1 /*
2 * Copyright (c) 2000-2019 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28 /*
29 * @OSF_COPYRIGHT@
30 */
31 /*
32 * Mach Operating System
33 * Copyright (c) 1991,1990,1989,1988,1987 Carnegie Mellon University
34 * All Rights Reserved.
35 *
36 * Permission to use, copy, modify and distribute this software and its
37 * documentation is hereby granted, provided that both the copyright
38 * notice and this permission notice appear in all copies of the
39 * software, derivative works or modified versions, and any portions
40 * thereof, and that both notices appear in supporting documentation.
41 *
42 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
43 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR
44 * ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
45 *
46 * Carnegie Mellon requests users of this software to return to
47 *
48 * Software Distribution Coordinator or [email protected]
49 * School of Computer Science
50 * Carnegie Mellon University
51 * Pittsburgh PA 15213-3890
52 *
53 * any improvements or extensions that they make and grant Carnegie Mellon
54 * the rights to redistribute these changes.
55 */
56 /*
57 */
58 /*
59 * File: priority.c
60 * Author: Avadis Tevanian, Jr.
61 * Date: 1986
62 *
63 * Priority related scheduler bits.
64 */
65
66 #include <mach/boolean.h>
67 #include <mach/kern_return.h>
68 #include <mach/machine.h>
69 #include <kern/host.h>
70 #include <kern/mach_param.h>
71 #include <kern/sched.h>
72 #include <sys/kdebug.h>
73 #include <kern/spl.h>
74 #include <kern/thread.h>
75 #include <kern/processor.h>
76 #include <kern/ledger.h>
77 #include <machine/machparam.h>
78 #include <kern/machine.h>
79 #include <kern/policy_internal.h>
80 #include <kern/sched_clutch.h>
81
82 #ifdef CONFIG_MACH_APPROXIMATE_TIME
83 #include <machine/commpage.h> /* for commpage_update_mach_approximate_time */
84 #endif
85
86 #if MONOTONIC
87 #include <kern/monotonic.h>
88 #endif /* MONOTONIC */
89
90 /*
91 * thread_quantum_expire:
92 *
93 * Recalculate the quantum and priority for a thread.
94 *
95 * Called at splsched.
96 */
97
98 void
thread_quantum_expire(timer_call_param_t p0,timer_call_param_t p1)99 thread_quantum_expire(
100 timer_call_param_t p0,
101 timer_call_param_t p1)
102 {
103 processor_t processor = p0;
104 thread_t thread = p1;
105 ast_t preempt;
106 uint64_t ctime;
107
108 assert(processor == current_processor());
109 assert(thread == current_thread());
110
111 KDBG_RELEASE(MACHDBG_CODE(
112 DBG_MACH_SCHED, MACH_SCHED_QUANTUM_EXPIRED) | DBG_FUNC_START);
113
114 SCHED_STATS_INC(quantum_timer_expirations);
115
116 /*
117 * We bill CPU time to both the individual thread and its task.
118 *
119 * Because this balance adjustment could potentially attempt to wake this
120 * very thread, we must credit the ledger before taking the thread lock.
121 * The ledger pointers are only manipulated by the thread itself at the ast
122 * boundary.
123 *
124 * TODO: This fails to account for the time between when the timer was
125 * armed and when it fired. It should be based on the system_timer and
126 * running a timer_update operation here.
127 */
128 ledger_credit(thread->t_ledger, task_ledgers.cpu_time, thread->quantum_remaining);
129 ledger_credit(thread->t_threadledger, thread_ledgers.cpu_time, thread->quantum_remaining);
130 if (thread->t_bankledger) {
131 ledger_credit(thread->t_bankledger, bank_ledgers.cpu_time,
132 (thread->quantum_remaining - thread->t_deduct_bank_ledger_time));
133 }
134 thread->t_deduct_bank_ledger_time = 0;
135
136 struct recount_snap snap = { 0 };
137 recount_snapshot(&snap);
138 ctime = snap.rsn_time_mach;
139 check_monotonic_time(ctime);
140 #ifdef CONFIG_MACH_APPROXIMATE_TIME
141 commpage_update_mach_approximate_time(ctime);
142 #endif /* CONFIG_MACH_APPROXIMATE_TIME */
143
144 sched_update_pset_avg_execution_time(processor->processor_set, thread->quantum_remaining, ctime, thread->th_sched_bucket);
145
146 recount_switch_thread(&snap, thread, get_threadtask(thread));
147 recount_log_switch_thread(&snap);
148
149 thread_lock(thread);
150
151 /*
152 * We've run up until our quantum expiration, and will (potentially)
153 * continue without re-entering the scheduler, so update this now.
154 */
155 processor->last_dispatch = ctime;
156 thread->last_run_time = ctime;
157
158 /*
159 * Check for fail-safe trip.
160 */
161 if ((thread->sched_mode == TH_MODE_REALTIME || thread->sched_mode == TH_MODE_FIXED) &&
162 !(thread->sched_flags & TH_SFLAG_PROMOTED) &&
163 !(thread->kern_promotion_schedpri != 0) &&
164 !(thread->sched_flags & TH_SFLAG_PROMOTE_REASON_MASK) &&
165 !(thread->options & TH_OPT_SYSTEM_CRITICAL)) {
166 uint64_t new_computation;
167
168 new_computation = ctime - thread->computation_epoch;
169 new_computation += thread->computation_metered;
170
171 bool demote = false;
172 switch (thread->sched_mode) {
173 case TH_MODE_REALTIME:
174 if (new_computation > max_unsafe_rt_computation) {
175 thread->safe_release = ctime + sched_safe_rt_duration;
176 demote = true;
177 }
178 break;
179 case TH_MODE_FIXED:
180 if (new_computation > max_unsafe_fixed_computation) {
181 thread->safe_release = ctime + sched_safe_fixed_duration;
182 demote = true;
183 }
184 break;
185 default:
186 panic("unexpected mode: %d", thread->sched_mode);
187 }
188
189 if (demote) {
190 KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_FAILSAFE) | DBG_FUNC_NONE,
191 (uintptr_t)thread->sched_pri, (uintptr_t)thread->sched_mode, 0, 0, 0);
192 sched_thread_mode_demote(thread, TH_SFLAG_FAILSAFE);
193 }
194 }
195
196 /*
197 * Recompute scheduled priority if appropriate.
198 */
199 if (SCHED(can_update_priority)(thread)) {
200 SCHED(update_priority)(thread);
201 } else {
202 SCHED(lightweight_update_priority)(thread);
203 }
204
205 if (thread->sched_mode != TH_MODE_REALTIME) {
206 SCHED(quantum_expire)(thread);
207 }
208
209 /*
210 * This quantum is up, give this thread another.
211 */
212 processor->first_timeslice = FALSE;
213
214 thread_quantum_init(thread);
215
216 timer_update(&thread->runnable_timer, ctime);
217
218 processor->quantum_end = ctime + thread->quantum_remaining;
219
220 /*
221 * Context switch check
222 *
223 * non-urgent flags don't affect kernel threads, so upgrade to urgent
224 * to ensure that rebalancing and non-recommendation kick in quickly.
225 */
226
227 ast_t check_reason = AST_QUANTUM;
228 if (get_threadtask(thread) == kernel_task) {
229 check_reason |= AST_URGENT;
230 }
231
232 if ((preempt = csw_check(thread, processor, check_reason)) != AST_NONE) {
233 ast_on(preempt);
234 }
235
236 /*
237 * AST_KEVENT does not send an IPI when setting the AST,
238 * to avoid waiting for the next context switch to propagate the AST,
239 * the AST is propagated here at quantum expiration.
240 */
241 ast_propagate(thread);
242
243 thread_unlock(thread);
244
245 /* Now that the processor->thread_timer has been updated, evaluate to see if
246 * the workqueue quantum expired and set AST_KEVENT if it has */
247 if (thread_get_tag(thread) & THREAD_TAG_WORKQUEUE) {
248 thread_evaluate_workqueue_quantum_expiry(thread);
249 }
250
251 running_timer_enter(processor, RUNNING_TIMER_QUANTUM, thread,
252 processor->quantum_end, ctime);
253
254 /* Tell platform layer that we are still running this thread */
255 thread_urgency_t urgency = thread_get_urgency(thread, NULL, NULL);
256 machine_thread_going_on_core(thread, urgency, 0, 0, ctime);
257 machine_switch_perfcontrol_state_update(QUANTUM_EXPIRY, ctime,
258 0, thread);
259
260 #if defined(CONFIG_SCHED_TIMESHARE_CORE)
261 sched_timeshare_consider_maintenance(ctime);
262 #endif /* CONFIG_SCHED_TIMESHARE_CORE */
263
264 #if __arm64__
265 if (thread->sched_mode == TH_MODE_REALTIME) {
266 sched_consider_recommended_cores(ctime, thread);
267 }
268 #endif /* __arm64__ */
269
270 KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_QUANTUM_EXPIRED) | DBG_FUNC_END, preempt, 0, 0, 0, 0);
271 }
272
273 /*
274 * sched_set_thread_base_priority:
275 *
276 * Set the base priority of the thread
277 * and reset its scheduled priority.
278 *
279 * This is the only path to change base_pri.
280 *
281 * Called with the thread locked.
282 */
283 void
sched_set_thread_base_priority(thread_t thread,int priority)284 sched_set_thread_base_priority(thread_t thread, int priority)
285 {
286 assert(priority >= MINPRI);
287 uint64_t ctime = 0;
288
289 if (thread->sched_mode == TH_MODE_REALTIME) {
290 assert((priority >= BASEPRI_RTQUEUES) && (priority <= MAXPRI));
291 } else {
292 assert(priority < BASEPRI_RTQUEUES);
293 }
294
295 int old_base_pri = thread->base_pri;
296 thread->req_base_pri = (int16_t)priority;
297 if (thread->sched_flags & TH_SFLAG_BASE_PRI_FROZEN) {
298 priority = MAX(priority, old_base_pri);
299 }
300 thread->base_pri = (int16_t)priority;
301
302 if ((thread->state & TH_RUN) == TH_RUN) {
303 assert(thread->last_made_runnable_time != THREAD_NOT_RUNNABLE);
304 ctime = mach_approximate_time();
305 thread->last_basepri_change_time = ctime;
306 } else {
307 assert(thread->last_basepri_change_time == THREAD_NOT_RUNNABLE);
308 assert(thread->last_made_runnable_time == THREAD_NOT_RUNNABLE);
309 }
310
311 /*
312 * Currently the perfcontrol_attr depends on the base pri of the
313 * thread. Therefore, we use this function as the hook for the
314 * perfcontrol callout.
315 */
316 if (thread == current_thread() && old_base_pri != priority) {
317 if (!ctime) {
318 ctime = mach_approximate_time();
319 }
320 machine_switch_perfcontrol_state_update(PERFCONTROL_ATTR_UPDATE,
321 ctime, PERFCONTROL_CALLOUT_WAKE_UNSAFE, thread);
322 }
323 #if !CONFIG_SCHED_CLUTCH
324 /* For the clutch scheduler, this operation is done in set_sched_pri() */
325 SCHED(update_thread_bucket)(thread);
326 #endif /* !CONFIG_SCHED_CLUTCH */
327
328 thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
329 }
330
331 /*
332 * sched_set_kernel_thread_priority:
333 *
334 * Set the absolute base priority of the thread
335 * and reset its scheduled priority.
336 *
337 * Called with the thread unlocked.
338 */
339 void
sched_set_kernel_thread_priority(thread_t thread,int new_priority)340 sched_set_kernel_thread_priority(thread_t thread, int new_priority)
341 {
342 spl_t s = splsched();
343
344 thread_lock(thread);
345
346 assert(thread->sched_mode != TH_MODE_REALTIME);
347 assert(thread->effective_policy.thep_qos == THREAD_QOS_UNSPECIFIED);
348
349 if (new_priority > thread->max_priority) {
350 new_priority = thread->max_priority;
351 }
352 #if !defined(XNU_TARGET_OS_OSX)
353 if (new_priority < MAXPRI_THROTTLE) {
354 new_priority = MAXPRI_THROTTLE;
355 }
356 #endif /* !defined(XNU_TARGET_OS_OSX) */
357
358 thread->importance = new_priority - thread->task_priority;
359
360 sched_set_thread_base_priority(thread, new_priority);
361
362 thread_unlock(thread);
363 splx(s);
364 }
365
366 /*
367 * thread_recompute_sched_pri:
368 *
369 * Reset the scheduled priority of the thread
370 * according to its base priority if the
371 * thread has not been promoted or depressed.
372 *
373 * This is the only way to push base_pri changes into sched_pri,
374 * or to recalculate the appropriate sched_pri after changing
375 * a promotion or depression.
376 *
377 * Called at splsched with the thread locked.
378 *
379 * TODO: Add an 'update urgency' flag to avoid urgency callouts on every rwlock operation
380 */
381 void
thread_recompute_sched_pri(thread_t thread,set_sched_pri_options_t options)382 thread_recompute_sched_pri(thread_t thread, set_sched_pri_options_t options)
383 {
384 uint32_t sched_flags = thread->sched_flags;
385 sched_mode_t sched_mode = thread->sched_mode;
386
387 int16_t priority = thread->base_pri;
388
389 if (sched_mode == TH_MODE_TIMESHARE) {
390 priority = (int16_t)SCHED(compute_timeshare_priority)(thread);
391 }
392
393 if (sched_flags & TH_SFLAG_DEPRESS) {
394 /* thread_yield_internal overrides kernel mutex promotion */
395 priority = DEPRESSPRI;
396 } else {
397 /* poll-depress is overridden by mutex promotion and promote-reasons */
398 if ((sched_flags & TH_SFLAG_POLLDEPRESS)) {
399 priority = DEPRESSPRI;
400 }
401
402 if (thread->kern_promotion_schedpri > 0) {
403 priority = MAX(priority, thread->kern_promotion_schedpri);
404
405 if (sched_mode != TH_MODE_REALTIME) {
406 priority = MIN(priority, MAXPRI_PROMOTE);
407 }
408 }
409
410 if (sched_flags & TH_SFLAG_PROMOTED) {
411 priority = MAX(priority, thread->promotion_priority);
412
413 if (sched_mode != TH_MODE_REALTIME) {
414 priority = MIN(priority, MAXPRI_PROMOTE);
415 }
416 }
417
418 if (sched_flags & TH_SFLAG_PROMOTE_REASON_MASK) {
419 if (sched_flags & TH_SFLAG_RW_PROMOTED) {
420 priority = MAX(priority, MINPRI_RWLOCK);
421 }
422
423 if (sched_flags & TH_SFLAG_WAITQ_PROMOTED) {
424 priority = MAX(priority, MINPRI_WAITQ);
425 }
426
427 if (sched_flags & TH_SFLAG_EXEC_PROMOTED) {
428 priority = MAX(priority, MINPRI_EXEC);
429 }
430
431 if (sched_flags & TH_SFLAG_FLOOR_PROMOTED) {
432 priority = MAX(priority, MINPRI_FLOOR);
433 }
434 }
435 }
436
437 set_sched_pri(thread, priority, options);
438 }
439
440 void
sched_default_quantum_expire(thread_t thread __unused)441 sched_default_quantum_expire(thread_t thread __unused)
442 {
443 /*
444 * No special behavior when a timeshare, fixed, or realtime thread
445 * uses up its entire quantum
446 */
447 }
448
449 int smt_timeshare_enabled = 1;
450 int smt_sched_bonus_16ths = 8;
451
452 #if defined(CONFIG_SCHED_TIMESHARE_CORE)
453
454 /*
455 * lightweight_update_priority:
456 *
457 * Update the scheduled priority for
458 * a timesharing thread.
459 *
460 * Only for use on the current thread.
461 *
462 * Called with the thread locked.
463 */
464 void
lightweight_update_priority(thread_t thread)465 lightweight_update_priority(thread_t thread)
466 {
467 assert(thread->runq == PROCESSOR_NULL);
468 assert(thread == current_thread());
469
470 if (thread->sched_mode == TH_MODE_TIMESHARE) {
471 int priority;
472 uint32_t delta;
473
474 sched_tick_delta(thread, delta);
475
476 /*
477 * Accumulate timesharing usage only
478 * during contention for processor
479 * resources.
480 */
481 if (thread->pri_shift < INT8_MAX) {
482 if (thread_no_smt(thread) && smt_timeshare_enabled) {
483 thread->sched_usage += (delta + ((delta * smt_sched_bonus_16ths) >> 4));
484 } else {
485 thread->sched_usage += delta;
486 }
487 }
488
489 thread->cpu_delta += delta;
490
491 #if CONFIG_SCHED_CLUTCH
492 /*
493 * Update the CPU usage for the thread group to which the thread belongs.
494 * The implementation assumes that the thread ran for the entire delta
495 * as part of the same thread group.
496 */
497 sched_clutch_cpu_usage_update(thread, delta);
498 #endif /* CONFIG_SCHED_CLUTCH */
499
500 priority = sched_compute_timeshare_priority(thread);
501
502 if (priority != thread->sched_pri) {
503 thread_recompute_sched_pri(thread, SETPRI_LAZY);
504 }
505 }
506 }
507
508 /*
509 * Define shifts for simulating (5/8) ** n
510 *
511 * Shift structures for holding update shifts. Actual computation
512 * is usage = (usage >> shift1) +/- (usage >> abs(shift2)) where the
513 * +/- is determined by the sign of shift 2.
514 */
515
516 const struct shift_data sched_decay_shifts[SCHED_DECAY_TICKS] = {
517 { .shift1 = 1, .shift2 = 1 },
518 { .shift1 = 1, .shift2 = 3 },
519 { .shift1 = 1, .shift2 = -3 },
520 { .shift1 = 2, .shift2 = -7 },
521 { .shift1 = 3, .shift2 = 5 },
522 { .shift1 = 3, .shift2 = -5 },
523 { .shift1 = 4, .shift2 = -8 },
524 { .shift1 = 5, .shift2 = 7 },
525 { .shift1 = 5, .shift2 = -7 },
526 { .shift1 = 6, .shift2 = -10 },
527 { .shift1 = 7, .shift2 = 10 },
528 { .shift1 = 7, .shift2 = -9 },
529 { .shift1 = 8, .shift2 = -11 },
530 { .shift1 = 9, .shift2 = 12 },
531 { .shift1 = 9, .shift2 = -11 },
532 { .shift1 = 10, .shift2 = -13 },
533 { .shift1 = 11, .shift2 = 14 },
534 { .shift1 = 11, .shift2 = -13 },
535 { .shift1 = 12, .shift2 = -15 },
536 { .shift1 = 13, .shift2 = 17 },
537 { .shift1 = 13, .shift2 = -15 },
538 { .shift1 = 14, .shift2 = -17 },
539 { .shift1 = 15, .shift2 = 19 },
540 { .shift1 = 16, .shift2 = 18 },
541 { .shift1 = 16, .shift2 = -19 },
542 { .shift1 = 17, .shift2 = 22 },
543 { .shift1 = 18, .shift2 = 20 },
544 { .shift1 = 18, .shift2 = -20 },
545 { .shift1 = 19, .shift2 = 26 },
546 { .shift1 = 20, .shift2 = 22 },
547 { .shift1 = 20, .shift2 = -22 },
548 { .shift1 = 21, .shift2 = -27 }
549 };
550
551 /*
552 * sched_compute_timeshare_priority:
553 *
554 * Calculate the timesharing priority based upon usage and load.
555 */
556 extern int sched_pri_decay_band_limit;
557
558
559 /* Only use the decay floor logic on non-macOS and non-clutch schedulers */
560 #if !defined(XNU_TARGET_OS_OSX) && !CONFIG_SCHED_CLUTCH
561
562 int
sched_compute_timeshare_priority(thread_t thread)563 sched_compute_timeshare_priority(thread_t thread)
564 {
565 int decay_amount;
566 int decay_limit = sched_pri_decay_band_limit;
567
568 if (thread->base_pri > BASEPRI_FOREGROUND) {
569 decay_limit += (thread->base_pri - BASEPRI_FOREGROUND);
570 }
571
572 if (thread->pri_shift == INT8_MAX) {
573 decay_amount = 0;
574 } else {
575 decay_amount = (thread->sched_usage >> thread->pri_shift);
576 }
577
578 if (decay_amount > decay_limit) {
579 decay_amount = decay_limit;
580 }
581
582 /* start with base priority */
583 int priority = thread->base_pri - decay_amount;
584
585 if (priority < MAXPRI_THROTTLE) {
586 if (get_threadtask(thread)->max_priority > MAXPRI_THROTTLE) {
587 priority = MAXPRI_THROTTLE;
588 } else if (priority < MINPRI_USER) {
589 priority = MINPRI_USER;
590 }
591 } else if (priority > MAXPRI_KERNEL) {
592 priority = MAXPRI_KERNEL;
593 }
594
595 return priority;
596 }
597
598 #else /* !defined(XNU_TARGET_OS_OSX) && !CONFIG_SCHED_CLUTCH */
599
600 int
sched_compute_timeshare_priority(thread_t thread)601 sched_compute_timeshare_priority(thread_t thread)
602 {
603 /* start with base priority */
604 int priority = thread->base_pri;
605
606 if (thread->pri_shift != INT8_MAX) {
607 priority -= (thread->sched_usage >> thread->pri_shift);
608 }
609
610 if (priority < MINPRI_USER) {
611 priority = MINPRI_USER;
612 } else if (priority > MAXPRI_KERNEL) {
613 priority = MAXPRI_KERNEL;
614 }
615
616 return priority;
617 }
618
619 #endif /* !defined(XNU_TARGET_OS_OSX) && !CONFIG_SCHED_CLUTCH */
620
621 /*
622 * can_update_priority
623 *
624 * Make sure we don't do re-dispatches more frequently than a scheduler tick.
625 *
626 * Called with the thread locked.
627 */
628 boolean_t
can_update_priority(thread_t thread)629 can_update_priority(
630 thread_t thread)
631 {
632 if (sched_tick == thread->sched_stamp) {
633 return FALSE;
634 } else {
635 return TRUE;
636 }
637 }
638
639 /*
640 * update_priority
641 *
642 * Perform housekeeping operations driven by scheduler tick.
643 *
644 * Called with the thread locked.
645 */
646 void
update_priority(thread_t thread)647 update_priority(
648 thread_t thread)
649 {
650 uint32_t ticks, delta;
651
652 ticks = sched_tick - thread->sched_stamp;
653 assert(ticks != 0);
654
655 thread->sched_stamp += ticks;
656
657 /* If requested, accelerate aging of sched_usage */
658 if (sched_decay_usage_age_factor > 1) {
659 ticks *= sched_decay_usage_age_factor;
660 }
661
662 /*
663 * Gather cpu usage data.
664 */
665 sched_tick_delta(thread, delta);
666 if (ticks < SCHED_DECAY_TICKS) {
667 /*
668 * Accumulate timesharing usage only during contention for processor
669 * resources. Use the pri_shift from the previous tick window to
670 * determine if the system was in a contended state.
671 */
672 if (thread->pri_shift < INT8_MAX) {
673 if (thread_no_smt(thread) && smt_timeshare_enabled) {
674 thread->sched_usage += (delta + ((delta * smt_sched_bonus_16ths) >> 4));
675 } else {
676 thread->sched_usage += delta;
677 }
678 }
679
680 thread->cpu_usage += delta + thread->cpu_delta;
681 thread->cpu_delta = 0;
682
683 #if CONFIG_SCHED_CLUTCH
684 /*
685 * Update the CPU usage for the thread group to which the thread belongs.
686 * The implementation assumes that the thread ran for the entire delta
687 * as part of the same thread group.
688 */
689 sched_clutch_cpu_usage_update(thread, delta);
690 #endif /* CONFIG_SCHED_CLUTCH */
691
692 const struct shift_data *shiftp = &sched_decay_shifts[ticks];
693
694 if (shiftp->shift2 > 0) {
695 thread->cpu_usage = (thread->cpu_usage >> shiftp->shift1) +
696 (thread->cpu_usage >> shiftp->shift2);
697 thread->sched_usage = (thread->sched_usage >> shiftp->shift1) +
698 (thread->sched_usage >> shiftp->shift2);
699 } else {
700 thread->cpu_usage = (thread->cpu_usage >> shiftp->shift1) -
701 (thread->cpu_usage >> -(shiftp->shift2));
702 thread->sched_usage = (thread->sched_usage >> shiftp->shift1) -
703 (thread->sched_usage >> -(shiftp->shift2));
704 }
705 } else {
706 thread->cpu_usage = thread->cpu_delta = 0;
707 thread->sched_usage = 0;
708 }
709
710 /*
711 * Check for fail-safe release.
712 */
713 if ((thread->sched_flags & TH_SFLAG_FAILSAFE) &&
714 mach_absolute_time() >= thread->safe_release) {
715 sched_thread_mode_undemote(thread, TH_SFLAG_FAILSAFE);
716 }
717
718 /*
719 * Now that the thread's CPU usage has been accumulated and aged
720 * based on contention of the previous tick window, update the
721 * pri_shift of the thread to match the current global load/shift
722 * values. The updated pri_shift would be used to calculate the
723 * new priority of the thread.
724 */
725 #if CONFIG_SCHED_CLUTCH
726 thread->pri_shift = sched_clutch_thread_pri_shift(thread, thread->th_sched_bucket);
727 #else /* CONFIG_SCHED_CLUTCH */
728 thread->pri_shift = sched_pri_shifts[thread->th_sched_bucket];
729 #endif /* CONFIG_SCHED_CLUTCH */
730
731 /* Recompute scheduled priority if appropriate. */
732 if (thread->sched_mode == TH_MODE_TIMESHARE) {
733 thread_recompute_sched_pri(thread, SETPRI_LAZY);
734 }
735 }
736
737 #endif /* CONFIG_SCHED_TIMESHARE_CORE */
738
739
740 /*
741 * TH_BUCKET_RUN is a count of *all* runnable non-idle threads.
742 * Each other bucket is a count of the runnable non-idle threads
743 * with that property. All updates to these counts should be
744 * performed with os_atomic_* operations.
745 *
746 * For the clutch scheduler, this global bucket is used only for
747 * keeping the total global run count.
748 */
749 uint32_t sched_run_buckets[TH_BUCKET_MAX];
750
751 static void
sched_incr_bucket(sched_bucket_t bucket)752 sched_incr_bucket(sched_bucket_t bucket)
753 {
754 assert(bucket >= TH_BUCKET_FIXPRI &&
755 bucket <= TH_BUCKET_SHARE_BG);
756
757 os_atomic_inc(&sched_run_buckets[bucket], relaxed);
758 }
759
760 static void
sched_decr_bucket(sched_bucket_t bucket)761 sched_decr_bucket(sched_bucket_t bucket)
762 {
763 assert(bucket >= TH_BUCKET_FIXPRI &&
764 bucket <= TH_BUCKET_SHARE_BG);
765
766 assert(os_atomic_load(&sched_run_buckets[bucket], relaxed) > 0);
767
768 os_atomic_dec(&sched_run_buckets[bucket], relaxed);
769 }
770
771 static void
sched_add_bucket(sched_bucket_t bucket,uint8_t run_weight)772 sched_add_bucket(sched_bucket_t bucket, uint8_t run_weight)
773 {
774 assert(bucket >= TH_BUCKET_FIXPRI &&
775 bucket <= TH_BUCKET_SHARE_BG);
776
777 os_atomic_add(&sched_run_buckets[bucket], run_weight, relaxed);
778 }
779
780 static void
sched_sub_bucket(sched_bucket_t bucket,uint8_t run_weight)781 sched_sub_bucket(sched_bucket_t bucket, uint8_t run_weight)
782 {
783 assert(bucket >= TH_BUCKET_FIXPRI &&
784 bucket <= TH_BUCKET_SHARE_BG);
785
786 assert(os_atomic_load(&sched_run_buckets[bucket], relaxed) > 0);
787
788 os_atomic_sub(&sched_run_buckets[bucket], run_weight, relaxed);
789 }
790
791 uint32_t
sched_run_incr(thread_t thread)792 sched_run_incr(thread_t thread)
793 {
794 assert((thread->state & (TH_RUN | TH_IDLE)) == TH_RUN);
795
796 uint32_t new_count = os_atomic_inc(&sched_run_buckets[TH_BUCKET_RUN], relaxed);
797
798 sched_incr_bucket(thread->th_sched_bucket);
799
800 return new_count;
801 }
802
803 uint32_t
sched_run_decr(thread_t thread)804 sched_run_decr(thread_t thread)
805 {
806 assert((thread->state & (TH_RUN | TH_IDLE)) != TH_RUN);
807
808 sched_decr_bucket(thread->th_sched_bucket);
809
810 uint32_t new_count = os_atomic_dec(&sched_run_buckets[TH_BUCKET_RUN], relaxed);
811
812 return new_count;
813 }
814
815 uint32_t
sched_smt_run_incr(thread_t thread)816 sched_smt_run_incr(thread_t thread)
817 {
818 assert((thread->state & (TH_RUN | TH_IDLE)) == TH_RUN);
819
820 uint8_t run_weight = (thread_no_smt(thread) && smt_timeshare_enabled) ? 2 : 1;
821 thread->sched_saved_run_weight = run_weight;
822
823 uint32_t new_count = os_atomic_add(&sched_run_buckets[TH_BUCKET_RUN], run_weight, relaxed);
824
825 sched_add_bucket(thread->th_sched_bucket, run_weight);
826
827 return new_count;
828 }
829
830 uint32_t
sched_smt_run_decr(thread_t thread)831 sched_smt_run_decr(thread_t thread)
832 {
833 assert((thread->state & (TH_RUN | TH_IDLE)) != TH_RUN);
834
835 uint8_t run_weight = thread->sched_saved_run_weight;
836
837 sched_sub_bucket(thread->th_sched_bucket, run_weight);
838
839 uint32_t new_count = os_atomic_sub(&sched_run_buckets[TH_BUCKET_RUN], run_weight, relaxed);
840
841 return new_count;
842 }
843
844 void
sched_update_thread_bucket(thread_t thread)845 sched_update_thread_bucket(thread_t thread)
846 {
847 sched_bucket_t old_bucket = thread->th_sched_bucket;
848 sched_bucket_t new_bucket = TH_BUCKET_RUN;
849
850 switch (thread->sched_mode) {
851 case TH_MODE_FIXED:
852 case TH_MODE_REALTIME:
853 new_bucket = TH_BUCKET_FIXPRI;
854 break;
855
856 case TH_MODE_TIMESHARE:
857 if (thread->base_pri > BASEPRI_DEFAULT) {
858 new_bucket = TH_BUCKET_SHARE_FG;
859 } else if (thread->base_pri > BASEPRI_UTILITY) {
860 new_bucket = TH_BUCKET_SHARE_DF;
861 } else if (thread->base_pri > MAXPRI_THROTTLE) {
862 new_bucket = TH_BUCKET_SHARE_UT;
863 } else {
864 new_bucket = TH_BUCKET_SHARE_BG;
865 }
866 break;
867
868 default:
869 panic("unexpected mode: %d", thread->sched_mode);
870 break;
871 }
872
873 if (old_bucket != new_bucket) {
874 thread->th_sched_bucket = new_bucket;
875 thread->pri_shift = sched_pri_shifts[new_bucket];
876
877 if ((thread->state & (TH_RUN | TH_IDLE)) == TH_RUN) {
878 sched_decr_bucket(old_bucket);
879 sched_incr_bucket(new_bucket);
880 }
881 }
882 }
883
884 void
sched_smt_update_thread_bucket(thread_t thread)885 sched_smt_update_thread_bucket(thread_t thread)
886 {
887 sched_bucket_t old_bucket = thread->th_sched_bucket;
888 sched_bucket_t new_bucket = TH_BUCKET_RUN;
889
890 switch (thread->sched_mode) {
891 case TH_MODE_FIXED:
892 case TH_MODE_REALTIME:
893 new_bucket = TH_BUCKET_FIXPRI;
894 break;
895
896 case TH_MODE_TIMESHARE:
897 if (thread->base_pri > BASEPRI_DEFAULT) {
898 new_bucket = TH_BUCKET_SHARE_FG;
899 } else if (thread->base_pri > BASEPRI_UTILITY) {
900 new_bucket = TH_BUCKET_SHARE_DF;
901 } else if (thread->base_pri > MAXPRI_THROTTLE) {
902 new_bucket = TH_BUCKET_SHARE_UT;
903 } else {
904 new_bucket = TH_BUCKET_SHARE_BG;
905 }
906 break;
907
908 default:
909 panic("unexpected mode: %d", thread->sched_mode);
910 break;
911 }
912
913 if (old_bucket != new_bucket) {
914 thread->th_sched_bucket = new_bucket;
915 thread->pri_shift = sched_pri_shifts[new_bucket];
916
917 if ((thread->state & (TH_RUN | TH_IDLE)) == TH_RUN) {
918 sched_sub_bucket(old_bucket, thread->sched_saved_run_weight);
919 sched_add_bucket(new_bucket, thread->sched_saved_run_weight);
920 }
921 }
922 }
923
924 static inline void
sched_validate_mode(sched_mode_t mode)925 sched_validate_mode(sched_mode_t mode)
926 {
927 switch (mode) {
928 case TH_MODE_FIXED:
929 case TH_MODE_REALTIME:
930 case TH_MODE_TIMESHARE:
931 break;
932
933 default:
934 panic("unexpected mode: %d", mode);
935 break;
936 }
937 }
938
939 /*
940 * Set the thread's true scheduling mode
941 * Called with thread mutex and thread locked
942 * The thread has already been removed from the runqueue.
943 *
944 * (saved_mode is handled before this point)
945 */
946 void
sched_set_thread_mode(thread_t thread,sched_mode_t new_mode)947 sched_set_thread_mode(thread_t thread, sched_mode_t new_mode)
948 {
949 assert(thread->runq == PROCESSOR_NULL);
950
951 sched_validate_mode(new_mode);
952
953 #if CONFIG_SCHED_AUTO_JOIN
954 /*
955 * Realtime threads might have auto-joined a work interval based on
956 * make runnable relationships. If such an RT thread is now being demoted
957 * to non-RT, unjoin the thread from the work interval.
958 */
959 if ((thread->sched_flags & TH_SFLAG_THREAD_GROUP_AUTO_JOIN) && (new_mode != TH_MODE_REALTIME)) {
960 assert((thread->sched_mode == TH_MODE_REALTIME) || (thread->th_work_interval_flags & TH_WORK_INTERVAL_FLAGS_AUTO_JOIN_LEAK));
961 work_interval_auto_join_demote(thread);
962 }
963 #endif /* CONFIG_SCHED_AUTO_JOIN */
964
965 thread->sched_mode = new_mode;
966
967 SCHED(update_thread_bucket)(thread);
968 }
969
970 /*
971 * TODO: Instead of having saved mode, have 'user mode' and 'true mode'.
972 * That way there's zero confusion over which the user wants
973 * and which the kernel wants.
974 */
975 void
sched_set_thread_mode_user(thread_t thread,sched_mode_t new_mode)976 sched_set_thread_mode_user(thread_t thread, sched_mode_t new_mode)
977 {
978 assert(thread->runq == PROCESSOR_NULL);
979
980 sched_validate_mode(new_mode);
981
982 /* If demoted, only modify the saved mode. */
983 if (thread->sched_flags & TH_SFLAG_DEMOTED_MASK) {
984 thread->saved_mode = new_mode;
985 } else {
986 sched_set_thread_mode(thread, new_mode);
987 }
988 }
989
990 /*
991 * Demote the true scheduler mode to timeshare (called with the thread locked)
992 */
993 void
sched_thread_mode_demote(thread_t thread,uint32_t reason)994 sched_thread_mode_demote(thread_t thread, uint32_t reason)
995 {
996 assert(reason & TH_SFLAG_DEMOTED_MASK);
997 assert((thread->sched_flags & reason) != reason);
998
999 if (thread->policy_reset) {
1000 return;
1001 }
1002
1003 switch (reason) {
1004 case TH_SFLAG_THROTTLED:
1005 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MODE_DEMOTE_THROTTLED),
1006 thread_tid(thread), thread->sched_flags);
1007 break;
1008 case TH_SFLAG_FAILSAFE:
1009 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MODE_DEMOTE_FAILSAFE),
1010 thread_tid(thread), thread->sched_flags);
1011 break;
1012 case TH_SFLAG_RT_RESTRICTED:
1013 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MODE_DEMOTE_RT_RESTRICTED),
1014 thread_tid(thread), thread->sched_flags);
1015 break;
1016 }
1017
1018 if (thread->sched_flags & TH_SFLAG_DEMOTED_MASK) {
1019 /* Another demotion reason is already active */
1020 thread->sched_flags |= reason;
1021 return;
1022 }
1023
1024 assert(thread->saved_mode == TH_MODE_NONE);
1025
1026 boolean_t removed = thread_run_queue_remove(thread);
1027
1028 thread->sched_flags |= reason;
1029
1030 thread->saved_mode = thread->sched_mode;
1031
1032 sched_set_thread_mode(thread, TH_MODE_TIMESHARE);
1033
1034 thread_recompute_priority(thread);
1035
1036 if (removed) {
1037 thread_run_queue_reinsert(thread, SCHED_TAILQ);
1038 }
1039 }
1040
1041 /*
1042 * Un-demote the true scheduler mode back to the saved mode (called with the thread locked)
1043 */
1044 void
sched_thread_mode_undemote(thread_t thread,uint32_t reason)1045 sched_thread_mode_undemote(thread_t thread, uint32_t reason)
1046 {
1047 assert(reason & TH_SFLAG_DEMOTED_MASK);
1048 assert((thread->sched_flags & reason) == reason);
1049 assert(thread->saved_mode != TH_MODE_NONE);
1050 assert(thread->sched_mode == TH_MODE_TIMESHARE);
1051 assert(thread->policy_reset == 0);
1052
1053 switch (reason) {
1054 case TH_SFLAG_THROTTLED:
1055 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MODE_UNDEMOTE_THROTTLED),
1056 thread_tid(thread), thread->sched_flags);
1057 break;
1058 case TH_SFLAG_FAILSAFE:
1059 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MODE_UNDEMOTE_FAILSAFE),
1060 thread_tid(thread), thread->sched_flags);
1061 break;
1062 case TH_SFLAG_RT_RESTRICTED:
1063 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MODE_UNDEMOTE_RT_RESTRICTED),
1064 thread_tid(thread), thread->sched_flags);
1065 break;
1066 }
1067
1068 thread->sched_flags &= ~reason;
1069
1070 if (thread->sched_flags & TH_SFLAG_DEMOTED_MASK) {
1071 /* Another demotion reason is still active */
1072 return;
1073 }
1074
1075 boolean_t removed = thread_run_queue_remove(thread);
1076
1077 sched_set_thread_mode(thread, thread->saved_mode);
1078
1079 thread->saved_mode = TH_MODE_NONE;
1080
1081 thread_recompute_priority(thread);
1082
1083 if (removed) {
1084 thread_run_queue_reinsert(thread, SCHED_TAILQ);
1085 }
1086 }
1087
1088 /*
1089 * Promote thread to have a sched pri floor for a specific reason
1090 *
1091 * Promotion must not last past syscall boundary
1092 * Clients must always pair promote and demote 1:1,
1093 * Handling nesting of the same promote reason is the client's responsibility
1094 *
1095 * Called at splsched with thread locked
1096 */
1097 void
sched_thread_promote_reason(thread_t thread,uint32_t reason,__kdebug_only uintptr_t trace_obj)1098 sched_thread_promote_reason(thread_t thread,
1099 uint32_t reason,
1100 __kdebug_only uintptr_t trace_obj /* already unslid */)
1101 {
1102 assert(reason & TH_SFLAG_PROMOTE_REASON_MASK);
1103 assert((thread->sched_flags & reason) != reason);
1104
1105 switch (reason) {
1106 case TH_SFLAG_RW_PROMOTED:
1107 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_RW_PROMOTE),
1108 thread_tid(thread), thread->sched_pri,
1109 thread->base_pri, trace_obj);
1110 break;
1111 case TH_SFLAG_WAITQ_PROMOTED:
1112 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAITQ_PROMOTE),
1113 thread_tid(thread), thread->sched_pri,
1114 thread->base_pri, trace_obj);
1115 break;
1116 case TH_SFLAG_EXEC_PROMOTED:
1117 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_EXEC_PROMOTE),
1118 thread_tid(thread), thread->sched_pri,
1119 thread->base_pri, trace_obj);
1120 break;
1121 case TH_SFLAG_FLOOR_PROMOTED:
1122 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_FLOOR_PROMOTE),
1123 thread_tid(thread), thread->sched_pri,
1124 thread->base_pri, trace_obj);
1125 break;
1126 }
1127
1128 thread->sched_flags |= reason;
1129 thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
1130 }
1131
1132 /*
1133 * End a specific promotion reason
1134 * Demotes a thread back to its expected priority without the promotion in place
1135 *
1136 * Called at splsched with thread locked
1137 */
1138 void
sched_thread_unpromote_reason(thread_t thread,uint32_t reason,__kdebug_only uintptr_t trace_obj)1139 sched_thread_unpromote_reason(thread_t thread,
1140 uint32_t reason,
1141 __kdebug_only uintptr_t trace_obj /* already unslid */)
1142 {
1143 assert(reason & TH_SFLAG_PROMOTE_REASON_MASK);
1144 assert((thread->sched_flags & reason) == reason);
1145
1146 switch (reason) {
1147 case TH_SFLAG_RW_PROMOTED:
1148 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_RW_DEMOTE),
1149 thread_tid(thread), thread->sched_pri,
1150 thread->base_pri, trace_obj);
1151 break;
1152 case TH_SFLAG_WAITQ_PROMOTED:
1153 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAITQ_DEMOTE),
1154 thread_tid(thread), thread->sched_pri,
1155 thread->base_pri, trace_obj);
1156 break;
1157 case TH_SFLAG_EXEC_PROMOTED:
1158 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_EXEC_DEMOTE),
1159 thread_tid(thread), thread->sched_pri,
1160 thread->base_pri, trace_obj);
1161 break;
1162 case TH_SFLAG_FLOOR_PROMOTED:
1163 KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_FLOOR_DEMOTE),
1164 thread_tid(thread), thread->sched_pri,
1165 thread->base_pri, trace_obj);
1166 break;
1167 }
1168
1169 thread->sched_flags &= ~reason;
1170
1171 thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
1172 }
1173