source: mainline/kernel/generic/src/proc/scheduler.c@ 663bb537

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since 663bb537 was 37c9fc8, checked in by Jakub Jermar <jakub@…>, 15 years ago

Cherry pick revision 404 from lp:~ersin/helenos/measure2.
This fixes the case when the CPU lock was taken while interrupts were enabled.
  • Property mode set to 100644
File size: 16.6 KB
Line 
1/*
2 * Copyright (c) 2001-2007 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
65static void before_task_runs(void);
66static void before_thread_runs(void);
67static void after_thread_ran(void);
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. */
73void 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 * tread is passed control.
83 *
84 * THREAD->lock is locked on entry
85 *
86 */
87void before_thread_runs(void)
88{
89 before_thread_runs_arch();
90#ifdef CONFIG_FPU_LAZY
91 if(THREAD == CPU->fpu_owner)
92 fpu_enable();
93 else
94 fpu_disable();
95#else
96 fpu_enable();
97 if (THREAD->fpu_context_exists)
98 fpu_context_restore(THREAD->saved_fpu_context);
99 else {
100 fpu_init();
101 THREAD->fpu_context_exists = 1;
102 }
103#endif
104}
105
106/** Take actions after THREAD had run.
107 *
108 * Perform actions that need to be
109 * taken after the running thread
110 * had been preempted by the scheduler.
111 *
112 * THREAD->lock is locked on entry
113 *
114 */
115void after_thread_ran(void)
116{
117 after_thread_ran_arch();
118}
119
120#ifdef CONFIG_FPU_LAZY
121void scheduler_fpu_lazy_request(void)
122{
123restart:
124 fpu_enable();
125 spinlock_lock(&CPU->lock);
126
127 /* Save old context */
128 if (CPU->fpu_owner != NULL) {
129 spinlock_lock(&CPU->fpu_owner->lock);
130 fpu_context_save(CPU->fpu_owner->saved_fpu_context);
131 /* don't prevent migration */
132 CPU->fpu_owner->fpu_context_engaged = 0;
133 spinlock_unlock(&CPU->fpu_owner->lock);
134 CPU->fpu_owner = NULL;
135 }
136
137 spinlock_lock(&THREAD->lock);
138 if (THREAD->fpu_context_exists) {
139 fpu_context_restore(THREAD->saved_fpu_context);
140 } else {
141 /* Allocate FPU context */
142 if (!THREAD->saved_fpu_context) {
143 /* Might sleep */
144 spinlock_unlock(&THREAD->lock);
145 spinlock_unlock(&CPU->lock);
146 THREAD->saved_fpu_context =
147 (fpu_context_t *) slab_alloc(fpu_context_slab, 0);
148 /* We may have switched CPUs during slab_alloc */
149 goto restart;
150 }
151 fpu_init();
152 THREAD->fpu_context_exists = 1;
153 }
154 CPU->fpu_owner = THREAD;
155 THREAD->fpu_context_engaged = 1;
156 spinlock_unlock(&THREAD->lock);
157
158 spinlock_unlock(&CPU->lock);
159}
160#endif
161
162/** Initialize scheduler
163 *
164 * Initialize kernel scheduler.
165 *
166 */
167void scheduler_init(void)
168{
169}
170
171/** Get thread to be scheduled
172 *
173 * Get the optimal thread to be scheduled
174 * according to thread accounting and scheduler
175 * policy.
176 *
177 * @return Thread to be scheduled.
178 *
179 */
180static thread_t *find_best_thread(void)
181{
182 thread_t *t;
183 runq_t *r;
184 int i;
185
186 ASSERT(CPU != NULL);
187
188loop:
189
190 if (atomic_get(&CPU->nrdy) == 0) {
191 /*
192 * For there was nothing to run, the CPU goes to sleep
193 * until a hardware interrupt or an IPI comes.
194 * This improves energy saving and hyperthreading.
195 */
196
197 /* Mark CPU as it was idle this clock tick */
198 spinlock_lock(&CPU->lock);
199 CPU->idle = true;
200 spinlock_unlock(&CPU->lock);
201
202 interrupts_enable();
203 /*
204 * An interrupt might occur right now and wake up a thread.
205 * In such case, the CPU will continue to go to sleep
206 * even though there is a runnable thread.
207 */
208 cpu_sleep();
209 interrupts_disable();
210 goto loop;
211 }
212
213 for (i = 0; i < RQ_COUNT; i++) {
214 r = &CPU->rq[i];
215 spinlock_lock(&r->lock);
216 if (r->n == 0) {
217 /*
218 * If this queue is empty, try a lower-priority queue.
219 */
220 spinlock_unlock(&r->lock);
221 continue;
222 }
223
224 atomic_dec(&CPU->nrdy);
225 atomic_dec(&nrdy);
226 r->n--;
227
228 /*
229 * Take the first thread from the queue.
230 */
231 t = list_get_instance(r->rq_head.next, thread_t, rq_link);
232 list_remove(&t->rq_link);
233
234 spinlock_unlock(&r->lock);
235
236 spinlock_lock(&t->lock);
237 t->cpu = CPU;
238
239 t->ticks = us2ticks((i + 1) * 10000);
240 t->priority = i; /* correct rq index */
241
242 /*
243 * Clear the THREAD_FLAG_STOLEN flag so that t can be migrated
244 * when load balancing needs emerge.
245 */
246 t->flags &= ~THREAD_FLAG_STOLEN;
247 spinlock_unlock(&t->lock);
248
249 return t;
250 }
251 goto loop;
252
253}
254
255/** Prevent rq starvation
256 *
257 * Prevent low priority threads from starving in rq's.
258 *
259 * When the function decides to relink rq's, it reconnects
260 * respective pointers so that in result threads with 'pri'
261 * greater or equal start are moved to a higher-priority queue.
262 *
263 * @param start Threshold priority.
264 *
265 */
266static void relink_rq(int start)
267{
268 link_t head;
269 runq_t *r;
270 int i, n;
271
272 list_initialize(&head);
273 spinlock_lock(&CPU->lock);
274 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
275 for (i = start; i < RQ_COUNT - 1; i++) {
276 /* remember and empty rq[i + 1] */
277 r = &CPU->rq[i + 1];
278 spinlock_lock(&r->lock);
279 list_concat(&head, &r->rq_head);
280 n = r->n;
281 r->n = 0;
282 spinlock_unlock(&r->lock);
283
284 /* append rq[i + 1] to rq[i] */
285 r = &CPU->rq[i];
286 spinlock_lock(&r->lock);
287 list_concat(&r->rq_head, &head);
288 r->n += n;
289 spinlock_unlock(&r->lock);
290 }
291 CPU->needs_relink = 0;
292 }
293 spinlock_unlock(&CPU->lock);
294
295}
296
297/** The scheduler
298 *
299 * The thread scheduling procedure.
300 * Passes control directly to
301 * scheduler_separated_stack().
302 *
303 */
304void scheduler(void)
305{
306 volatile ipl_t ipl;
307
308 ASSERT(CPU != NULL);
309
310 ipl = interrupts_disable();
311
312 if (atomic_get(&haltstate))
313 halt();
314
315 if (THREAD) {
316 spinlock_lock(&THREAD->lock);
317
318 /* Update thread kernel accounting */
319 THREAD->kcycles += get_cycle() - THREAD->last_cycle;
320
321#ifndef CONFIG_FPU_LAZY
322 fpu_context_save(THREAD->saved_fpu_context);
323#endif
324 if (!context_save(&THREAD->saved_context)) {
325 /*
326 * This is the place where threads leave scheduler();
327 */
328
329 /* Save current CPU cycle */
330 THREAD->last_cycle = get_cycle();
331
332 spinlock_unlock(&THREAD->lock);
333 interrupts_restore(THREAD->saved_context.ipl);
334
335 return;
336 }
337
338 /*
339 * Interrupt priority level of preempted thread is recorded
340 * here to facilitate scheduler() invocations from
341 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
342 */
343 THREAD->saved_context.ipl = ipl;
344 }
345
346 /*
347 * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, VM
348 * and preemption counter. At this point THE could be coming either
349 * from THREAD's or CPU's stack.
350 */
351 the_copy(THE, (the_t *) CPU->stack);
352
353 /*
354 * We may not keep the old stack.
355 * Reason: If we kept the old stack and got blocked, for instance, in
356 * find_best_thread(), the old thread could get rescheduled by another
357 * CPU and overwrite the part of its own stack that was also used by
358 * the scheduler on this CPU.
359 *
360 * Moreover, we have to bypass the compiler-generated POP sequence
361 * which is fooled by SP being set to the very top of the stack.
362 * Therefore the scheduler() function continues in
363 * scheduler_separated_stack().
364 */
365 context_save(&CPU->saved_context);
366 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack),
367 (uintptr_t) CPU->stack, CPU_STACK_SIZE);
368 context_restore(&CPU->saved_context);
369 /* not reached */
370}
371
372/** Scheduler stack switch wrapper
373 *
374 * Second part of the scheduler() function
375 * using new stack. Handling the actual context
376 * switch to a new thread.
377 *
378 * Assume THREAD->lock is held.
379 */
380void scheduler_separated_stack(void)
381{
382 int priority;
383 DEADLOCK_PROBE_INIT(p_joinwq);
384
385 ASSERT(CPU != NULL);
386
387 if (THREAD) {
388 /* must be run after the switch to scheduler stack */
389 after_thread_ran();
390
391 switch (THREAD->state) {
392 case Running:
393 spinlock_unlock(&THREAD->lock);
394 thread_ready(THREAD);
395 break;
396
397 case Exiting:
398repeat:
399 if (THREAD->detached) {
400 thread_destroy(THREAD);
401 } else {
402 /*
403 * The thread structure is kept allocated until
404 * somebody calls thread_detach() on it.
405 */
406 if (!spinlock_trylock(&THREAD->join_wq.lock)) {
407 /*
408 * Avoid deadlock.
409 */
410 spinlock_unlock(&THREAD->lock);
411 delay(HZ);
412 spinlock_lock(&THREAD->lock);
413 DEADLOCK_PROBE(p_joinwq,
414 DEADLOCK_THRESHOLD);
415 goto repeat;
416 }
417 _waitq_wakeup_unsafe(&THREAD->join_wq,
418 WAKEUP_FIRST);
419 spinlock_unlock(&THREAD->join_wq.lock);
420
421 THREAD->state = Lingering;
422 spinlock_unlock(&THREAD->lock);
423 }
424 break;
425
426 case Sleeping:
427 /*
428 * Prefer the thread after it's woken up.
429 */
430 THREAD->priority = -1;
431
432 /*
433 * We need to release wq->lock which we locked in
434 * waitq_sleep(). Address of wq->lock is kept in
435 * THREAD->sleep_queue.
436 */
437 spinlock_unlock(&THREAD->sleep_queue->lock);
438
439 /*
440 * Check for possible requests for out-of-context
441 * invocation.
442 */
443 if (THREAD->call_me) {
444 THREAD->call_me(THREAD->call_me_with);
445 THREAD->call_me = NULL;
446 THREAD->call_me_with = NULL;
447 }
448
449 spinlock_unlock(&THREAD->lock);
450
451 break;
452
453 default:
454 /*
455 * Entering state is unexpected.
456 */
457 panic("tid%" PRIu64 ": unexpected state %s.",
458 THREAD->tid, thread_states[THREAD->state]);
459 break;
460 }
461
462 THREAD = NULL;
463 }
464
465 THREAD = find_best_thread();
466
467 spinlock_lock(&THREAD->lock);
468 priority = THREAD->priority;
469 spinlock_unlock(&THREAD->lock);
470
471 relink_rq(priority);
472
473 /*
474 * If both the old and the new task are the same, lots of work is
475 * avoided.
476 */
477 if (TASK != THREAD->task) {
478 as_t *as1 = NULL;
479 as_t *as2;
480
481 if (TASK) {
482 spinlock_lock(&TASK->lock);
483 as1 = TASK->as;
484 spinlock_unlock(&TASK->lock);
485 }
486
487 spinlock_lock(&THREAD->task->lock);
488 as2 = THREAD->task->as;
489 spinlock_unlock(&THREAD->task->lock);
490
491 /*
492 * Note that it is possible for two tasks to share one address
493 * space.
494 */
495 if (as1 != as2) {
496 /*
497 * Both tasks and address spaces are different.
498 * Replace the old one with the new one.
499 */
500 as_switch(as1, as2);
501 }
502 TASK = THREAD->task;
503 before_task_runs();
504 }
505
506 spinlock_lock(&THREAD->lock);
507 THREAD->state = Running;
508
509#ifdef SCHEDULER_VERBOSE
510 printf("cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
511 ", nrdy=%ld)\n", CPU->id, THREAD->tid, THREAD->priority,
512 THREAD->ticks, atomic_get(&CPU->nrdy));
513#endif
514
515 /*
516 * Some architectures provide late kernel PA2KA(identity)
517 * mapping in a page fault handler. However, the page fault
518 * handler uses the kernel stack of the running thread and
519 * therefore cannot be used to map it. The kernel stack, if
520 * necessary, is to be mapped in before_thread_runs(). This
521 * function must be executed before the switch to the new stack.
522 */
523 before_thread_runs();
524
525 /*
526 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
527 * thread's stack.
528 */
529 the_copy(THE, (the_t *) THREAD->kstack);
530
531 context_restore(&THREAD->saved_context);
532 /* not reached */
533}
534
535#ifdef CONFIG_SMP
536/** Load balancing thread
537 *
538 * SMP load balancing thread, supervising thread supplies
539 * for the CPU it's wired to.
540 *
541 * @param arg Generic thread argument (unused).
542 *
543 */
544void kcpulb(void *arg)
545{
546 thread_t *t;
547 int count;
548 atomic_count_t average;
549 unsigned int i;
550 int j;
551 int k = 0;
552 ipl_t ipl;
553
554 /*
555 * Detach kcpulb as nobody will call thread_join_timeout() on it.
556 */
557 thread_detach(THREAD);
558
559loop:
560 /*
561 * Work in 1s intervals.
562 */
563 thread_sleep(1);
564
565not_satisfied:
566 /*
567 * Calculate the number of threads that will be migrated/stolen from
568 * other CPU's. Note that situation can have changed between two
569 * passes. Each time get the most up to date counts.
570 */
571 average = atomic_get(&nrdy) / config.cpu_active + 1;
572 count = average - atomic_get(&CPU->nrdy);
573
574 if (count <= 0)
575 goto satisfied;
576
577 /*
578 * Searching least priority queues on all CPU's first and most priority
579 * queues on all CPU's last.
580 */
581 for (j = RQ_COUNT - 1; j >= 0; j--) {
582 for (i = 0; i < config.cpu_active; i++) {
583 link_t *l;
584 runq_t *r;
585 cpu_t *cpu;
586
587 cpu = &cpus[(i + k) % config.cpu_active];
588
589 /*
590 * Not interested in ourselves.
591 * Doesn't require interrupt disabling for kcpulb has
592 * THREAD_FLAG_WIRED.
593 */
594 if (CPU == cpu)
595 continue;
596 if (atomic_get(&cpu->nrdy) <= average)
597 continue;
598
599 ipl = interrupts_disable();
600 r = &cpu->rq[j];
601 spinlock_lock(&r->lock);
602 if (r->n == 0) {
603 spinlock_unlock(&r->lock);
604 interrupts_restore(ipl);
605 continue;
606 }
607
608 t = NULL;
609 l = r->rq_head.prev; /* search rq from the back */
610 while (l != &r->rq_head) {
611 t = list_get_instance(l, thread_t, rq_link);
612 /*
613 * We don't want to steal CPU-wired threads
614 * neither threads already stolen. The latter
615 * prevents threads from migrating between CPU's
616 * without ever being run. We don't want to
617 * steal threads whose FPU context is still in
618 * CPU.
619 */
620 spinlock_lock(&t->lock);
621 if ((!(t->flags & (THREAD_FLAG_WIRED |
622 THREAD_FLAG_STOLEN))) &&
623 (!(t->fpu_context_engaged))) {
624 /*
625 * Remove t from r.
626 */
627 spinlock_unlock(&t->lock);
628
629 atomic_dec(&cpu->nrdy);
630 atomic_dec(&nrdy);
631
632 r->n--;
633 list_remove(&t->rq_link);
634
635 break;
636 }
637 spinlock_unlock(&t->lock);
638 l = l->prev;
639 t = NULL;
640 }
641 spinlock_unlock(&r->lock);
642
643 if (t) {
644 /*
645 * Ready t on local CPU
646 */
647 spinlock_lock(&t->lock);
648#ifdef KCPULB_VERBOSE
649 printf("kcpulb%u: TID %" PRIu64 " -> cpu%u, "
650 "nrdy=%ld, avg=%ld\n", CPU->id, t->tid,
651 CPU->id, atomic_get(&CPU->nrdy),
652 atomic_get(&nrdy) / config.cpu_active);
653#endif
654 t->flags |= THREAD_FLAG_STOLEN;
655 t->state = Entering;
656 spinlock_unlock(&t->lock);
657
658 thread_ready(t);
659
660 interrupts_restore(ipl);
661
662 if (--count == 0)
663 goto satisfied;
664
665 /*
666 * We are not satisfied yet, focus on another
667 * CPU next time.
668 */
669 k++;
670
671 continue;
672 }
673 interrupts_restore(ipl);
674 }
675 }
676
677 if (atomic_get(&CPU->nrdy)) {
678 /*
679 * Be a little bit light-weight and let migrated threads run.
680 */
681 scheduler();
682 } else {
683 /*
684 * We failed to migrate a single thread.
685 * Give up this turn.
686 */
687 goto loop;
688 }
689
690 goto not_satisfied;
691
692satisfied:
693 goto loop;
694}
695
696#endif /* CONFIG_SMP */
697
698
699/** Print information about threads & scheduler queues */
700void sched_print_list(void)
701{
702 ipl_t ipl;
703 unsigned int cpu, i;
704 runq_t *r;
705 thread_t *t;
706 link_t *cur;
707
708 /* We are going to mess with scheduler structures,
709 * let's not be interrupted */
710 ipl = interrupts_disable();
711 for (cpu = 0; cpu < config.cpu_count; cpu++) {
712
713 if (!cpus[cpu].active)
714 continue;
715
716 spinlock_lock(&cpus[cpu].lock);
717 printf("cpu%u: address=%p, nrdy=%ld, needs_relink=%" PRIs "\n",
718 cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy),
719 cpus[cpu].needs_relink);
720
721 for (i = 0; i < RQ_COUNT; i++) {
722 r = &cpus[cpu].rq[i];
723 spinlock_lock(&r->lock);
724 if (!r->n) {
725 spinlock_unlock(&r->lock);
726 continue;
727 }
728 printf("\trq[%u]: ", i);
729 for (cur = r->rq_head.next; cur != &r->rq_head;
730 cur = cur->next) {
731 t = list_get_instance(cur, thread_t, rq_link);
732 printf("%" PRIu64 "(%s) ", t->tid,
733 thread_states[t->state]);
734 }
735 printf("\n");
736 spinlock_unlock(&r->lock);
737 }
738 spinlock_unlock(&cpus[cpu].lock);
739 }
740
741 interrupts_restore(ipl);
742}
743
744/** @}
745 */
Note: See TracBrowser for help on using the repository browser.