source: mainline/kernel/generic/src/proc/scheduler.c@ a2a00e8

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since a2a00e8 was a2a00e8, checked in by Stanislav Kozina <stanislav.kozina@…>, 15 years ago

Accounting separated to kernel and user time.
  • Property mode set to 100644
File size: 16.5 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 interrupts_enable();
190
191 if (atomic_get(&CPU->nrdy) == 0) {
192 /*
193 * For there was nothing to run, the CPU goes to sleep
194 * until a hardware interrupt or an IPI comes.
195 * This improves energy saving and hyperthreading.
196 */
197
198 /*
199 * An interrupt might occur right now and wake up a thread.
200 * In such case, the CPU will continue to go to sleep
201 * even though there is a runnable thread.
202 */
203
204 cpu_sleep();
205 goto loop;
206 }
207
208 interrupts_disable();
209
210 for (i = 0; i < RQ_COUNT; i++) {
211 r = &CPU->rq[i];
212 spinlock_lock(&r->lock);
213 if (r->n == 0) {
214 /*
215 * If this queue is empty, try a lower-priority queue.
216 */
217 spinlock_unlock(&r->lock);
218 continue;
219 }
220
221 atomic_dec(&CPU->nrdy);
222 atomic_dec(&nrdy);
223 r->n--;
224
225 /*
226 * Take the first thread from the queue.
227 */
228 t = list_get_instance(r->rq_head.next, thread_t, rq_link);
229 list_remove(&t->rq_link);
230
231 spinlock_unlock(&r->lock);
232
233 spinlock_lock(&t->lock);
234 t->cpu = CPU;
235
236 t->ticks = us2ticks((i + 1) * 10000);
237 t->priority = i; /* correct rq index */
238
239 /*
240 * Clear the THREAD_FLAG_STOLEN flag so that t can be migrated
241 * when load balancing needs emerge.
242 */
243 t->flags &= ~THREAD_FLAG_STOLEN;
244 spinlock_unlock(&t->lock);
245
246 return t;
247 }
248 goto loop;
249
250}
251
252/** Prevent rq starvation
253 *
254 * Prevent low priority threads from starving in rq's.
255 *
256 * When the function decides to relink rq's, it reconnects
257 * respective pointers so that in result threads with 'pri'
258 * greater or equal start are moved to a higher-priority queue.
259 *
260 * @param start Threshold priority.
261 *
262 */
263static void relink_rq(int start)
264{
265 link_t head;
266 runq_t *r;
267 int i, n;
268
269 list_initialize(&head);
270 spinlock_lock(&CPU->lock);
271 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
272 for (i = start; i < RQ_COUNT - 1; i++) {
273 /* remember and empty rq[i + 1] */
274 r = &CPU->rq[i + 1];
275 spinlock_lock(&r->lock);
276 list_concat(&head, &r->rq_head);
277 n = r->n;
278 r->n = 0;
279 spinlock_unlock(&r->lock);
280
281 /* append rq[i + 1] to rq[i] */
282 r = &CPU->rq[i];
283 spinlock_lock(&r->lock);
284 list_concat(&r->rq_head, &head);
285 r->n += n;
286 spinlock_unlock(&r->lock);
287 }
288 CPU->needs_relink = 0;
289 }
290 spinlock_unlock(&CPU->lock);
291
292}
293
294/** The scheduler
295 *
296 * The thread scheduling procedure.
297 * Passes control directly to
298 * scheduler_separated_stack().
299 *
300 */
301void scheduler(void)
302{
303 volatile ipl_t ipl;
304
305 ASSERT(CPU != NULL);
306
307 ipl = interrupts_disable();
308
309 if (atomic_get(&haltstate))
310 halt();
311
312 if (THREAD) {
313 spinlock_lock(&THREAD->lock);
314
315 /* Update thread accounting */
316 THREAD->cycles += get_cycle() - THREAD->last_cycle;
317 THREAD->kcycles += get_cycle() - THREAD->last_cycle;
318
319#ifndef CONFIG_FPU_LAZY
320 fpu_context_save(THREAD->saved_fpu_context);
321#endif
322 if (!context_save(&THREAD->saved_context)) {
323 /*
324 * This is the place where threads leave scheduler();
325 */
326
327 /* Save current CPU cycle */
328 THREAD->last_cycle = get_cycle();
329
330 spinlock_unlock(&THREAD->lock);
331 interrupts_restore(THREAD->saved_context.ipl);
332
333 return;
334 }
335
336 /*
337 * Interrupt priority level of preempted thread is recorded
338 * here to facilitate scheduler() invocations from
339 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
340 */
341 THREAD->saved_context.ipl = ipl;
342 }
343
344 /*
345 * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, VM
346 * and preemption counter. At this point THE could be coming either
347 * from THREAD's or CPU's stack.
348 */
349 the_copy(THE, (the_t *) CPU->stack);
350
351 /*
352 * We may not keep the old stack.
353 * Reason: If we kept the old stack and got blocked, for instance, in
354 * find_best_thread(), the old thread could get rescheduled by another
355 * CPU and overwrite the part of its own stack that was also used by
356 * the scheduler on this CPU.
357 *
358 * Moreover, we have to bypass the compiler-generated POP sequence
359 * which is fooled by SP being set to the very top of the stack.
360 * Therefore the scheduler() function continues in
361 * scheduler_separated_stack().
362 */
363 context_save(&CPU->saved_context);
364 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack),
365 (uintptr_t) CPU->stack, CPU_STACK_SIZE);
366 context_restore(&CPU->saved_context);
367 /* not reached */
368}
369
370/** Scheduler stack switch wrapper
371 *
372 * Second part of the scheduler() function
373 * using new stack. Handling the actual context
374 * switch to a new thread.
375 *
376 * Assume THREAD->lock is held.
377 */
378void scheduler_separated_stack(void)
379{
380 int priority;
381 DEADLOCK_PROBE_INIT(p_joinwq);
382
383 ASSERT(CPU != NULL);
384
385 if (THREAD) {
386 /* must be run after the switch to scheduler stack */
387 after_thread_ran();
388
389 switch (THREAD->state) {
390 case Running:
391 spinlock_unlock(&THREAD->lock);
392 thread_ready(THREAD);
393 break;
394
395 case Exiting:
396repeat:
397 if (THREAD->detached) {
398 thread_destroy(THREAD);
399 } else {
400 /*
401 * The thread structure is kept allocated until
402 * somebody calls thread_detach() on it.
403 */
404 if (!spinlock_trylock(&THREAD->join_wq.lock)) {
405 /*
406 * Avoid deadlock.
407 */
408 spinlock_unlock(&THREAD->lock);
409 delay(HZ);
410 spinlock_lock(&THREAD->lock);
411 DEADLOCK_PROBE(p_joinwq,
412 DEADLOCK_THRESHOLD);
413 goto repeat;
414 }
415 _waitq_wakeup_unsafe(&THREAD->join_wq,
416 WAKEUP_FIRST);
417 spinlock_unlock(&THREAD->join_wq.lock);
418
419 THREAD->state = Lingering;
420 spinlock_unlock(&THREAD->lock);
421 }
422 break;
423
424 case Sleeping:
425 /*
426 * Prefer the thread after it's woken up.
427 */
428 THREAD->priority = -1;
429
430 /*
431 * We need to release wq->lock which we locked in
432 * waitq_sleep(). Address of wq->lock is kept in
433 * THREAD->sleep_queue.
434 */
435 spinlock_unlock(&THREAD->sleep_queue->lock);
436
437 /*
438 * Check for possible requests for out-of-context
439 * invocation.
440 */
441 if (THREAD->call_me) {
442 THREAD->call_me(THREAD->call_me_with);
443 THREAD->call_me = NULL;
444 THREAD->call_me_with = NULL;
445 }
446
447 spinlock_unlock(&THREAD->lock);
448
449 break;
450
451 default:
452 /*
453 * Entering state is unexpected.
454 */
455 panic("tid%" PRIu64 ": unexpected state %s.",
456 THREAD->tid, thread_states[THREAD->state]);
457 break;
458 }
459
460 THREAD = NULL;
461 }
462
463 THREAD = find_best_thread();
464
465 spinlock_lock(&THREAD->lock);
466 priority = THREAD->priority;
467 spinlock_unlock(&THREAD->lock);
468
469 relink_rq(priority);
470
471 /*
472 * If both the old and the new task are the same, lots of work is
473 * avoided.
474 */
475 if (TASK != THREAD->task) {
476 as_t *as1 = NULL;
477 as_t *as2;
478
479 if (TASK) {
480 spinlock_lock(&TASK->lock);
481 as1 = TASK->as;
482 spinlock_unlock(&TASK->lock);
483 }
484
485 spinlock_lock(&THREAD->task->lock);
486 as2 = THREAD->task->as;
487 spinlock_unlock(&THREAD->task->lock);
488
489 /*
490 * Note that it is possible for two tasks to share one address
491 * space.
492 */
493 if (as1 != as2) {
494 /*
495 * Both tasks and address spaces are different.
496 * Replace the old one with the new one.
497 */
498 as_switch(as1, as2);
499 }
500 TASK = THREAD->task;
501 before_task_runs();
502 }
503
504 spinlock_lock(&THREAD->lock);
505 THREAD->state = Running;
506
507#ifdef SCHEDULER_VERBOSE
508 printf("cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
509 ", nrdy=%ld)\n", CPU->id, THREAD->tid, THREAD->priority,
510 THREAD->ticks, atomic_get(&CPU->nrdy));
511#endif
512
513 /*
514 * Some architectures provide late kernel PA2KA(identity)
515 * mapping in a page fault handler. However, the page fault
516 * handler uses the kernel stack of the running thread and
517 * therefore cannot be used to map it. The kernel stack, if
518 * necessary, is to be mapped in before_thread_runs(). This
519 * function must be executed before the switch to the new stack.
520 */
521 before_thread_runs();
522
523 /*
524 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
525 * thread's stack.
526 */
527 the_copy(THE, (the_t *) THREAD->kstack);
528
529 context_restore(&THREAD->saved_context);
530 /* not reached */
531}
532
533#ifdef CONFIG_SMP
534/** Load balancing thread
535 *
536 * SMP load balancing thread, supervising thread supplies
537 * for the CPU it's wired to.
538 *
539 * @param arg Generic thread argument (unused).
540 *
541 */
542void kcpulb(void *arg)
543{
544 thread_t *t;
545 int count;
546 atomic_count_t average;
547 unsigned int i;
548 int j;
549 int k = 0;
550 ipl_t ipl;
551
552 /*
553 * Detach kcpulb as nobody will call thread_join_timeout() on it.
554 */
555 thread_detach(THREAD);
556
557loop:
558 /*
559 * Work in 1s intervals.
560 */
561 thread_sleep(1);
562
563not_satisfied:
564 /*
565 * Calculate the number of threads that will be migrated/stolen from
566 * other CPU's. Note that situation can have changed between two
567 * passes. Each time get the most up to date counts.
568 */
569 average = atomic_get(&nrdy) / config.cpu_active + 1;
570 count = average - atomic_get(&CPU->nrdy);
571
572 if (count <= 0)
573 goto satisfied;
574
575 /*
576 * Searching least priority queues on all CPU's first and most priority
577 * queues on all CPU's last.
578 */
579 for (j = RQ_COUNT - 1; j >= 0; j--) {
580 for (i = 0; i < config.cpu_active; i++) {
581 link_t *l;
582 runq_t *r;
583 cpu_t *cpu;
584
585 cpu = &cpus[(i + k) % config.cpu_active];
586
587 /*
588 * Not interested in ourselves.
589 * Doesn't require interrupt disabling for kcpulb has
590 * THREAD_FLAG_WIRED.
591 */
592 if (CPU == cpu)
593 continue;
594 if (atomic_get(&cpu->nrdy) <= average)
595 continue;
596
597 ipl = interrupts_disable();
598 r = &cpu->rq[j];
599 spinlock_lock(&r->lock);
600 if (r->n == 0) {
601 spinlock_unlock(&r->lock);
602 interrupts_restore(ipl);
603 continue;
604 }
605
606 t = NULL;
607 l = r->rq_head.prev; /* search rq from the back */
608 while (l != &r->rq_head) {
609 t = list_get_instance(l, thread_t, rq_link);
610 /*
611 * We don't want to steal CPU-wired threads
612 * neither threads already stolen. The latter
613 * prevents threads from migrating between CPU's
614 * without ever being run. We don't want to
615 * steal threads whose FPU context is still in
616 * CPU.
617 */
618 spinlock_lock(&t->lock);
619 if ((!(t->flags & (THREAD_FLAG_WIRED |
620 THREAD_FLAG_STOLEN))) &&
621 (!(t->fpu_context_engaged))) {
622 /*
623 * Remove t from r.
624 */
625 spinlock_unlock(&t->lock);
626
627 atomic_dec(&cpu->nrdy);
628 atomic_dec(&nrdy);
629
630 r->n--;
631 list_remove(&t->rq_link);
632
633 break;
634 }
635 spinlock_unlock(&t->lock);
636 l = l->prev;
637 t = NULL;
638 }
639 spinlock_unlock(&r->lock);
640
641 if (t) {
642 /*
643 * Ready t on local CPU
644 */
645 spinlock_lock(&t->lock);
646#ifdef KCPULB_VERBOSE
647 printf("kcpulb%u: TID %" PRIu64 " -> cpu%u, "
648 "nrdy=%ld, avg=%ld\n", CPU->id, t->tid,
649 CPU->id, atomic_get(&CPU->nrdy),
650 atomic_get(&nrdy) / config.cpu_active);
651#endif
652 t->flags |= THREAD_FLAG_STOLEN;
653 t->state = Entering;
654 spinlock_unlock(&t->lock);
655
656 thread_ready(t);
657
658 interrupts_restore(ipl);
659
660 if (--count == 0)
661 goto satisfied;
662
663 /*
664 * We are not satisfied yet, focus on another
665 * CPU next time.
666 */
667 k++;
668
669 continue;
670 }
671 interrupts_restore(ipl);
672 }
673 }
674
675 if (atomic_get(&CPU->nrdy)) {
676 /*
677 * Be a little bit light-weight and let migrated threads run.
678 */
679 scheduler();
680 } else {
681 /*
682 * We failed to migrate a single thread.
683 * Give up this turn.
684 */
685 goto loop;
686 }
687
688 goto not_satisfied;
689
690satisfied:
691 goto loop;
692}
693
694#endif /* CONFIG_SMP */
695
696
697/** Print information about threads & scheduler queues */
698void sched_print_list(void)
699{
700 ipl_t ipl;
701 unsigned int cpu, i;
702 runq_t *r;
703 thread_t *t;
704 link_t *cur;
705
706 /* We are going to mess with scheduler structures,
707 * let's not be interrupted */
708 ipl = interrupts_disable();
709 for (cpu = 0; cpu < config.cpu_count; cpu++) {
710
711 if (!cpus[cpu].active)
712 continue;
713
714 spinlock_lock(&cpus[cpu].lock);
715 printf("cpu%u: address=%p, nrdy=%ld, needs_relink=%" PRIs "\n",
716 cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy),
717 cpus[cpu].needs_relink);
718
719 for (i = 0; i < RQ_COUNT; i++) {
720 r = &cpus[cpu].rq[i];
721 spinlock_lock(&r->lock);
722 if (!r->n) {
723 spinlock_unlock(&r->lock);
724 continue;
725 }
726 printf("\trq[%u]: ", i);
727 for (cur = r->rq_head.next; cur != &r->rq_head;
728 cur = cur->next) {
729 t = list_get_instance(cur, thread_t, rq_link);
730 printf("%" PRIu64 "(%s) ", t->tid,
731 thread_states[t->state]);
732 }
733 printf("\n");
734 spinlock_unlock(&r->lock);
735 }
736 spinlock_unlock(&cpus[cpu].lock);
737 }
738
739 interrupts_restore(ipl);
740}
741
742/** @}
743 */
Note: See TracBrowser for help on using the repository browser.