_lock_pair() only catches
1018 * this situation about 20% of the time on an 8 core
1019 * 16 thread Ryzen 7, but it still helps performance.
1020 */
1021 if (steal->tdq_load < steal_thresh ||
1022 steal->tdq_transferable == 0)
1023 goto restart;
1024 tdq_lock_pair(tdq, steal);
1025 /*
1026 * We were assigned a thread while waiting for the locks.
1027 * Switch to it now instead of stealing a thread.
1028 */
1029 if (tdq->tdq_load)
1030 break;
1031 /*
1032 * The data returned by sched_highest() is stale and
1033 * the chosen CPU no longer has an eligible thread, or
1034 * we were preempted and the CPU loading info may be out
1035 * of date. The latter is rare. In either case restart
1036 * the search.
1037 */
1038 if (steal->tdq_load < steal_thresh ||
1039 steal->tdq_transferable == 0 ||
1040 switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1041 tdq_unlock_pair(tdq, steal);
1042 goto restart;
1043 }
1044 /*
1045 * Steal the thread and switch to it.
1046 */
1047 if (tdq_move(steal, tdq) != NULL)
1048 break;
1049 /*
1050 * We failed to acquire a thread even though it looked
1051 * like one was available. This could be due to affinity
1052 * restrictions or for other reasons. Loop again after
1053 * removing this CPU from the set. The restart logic
1054 * above does not restore this CPU to the set due to the
1055 * likelyhood of failing here again.
1056 */
1057 CPU_CLR(cpu, &mask);
1058 tdq_unlock_pair(tdq, steal);
1059 }
1060 TDQ_UNLOCK(steal);
1061 mi_switch(SW_VOL | SWT_IDLE, NULL);
1062 thread_unlock(curthread);
1063 return (0);
1064 }
1065
1066 /*
1067 * Notify a remote cpu of new work. Sends an IPI if criteria are met.
1068 */
1069 static void
1070 tdq_notify(struct tdq *tdq, struct thread *td)
1071 {
1072 struct thread *ctd;
1073 int pri;
1074 int cpu;
1075
1076 if (tdq->tdq_ipipending)
1077 return;
1078 cpu = td_get_sched(td)->ts_cpu;
1079 pri = td->td_priority;
1080 ctd = pcpu_find(cpu)->pc_curthread;
1081 if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1082 return;
1083
1084 /*
1085 * Make sure that our caller's earlier update to tdq_load is
1086 * globally visible before we read tdq_cpu_idle. Idle thread
1087 * accesses both of them without locks, and the order is important.
1088 */
1089 atomic_thread_fence_seq_cst();
1090
1091 if (TD_IS_IDLETHREAD(ctd)) {
1092 /*
1093 * If the MD code has an idle wakeup routine try that before
1094 * falling back to IPI.
1095 */
1096 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1097 return;
1098 }
1099 tdq->tdq_ipipending = 1;
1100 ipi_cpu(cpu, IPI_PREEMPT);
1101 }
1102
1103 /*
1104 * Steals load from a timeshare queue. Honors the rotating queue head
1105 * index.
1106 */
1107 static struct thread *
1108 runq_steal_from(struct runq *rq, int cpu, u_char start)
1109 {
1110 struct rqbits *rqb;
1111 struct rqhead *rqh;
1112 struct thread *td, *first;
1113 int bit;
1114 int i;
1115
1116 rqb = &rq->rq_status;
1117 bit = start & (RQB_BPW -1);
1118 first = NULL;
1119 again:
1120 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1121 if (rqb->rqb_bits[i] == 0)
1122 continue;
1123 if (bit == 0)
1124 bit = RQB_FFS(rqb->rqb_bits[i]);
1125 for (; bit < RQB_BPW; bit++) {
1126 if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1127 continue;
1128 rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1129 TAILQ_FOREACH(td, rqh, td_runq) {
1130 if (first) {
1131 if (THREAD_CAN_MIGRATE(td) &&
1132 THREAD_CAN_SCHED(td, cpu))
1133 return (td);
1134 } else
1135 first = td;
1136 }
1137 }
1138 }
1139 if (start != 0) {
1140 start = 0;
1141 goto again;
1142 }
1143
1144 if (first && THREAD_CAN_MIGRATE(first) &&
1145 THREAD_CAN_SCHED(first, cpu))
1146 return (first);
1147 return (NULL);
1148 }
1149
1150 /*
1151 * Steals load from a standard linear queue.
1152 */
1153 static struct thread *
1154 runq_steal(struct runq *rq, int cpu)
1155 {
1156 struct rqhead *rqh;
1157 struct rqbits *rqb;
1158 struct thread *td;
1159 int word;
1160 int bit;
1161
1162 rqb = &rq->rq_status;
1163 for (word = 0; word < RQB_LEN; word++) {
1164 if (rqb->rqb_bits[word] == 0)
1165 continue;
1166 for (bit = 0; bit < RQB_BPW; bit++) {
1167 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1168 continue;
1169 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1170 TAILQ_FOREACH(td, rqh, td_runq)
1171 if (THREAD_CAN_MIGRATE(td) &&
1172 THREAD_CAN_SCHED(td, cpu))
1173 return (td);
1174 }
1175 }
1176 return (NULL);
1177 }
1178
1179 /*
1180 * Attempt to steal a thread in priority order from a thread queue.
1181 */
1182 static struct thread *
1183 tdq_steal(struct tdq *tdq, int cpu)
1184 {
1185 struct thread *td;
1186
1187 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1188 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1189 return (td);
1190 if ((td = runq_steal_from(&tdq->tdq_timeshare,
1191 cpu, tdq->tdq_ridx)) != NULL)
1192 return (td);
1193 return (runq_steal(&tdq->tdq_idle, cpu));
1194 }
1195
1196 /*
1197 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the
1198 * current lock and returns with the assigned queue locked.
1199 */
1200 static inline struct tdq *
1201 sched_setcpu(struct thread *td, int cpu, int flags)
1202 {
1203
1204 struct tdq *tdq;
1205
1206 THREAD_LOCK_ASSERT(td, MA_OWNED);
1207 tdq = TDQ_CPU(cpu);
1208 td_get_sched(td)->ts_cpu = cpu;
1209 /*
1210 * If the lock matches just return the queue.
1211 */
1212 if (td->td_lock == TDQ_LOCKPTR(tdq))
1213 return (tdq);
1214 #ifdef notyet
1215 /*
1216 * If the thread isn't running its lockptr is a
1217 * turnstile or a sleepqueue. We can just lock_set without
1218 * blocking.
1219 */
1220 if (TD_CAN_RUN(td)) {
1221 TDQ_LOCK(tdq);
1222 thread_lock_set(td, TDQ_LOCKPTR(tdq));
1223 return (tdq);
1224 }
1225 #endif
1226 /*
1227 * The hard case, migration, we need to block the thread first to
1228 * prevent order reversals with other cpus locks.
1229 */
1230 spinlock_enter();
1231 thread_lock_block(td);
1232 TDQ_LOCK(tdq);
1233 thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1234 spinlock_exit();
1235 return (tdq);
1236 }
1237
1238 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1239 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1240 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1241 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1242 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1243 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1244
1245 static int
1246 sched_pickcpu(struct thread *td, int flags)
1247 {
1248 struct cpu_group *cg, *ccg;
1249 struct td_sched *ts;
1250 struct tdq *tdq;
1251 cpuset_t mask;
1252 int cpu, pri, self, intr;
1253
1254 self = PCPU_GET(cpuid);
1255 ts = td_get_sched(td);
1256 KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1257 "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1258 if (smp_started == 0)
1259 return (self);
1260 /*
1261 * Don't migrate a running thread from sched_switch().
1262 */
1263 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1264 return (ts->ts_cpu);
1265 /*
1266 * Prefer to run interrupt threads on the processors that generate
1267 * the interrupt.
1268 */
1269 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1270 curthread->td_intr_nesting_level) {
1271 tdq = TDQ_SELF();
1272 if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1273 SCHED_STAT_INC(pickcpu_idle_affinity);
1274 return (self);
1275 }
1276 ts->ts_cpu = self;
1277 intr = 1;
1278 cg = tdq->tdq_cg;
1279 goto llc;
1280 } else {
1281 intr = 0;
1282 tdq = TDQ_CPU(ts->ts_cpu);
1283 cg = tdq->tdq_cg;
1284 }
1285 /*
1286 * If the thread can run on the last cpu and the affinity has not
1287 * expired and it is idle, run it there.
1288 */
1289 if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1290 tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1291 SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1292 if (cg->cg_flags & CG_FLAG_THREAD) {
1293 /* Check all SMT threads for being idle. */
1294 for (cpu = CPU_FFS(&cg->cg_mask) - 1; ; cpu++) {
1295 if (CPU_ISSET(cpu, &cg->cg_mask) &&
1296 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1297 break;
1298 if (cpu >= mp_maxid) {
1299 SCHED_STAT_INC(pickcpu_idle_affinity);
1300 return (ts->ts_cpu);
1301 }
1302 }
1303 } else {
1304 SCHED_STAT_INC(pickcpu_idle_affinity);
1305 return (ts->ts_cpu);
1306 }
1307 }
1308 llc:
1309 /*
1310 * Search for the last level cache CPU group in the tree.
1311 * Skip SMT, identical groups and caches with expired affinity.
1312 * Interrupt threads affinity is explicit and never expires.
1313 */
1314 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1315 if (cg->cg_flags & CG_FLAG_THREAD)
1316 continue;
1317 if (cg->cg_children == 1 || cg->cg_count == 1)
1318 continue;
1319 if (cg->cg_level == CG_SHARE_NONE ||
1320 (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1321 continue;
1322 ccg = cg;
1323 }
1324 /* Found LLC shared by all CPUs, so do a global search. */
1325 if (ccg == cpu_top)
1326 ccg = NULL;
1327 cpu = -1;
1328 mask = td->td_cpuset->cs_mask;
1329 pri = td->td_priority;
1330 /*
1331 * Try hard to keep interrupts within found LLC. Search the LLC for
1332 * the least loaded CPU we can run now. For NUMA systems it should
1333 * be within target domain, and it also reduces scheduling overhead.
1334 */
1335 if (ccg != NULL && intr) {
1336 cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu);
1337 if (cpu >= 0)
1338 SCHED_STAT_INC(pickcpu_intrbind);
1339 } else
1340 /* Search the LLC for the least loaded idle CPU we can run now. */
1341 if (ccg != NULL) {
1342 cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1343 INT_MAX, ts->ts_cpu);
1344 if (cpu >= 0)
1345 SCHED_STAT_INC(pickcpu_affinity);
1346 }
1347 /* Search globally for the least loaded CPU we can run now. */
1348 if (cpu < 0) {
1349 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1350 if (cpu >= 0)
1351 SCHED_STAT_INC(pickcpu_lowest);
1352 }
1353 /* Search globally for the least loaded CPU. */
1354 if (cpu < 0) {
1355 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1356 if (cpu >= 0)
1357 SCHED_STAT_INC(pickcpu_lowest);
1358 }
1359 KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1360 KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1361 /*
1362 * Compare the lowest loaded cpu to current cpu.
1363 */
1364 tdq = TDQ_CPU(cpu);
1365 if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1366 tdq->tdq_lowpri < PRI_MIN_IDLE &&
1367 TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1368 SCHED_STAT_INC(pickcpu_local);
1369 cpu = self;
1370 }
1371 if (cpu != ts->ts_cpu)
1372 SCHED_STAT_INC(pickcpu_migration);
1373 return (cpu);
1374 }
1375 #endif
1376
1377 /*
1378 * Pick the highest priority task we have and return it.
1379 */
1380 static struct thread *
1381 tdq_choose(struct tdq *tdq)
1382 {
1383 struct thread *td;
1384
1385 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1386 td = runq_choose(&tdq->tdq_realtime);
1387 if (td != NULL)
1388 return (td);
1389 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1390 if (td != NULL) {
1391 KASSERT(td->td_priority >= PRI_MIN_BATCH,
1392 ("tdq_choose: Invalid priority on timeshare queue %d",
1393 td->td_priority));
1394 return (td);
1395 }
1396 td = runq_choose(&tdq->tdq_idle);
1397 if (td != NULL) {
1398 KASSERT(td->td_priority >= PRI_MIN_IDLE,
1399 ("tdq_choose: Invalid priority on idle queue %d",
1400 td->td_priority));
1401 return (td);
1402 }
1403
1404 return (NULL);
1405 }
1406
1407 /*
1408 * Initialize a thread queue.
1409 */
1410 static void
1411 tdq_setup(struct tdq *tdq, int id)
1412 {
1413
1414 if (bootverbose)
1415 printf("ULE: setup cpu %d\n", id);
1416 runq_init(&tdq->tdq_realtime);
1417 runq_init(&tdq->tdq_timeshare);
1418 runq_init(&tdq->tdq_idle);
1419 tdq->tdq_id = id;
1420 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1421 "sched lock %d", (int)TDQ_ID(tdq));
1422 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1423 MTX_SPIN | MTX_RECURSE);
1424 #ifdef KTR
1425 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1426 "CPU %d load", (int)TDQ_ID(tdq));
1427 #endif
1428 }
1429
1430 #ifdef SMP
1431 static void
1432 sched_setup_smp(void)
1433 {
1434 struct tdq *tdq;
1435 int i;
1436
1437 cpu_top = smp_topo();
1438 CPU_FOREACH(i) {
1439 tdq = DPCPU_ID_PTR(i, tdq);
1440 tdq_setup(tdq, i);
1441 tdq->tdq_cg = smp_topo_find(cpu_top, i);
1442 if (tdq->tdq_cg == NULL)
1443 panic("Can't find cpu group for %d\n", i);
1444 }
1445 balance_tdq = TDQ_SELF();
1446 }
1447 #endif
1448
1449 /*
1450 * Setup the thread queues and initialize the topology based on MD
1451 * information.
1452 */
1453 static void
1454 sched_setup(void *dummy)
1455 {
1456 struct tdq *tdq;
1457
1458 #ifdef SMP
1459 sched_setup_smp();
1460 #else
1461 tdq_setup(TDQ_SELF(), 0);
1462 #endif
1463 tdq = TDQ_SELF();
1464
1465 /* Add thread0's load since it's running. */
1466 TDQ_LOCK(tdq);
1467 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1468 tdq_load_add(tdq, &thread0);
1469 tdq->tdq_lowpri = thread0.td_priority;
1470 TDQ_UNLOCK(tdq);
1471 }
1472
1473 /*
1474 * This routine determines time constants after stathz and hz are setup.
1475 */
1476 /* ARGSUSED */
1477 static void
1478 sched_initticks(void *dummy)
1479 {
1480 int incr;
1481
1482 realstathz = stathz ? stathz : hz;
1483 sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1484 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1485 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1486 realstathz);
1487
1488 /*
1489 * tickincr is shifted out by 10 to avoid rounding errors due to
1490 * hz not being evenly divisible by stathz on all platforms.
1491 */
1492 incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1493 /*
1494 * This does not work for values of stathz that are more than
1495 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
1496 */
1497 if (incr == 0)
1498 incr = 1;
1499 tickincr = incr;
1500 #ifdef SMP
1501 /*
1502 * Set the default balance interval now that we know
1503 * what realstathz is.
1504 */
1505 balance_interval = realstathz;
1506 balance_ticks = balance_interval;
1507 affinity = SCHED_AFFINITY_DEFAULT;
1508 #endif
1509 if (sched_idlespinthresh < 0)
1510 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1511 }
1512
1513
1514 /*
1515 * This is the core of the interactivity algorithm. Determines a score based
1516 * on past behavior. It is the ratio of sleep time to run time scaled to
1517 * a [0, 100] integer. This is the voluntary sleep time of a process, which
1518 * differs from the cpu usage because it does not account for time spent
1519 * waiting on a run-queue. Would be prettier if we had floating point.
1520 *
1521 * When a thread's sleep time is greater than its run time the
1522 * calculation is:
1523 *
1524 * scaling factor
1525 * interactivity score = ---------------------
1526 * sleep time / run time
1527 *
1528 *
1529 * When a thread's run time is greater than its sleep time the
1530 * calculation is:
1531 *
1532 * scaling factor
1533 * interactivity score = --------------------- + scaling factor
1534 * run time / sleep time
1535 */
1536 static int
1537 sched_interact_score(struct thread *td)
1538 {
1539 struct td_sched *ts;
1540 int div;
1541
1542 ts = td_get_sched(td);
1543 /*
1544 * The score is only needed if this is likely to be an interactive
1545 * task. Don't go through the expense of computing it if there's
1546 * no chance.
1547 */
1548 if (sched_interact <= SCHED_INTERACT_HALF &&
1549 ts->ts_runtime >= ts->ts_slptime)
1550 return (SCHED_INTERACT_HALF);
1551
1552 if (ts->ts_runtime > ts->ts_slptime) {
1553 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1554 return (SCHED_INTERACT_HALF +
1555 (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1556 }
1557 if (ts->ts_slptime > ts->ts_runtime) {
1558 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1559 return (ts->ts_runtime / div);
1560 }
1561 /* runtime == slptime */
1562 if (ts->ts_runtime)
1563 return (SCHED_INTERACT_HALF);
1564
1565 /*
1566 * This can happen if slptime and runtime are 0.
1567 */
1568 return (0);
1569
1570 }
1571
1572 /*
1573 * Scale the scheduling priority according to the "interactivity" of this
1574 * process.
1575 */
1576 static void
1577 sched_priority(struct thread *td)
1578 {
1579 u_int pri, score;
1580
1581 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1582 return;
1583 /*
1584 * If the score is interactive we place the thread in the realtime
1585 * queue with a priority that is less than kernel and interrupt
1586 * priorities. These threads are not subject to nice restrictions.
1587 *
1588 * Scores greater than this are placed on the normal timeshare queue
1589 * where the priority is partially decided by the most recent cpu
1590 * utilization and the rest is decided by nice value.
1591 *
1592 * The nice value of the process has a linear effect on the calculated
1593 * score. Negative nice values make it easier for a thread to be
1594 * considered interactive.
1595 */
1596 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1597 if (score < sched_interact) {
1598 pri = PRI_MIN_INTERACT;
1599 pri += (PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) * score /
1600 sched_interact;
1601 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1602 ("sched_priority: invalid interactive priority %u score %u",
1603 pri, score));
1604 } else {
1605 pri = SCHED_PRI_MIN;
1606 if (td_get_sched(td)->ts_ticks)
1607 pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1608 SCHED_PRI_RANGE - 1);
1609 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1610 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1611 ("sched_priority: invalid priority %u: nice %d, "
1612 "ticks %d ftick %d ltick %d tick pri %d",
1613 pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1614 td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1615 SCHED_PRI_TICKS(td_get_sched(td))));
1616 }
1617 sched_user_prio(td, pri);
1618
1619 return;
1620 }
1621
1622 /*
1623 * This routine enforces a maximum limit on the amount of scheduling history
1624 * kept. It is called after either the slptime or runtime is adjusted. This
1625 * function is ugly due to integer math.
1626 */
1627 static void
1628 sched_interact_update(struct thread *td)
1629 {
1630 struct td_sched *ts;
1631 u_int sum;
1632
1633 ts = td_get_sched(td);
1634 sum = ts->ts_runtime + ts->ts_slptime;
1635 if (sum < SCHED_SLP_RUN_MAX)
1636 return;
1637 /*
1638 * This only happens from two places:
1639 * 1) We have added an unusual amount of run time from fork_exit.
1640 * 2) We have added an unusual amount of sleep time from sched_sleep().
1641 */
1642 if (sum > SCHED_SLP_RUN_MAX * 2) {
1643 if (ts->ts_runtime > ts->ts_slptime) {
1644 ts->ts_runtime = SCHED_SLP_RUN_MAX;
1645 ts->ts_slptime = 1;
1646 } else {
1647 ts->ts_slptime = SCHED_SLP_RUN_MAX;
1648 ts->ts_runtime = 1;
1649 }
1650 return;
1651 }
1652 /*
1653 * If we have exceeded by more than 1/5th then the algorithm below
1654 * will not bring us back into range. Dividing by two here forces
1655 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1656 */
1657 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1658 ts->ts_runtime /= 2;
1659 ts->ts_slptime /= 2;
1660 return;
1661 }
1662 ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1663 ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1664 }
1665
1666 /*
1667 * Scale back the interactivity history when a child thread is created. The
1668 * history is inherited from the parent but the thread may behave totally
1669 * differently. For example, a shell spawning a compiler process. We want
1670 * to learn that the compiler is behaving badly very quickly.
1671 */
1672 static void
1673 sched_interact_fork(struct thread *td)
1674 {
1675 struct td_sched *ts;
1676 int ratio;
1677 int sum;
1678
1679 ts = td_get_sched(td);
1680 sum = ts->ts_runtime + ts->ts_slptime;
1681 if (sum > SCHED_SLP_RUN_FORK) {
1682 ratio = sum / SCHED_SLP_RUN_FORK;
1683 ts->ts_runtime /= ratio;
1684 ts->ts_slptime /= ratio;
1685 }
1686 }
1687
1688 /*
1689 * Called from proc0_init() to setup the scheduler fields.
1690 */
1691 void
1692 schedinit(void)
1693 {
1694 struct td_sched *ts0;
1695
1696 /*
1697 * Set up the scheduler specific parts of thread0.
1698 */
1699 ts0 = td_get_sched(&thread0);
1700 ts0->ts_ltick = ticks;
1701 ts0->ts_ftick = ticks;
1702 ts0->ts_slice = 0;
1703 ts0->ts_cpu = curcpu; /* set valid CPU number */
1704 }
1705
1706 /*
1707 * This is only somewhat accurate since given many processes of the same
1708 * priority they will switch when their slices run out, which will be
1709 * at most sched_slice stathz ticks.
1710 */
1711 int
1712 sched_rr_interval(void)
1713 {
1714
1715 /* Convert sched_slice from stathz to hz. */
1716 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1717 }
1718
1719 /*
1720 * Update the percent cpu tracking information when it is requested or
1721 * the total history exceeds the maximum. We keep a sliding history of
1722 * tick counts that slowly decays. This is less precise than the 4BSD
1723 * mechanism since it happens with less regular and frequent events.
1724 */
1725 static void
1726 sched_pctcpu_update(struct td_sched *ts, int run)
1727 {
1728 int t = ticks;
1729
1730 /*
1731 * The signed difference may be negative if the thread hasn't run for
1732 * over half of the ticks rollover period.
1733 */
1734 if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1735 ts->ts_ticks = 0;
1736 ts->ts_ftick = t - SCHED_TICK_TARG;
1737 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1738 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1739 (ts->ts_ltick - (t - SCHED_TICK_TARG));
1740 ts->ts_ftick = t - SCHED_TICK_TARG;
1741 }
1742 if (run)
1743 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1744 ts->ts_ltick = t;
1745 }
1746
1747 /*
1748 * Adjust the priority of a thread. Move it to the appropriate run-queue
1749 * if necessary. This is the back-end for several priority related
1750 * functions.
1751 */
1752 static void
1753 sched_thread_priority(struct thread *td, u_char prio)
1754 {
1755 struct td_sched *ts;
1756 struct tdq *tdq;
1757 int oldpri;
1758
1759 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1760 "prio:%d", td->td_priority, "new prio:%d", prio,
1761 KTR_ATTR_LINKED, sched_tdname(curthread));
1762 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1763 if (td != curthread && prio < td->td_priority) {
1764 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1765 "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1766 prio, KTR_ATTR_LINKED, sched_tdname(td));
1767 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1768 curthread);
1769 }
1770 ts = td_get_sched(td);
1771 THREAD_LOCK_ASSERT(td, MA_OWNED);
1772 if (td->td_priority == prio)
1773 return;
1774 /*
1775 * If the priority has been elevated due to priority
1776 * propagation, we may have to move ourselves to a new
1777 * queue. This could be optimized to not re-add in some
1778 * cases.
1779 */
1780 if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1781 sched_rem(td);
1782 td->td_priority = prio;
1783 sched_add(td, SRQ_BORROWING);
1784 return;
1785 }
1786 /*
1787 * If the thread is currently running we may have to adjust the lowpri
1788 * information so other cpus are aware of our current priority.
1789 */
1790 if (TD_IS_RUNNING(td)) {
1791 tdq = TDQ_CPU(ts->ts_cpu);
1792 oldpri = td->td_priority;
1793 td->td_priority = prio;
1794 if (prio < tdq->tdq_lowpri)
1795 tdq->tdq_lowpri = prio;
1796 else if (tdq->tdq_lowpri == oldpri)
1797 tdq_setlowpri(tdq, td);
1798 return;
1799 }
1800 td->td_priority = prio;
1801 }
1802
1803 /*
1804 * Update a thread's priority when it is lent another thread's
1805 * priority.
1806 */
1807 void
1808 sched_lend_prio(struct thread *td, u_char prio)
1809 {
1810
1811 td->td_flags |= TDF_BORROWING;
1812 sched_thread_priority(td, prio);
1813 }
1814
1815 /*
1816 * Restore a thread's priority when priority propagation is
1817 * over. The prio argument is the minimum priority the thread
1818 * needs to have to satisfy other possible priority lending
1819 * requests. If the thread's regular priority is less
1820 * important than prio, the thread will keep a priority boost
1821 * of prio.
1822 */
1823 void
1824 sched_unlend_prio(struct thread *td, u_char prio)
1825 {
1826 u_char base_pri;
1827
1828 if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1829 td->td_base_pri <= PRI_MAX_TIMESHARE)
1830 base_pri = td->td_user_pri;
1831 else
1832 base_pri = td->td_base_pri;
1833 if (prio >= base_pri) {
1834 td->td_flags &= ~TDF_BORROWING;
1835 sched_thread_priority(td, base_pri);
1836 } else
1837 sched_lend_prio(td, prio);
1838 }
1839
1840 /*
1841 * Standard entry for setting the priority to an absolute value.
1842 */
1843 void
1844 sched_prio(struct thread *td, u_char prio)
1845 {
1846 u_char oldprio;
1847
1848 /* First, update the base priority. */
1849 td->td_base_pri = prio;
1850
1851 /*
1852 * If the thread is borrowing another thread's priority, don't
1853 * ever lower the priority.
1854 */
1855 if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1856 return;
1857
1858 /* Change the real priority. */
1859 oldprio = td->td_priority;
1860 sched_thread_priority(td, prio);
1861
1862 /*
1863 * If the thread is on a turnstile, then let the turnstile update
1864 * its state.
1865 */
1866 if (TD_ON_LOCK(td) && oldprio != prio)
1867 turnstile_adjust(td, oldprio);
1868 }
1869
1870 /*
1871 * Set the base user priority, does not effect current running priority.
1872 */
1873 void
1874 sched_user_prio(struct thread *td, u_char prio)
1875 {
1876
1877 td->td_base_user_pri = prio;
1878 if (td->td_lend_user_pri <= prio)
1879 return;
1880 td->td_user_pri = prio;
1881 }
1882
1883 void
1884 sched_lend_user_prio(struct thread *td, u_char prio)
1885 {
1886
1887 THREAD_LOCK_ASSERT(td, MA_OWNED);
1888 td->td_lend_user_pri = prio;
1889 td->td_user_pri = min(prio, td->td_base_user_pri);
1890 if (td->td_priority > td->td_user_pri)
1891 sched_prio(td, td->td_user_pri);
1892 else if (td->td_priority != td->td_user_pri)
1893 td->td_flags |= TDF_NEEDRESCHED;
1894 }
1895
1896 #ifdef SMP
1897 /*
1898 * This tdq is about to idle. Try to steal a thread from another CPU before
1899 * choosing the idle thread.
1900 */
1901 static void
1902 tdq_trysteal(struct tdq *tdq)
1903 {
1904 struct cpu_group *cg;
1905 struct tdq *steal;
1906 cpuset_t mask;
1907 int cpu, i;
1908
1909 if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1910 return;
1911 CPU_FILL(&mask);
1912 CPU_CLR(PCPU_GET(cpuid), &mask);
1913 /* We don't want to be preempted while we're iterating. */
1914 spinlock_enter();
1915 TDQ_UNLOCK(tdq);
1916 for (i = 1, cg = tdq->tdq_cg; ; ) {
1917 cpu = sched_highest(cg, mask, steal_thresh);
1918 /*
1919 * If a thread was added while interrupts were disabled don't
1920 * steal one here.
1921 */
1922 if (tdq->tdq_load > 0) {
1923 TDQ_LOCK(tdq);
1924 break;
1925 }
1926 if (cpu == -1) {
1927 i++;
1928 cg = cg->cg_parent;
1929 if (cg == NULL || i > trysteal_limit) {
1930 TDQ_LOCK(tdq);
1931 break;
1932 }
1933 continue;
1934 }
1935 steal = TDQ_CPU(cpu);
1936 /*
1937 * The data returned by sched_highest() is stale and
1938 * the chosen CPU no longer has an eligible thread.
1939 */
1940 if (steal->tdq_load < steal_thresh ||
1941 steal->tdq_transferable == 0)
1942 continue;
1943 tdq_lock_pair(tdq, steal);
1944 /*
1945 * If we get to this point, unconditonally exit the loop
1946 * to bound the time spent in the critcal section.
1947 *
1948 * If a thread was added while interrupts were disabled don't
1949 * steal one here.
1950 */
1951 if (tdq->tdq_load > 0) {
1952 TDQ_UNLOCK(steal);
1953 break;
1954 }
1955 /*
1956 * The data returned by sched_highest() is stale and
1957 * the chosen CPU no longer has an eligible thread.
1958 */
1959 if (steal->tdq_load < steal_thresh ||
1960 steal->tdq_transferable == 0) {
1961 TDQ_UNLOCK(steal);
1962 break;
1963 }
1964 /*
1965 * If we fail to acquire one due to affinity restrictions,
1966 * bail out and let the idle thread to a more complete search
1967 * outside of a critical section.
1968 */
1969 if (tdq_move(steal, tdq) == NULL) {
1970 TDQ_UNLOCK(steal);
1971 break;
1972 }
1973 TDQ_UNLOCK(steal);
1974 break;
1975 }
1976 spinlock_exit();
1977 }
1978 #endif
1979
1980 /*
1981 * Handle migration from sched_switch(). This happens only for
1982 * cpu binding.
1983 */
1984 static struct mtx *
1985 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1986 {
1987 struct tdq *tdn;
1988
1989 KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
1990 "thread %s queued on absent CPU %d.", td->td_name,
1991 td_get_sched(td)->ts_cpu));
1992 tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1993 #ifdef SMP
1994 tdq_load_rem(tdq, td);
1995 /*
1996 * Do the lock dance required to avoid LOR. We grab an extra
1997 * spinlock nesting to prevent preemption while we're
1998 * not holding either run-queue lock.
1999 */
2000 spinlock_enter();
2001 thread_lock_block(td); /* This releases the lock on tdq. */
2002
2003 /*
2004 * Acquire both run-queue locks before placing the thread on the new
2005 * run-queue to avoid deadlocks created by placing a thread with a
2006 * blocked lock on the run-queue of a remote processor. The deadlock
2007 * occurs when a third processor attempts to lock the two queues in
2008 * question while the target processor is spinning with its own
2009 * run-queue lock held while waiting for the blocked lock to clear.
2010 */
2011 tdq_lock_pair(tdn, tdq);
2012 tdq_add(tdn, td, flags);
2013 tdq_notify(tdn, td);
2014 TDQ_UNLOCK(tdn);
2015 spinlock_exit();
2016 #endif
2017 return (TDQ_LOCKPTR(tdn));
2018 }
2019
2020 /*
2021 * Variadic version of thread_lock_unblock() that does not assume td_lock
2022 * is blocked.
2023 */
2024 static inline void
2025 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2026 {
2027 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2028 (uintptr_t)mtx);
2029 }
2030
2031 /*
2032 * Switch threads. This function has to handle threads coming in while
2033 * blocked for some reason, running, or idle. It also must deal with
2034 * migrating a thread from one queue to another as running threads may
2035 * be assigned elsewhere via binding.
2036 */
2037 void
2038 sched_switch(struct thread *td, struct thread *newtd, int flags)
2039 {
2040 struct tdq *tdq;
2041 struct td_sched *ts;
2042 struct mtx *mtx;
2043 int srqflag;
2044 int cpuid, preempted;
2045
2046 THREAD_LOCK_ASSERT(td, MA_OWNED);
2047 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2048
2049 cpuid = PCPU_GET(cpuid);
2050 tdq = TDQ_SELF();
2051 ts = td_get_sched(td);
2052 mtx = td->td_lock;
2053 sched_pctcpu_update(ts, 1);
2054 ts->ts_rltick = ticks;
2055 td->td_lastcpu = td->td_oncpu;
2056 td->td_oncpu = NOCPU;
2057 preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2058 (flags & SW_PREEMPT) != 0;
2059 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2060 td->td_owepreempt = 0;
2061 if (!TD_IS_IDLETHREAD(td))
2062 tdq->tdq_switchcnt++;
2063 /*
2064 * The lock pointer in an idle thread should never change. Reset it
2065 * to CAN_RUN as well.
2066 */
2067 if (TD_IS_IDLETHREAD(td)) {
2068 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2069 TD_SET_CAN_RUN(td);
2070 } else if (TD_IS_RUNNING(td)) {
2071 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2072 srqflag = preempted ?
2073 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2074 SRQ_OURSELF|SRQ_YIELDING;
2075 #ifdef SMP
2076 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2077 ts->ts_cpu = sched_pickcpu(td, 0);
2078 #endif
2079 if (ts->ts_cpu == cpuid)
2080 tdq_runq_add(tdq, td, srqflag);
2081 else {
2082 KASSERT(THREAD_CAN_MIGRATE(td) ||
2083 (ts->ts_flags & TSF_BOUND) != 0,
2084 ("Thread %p shouldn't migrate", td));
2085 mtx = sched_switch_migrate(tdq, td, srqflag);
2086 }
2087 } else {
2088 /* This thread must be going to sleep. */
2089 TDQ_LOCK(tdq);
2090 mtx = thread_lock_block(td);
2091 tdq_load_rem(tdq, td);
2092 #ifdef SMP
2093 if (tdq->tdq_load == 0)
2094 tdq_trysteal(tdq);
2095 #endif
2096 }
2097
2098 #if (KTR_COMPILE & KTR_SCHED) != 0
2099 if (TD_IS_IDLETHREAD(td))
2100 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2101 "prio:%d", td->td_priority);
2102 else
2103 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2104 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2105 "lockname:\"%s\"", td->td_lockname);
2106 #endif
2107
2108 /*
2109 * We enter here with the thread blocked and assigned to the
2110 * appropriate cpu run-queue or sleep-queue and with the current
2111 * thread-queue locked.
2112 */
2113 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2114 newtd = choosethread();
2115 /*
2116 * Call the MD code to switch contexts if necessary.
2117 */
2118 if (td != newtd) {
2119 #ifdef HWPMC_HOOKS
2120 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2121 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2122 #endif
2123 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2124 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2125 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2126 sched_pctcpu_update(td_get_sched(newtd), 0);
2127
2128 #ifdef KDTRACE_HOOKS
2129 /*
2130 * If DTrace has set the active vtime enum to anything
2131 * other than INACTIVE (0), then it should have set the
2132 * function to call.
2133 */
2134 if (dtrace_vtime_active)
2135 (*dtrace_vtime_switch_func)(newtd);
2136 #endif
2137
2138 cpu_switch(td, newtd, mtx);
2139 /*
2140 * We may return from cpu_switch on a different cpu. However,
2141 * we always return with td_lock pointing to the current cpu's
2142 * run queue lock.
2143 */
2144 cpuid = PCPU_GET(cpuid);
2145 tdq = TDQ_SELF();
2146 lock_profile_obtain_lock_success(
2147 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2148
2149 SDT_PROBE0(sched, , , on__cpu);
2150 #ifdef HWPMC_HOOKS
2151 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2152 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2153 #endif
2154 } else {
2155 thread_unblock_switch(td, mtx);
2156 SDT_PROBE0(sched, , , remain__cpu);
2157 }
2158
2159 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2160 "prio:%d", td->td_priority);
2161
2162 /*
2163 * Assert that all went well and return.
2164 */
2165 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2166 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2167 td->td_oncpu = cpuid;
2168 }
2169
2170 /*
2171 * Adjust thread priorities as a result of a nice request.
2172 */
2173 void
2174 sched_nice(struct proc *p, int nice)
2175 {
2176 struct thread *td;
2177
2178 PROC_LOCK_ASSERT(p, MA_OWNED);
2179
2180 p->p_nice = nice;
2181 FOREACH_THREAD_IN_PROC(p, td) {
2182 thread_lock(td);
2183 sched_priority(td);
2184 sched_prio(td, td->td_base_user_pri);
2185 thread_unlock(td);
2186 }
2187 }
2188
2189 /*
2190 * Record the sleep time for the interactivity scorer.
2191 */
2192 void
2193 sched_sleep(struct thread *td, int prio)
2194 {
2195
2196 THREAD_LOCK_ASSERT(td, MA_OWNED);
2197
2198 td->td_slptick = ticks;
2199 if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2200 td->td_flags |= TDF_CANSWAP;
2201 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2202 return;
2203 if (static_boost == 1 && prio)
2204 sched_prio(td, prio);
2205 else if (static_boost && td->td_priority > static_boost)
2206 sched_prio(td, static_boost);
2207 }
2208
2209 /*
2210 * Schedule a thread to resume execution and record how long it voluntarily
2211 * slept. We also update the pctcpu, interactivity, and priority.
2212 */
2213 void
2214 sched_wakeup(struct thread *td)
2215 {
2216 struct td_sched *ts;
2217 int slptick;
2218
2219 THREAD_LOCK_ASSERT(td, MA_OWNED);
2220 ts = td_get_sched(td);
2221 td->td_flags &= ~TDF_CANSWAP;
2222 /*
2223 * If we slept for more than a tick update our interactivity and
2224 * priority.
2225 */
2226 slptick = td->td_slptick;
2227 td->td_slptick = 0;
2228 if (slptick && slptick != ticks) {
2229 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2230 sched_interact_update(td);
2231 sched_pctcpu_update(ts, 0);
2232 }
2233 /*
2234 * Reset the slice value since we slept and advanced the round-robin.
2235 */
2236 ts->ts_slice = 0;
2237 sched_add(td, SRQ_BORING);
2238 }
2239
2240 /*
2241 * Penalize the parent for creating a new child and initialize the child's
2242 * priority.
2243 */
2244 void
2245 sched_fork(struct thread *td, struct thread *child)
2246 {
2247 THREAD_LOCK_ASSERT(td, MA_OWNED);
2248 sched_pctcpu_update(td_get_sched(td), 1);
2249 sched_fork_thread(td, child);
2250 /*
2251 * Penalize the parent and child for forking.
2252 */
2253 sched_interact_fork(child);
2254 sched_priority(child);
2255 td_get_sched(td)->ts_runtime += tickincr;
2256 sched_interact_update(td);
2257 sched_priority(td);
2258 }
2259
2260 /*
2261 * Fork a new thread, may be within the same process.
2262 */
2263 void
2264 sched_fork_thread(struct thread *td, struct thread *child)
2265 {
2266 struct td_sched *ts;
2267 struct td_sched *ts2;
2268 struct tdq *tdq;
2269
2270 tdq = TDQ_SELF();
2271 THREAD_LOCK_ASSERT(td, MA_OWNED);
2272 /*
2273 * Initialize child.
2274 */
2275 ts = td_get_sched(td);
2276 ts2 = td_get_sched(child);
2277 child->td_oncpu = NOCPU;
2278 child->td_lastcpu = NOCPU;
2279 child->td_lock = TDQ_LOCKPTR(tdq);
2280 child->td_cpuset = cpuset_ref(td->td_cpuset);
2281 child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2282 ts2->ts_cpu = ts->ts_cpu;
2283 ts2->ts_flags = 0;
2284 /*
2285 * Grab our parents cpu estimation information.
2286 */
2287 ts2->ts_ticks = ts->ts_ticks;
2288 ts2->ts_ltick = ts->ts_ltick;
2289 ts2->ts_ftick = ts->ts_ftick;
2290 /*
2291 * Do not inherit any borrowed priority from the parent.
2292 */
2293 child->td_priority = child->td_base_pri;
2294 /*
2295 * And update interactivity score.
2296 */
2297 ts2->ts_slptime = ts->ts_slptime;
2298 ts2->ts_runtime = ts->ts_runtime;
2299 /* Attempt to quickly learn interactivity. */
2300 ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2301 #ifdef KTR
2302 bzero(ts2->ts_name, sizeof(ts2->ts_name));
2303 #endif
2304 }
2305
2306 /*
2307 * Adjust the priority class of a thread.
2308 */
2309 void
2310 sched_class(struct thread *td, int class)
2311 {
2312
2313 THREAD_LOCK_ASSERT(td, MA_OWNED);
2314 if (td->td_pri_class == class)
2315 return;
2316 td->td_pri_class = class;
2317 }
2318
2319 /*
2320 * Return some of the child's priority and interactivity to the parent.
2321 */
2322 void
2323 sched_exit(struct proc *p, struct thread *child)
2324 {
2325 struct thread *td;
2326
2327 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2328 "prio:%d", child->td_priority);
2329 PROC_LOCK_ASSERT(p, MA_OWNED);
2330 td = FIRST_THREAD_IN_PROC(p);
2331 sched_exit_thread(td, child);
2332 }
2333
2334 /*
2335 * Penalize another thread for the time spent on this one. This helps to
2336 * worsen the priority and interactivity of processes which schedule batch
2337 * jobs such as make. This has little effect on the make process itself but
2338 * causes new processes spawned by it to receive worse scores immediately.
2339 */
2340 void
2341 sched_exit_thread(struct thread *td, struct thread *child)
2342 {
2343
2344 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2345 "prio:%d", child->td_priority);
2346 /*
2347 * Give the child's runtime to the parent without returning the
2348 * sleep time as a penalty to the parent. This causes shells that
2349 * launch expensive things to mark their children as expensive.
2350 */
2351 thread_lock(td);
2352 td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2353 sched_interact_update(td);
2354 sched_priority(td);
2355 thread_unlock(td);
2356 }
2357
2358 void
2359 sched_preempt(struct thread *td)
2360 {
2361 struct tdq *tdq;
2362
2363 SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2364
2365 thread_lock(td);
2366 tdq = TDQ_SELF();
2367 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2368 tdq->tdq_ipipending = 0;
2369 if (td->td_priority > tdq->tdq_lowpri) {
2370 int flags;
2371
2372 flags = SW_INVOL | SW_PREEMPT;
2373 if (td->td_critnest > 1)
2374 td->td_owepreempt = 1;
2375 else if (TD_IS_IDLETHREAD(td))
2376 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2377 else
2378 mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2379 }
2380 thread_unlock(td);
2381 }
2382
2383 /*
2384 * Fix priorities on return to user-space. Priorities may be elevated due
2385 * to static priorities in msleep() or similar.
2386 */
2387 void
2388 sched_userret_slowpath(struct thread *td)
2389 {
2390
2391 thread_lock(td);
2392 td->td_priority = td->td_user_pri;
2393 td->td_base_pri = td->td_user_pri;
2394 tdq_setlowpri(TDQ_SELF(), td);
2395 thread_unlock(td);
2396 }
2397
2398 /*
2399 * Handle a stathz tick. This is really only relevant for timeshare
2400 * threads.
2401 */
2402 void
2403 sched_clock(struct thread *td)
2404 {
2405 struct tdq *tdq;
2406 struct td_sched *ts;
2407
2408 THREAD_LOCK_ASSERT(td, MA_OWNED);
2409 tdq = TDQ_SELF();
2410 #ifdef SMP
2411 /*
2412 * We run the long term load balancer infrequently on the first cpu.
2413 */
2414 if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2415 if (balance_ticks && --balance_ticks == 0)
2416 sched_balance();
2417 }
2418 #endif
2419 /*
2420 * Save the old switch count so we have a record of the last ticks
2421 * activity. Initialize the new switch count based on our load.
2422 * If there is some activity seed it to reflect that.
2423 */
2424 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2425 tdq->tdq_switchcnt = tdq->tdq_load;
2426 /*
2427 * Advance the insert index once for each tick to ensure that all
2428 * threads get a chance to run.
2429 */
2430 if (tdq->tdq_idx == tdq->tdq_ridx) {
2431 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2432 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2433 tdq->tdq_ridx = tdq->tdq_idx;
2434 }
2435 ts = td_get_sched(td);
2436 sched_pctcpu_update(ts, 1);
2437 if (td->td_pri_class & PRI_FIFO_BIT)
2438 return;
2439 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2440 /*
2441 * We used a tick; charge it to the thread so
2442 * that we can compute our interactivity.
2443 */
2444 td_get_sched(td)->ts_runtime += tickincr;
2445 sched_interact_update(td);
2446 sched_priority(td);
2447 }
2448
2449 /*
2450 * Force a context switch if the current thread has used up a full
2451 * time slice (default is 100ms).
2452 */
2453 if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2454 ts->ts_slice = 0;
2455 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2456 }
2457 }
2458
2459 u_int
2460 sched_estcpu(struct thread *td __unused)
2461 {
2462
2463 return (0);
2464 }
2465
2466 /*
2467 * Return whether the current CPU has runnable tasks. Used for in-kernel
2468 * cooperative idle threads.
2469 */
2470 int
2471 sched_runnable(void)
2472 {
2473 struct tdq *tdq;
2474 int load;
2475
2476 load = 1;
2477
2478 tdq = TDQ_SELF();
2479 if ((curthread->td_flags & TDF_IDLETD) != 0) {
2480 if (tdq->tdq_load > 0)
2481 goto out;
2482 } else
2483 if (tdq->tdq_load - 1 > 0)
2484 goto out;
2485 load = 0;
2486 out:
2487 return (load);
2488 }
2489
2490 /*
2491 * Choose the highest priority thread to run. The thread is removed from
2492 * the run-queue while running however the load remains. For SMP we set
2493 * the tdq in the global idle bitmask if it idles here.
2494 */
2495 struct thread *
2496 sched_choose(void)
2497 {
2498 struct thread *td;
2499 struct tdq *tdq;
2500
2501 tdq = TDQ_SELF();
2502 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2503 td = tdq_choose(tdq);
2504 if (td) {
2505 tdq_runq_rem(tdq, td);
2506 tdq->tdq_lowpri = td->td_priority;
2507 return (td);
2508 }
2509 tdq->tdq_lowpri = PRI_MAX_IDLE;
2510 return (PCPU_GET(idlethread));
2511 }
2512
2513 /*
2514 * Set owepreempt if necessary. Preemption never happens directly in ULE,
2515 * we always request it once we exit a critical section.
2516 */
2517 static inline void
2518 sched_setpreempt(struct thread *td)
2519 {
2520 struct thread *ctd;
2521 int cpri;
2522 int pri;
2523
2524 THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2525
2526 ctd = curthread;
2527 pri = td->td_priority;
2528 cpri = ctd->td_priority;
2529 if (pri < cpri)
2530 ctd->td_flags |= TDF_NEEDRESCHED;
2531 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2532 return;
2533 if (!sched_shouldpreempt(pri, cpri, 0))
2534 return;
2535 ctd->td_owepreempt = 1;
2536 }
2537
2538 /*
2539 * Add a thread to a thread queue. Select the appropriate runq and add the
2540 * thread to it. This is the internal function called when the tdq is
2541 * predetermined.
2542 */
2543 void
2544 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2545 {
2546
2547 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2548 KASSERT((td->td_inhibitors == 0),
2549 ("sched_add: trying to run inhibited thread"));
2550 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2551 ("sched_add: bad thread state"));
2552 KASSERT(td->td_flags & TDF_INMEM,
2553 ("sched_add: thread swapped out"));
2554
2555 if (td->td_priority < tdq->tdq_lowpri)
2556 tdq->tdq_lowpri = td->td_priority;
2557 tdq_runq_add(tdq, td, flags);
2558 tdq_load_add(tdq, td);
2559 }
2560
2561 /*
2562 * Select the target thread queue and add a thread to it. Request
2563 * preemption or IPI a remote processor if required.
2564 */
2565 void
2566 sched_add(struct thread *td, int flags)
2567 {
2568 struct tdq *tdq;
2569 #ifdef SMP
2570 int cpu;
2571 #endif
2572
2573 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2574 "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2575 sched_tdname(curthread));
2576 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2577 KTR_ATTR_LINKED, sched_tdname(td));
2578 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2579 flags & SRQ_PREEMPTED);
2580 THREAD_LOCK_ASSERT(td, MA_OWNED);
2581 /*
2582 * Recalculate the priority before we select the target cpu or
2583 * run-queue.
2584 */
2585 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2586 sched_priority(td);
2587 #ifdef SMP
2588 /*
2589 * Pick the destination cpu and if it isn't ours transfer to the
2590 * target cpu.
2591 */
2592 cpu = sched_pickcpu(td, flags);
2593 tdq = sched_setcpu(td, cpu, flags);
2594 tdq_add(tdq, td, flags);
2595 if (cpu != PCPU_GET(cpuid)) {
2596 tdq_notify(tdq, td);
2597 return;
2598 }
2599 #else
2600 tdq = TDQ_SELF();
2601 TDQ_LOCK(tdq);
2602 /*
2603 * Now that the thread is moving to the run-queue, set the lock
2604 * to the scheduler's lock.
2605 */
2606 thread_lock_set(td, TDQ_LOCKPTR(tdq));
2607 tdq_add(tdq, td, flags);
2608 #endif
2609 if (!(flags & SRQ_YIELDING))
2610 sched_setpreempt(td);
2611 }
2612
2613 /*
2614 * Remove a thread from a run-queue without running it. This is used
2615 * when we're stealing a thread from a remote queue. Otherwise all threads
2616 * exit by calling sched_exit_thread() and sched_throw() themselves.
2617 */
2618 void
2619 sched_rem(struct thread *td)
2620 {
2621 struct tdq *tdq;
2622
2623 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2624 "prio:%d", td->td_priority);
2625 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2626 tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2627 TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2628 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2629 KASSERT(TD_ON_RUNQ(td),
2630 ("sched_rem: thread not on run queue"));
2631 tdq_runq_rem(tdq, td);
2632 tdq_load_rem(tdq, td);
2633 TD_SET_CAN_RUN(td);
2634 if (td->td_priority == tdq->tdq_lowpri)
2635 tdq_setlowpri(tdq, NULL);
2636 }
2637
2638 /*
2639 * Fetch cpu utilization information. Updates on demand.
2640 */
2641 fixpt_t
2642 sched_pctcpu(struct thread *td)
2643 {
2644 fixpt_t pctcpu;
2645 struct td_sched *ts;
2646
2647 pctcpu = 0;
2648 ts = td_get_sched(td);
2649
2650 THREAD_LOCK_ASSERT(td, MA_OWNED);
2651 sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2652 if (ts->ts_ticks) {
2653 int rtick;
2654
2655 /* How many rtick per second ? */
2656 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2657 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2658 }
2659
2660 return (pctcpu);
2661 }
2662
2663 /*
2664 * Enforce affinity settings for a thread. Called after adjustments to
2665 * cpumask.
2666 */
2667 void
2668 sched_affinity(struct thread *td)
2669 {
2670 #ifdef SMP
2671 struct td_sched *ts;
2672
2673 THREAD_LOCK_ASSERT(td, MA_OWNED);
2674 ts = td_get_sched(td);
2675 if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2676 return;
2677 if (TD_ON_RUNQ(td)) {
2678 sched_rem(td);
2679 sched_add(td, SRQ_BORING);
2680 return;
2681 }
2682 if (!TD_IS_RUNNING(td))
2683 return;
2684 /*
2685 * Force a switch before returning to userspace. If the
2686 * target thread is not running locally send an ipi to force
2687 * the issue.
2688 */
2689 td->td_flags |= TDF_NEEDRESCHED;
2690 if (td != curthread)
2691 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2692 #endif
2693 }
2694
2695 /*
2696 * Bind a thread to a target cpu.
2697 */
2698 void
2699 sched_bind(struct thread *td, int cpu)
2700 {
2701 struct td_sched *ts;
2702
2703 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2704 KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2705 ts = td_get_sched(td);
2706 if (ts->ts_flags & TSF_BOUND)
2707 sched_unbind(td);
2708 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2709 ts->ts_flags |= TSF_BOUND;
2710 sched_pin();
2711 if (PCPU_GET(cpuid) == cpu)
2712 return;
2713 ts->ts_cpu = cpu;
2714 /* When we return from mi_switch we'll be on the correct cpu. */
2715 mi_switch(SW_VOL, NULL);
2716 }
2717
2718 /*
2719 * Release a bound thread.
2720 */
2721 void
2722 sched_unbind(struct thread *td)
2723 {
2724 struct td_sched *ts;
2725
2726 THREAD_LOCK_ASSERT(td, MA_OWNED);
2727 KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2728 ts = td_get_sched(td);
2729 if ((ts->ts_flags & TSF_BOUND) == 0)
2730 return;
2731 ts->ts_flags &= ~TSF_BOUND;
2732 sched_unpin();
2733 }
2734
2735 int
2736 sched_is_bound(struct thread *td)
2737 {
2738 THREAD_LOCK_ASSERT(td, MA_OWNED);
2739 return (td_get_sched(td)->ts_flags & TSF_BOUND);
2740 }
2741
2742 /*
2743 * Basic yield call.
2744 */
2745 void
2746 sched_relinquish(struct thread *td)
2747 {
2748 thread_lock(td);
2749 mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2750 thread_unlock(td);
2751 }
2752
2753 /*
2754 * Return the total system load.
2755 */
2756 int
2757 sched_load(void)
2758 {
2759 #ifdef SMP
2760 int total;
2761 int i;
2762
2763 total = 0;
2764 CPU_FOREACH(i)
2765 total += TDQ_CPU(i)->tdq_sysload;
2766 return (total);
2767 #else
2768 return (TDQ_SELF()->tdq_sysload);
2769 #endif
2770 }
2771
2772 int
2773 sched_sizeof_proc(void)
2774 {
2775 return (sizeof(struct proc));
2776 }
2777
2778 int
2779 sched_sizeof_thread(void)
2780 {
2781 return (sizeof(struct thread) + sizeof(struct td_sched));
2782 }
2783
2784 #ifdef SMP
2785 #define TDQ_IDLESPIN(tdq) \
2786 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2787 #else
2788 #define TDQ_IDLESPIN(tdq) 1
2789 #endif
2790
2791 /*
2792 * The actual idle process.
2793 */
2794 void
2795 sched_idletd(void *dummy)
2796 {
2797 struct thread *td;
2798 struct tdq *tdq;
2799 int oldswitchcnt, switchcnt;
2800 int i;
2801
2802 mtx_assert(&Giant, MA_NOTOWNED);
2803 td = curthread;
2804 tdq = TDQ_SELF();
2805 THREAD_NO_SLEEPING();
2806 oldswitchcnt = -1;
2807 for (;;) {
2808 if (tdq->tdq_load) {
2809 thread_lock(td);
2810 mi_switch(SW_VOL | SWT_IDLE, NULL);
2811 thread_unlock(td);
2812 }
2813 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2814 #ifdef SMP
2815 if (always_steal || switchcnt != oldswitchcnt) {
2816 oldswitchcnt = switchcnt;
2817 if (tdq_idled(tdq) == 0)
2818 continue;
2819 }
2820 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2821 #else
2822 oldswitchcnt = switchcnt;
2823 #endif
2824 /*
2825 * If we're switching very frequently, spin while checking
2826 * for load rather than entering a low power state that
2827 * may require an IPI. However, don't do any busy
2828 * loops while on SMT machines as this simply steals
2829 * cycles from cores doing useful work.
2830 */
2831 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2832 for (i = 0; i < sched_idlespins; i++) {
2833 if (tdq->tdq_load)
2834 break;
2835 cpu_spinwait();
2836 }
2837 }
2838
2839 /* If there was context switch during spin, restart it. */
2840 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2841 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2842 continue;
2843
2844 /* Run main MD idle handler. */
2845 tdq->tdq_cpu_idle = 1;
2846 /*
2847 * Make sure that tdq_cpu_idle update is globally visible
2848 * before cpu_idle() read tdq_load. The order is important
2849 * to avoid race with tdq_notify.
2850 */
2851 atomic_thread_fence_seq_cst();
2852 /*
2853 * Checking for again after the fence picks up assigned
2854 * threads often enough to make it worthwhile to do so in
2855 * order to avoid calling cpu_idle().
2856 */
2857 if (tdq->tdq_load != 0) {
2858 tdq->tdq_cpu_idle = 0;
2859 continue;
2860 }
2861 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2862 tdq->tdq_cpu_idle = 0;
2863
2864 /*
2865 * Account thread-less hardware interrupts and
2866 * other wakeup reasons equal to context switches.
2867 */
2868 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2869 if (switchcnt != oldswitchcnt)
2870 continue;
2871 tdq->tdq_switchcnt++;
2872 oldswitchcnt++;
2873 }
2874 }
2875
2876 /*
2877 * A CPU is entering for the first time or a thread is exiting.
2878 */
2879 void
2880 sched_throw(struct thread *td)
2881 {
2882 struct thread *newtd;
2883 struct tdq *tdq;
2884
2885 if (td == NULL) {
2886 /* Correct spinlock nesting and acquire the correct lock. */
2887 tdq = TDQ_SELF();
2888 TDQ_LOCK(tdq);
2889 spinlock_exit();
2890 PCPU_SET(switchtime, cpu_ticks());
2891 PCPU_SET(switchticks, ticks);
2892 } else {
2893 tdq = TDQ_SELF();
2894 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2895 tdq_load_rem(tdq, td);
2896 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2897 td->td_lastcpu = td->td_oncpu;
2898 td->td_oncpu = NOCPU;
2899 }
2900 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2901 newtd = choosethread();
2902 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2903 cpu_throw(td, newtd); /* doesn't return */
2904 }
2905
2906 /*
2907 * This is called from fork_exit(). Just acquire the correct locks and
2908 * let fork do the rest of the work.
2909 */
2910 void
2911 sched_fork_exit(struct thread *td)
2912 {
2913 struct tdq *tdq;
2914 int cpuid;
2915
2916 /*
2917 * Finish setting up thread glue so that it begins execution in a
2918 * non-nested critical section with the scheduler lock held.
2919 */
2920 cpuid = PCPU_GET(cpuid);
2921 tdq = TDQ_SELF();
2922 if (TD_IS_IDLETHREAD(td))
2923 td->td_lock = TDQ_LOCKPTR(tdq);
2924 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2925 td->td_oncpu = cpuid;
2926 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2927 lock_profile_obtain_lock_success(
2928 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2929
2930 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2931 "prio:%d", td->td_priority);
2932 SDT_PROBE0(sched, , , on__cpu);
2933 }
2934
2935 /*
2936 * Create on first use to catch odd startup conditons.
2937 */
2938 char *
2939 sched_tdname(struct thread *td)
2940 {
2941 #ifdef KTR
2942 struct td_sched *ts;
2943
2944 ts = td_get_sched(td);
2945 if (ts->ts_name[0] == '\0')
2946 snprintf(ts->ts_name, sizeof(ts->ts_name),
2947 "%s tid %d", td->td_name, td->td_tid);
2948 return (ts->ts_name);
2949 #else
2950 return (td->td_name);
2951 #endif
2952 }
2953
2954 #ifdef KTR
2955 void
2956 sched_clear_tdname(struct thread *td)
2957 {
2958 struct td_sched *ts;
2959
2960 ts = td_get_sched(td);
2961 ts->ts_name[0] = '\0';
2962 }
2963 #endif
2964
2965 #ifdef SMP
2966
2967 /*
2968 * Build the CPU topology dump string. Is recursively called to collect
2969 * the topology tree.
2970 */
2971 static int
2972 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2973 int indent)
2974 {
2975 char cpusetbuf[CPUSETBUFSIZ];
2976 int i, first;
2977
2978 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2979 "", 1 + indent / 2, cg->cg_level);
2980 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2981 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2982 first = TRUE;
2983 for (i = 0; i < MAXCPU; i++) {
2984 if (CPU_ISSET(i, &cg->cg_mask)) {
2985 if (!first)
2986 sbuf_printf(sb, ", ");
2987 else
2988 first = FALSE;
2989 sbuf_printf(sb, "%d", i);
2990 }
2991 }
2992 sbuf_printf(sb, "</cpu>\n");
2993
2994 if (cg->cg_flags != 0) {
2995 sbuf_printf(sb, "%*s <flags>", indent, "");
2996 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2997 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2998 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2999 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3000 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3001 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3002 sbuf_printf(sb, "</flags>\n");
3003 }
3004
3005 if (cg->cg_children > 0) {
3006 sbuf_printf(sb, "%*s <children>\n", indent, "");
3007 for (i = 0; i < cg->cg_children; i++)
3008 sysctl_kern_sched_topology_spec_internal(sb,
3009 &cg->cg_child[i], indent+2);
3010 sbuf_printf(sb, "%*s </children>\n", indent, "");
3011 }
3012 sbuf_printf(sb, "%*s</group>\n", indent, "");
3013 return (0);
3014 }
3015
3016 /*
3017 * Sysctl handler for retrieving topology dump. It's a wrapper for
3018 * the recursive sysctl_kern_smp_topology_spec_internal().
3019 */
3020 static int
3021 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3022 {
3023 struct sbuf *topo;
3024 int err;
3025
3026 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3027
3028 topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3029 if (topo == NULL)
3030 return (ENOMEM);
3031
3032 sbuf_printf(topo, "<groups>\n");
3033 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3034 sbuf_printf(topo, "</groups>\n");
3035
3036 if (err == 0) {
3037 err = sbuf_finish(topo);
3038 }
3039 sbuf_delete(topo);
3040 return (err);
3041 }
3042
3043 #endif
3044
3045 static int
3046 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3047 {
3048 int error, new_val, period;
3049
3050 period = 1000000 / realstathz;
3051 new_val = period * sched_slice;
3052 error = sysctl_handle_int(oidp, &new_val, 0, req);
3053 if (error != 0 || req->newptr == NULL)
3054 return (error);
3055 if (new_val <= 0)
3056 return (EINVAL);
3057 sched_slice = imax(1, (new_val + period / 2) / period);
3058 sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3059 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3060 realstathz);
3061 return (0);
3062 }
3063
3064 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3065 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3066 "Scheduler name");
3067 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3068 NULL, 0, sysctl_kern_quantum, "I",
3069 "Quantum for timeshare threads in microseconds");
3070 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3071 "Quantum for timeshare threads in stathz ticks");
3072 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3073 "Interactivity score threshold");
3074 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3075 &preempt_thresh, 0,
3076 "Maximal (lowest) priority for preemption");
3077 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3078 "Assign static kernel priorities to sleeping threads");
3079 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3080 "Number of times idle thread will spin waiting for new work");
3081 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3082 &sched_idlespinthresh, 0,
3083 "Threshold before we will permit idle thread spinning");
3084 #ifdef SMP
3085 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3086 "Number of hz ticks to keep thread affinity for");
3087 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3088 "Enables the long-term load balancer");
3089 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3090 &balance_interval, 0,
3091 "Average period in stathz ticks to run the long-term balancer");
3092 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3093 "Attempts to steal work from other cores before idling");
3094 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3095 "Minimum load on remote CPU before we'll steal");
3096 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3097 0, "Topological distance limit for stealing threads in sched_switch()");
3098 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3099 "Always run the stealer from the idle thread");
3100 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3101 CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3102 "XML dump of detected CPU topology");
3103 #endif
3104
3105 /* ps compat. All cpu percentages from ULE are weighted. */
3106 static int ccpu = 0;
3107 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
Cache object: ac7238e8ed9bbab9c72509c903dcfe2b
|