source: mainline/kernel/generic/src/proc/scheduler.c@ 550523f5

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since 550523f5 was 550523f5, checked in by Martin Decky <martin@…>, 12 years ago

fix debugging output
  • Property mode set to 100644
File size: 17.5 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 <config.h>
55#include <context.h>
56#include <fpu_context.h>
57#include <func.h>
58#include <arch.h>
59#include <adt/list.h>
60#include <panic.h>
61#include <cpu.h>
62#include <print.h>
63#include <debug.h>
64#include <stacktrace.h>
65
66static void scheduler_separated_stack(void);
67
68atomic_t nrdy; /**< Number of ready threads in the system. */
69
70/** Carry out actions before new task runs. */
71static void before_task_runs(void)
72{
73 before_task_runs_arch();
74}
75
76/** Take actions before new thread runs.
77 *
78 * Perform actions that need to be
79 * taken before the newly selected
80 * thread is passed control.
81 *
82 * THREAD->lock is locked on entry
83 *
84 */
85static void before_thread_runs(void)
86{
87 before_thread_runs_arch();
88
89#ifdef CONFIG_FPU_LAZY
90 if (THREAD == CPU->fpu_owner)
91 fpu_enable();
92 else
93 fpu_disable();
94#elif defined CONFIG_FPU
95 fpu_enable();
96 if (THREAD->fpu_context_exists)
97 fpu_context_restore(THREAD->saved_fpu_context);
98 else {
99 fpu_init();
100 THREAD->fpu_context_exists = true;
101 }
102#endif
103
104#ifdef CONFIG_UDEBUG
105 if (THREAD->btrace) {
106 istate_t *istate = THREAD->udebug.uspace_state;
107 if (istate != NULL) {
108 printf("Thread %" PRIu64 " stack trace:\n", THREAD->tid);
109 stack_trace_istate(istate);
110 }
111
112 THREAD->btrace = false;
113 }
114#endif
115}
116
117/** Take actions after THREAD had run.
118 *
119 * Perform actions that need to be
120 * taken after the running thread
121 * had been preempted by the scheduler.
122 *
123 * THREAD->lock is locked on entry
124 *
125 */
126static void after_thread_ran(void)
127{
128 after_thread_ran_arch();
129}
130
131#ifdef CONFIG_FPU_LAZY
132void scheduler_fpu_lazy_request(void)
133{
134restart:
135 fpu_enable();
136 irq_spinlock_lock(&CPU->lock, false);
137
138 /* Save old context */
139 if (CPU->fpu_owner != NULL) {
140 irq_spinlock_lock(&CPU->fpu_owner->lock, false);
141 fpu_context_save(CPU->fpu_owner->saved_fpu_context);
142
143 /* Don't prevent migration */
144 CPU->fpu_owner->fpu_context_engaged = false;
145 irq_spinlock_unlock(&CPU->fpu_owner->lock, false);
146 CPU->fpu_owner = NULL;
147 }
148
149 irq_spinlock_lock(&THREAD->lock, false);
150 if (THREAD->fpu_context_exists) {
151 fpu_context_restore(THREAD->saved_fpu_context);
152 } else {
153 /* Allocate FPU context */
154 if (!THREAD->saved_fpu_context) {
155 /* Might sleep */
156 irq_spinlock_unlock(&THREAD->lock, false);
157 irq_spinlock_unlock(&CPU->lock, false);
158 THREAD->saved_fpu_context =
159 (fpu_context_t *) slab_alloc(fpu_context_slab, 0);
160
161 /* We may have switched CPUs during slab_alloc */
162 goto restart;
163 }
164 fpu_init();
165 THREAD->fpu_context_exists = true;
166 }
167
168 CPU->fpu_owner = THREAD;
169 THREAD->fpu_context_engaged = true;
170 irq_spinlock_unlock(&THREAD->lock, false);
171
172 irq_spinlock_unlock(&CPU->lock, false);
173}
174#endif /* CONFIG_FPU_LAZY */
175
176/** Initialize scheduler
177 *
178 * Initialize kernel scheduler.
179 *
180 */
181void scheduler_init(void)
182{
183}
184
185/** Get thread to be scheduled
186 *
187 * Get the optimal thread to be scheduled
188 * according to thread accounting and scheduler
189 * policy.
190 *
191 * @return Thread to be scheduled.
192 *
193 */
194static thread_t *find_best_thread(void)
195{
196 ASSERT(CPU != NULL);
197
198loop:
199
200 if (atomic_get(&CPU->nrdy) == 0) {
201 /*
202 * For there was nothing to run, the CPU goes to sleep
203 * until a hardware interrupt or an IPI comes.
204 * This improves energy saving and hyperthreading.
205 */
206 irq_spinlock_lock(&CPU->lock, false);
207 CPU->idle = true;
208 irq_spinlock_unlock(&CPU->lock, false);
209 interrupts_enable();
210
211 /*
212 * An interrupt might occur right now and wake up a thread.
213 * In such case, the CPU will continue to go to sleep
214 * even though there is a runnable thread.
215 */
216 cpu_sleep();
217 interrupts_disable();
218 goto loop;
219 }
220
221 unsigned int i;
222 for (i = 0; i < RQ_COUNT; i++) {
223 irq_spinlock_lock(&(CPU->rq[i].lock), false);
224 if (CPU->rq[i].n == 0) {
225 /*
226 * If this queue is empty, try a lower-priority queue.
227 */
228 irq_spinlock_unlock(&(CPU->rq[i].lock), false);
229 continue;
230 }
231
232 atomic_dec(&CPU->nrdy);
233 atomic_dec(&nrdy);
234 CPU->rq[i].n--;
235
236 /*
237 * Take the first thread from the queue.
238 */
239 thread_t *thread = list_get_instance(
240 list_first(&CPU->rq[i].rq), thread_t, rq_link);
241 list_remove(&thread->rq_link);
242
243 irq_spinlock_pass(&(CPU->rq[i].lock), &thread->lock);
244
245 thread->cpu = CPU;
246 thread->ticks = us2ticks((i + 1) * 10000);
247 thread->priority = i; /* Correct rq index */
248
249 /*
250 * Clear the stolen flag so that it can be migrated
251 * when load balancing needs emerge.
252 */
253 thread->stolen = false;
254 irq_spinlock_unlock(&thread->lock, false);
255
256 return thread;
257 }
258
259 goto loop;
260}
261
262/** Prevent rq starvation
263 *
264 * Prevent low priority threads from starving in rq's.
265 *
266 * When the function decides to relink rq's, it reconnects
267 * respective pointers so that in result threads with 'pri'
268 * greater or equal start are moved to a higher-priority queue.
269 *
270 * @param start Threshold priority.
271 *
272 */
273static void relink_rq(int start)
274{
275 list_t list;
276
277 list_initialize(&list);
278 irq_spinlock_lock(&CPU->lock, false);
279
280 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
281 int i;
282 for (i = start; i < RQ_COUNT - 1; i++) {
283 /* Remember and empty rq[i + 1] */
284
285 irq_spinlock_lock(&CPU->rq[i + 1].lock, false);
286 list_concat(&list, &CPU->rq[i + 1].rq);
287 size_t n = CPU->rq[i + 1].n;
288 CPU->rq[i + 1].n = 0;
289 irq_spinlock_unlock(&CPU->rq[i + 1].lock, false);
290
291 /* Append rq[i + 1] to rq[i] */
292
293 irq_spinlock_lock(&CPU->rq[i].lock, false);
294 list_concat(&CPU->rq[i].rq, &list);
295 CPU->rq[i].n += n;
296 irq_spinlock_unlock(&CPU->rq[i].lock, false);
297 }
298
299 CPU->needs_relink = 0;
300 }
301
302 irq_spinlock_unlock(&CPU->lock, false);
303}
304
305/** The scheduler
306 *
307 * The thread scheduling procedure.
308 * Passes control directly to
309 * scheduler_separated_stack().
310 *
311 */
312void scheduler(void)
313{
314 volatile ipl_t ipl;
315
316 ASSERT(CPU != NULL);
317
318 ipl = interrupts_disable();
319
320 if (atomic_get(&haltstate))
321 halt();
322
323 if (THREAD) {
324 irq_spinlock_lock(&THREAD->lock, false);
325
326 /* Update thread kernel accounting */
327 THREAD->kcycles += get_cycle() - THREAD->last_cycle;
328
329#if (defined CONFIG_FPU) && (!defined CONFIG_FPU_LAZY)
330 fpu_context_save(THREAD->saved_fpu_context);
331#endif
332 if (!context_save(&THREAD->saved_context)) {
333 /*
334 * This is the place where threads leave scheduler();
335 */
336
337 /* Save current CPU cycle */
338 THREAD->last_cycle = get_cycle();
339
340 irq_spinlock_unlock(&THREAD->lock, false);
341 interrupts_restore(THREAD->saved_context.ipl);
342
343 return;
344 }
345
346 /*
347 * Interrupt priority level of preempted thread is recorded
348 * here to facilitate scheduler() invocations from
349 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
350 *
351 */
352 THREAD->saved_context.ipl = ipl;
353 }
354
355 /*
356 * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, AS
357 * and preemption counter. At this point THE could be coming either
358 * from THREAD's or CPU's stack.
359 *
360 */
361 the_copy(THE, (the_t *) CPU->stack);
362
363 /*
364 * We may not keep the old stack.
365 * Reason: If we kept the old stack and got blocked, for instance, in
366 * find_best_thread(), the old thread could get rescheduled by another
367 * CPU and overwrite the part of its own stack that was also used by
368 * the scheduler on this CPU.
369 *
370 * Moreover, we have to bypass the compiler-generated POP sequence
371 * which is fooled by SP being set to the very top of the stack.
372 * Therefore the scheduler() function continues in
373 * scheduler_separated_stack().
374 *
375 */
376 context_save(&CPU->saved_context);
377 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack),
378 (uintptr_t) CPU->stack, STACK_SIZE);
379 context_restore(&CPU->saved_context);
380
381 /* Not reached */
382}
383
384/** Scheduler stack switch wrapper
385 *
386 * Second part of the scheduler() function
387 * using new stack. Handling the actual context
388 * switch to a new thread.
389 *
390 */
391void scheduler_separated_stack(void)
392{
393 DEADLOCK_PROBE_INIT(p_joinwq);
394 task_t *old_task = TASK;
395 as_t *old_as = AS;
396
397 ASSERT((!THREAD) || (irq_spinlock_locked(&THREAD->lock)));
398 ASSERT(CPU != NULL);
399
400 /*
401 * Hold the current task and the address space to prevent their
402 * possible destruction should thread_destroy() be called on this or any
403 * other processor while the scheduler is still using them.
404 */
405 if (old_task)
406 task_hold(old_task);
407
408 if (old_as)
409 as_hold(old_as);
410
411 if (THREAD) {
412 /* Must be run after the switch to scheduler stack */
413 after_thread_ran();
414
415 switch (THREAD->state) {
416 case Running:
417 irq_spinlock_unlock(&THREAD->lock, false);
418 thread_ready(THREAD);
419 break;
420
421 case Exiting:
422repeat:
423 if (THREAD->detached) {
424 thread_destroy(THREAD, false);
425 } else {
426 /*
427 * The thread structure is kept allocated until
428 * somebody calls thread_detach() on it.
429 */
430 if (!irq_spinlock_trylock(&THREAD->join_wq.lock)) {
431 /*
432 * Avoid deadlock.
433 */
434 irq_spinlock_unlock(&THREAD->lock, false);
435 delay(HZ);
436 irq_spinlock_lock(&THREAD->lock, false);
437 DEADLOCK_PROBE(p_joinwq,
438 DEADLOCK_THRESHOLD);
439 goto repeat;
440 }
441 _waitq_wakeup_unsafe(&THREAD->join_wq,
442 WAKEUP_FIRST);
443 irq_spinlock_unlock(&THREAD->join_wq.lock, false);
444
445 THREAD->state = Lingering;
446 irq_spinlock_unlock(&THREAD->lock, false);
447 }
448 break;
449
450 case Sleeping:
451 /*
452 * Prefer the thread after it's woken up.
453 */
454 THREAD->priority = -1;
455
456 /*
457 * We need to release wq->lock which we locked in
458 * waitq_sleep(). Address of wq->lock is kept in
459 * THREAD->sleep_queue.
460 */
461 irq_spinlock_unlock(&THREAD->sleep_queue->lock, false);
462
463 irq_spinlock_unlock(&THREAD->lock, false);
464 break;
465
466 default:
467 /*
468 * Entering state is unexpected.
469 */
470 panic("tid%" PRIu64 ": unexpected state %s.",
471 THREAD->tid, thread_states[THREAD->state]);
472 break;
473 }
474
475 THREAD = NULL;
476 }
477
478 THREAD = find_best_thread();
479
480 irq_spinlock_lock(&THREAD->lock, false);
481 int priority = THREAD->priority;
482 irq_spinlock_unlock(&THREAD->lock, false);
483
484 relink_rq(priority);
485
486 /*
487 * If both the old and the new task are the same,
488 * lots of work is avoided.
489 */
490 if (TASK != THREAD->task) {
491 as_t *new_as = THREAD->task->as;
492
493 /*
494 * Note that it is possible for two tasks
495 * to share one address space.
496 */
497 if (old_as != new_as) {
498 /*
499 * Both tasks and address spaces are different.
500 * Replace the old one with the new one.
501 */
502 as_switch(old_as, new_as);
503 }
504
505 TASK = THREAD->task;
506 before_task_runs();
507 }
508
509 if (old_task)
510 task_release(old_task);
511
512 if (old_as)
513 as_release(old_as);
514
515 irq_spinlock_lock(&THREAD->lock, false);
516 THREAD->state = Running;
517
518#ifdef SCHEDULER_VERBOSE
519 printf("cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
520 ", nrdy=%" PRIua ")\n", CPU->id, THREAD->tid, THREAD->priority,
521 THREAD->ticks, atomic_get(&CPU->nrdy));
522#endif
523
524 /*
525 * Some architectures provide late kernel PA2KA(identity)
526 * mapping in a page fault handler. However, the page fault
527 * handler uses the kernel stack of the running thread and
528 * therefore cannot be used to map it. The kernel stack, if
529 * necessary, is to be mapped in before_thread_runs(). This
530 * function must be executed before the switch to the new stack.
531 */
532 before_thread_runs();
533
534 /*
535 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
536 * thread's stack.
537 */
538 the_copy(THE, (the_t *) THREAD->kstack);
539
540 context_restore(&THREAD->saved_context);
541
542 /* Not reached */
543}
544
545#ifdef CONFIG_SMP
546/** Load balancing thread
547 *
548 * SMP load balancing thread, supervising thread supplies
549 * for the CPU it's wired to.
550 *
551 * @param arg Generic thread argument (unused).
552 *
553 */
554void kcpulb(void *arg)
555{
556 atomic_count_t average;
557 atomic_count_t rdy;
558
559 /*
560 * Detach kcpulb as nobody will call thread_join_timeout() on it.
561 */
562 thread_detach(THREAD);
563
564loop:
565 /*
566 * Work in 1s intervals.
567 */
568 thread_sleep(1);
569
570not_satisfied:
571 /*
572 * Calculate the number of threads that will be migrated/stolen from
573 * other CPU's. Note that situation can have changed between two
574 * passes. Each time get the most up to date counts.
575 *
576 */
577 average = atomic_get(&nrdy) / config.cpu_active + 1;
578 rdy = atomic_get(&CPU->nrdy);
579
580 if (average <= rdy)
581 goto satisfied;
582
583 atomic_count_t count = average - rdy;
584
585 /*
586 * Searching least priority queues on all CPU's first and most priority
587 * queues on all CPU's last.
588 */
589 size_t acpu;
590 size_t acpu_bias = 0;
591 int rq;
592
593 for (rq = RQ_COUNT - 1; rq >= 0; rq--) {
594 for (acpu = 0; acpu < config.cpu_active; acpu++) {
595 cpu_t *cpu = &cpus[(acpu + acpu_bias) % config.cpu_active];
596
597 /*
598 * Not interested in ourselves.
599 * Doesn't require interrupt disabling for kcpulb has
600 * THREAD_FLAG_WIRED.
601 *
602 */
603 if (CPU == cpu)
604 continue;
605
606 if (atomic_get(&cpu->nrdy) <= average)
607 continue;
608
609 irq_spinlock_lock(&(cpu->rq[rq].lock), true);
610 if (cpu->rq[rq].n == 0) {
611 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
612 continue;
613 }
614
615 thread_t *thread = NULL;
616
617 /* Search rq from the back */
618 link_t *link = cpu->rq[rq].rq.head.prev;
619
620 while (link != &(cpu->rq[rq].rq.head)) {
621 thread = (thread_t *) list_get_instance(link,
622 thread_t, rq_link);
623
624 /*
625 * Do not steal CPU-wired threads, threads
626 * already stolen, threads for which migration
627 * was temporarily disabled or threads whose
628 * FPU context is still in the CPU.
629 */
630 irq_spinlock_lock(&thread->lock, false);
631
632 if ((!thread->wired) && (!thread->stolen) &&
633 (!thread->nomigrate) &&
634 (!thread->fpu_context_engaged)) {
635 /*
636 * Remove thread from ready queue.
637 */
638 irq_spinlock_unlock(&thread->lock,
639 false);
640
641 atomic_dec(&cpu->nrdy);
642 atomic_dec(&nrdy);
643
644 cpu->rq[rq].n--;
645 list_remove(&thread->rq_link);
646
647 break;
648 }
649
650 irq_spinlock_unlock(&thread->lock, false);
651
652 link = link->prev;
653 thread = NULL;
654 }
655
656 if (thread) {
657 /*
658 * Ready thread on local CPU
659 */
660
661 irq_spinlock_pass(&(cpu->rq[rq].lock),
662 &thread->lock);
663
664#ifdef KCPULB_VERBOSE
665 printf("kcpulb%u: TID %" PRIu64 " -> cpu%u, "
666 "nrdy=%ld, avg=%ld\n", CPU->id, t->tid,
667 CPU->id, atomic_get(&CPU->nrdy),
668 atomic_get(&nrdy) / config.cpu_active);
669#endif
670
671 thread->stolen = true;
672 thread->state = Entering;
673
674 irq_spinlock_unlock(&thread->lock, true);
675 thread_ready(thread);
676
677 if (--count == 0)
678 goto satisfied;
679
680 /*
681 * We are not satisfied yet, focus on another
682 * CPU next time.
683 *
684 */
685 acpu_bias++;
686
687 continue;
688 } else
689 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
690
691 }
692 }
693
694 if (atomic_get(&CPU->nrdy)) {
695 /*
696 * Be a little bit light-weight and let migrated threads run.
697 *
698 */
699 scheduler();
700 } else {
701 /*
702 * We failed to migrate a single thread.
703 * Give up this turn.
704 *
705 */
706 goto loop;
707 }
708
709 goto not_satisfied;
710
711satisfied:
712 goto loop;
713}
714#endif /* CONFIG_SMP */
715
716/** Print information about threads & scheduler queues
717 *
718 */
719void sched_print_list(void)
720{
721 size_t cpu;
722 for (cpu = 0; cpu < config.cpu_count; cpu++) {
723 if (!cpus[cpu].active)
724 continue;
725
726 irq_spinlock_lock(&cpus[cpu].lock, true);
727
728 printf("cpu%u: address=%p, nrdy=%" PRIua ", needs_relink=%zu\n",
729 cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy),
730 cpus[cpu].needs_relink);
731
732 unsigned int i;
733 for (i = 0; i < RQ_COUNT; i++) {
734 irq_spinlock_lock(&(cpus[cpu].rq[i].lock), false);
735 if (cpus[cpu].rq[i].n == 0) {
736 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
737 continue;
738 }
739
740 printf("\trq[%u]: ", i);
741 list_foreach(cpus[cpu].rq[i].rq, rq_link, thread_t,
742 thread) {
743 printf("%" PRIu64 "(%s) ", thread->tid,
744 thread_states[thread->state]);
745 }
746 printf("\n");
747
748 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
749 }
750
751 irq_spinlock_unlock(&cpus[cpu].lock, true);
752 }
753}
754
755/** @}
756 */
Note: See TracBrowser for help on using the repository browser.