source: mainline/kernel/generic/src/proc/scheduler.c@ b23c88e

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since b23c88e was b23c88e, checked in by Adam Hraska <adam.hraska+hos@…>, 13 years ago

preemption_disable: Replaced memory barriers with compiler barriers. Added checks if reschedule is needed once preemption is enabled.
  • Property mode set to 100644
File size: 17.7 KB
Line 
1/*
2 * Copyright (c) 2010 Jakub Jermar
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *
9 * - Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * - Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * - The name of the author may not be used to endorse or promote products
15 * derived from this software without specific prior written permission.
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/** @addtogroup genericproc
30 * @{
31 */
32
33/**
34 * @file
35 * @brief Scheduler and load balancing.
36 *
37 * This file contains the scheduler and kcpulb kernel thread which
38 * performs load-balancing of per-CPU run queues.
39 */
40
41#include <proc/scheduler.h>
42#include <proc/thread.h>
43#include <proc/task.h>
44#include <mm/frame.h>
45#include <mm/page.h>
46#include <mm/as.h>
47#include <time/timeout.h>
48#include <time/delay.h>
49#include <arch/asm.h>
50#include <arch/faddr.h>
51#include <arch/cycle.h>
52#include <atomic.h>
53#include <synch/spinlock.h>
54#include <synch/workqueue.h>
55#include <synch/rcu.h>
56#include <config.h>
57#include <context.h>
58#include <fpu_context.h>
59#include <func.h>
60#include <arch.h>
61#include <adt/list.h>
62#include <panic.h>
63#include <cpu.h>
64#include <print.h>
65#include <debug.h>
66#include <stacktrace.h>
67
68static void scheduler_separated_stack(void);
69
70atomic_t nrdy; /**< Number of ready threads in the system. */
71
72/** Carry out actions before new task runs. */
73static void before_task_runs(void)
74{
75 before_task_runs_arch();
76}
77
78/** Take actions before new thread runs.
79 *
80 * Perform actions that need to be
81 * taken before the newly selected
82 * thread is passed control.
83 *
84 * THREAD->lock is locked on entry
85 *
86 */
87static void before_thread_runs(void)
88{
89 before_thread_runs_arch();
90 rcu_before_thread_runs();
91
92#ifdef CONFIG_FPU_LAZY
93 if (THREAD == CPU->fpu_owner)
94 fpu_enable();
95 else
96 fpu_disable();
97#else
98 fpu_enable();
99 if (THREAD->fpu_context_exists)
100 fpu_context_restore(THREAD->saved_fpu_context);
101 else {
102 fpu_init();
103 THREAD->fpu_context_exists = true;
104 }
105#endif
106
107#ifdef CONFIG_UDEBUG
108 if (THREAD->btrace) {
109 istate_t *istate = THREAD->udebug.uspace_state;
110 if (istate != NULL) {
111 printf("Thread %" PRIu64 " stack trace:\n", THREAD->tid);
112 stack_trace_istate(istate);
113 }
114
115 THREAD->btrace = false;
116 }
117#endif
118}
119
120/** Take actions after THREAD had run.
121 *
122 * Perform actions that need to be
123 * taken after the running thread
124 * had been preempted by the scheduler.
125 *
126 * THREAD->lock is locked on entry
127 *
128 */
129static void after_thread_ran(void)
130{
131 workq_after_thread_ran();
132 rcu_after_thread_ran();
133 after_thread_ran_arch();
134}
135
136#ifdef CONFIG_FPU_LAZY
137void scheduler_fpu_lazy_request(void)
138{
139restart:
140 fpu_enable();
141 irq_spinlock_lock(&CPU->lock, false);
142
143 /* Save old context */
144 if (CPU->fpu_owner != NULL) {
145 irq_spinlock_lock(&CPU->fpu_owner->lock, false);
146 fpu_context_save(CPU->fpu_owner->saved_fpu_context);
147
148 /* Don't prevent migration */
149 CPU->fpu_owner->fpu_context_engaged = false;
150 irq_spinlock_unlock(&CPU->fpu_owner->lock, false);
151 CPU->fpu_owner = NULL;
152 }
153
154 irq_spinlock_lock(&THREAD->lock, false);
155 if (THREAD->fpu_context_exists) {
156 fpu_context_restore(THREAD->saved_fpu_context);
157 } else {
158 /* Allocate FPU context */
159 if (!THREAD->saved_fpu_context) {
160 /* Might sleep */
161 irq_spinlock_unlock(&THREAD->lock, false);
162 irq_spinlock_unlock(&CPU->lock, false);
163 THREAD->saved_fpu_context =
164 (fpu_context_t *) slab_alloc(fpu_context_slab, 0);
165
166 /* We may have switched CPUs during slab_alloc */
167 goto restart;
168 }
169 fpu_init();
170 THREAD->fpu_context_exists = true;
171 }
172
173 CPU->fpu_owner = THREAD;
174 THREAD->fpu_context_engaged = true;
175 irq_spinlock_unlock(&THREAD->lock, false);
176
177 irq_spinlock_unlock(&CPU->lock, false);
178}
179#endif /* CONFIG_FPU_LAZY */
180
181/** Initialize scheduler
182 *
183 * Initialize kernel scheduler.
184 *
185 */
186void scheduler_init(void)
187{
188}
189
190/** Get thread to be scheduled
191 *
192 * Get the optimal thread to be scheduled
193 * according to thread accounting and scheduler
194 * policy.
195 *
196 * @return Thread to be scheduled.
197 *
198 */
199static thread_t *find_best_thread(void)
200{
201 ASSERT(CPU != NULL);
202
203loop:
204
205 if (atomic_get(&CPU->nrdy) == 0) {
206 /*
207 * For there was nothing to run, the CPU goes to sleep
208 * until a hardware interrupt or an IPI comes.
209 * This improves energy saving and hyperthreading.
210 */
211 irq_spinlock_lock(&CPU->lock, false);
212 CPU->idle = true;
213 irq_spinlock_unlock(&CPU->lock, false);
214 interrupts_enable();
215
216 /*
217 * An interrupt might occur right now and wake up a thread.
218 * In such case, the CPU will continue to go to sleep
219 * even though there is a runnable thread.
220 */
221 cpu_sleep();
222 interrupts_disable();
223 goto loop;
224 }
225
226 ASSERT(!CPU->idle);
227
228 unsigned int i;
229 for (i = 0; i < RQ_COUNT; i++) {
230 irq_spinlock_lock(&(CPU->rq[i].lock), false);
231 if (CPU->rq[i].n == 0) {
232 /*
233 * If this queue is empty, try a lower-priority queue.
234 */
235 irq_spinlock_unlock(&(CPU->rq[i].lock), false);
236 continue;
237 }
238
239 atomic_dec(&CPU->nrdy);
240 atomic_dec(&nrdy);
241 CPU->rq[i].n--;
242
243 /*
244 * Take the first thread from the queue.
245 */
246 thread_t *thread = list_get_instance(
247 list_first(&CPU->rq[i].rq), thread_t, rq_link);
248 list_remove(&thread->rq_link);
249
250 irq_spinlock_pass(&(CPU->rq[i].lock), &thread->lock);
251
252 thread->cpu = CPU;
253 thread->ticks = us2ticks((i + 1) * 10000);
254 thread->priority = i; /* Correct rq index */
255
256 /*
257 * Clear the stolen flag so that it can be migrated
258 * when load balancing needs emerge.
259 */
260 thread->stolen = false;
261 irq_spinlock_unlock(&thread->lock, false);
262
263 return thread;
264 }
265
266 goto loop;
267}
268
269/** Prevent rq starvation
270 *
271 * Prevent low priority threads from starving in rq's.
272 *
273 * When the function decides to relink rq's, it reconnects
274 * respective pointers so that in result threads with 'pri'
275 * greater or equal start are moved to a higher-priority queue.
276 *
277 * @param start Threshold priority.
278 *
279 */
280static void relink_rq(int start)
281{
282 list_t list;
283
284 list_initialize(&list);
285 irq_spinlock_lock(&CPU->lock, false);
286
287 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
288 int i;
289 for (i = start; i < RQ_COUNT - 1; i++) {
290 /* Remember and empty rq[i + 1] */
291
292 irq_spinlock_lock(&CPU->rq[i + 1].lock, false);
293 list_concat(&list, &CPU->rq[i + 1].rq);
294 size_t n = CPU->rq[i + 1].n;
295 CPU->rq[i + 1].n = 0;
296 irq_spinlock_unlock(&CPU->rq[i + 1].lock, false);
297
298 /* Append rq[i + 1] to rq[i] */
299
300 irq_spinlock_lock(&CPU->rq[i].lock, false);
301 list_concat(&CPU->rq[i].rq, &list);
302 CPU->rq[i].n += n;
303 irq_spinlock_unlock(&CPU->rq[i].lock, false);
304 }
305
306 CPU->needs_relink = 0;
307 }
308
309 irq_spinlock_unlock(&CPU->lock, false);
310}
311
312/** The scheduler
313 *
314 * The thread scheduling procedure.
315 * Passes control directly to
316 * scheduler_separated_stack().
317 *
318 */
319void scheduler(void)
320{
321 volatile ipl_t ipl;
322
323 ASSERT(CPU != NULL);
324
325 ipl = interrupts_disable();
326
327 if (atomic_get(&haltstate))
328 halt();
329
330 if (THREAD) {
331 irq_spinlock_lock(&THREAD->lock, false);
332
333 /* Update thread kernel accounting */
334 THREAD->kcycles += get_cycle() - THREAD->last_cycle;
335
336#ifndef CONFIG_FPU_LAZY
337 fpu_context_save(THREAD->saved_fpu_context);
338#endif
339 if (!context_save(&THREAD->saved_context)) {
340 /*
341 * This is the place where threads leave scheduler();
342 */
343
344 /* Save current CPU cycle */
345 THREAD->last_cycle = get_cycle();
346
347 irq_spinlock_unlock(&THREAD->lock, false);
348 interrupts_restore(THREAD->saved_context.ipl);
349
350 return;
351 }
352
353 /*
354 * Interrupt priority level of preempted thread is recorded
355 * here to facilitate scheduler() invocations from
356 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
357 *
358 */
359 THREAD->saved_context.ipl = ipl;
360 }
361
362 /*
363 * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, AS
364 * and preemption counter. At this point THE could be coming either
365 * from THREAD's or CPU's stack.
366 *
367 */
368 the_copy(THE, (the_t *) CPU->stack);
369
370 /*
371 * We may not keep the old stack.
372 * Reason: If we kept the old stack and got blocked, for instance, in
373 * find_best_thread(), the old thread could get rescheduled by another
374 * CPU and overwrite the part of its own stack that was also used by
375 * the scheduler on this CPU.
376 *
377 * Moreover, we have to bypass the compiler-generated POP sequence
378 * which is fooled by SP being set to the very top of the stack.
379 * Therefore the scheduler() function continues in
380 * scheduler_separated_stack().
381 *
382 */
383 context_save(&CPU->saved_context);
384 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack),
385 (uintptr_t) CPU->stack, STACK_SIZE);
386 context_restore(&CPU->saved_context);
387
388 /* Not reached */
389}
390
391/** Scheduler stack switch wrapper
392 *
393 * Second part of the scheduler() function
394 * using new stack. Handling the actual context
395 * switch to a new thread.
396 *
397 */
398void scheduler_separated_stack(void)
399{
400 DEADLOCK_PROBE_INIT(p_joinwq);
401 task_t *old_task = TASK;
402 as_t *old_as = AS;
403
404 ASSERT((!THREAD) || (irq_spinlock_locked(&THREAD->lock)));
405 ASSERT(CPU != NULL);
406 ASSERT(interrupts_disabled());
407
408 /*
409 * Hold the current task and the address space to prevent their
410 * possible destruction should thread_destroy() be called on this or any
411 * other processor while the scheduler is still using them.
412 */
413 if (old_task)
414 task_hold(old_task);
415
416 if (old_as)
417 as_hold(old_as);
418
419 if (THREAD) {
420 /* Must be run after the switch to scheduler stack */
421 after_thread_ran();
422
423 THREAD->need_resched = false;
424
425 switch (THREAD->state) {
426 case Running:
427 irq_spinlock_unlock(&THREAD->lock, false);
428 thread_ready(THREAD);
429 break;
430
431 case Exiting:
432 rcu_thread_exiting();
433repeat:
434 if (THREAD->detached) {
435 thread_destroy(THREAD, false);
436 } else {
437 /*
438 * The thread structure is kept allocated until
439 * somebody calls thread_detach() on it.
440 */
441 if (!irq_spinlock_trylock(&THREAD->join_wq.lock)) {
442 /*
443 * Avoid deadlock.
444 */
445 irq_spinlock_unlock(&THREAD->lock, false);
446 delay(HZ);
447 irq_spinlock_lock(&THREAD->lock, false);
448 DEADLOCK_PROBE(p_joinwq,
449 DEADLOCK_THRESHOLD);
450 goto repeat;
451 }
452 _waitq_wakeup_unsafe(&THREAD->join_wq,
453 WAKEUP_FIRST);
454 irq_spinlock_unlock(&THREAD->join_wq.lock, false);
455
456 THREAD->state = Lingering;
457 irq_spinlock_unlock(&THREAD->lock, false);
458 }
459 break;
460
461 case Sleeping:
462 /*
463 * Prefer the thread after it's woken up.
464 */
465 THREAD->priority = -1;
466
467 /*
468 * We need to release wq->lock which we locked in
469 * waitq_sleep(). Address of wq->lock is kept in
470 * THREAD->sleep_queue.
471 */
472 irq_spinlock_unlock(&THREAD->sleep_queue->lock, false);
473
474 irq_spinlock_unlock(&THREAD->lock, false);
475 break;
476
477 default:
478 /*
479 * Entering state is unexpected.
480 */
481 panic("tid%" PRIu64 ": unexpected state %s.",
482 THREAD->tid, thread_states[THREAD->state]);
483 break;
484 }
485
486 THREAD = NULL;
487 }
488
489 THREAD = find_best_thread();
490
491 irq_spinlock_lock(&THREAD->lock, false);
492 int priority = THREAD->priority;
493 irq_spinlock_unlock(&THREAD->lock, false);
494
495 relink_rq(priority);
496
497 /*
498 * If both the old and the new task are the same,
499 * lots of work is avoided.
500 */
501 if (TASK != THREAD->task) {
502 as_t *new_as = THREAD->task->as;
503
504 /*
505 * Note that it is possible for two tasks
506 * to share one address space.
507 */
508 if (old_as != new_as) {
509 /*
510 * Both tasks and address spaces are different.
511 * Replace the old one with the new one.
512 */
513 as_switch(old_as, new_as);
514 }
515
516 TASK = THREAD->task;
517 before_task_runs();
518 }
519
520 if (old_task)
521 task_release(old_task);
522
523 if (old_as)
524 as_release(old_as);
525
526 irq_spinlock_lock(&THREAD->lock, false);
527 THREAD->state = Running;
528
529#ifdef SCHEDULER_VERBOSE
530 printf("cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
531 ", nrdy=%ld)\n", CPU->id, THREAD->tid, THREAD->priority,
532 THREAD->ticks, atomic_get(&CPU->nrdy));
533#endif
534
535 /*
536 * Some architectures provide late kernel PA2KA(identity)
537 * mapping in a page fault handler. However, the page fault
538 * handler uses the kernel stack of the running thread and
539 * therefore cannot be used to map it. The kernel stack, if
540 * necessary, is to be mapped in before_thread_runs(). This
541 * function must be executed before the switch to the new stack.
542 */
543 before_thread_runs();
544
545 /*
546 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
547 * thread's stack.
548 */
549 the_copy(THE, (the_t *) THREAD->kstack);
550
551 context_restore(&THREAD->saved_context);
552
553 /* Not reached */
554}
555
556#ifdef CONFIG_SMP
557/** Load balancing thread
558 *
559 * SMP load balancing thread, supervising thread supplies
560 * for the CPU it's wired to.
561 *
562 * @param arg Generic thread argument (unused).
563 *
564 */
565void kcpulb(void *arg)
566{
567 atomic_count_t average;
568 atomic_count_t rdy;
569
570 /*
571 * Detach kcpulb as nobody will call thread_join_timeout() on it.
572 */
573 thread_detach(THREAD);
574
575loop:
576 /*
577 * Work in 1s intervals.
578 */
579 thread_sleep(1);
580
581not_satisfied:
582 /*
583 * Calculate the number of threads that will be migrated/stolen from
584 * other CPU's. Note that situation can have changed between two
585 * passes. Each time get the most up to date counts.
586 *
587 */
588 average = atomic_get(&nrdy) / config.cpu_active + 1;
589 rdy = atomic_get(&CPU->nrdy);
590
591 if (average <= rdy)
592 goto satisfied;
593
594 atomic_count_t count = average - rdy;
595
596 /*
597 * Searching least priority queues on all CPU's first and most priority
598 * queues on all CPU's last.
599 */
600 size_t acpu;
601 size_t acpu_bias = 0;
602 int rq;
603
604 for (rq = RQ_COUNT - 1; rq >= 0; rq--) {
605 for (acpu = 0; acpu < config.cpu_active; acpu++) {
606 cpu_t *cpu = &cpus[(acpu + acpu_bias) % config.cpu_active];
607
608 /*
609 * Not interested in ourselves.
610 * Doesn't require interrupt disabling for kcpulb has
611 * THREAD_FLAG_WIRED.
612 *
613 */
614 if (CPU == cpu)
615 continue;
616
617 if (atomic_get(&cpu->nrdy) <= average)
618 continue;
619
620 irq_spinlock_lock(&(cpu->rq[rq].lock), true);
621 if (cpu->rq[rq].n == 0) {
622 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
623 continue;
624 }
625
626 thread_t *thread = NULL;
627
628 /* Search rq from the back */
629 link_t *link = cpu->rq[rq].rq.head.prev;
630
631 while (link != &(cpu->rq[rq].rq.head)) {
632 thread = (thread_t *) list_get_instance(link,
633 thread_t, rq_link);
634
635 /*
636 * Do not steal CPU-wired threads, threads
637 * already stolen, threads for which migration
638 * was temporarily disabled or threads whose
639 * FPU context is still in the CPU.
640 */
641 irq_spinlock_lock(&thread->lock, false);
642
643 if ((!thread->wired) && (!thread->stolen) &&
644 (!thread->nomigrate) &&
645 (!thread->fpu_context_engaged)) {
646 /*
647 * Remove thread from ready queue.
648 */
649 irq_spinlock_unlock(&thread->lock,
650 false);
651
652 atomic_dec(&cpu->nrdy);
653 atomic_dec(&nrdy);
654
655 cpu->rq[rq].n--;
656 list_remove(&thread->rq_link);
657
658 break;
659 }
660
661 irq_spinlock_unlock(&thread->lock, false);
662
663 link = link->prev;
664 thread = NULL;
665 }
666
667 if (thread) {
668 /*
669 * Ready thread on local CPU
670 */
671
672 irq_spinlock_pass(&(cpu->rq[rq].lock),
673 &thread->lock);
674
675#ifdef KCPULB_VERBOSE
676 printf("kcpulb%u: TID %" PRIu64 " -> cpu%u, "
677 "nrdy=%ld, avg=%ld\n", CPU->id, t->tid,
678 CPU->id, atomic_get(&CPU->nrdy),
679 atomic_get(&nrdy) / config.cpu_active);
680#endif
681
682 thread->stolen = true;
683 thread->state = Entering;
684
685 irq_spinlock_unlock(&thread->lock, true);
686 thread_ready(thread);
687
688 if (--count == 0)
689 goto satisfied;
690
691 /*
692 * We are not satisfied yet, focus on another
693 * CPU next time.
694 *
695 */
696 acpu_bias++;
697
698 continue;
699 } else
700 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
701
702 }
703 }
704
705 if (atomic_get(&CPU->nrdy)) {
706 /*
707 * Be a little bit light-weight and let migrated threads run.
708 *
709 */
710 scheduler();
711 } else {
712 /*
713 * We failed to migrate a single thread.
714 * Give up this turn.
715 *
716 */
717 goto loop;
718 }
719
720 goto not_satisfied;
721
722satisfied:
723 goto loop;
724}
725#endif /* CONFIG_SMP */
726
727/** Print information about threads & scheduler queues
728 *
729 */
730void sched_print_list(void)
731{
732 size_t cpu;
733 for (cpu = 0; cpu < config.cpu_count; cpu++) {
734 if (!cpus[cpu].active)
735 continue;
736
737 irq_spinlock_lock(&cpus[cpu].lock, true);
738
739 printf("cpu%u: address=%p, nrdy=%" PRIua ", needs_relink=%zu\n",
740 cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy),
741 cpus[cpu].needs_relink);
742
743 unsigned int i;
744 for (i = 0; i < RQ_COUNT; i++) {
745 irq_spinlock_lock(&(cpus[cpu].rq[i].lock), false);
746 if (cpus[cpu].rq[i].n == 0) {
747 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
748 continue;
749 }
750
751 printf("\trq[%u]: ", i);
752 list_foreach(cpus[cpu].rq[i].rq, cur) {
753 thread_t *thread = list_get_instance(cur,
754 thread_t, rq_link);
755 printf("%" PRIu64 "(%s) ", thread->tid,
756 thread_states[thread->state]);
757 }
758 printf("\n");
759
760 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
761 }
762
763 irq_spinlock_unlock(&cpus[cpu].lock, true);
764 }
765}
766
767/** @}
768 */
Note: See TracBrowser for help on using the repository browser.