source: mainline/generic/src/proc/scheduler.c@ 7257021e

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since 7257021e was b45c443, checked in by Josef Cejka <malyzelenyhnus@…>, 19 years ago

Kernel doxygen comments updated.
  • Property mode set to 100644
File size: 15.9 KB
Line 
1/*
2 * Copyright (C) 2001-2004 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 /**
30 * @defgroup proc Proc
31 * @ingroup kernel
32 * @{
33 * @}
34 */
35
36 /** @addtogroup genericproc generic
37 * @ingroup proc
38 * @{
39 */
40
41/**
42 * @file
43 * @brief Scheduler and load balancing.
44 *
45 * This file contains the scheduler and kcpulb kernel thread which
46 * performs load-balancing of per-CPU run queues.
47 */
48
49#include <proc/scheduler.h>
50#include <proc/thread.h>
51#include <proc/task.h>
52#include <mm/frame.h>
53#include <mm/page.h>
54#include <mm/as.h>
55#include <time/delay.h>
56#include <arch/asm.h>
57#include <arch/faddr.h>
58#include <atomic.h>
59#include <synch/spinlock.h>
60#include <config.h>
61#include <context.h>
62#include <func.h>
63#include <arch.h>
64#include <adt/list.h>
65#include <panic.h>
66#include <typedefs.h>
67#include <cpu.h>
68#include <print.h>
69#include <debug.h>
70
71static void before_task_runs(void);
72static void before_thread_runs(void);
73static void after_thread_ran(void);
74static void scheduler_separated_stack(void);
75
76atomic_t nrdy; /**< Number of ready threads in the system. */
77
78/** Carry out actions before new task runs. */
79void before_task_runs(void)
80{
81 before_task_runs_arch();
82}
83
84/** Take actions before new thread runs.
85 *
86 * Perform actions that need to be
87 * taken before the newly selected
88 * tread is passed control.
89 *
90 * THREAD->lock is locked on entry
91 *
92 */
93void before_thread_runs(void)
94{
95 before_thread_runs_arch();
96#ifdef CONFIG_FPU_LAZY
97 if(THREAD==CPU->fpu_owner)
98 fpu_enable();
99 else
100 fpu_disable();
101#else
102 fpu_enable();
103 if (THREAD->fpu_context_exists)
104 fpu_context_restore(THREAD->saved_fpu_context);
105 else {
106 fpu_init();
107 THREAD->fpu_context_exists=1;
108 }
109#endif
110}
111
112/** Take actions after THREAD had run.
113 *
114 * Perform actions that need to be
115 * taken after the running thread
116 * had been preempted by the scheduler.
117 *
118 * THREAD->lock is locked on entry
119 *
120 */
121void after_thread_ran(void)
122{
123 after_thread_ran_arch();
124}
125
126#ifdef CONFIG_FPU_LAZY
127void scheduler_fpu_lazy_request(void)
128{
129restart:
130 fpu_enable();
131 spinlock_lock(&CPU->lock);
132
133 /* Save old context */
134 if (CPU->fpu_owner != NULL) {
135 spinlock_lock(&CPU->fpu_owner->lock);
136 fpu_context_save(CPU->fpu_owner->saved_fpu_context);
137 /* don't prevent migration */
138 CPU->fpu_owner->fpu_context_engaged=0;
139 spinlock_unlock(&CPU->fpu_owner->lock);
140 CPU->fpu_owner = NULL;
141 }
142
143 spinlock_lock(&THREAD->lock);
144 if (THREAD->fpu_context_exists) {
145 fpu_context_restore(THREAD->saved_fpu_context);
146 } else {
147 /* Allocate FPU context */
148 if (!THREAD->saved_fpu_context) {
149 /* Might sleep */
150 spinlock_unlock(&THREAD->lock);
151 spinlock_unlock(&CPU->lock);
152 THREAD->saved_fpu_context = slab_alloc(fpu_context_slab,
153 0);
154 /* We may have switched CPUs during slab_alloc */
155 goto restart;
156 }
157 fpu_init();
158 THREAD->fpu_context_exists=1;
159 }
160 CPU->fpu_owner=THREAD;
161 THREAD->fpu_context_engaged = 1;
162 spinlock_unlock(&THREAD->lock);
163
164 spinlock_unlock(&CPU->lock);
165}
166#endif
167
168/** Initialize scheduler
169 *
170 * Initialize kernel scheduler.
171 *
172 */
173void scheduler_init(void)
174{
175}
176
177/** Get thread to be scheduled
178 *
179 * Get the optimal thread to be scheduled
180 * according to thread accounting and scheduler
181 * policy.
182 *
183 * @return Thread to be scheduled.
184 *
185 */
186static thread_t *find_best_thread(void)
187{
188 thread_t *t;
189 runq_t *r;
190 int i;
191
192 ASSERT(CPU != NULL);
193
194loop:
195 interrupts_enable();
196
197 if (atomic_get(&CPU->nrdy) == 0) {
198 /*
199 * For there was nothing to run, the CPU goes to sleep
200 * until a hardware interrupt or an IPI comes.
201 * This improves energy saving and hyperthreading.
202 */
203
204 /*
205 * An interrupt might occur right now and wake up a thread.
206 * In such case, the CPU will continue to go to sleep
207 * even though there is a runnable thread.
208 */
209
210 cpu_sleep();
211 goto loop;
212 }
213
214 interrupts_disable();
215
216 for (i = 0; i<RQ_COUNT; i++) {
217 r = &CPU->rq[i];
218 spinlock_lock(&r->lock);
219 if (r->n == 0) {
220 /*
221 * If this queue is empty, try a lower-priority queue.
222 */
223 spinlock_unlock(&r->lock);
224 continue;
225 }
226
227 atomic_dec(&CPU->nrdy);
228 atomic_dec(&nrdy);
229 r->n--;
230
231 /*
232 * Take the first thread from the queue.
233 */
234 t = list_get_instance(r->rq_head.next, thread_t, rq_link);
235 list_remove(&t->rq_link);
236
237 spinlock_unlock(&r->lock);
238
239 spinlock_lock(&t->lock);
240 t->cpu = CPU;
241
242 t->ticks = us2ticks((i+1)*10000);
243 t->priority = i; /* correct rq index */
244
245 /*
246 * Clear the X_STOLEN flag so that t can be migrated when load balancing needs emerge.
247 */
248 t->flags &= ~X_STOLEN;
249 spinlock_unlock(&t->lock);
250
251 return t;
252 }
253 goto loop;
254
255}
256
257/** Prevent rq starvation
258 *
259 * Prevent low priority threads from starving in rq's.
260 *
261 * When the function decides to relink rq's, it reconnects
262 * respective pointers so that in result threads with 'pri'
263 * greater or equal @start are moved to a higher-priority queue.
264 *
265 * @param start Threshold priority.
266 *
267 */
268static void relink_rq(int start)
269{
270 link_t head;
271 runq_t *r;
272 int i, n;
273
274 list_initialize(&head);
275 spinlock_lock(&CPU->lock);
276 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
277 for (i = start; i<RQ_COUNT-1; i++) {
278 /* remember and empty rq[i + 1] */
279 r = &CPU->rq[i + 1];
280 spinlock_lock(&r->lock);
281 list_concat(&head, &r->rq_head);
282 n = r->n;
283 r->n = 0;
284 spinlock_unlock(&r->lock);
285
286 /* append rq[i + 1] to rq[i] */
287 r = &CPU->rq[i];
288 spinlock_lock(&r->lock);
289 list_concat(&r->rq_head, &head);
290 r->n += n;
291 spinlock_unlock(&r->lock);
292 }
293 CPU->needs_relink = 0;
294 }
295 spinlock_unlock(&CPU->lock);
296
297}
298
299/** The scheduler
300 *
301 * The thread scheduling procedure.
302 * Passes control directly to
303 * scheduler_separated_stack().
304 *
305 */
306void scheduler(void)
307{
308 volatile ipl_t ipl;
309
310 ASSERT(CPU != NULL);
311
312 ipl = interrupts_disable();
313
314 if (atomic_get(&haltstate))
315 halt();
316
317 if (THREAD) {
318 spinlock_lock(&THREAD->lock);
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 spinlock_unlock(&THREAD->lock);
327 interrupts_restore(THREAD->saved_context.ipl);
328
329 return;
330 }
331
332 /*
333 * Interrupt priority level of preempted thread is recorded here
334 * to facilitate scheduler() invocations from interrupts_disable()'d
335 * code (e.g. waitq_sleep_timeout()).
336 */
337 THREAD->saved_context.ipl = ipl;
338 }
339
340 /*
341 * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, VM
342 * and preemption counter. At this point THE could be coming either
343 * from THREAD's or CPU's stack.
344 */
345 the_copy(THE, (the_t *) CPU->stack);
346
347 /*
348 * We may not keep the old stack.
349 * Reason: If we kept the old stack and got blocked, for instance, in
350 * find_best_thread(), the old thread could get rescheduled by another
351 * CPU and overwrite the part of its own stack that was also used by
352 * the scheduler on this CPU.
353 *
354 * Moreover, we have to bypass the compiler-generated POP sequence
355 * which is fooled by SP being set to the very top of the stack.
356 * Therefore the scheduler() function continues in
357 * scheduler_separated_stack().
358 */
359 context_save(&CPU->saved_context);
360 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack), (__address) CPU->stack, CPU_STACK_SIZE);
361 context_restore(&CPU->saved_context);
362 /* not reached */
363}
364
365/** Scheduler stack switch wrapper
366 *
367 * Second part of the scheduler() function
368 * using new stack. Handling the actual context
369 * switch to a new thread.
370 *
371 * Assume THREAD->lock is held.
372 */
373void scheduler_separated_stack(void)
374{
375 int priority;
376
377 ASSERT(CPU != NULL);
378
379 if (THREAD) {
380 /* must be run after the switch to scheduler stack */
381 after_thread_ran();
382
383 switch (THREAD->state) {
384 case Running:
385 spinlock_unlock(&THREAD->lock);
386 thread_ready(THREAD);
387 break;
388
389 case Exiting:
390repeat:
391 if (THREAD->detached) {
392 thread_destroy(THREAD);
393 } else {
394 /*
395 * The thread structure is kept allocated until somebody
396 * calls thread_detach() on it.
397 */
398 if (!spinlock_trylock(&THREAD->join_wq.lock)) {
399 /*
400 * Avoid deadlock.
401 */
402 spinlock_unlock(&THREAD->lock);
403 delay(10);
404 spinlock_lock(&THREAD->lock);
405 goto repeat;
406 }
407 _waitq_wakeup_unsafe(&THREAD->join_wq, false);
408 spinlock_unlock(&THREAD->join_wq.lock);
409
410 THREAD->state = Undead;
411 spinlock_unlock(&THREAD->lock);
412 }
413 break;
414
415 case Sleeping:
416 /*
417 * Prefer the thread after it's woken up.
418 */
419 THREAD->priority = -1;
420
421 /*
422 * We need to release wq->lock which we locked in waitq_sleep().
423 * Address of wq->lock is kept in THREAD->sleep_queue.
424 */
425 spinlock_unlock(&THREAD->sleep_queue->lock);
426
427 /*
428 * Check for possible requests for out-of-context invocation.
429 */
430 if (THREAD->call_me) {
431 THREAD->call_me(THREAD->call_me_with);
432 THREAD->call_me = NULL;
433 THREAD->call_me_with = NULL;
434 }
435
436 spinlock_unlock(&THREAD->lock);
437
438 break;
439
440 default:
441 /*
442 * Entering state is unexpected.
443 */
444 panic("tid%d: unexpected state %s\n", THREAD->tid, thread_states[THREAD->state]);
445 break;
446 }
447
448 THREAD = NULL;
449 }
450
451 THREAD = find_best_thread();
452
453 spinlock_lock(&THREAD->lock);
454 priority = THREAD->priority;
455 spinlock_unlock(&THREAD->lock);
456
457 relink_rq(priority);
458
459 /*
460 * If both the old and the new task are the same, lots of work is avoided.
461 */
462 if (TASK != THREAD->task) {
463 as_t *as1 = NULL;
464 as_t *as2;
465
466 if (TASK) {
467 spinlock_lock(&TASK->lock);
468 as1 = TASK->as;
469 spinlock_unlock(&TASK->lock);
470 }
471
472 spinlock_lock(&THREAD->task->lock);
473 as2 = THREAD->task->as;
474 spinlock_unlock(&THREAD->task->lock);
475
476 /*
477 * Note that it is possible for two tasks to share one address space.
478 */
479 if (as1 != as2) {
480 /*
481 * Both tasks and address spaces are different.
482 * Replace the old one with the new one.
483 */
484 as_switch(as1, as2);
485 }
486 TASK = THREAD->task;
487 before_task_runs();
488 }
489
490 spinlock_lock(&THREAD->lock);
491 THREAD->state = Running;
492
493#ifdef SCHEDULER_VERBOSE
494 printf("cpu%d: tid %d (priority=%d,ticks=%lld,nrdy=%ld)\n", CPU->id, THREAD->tid, THREAD->priority, THREAD->ticks, atomic_get(&CPU->nrdy));
495#endif
496
497 /*
498 * Some architectures provide late kernel PA2KA(identity)
499 * mapping in a page fault handler. However, the page fault
500 * handler uses the kernel stack of the running thread and
501 * therefore cannot be used to map it. The kernel stack, if
502 * necessary, is to be mapped in before_thread_runs(). This
503 * function must be executed before the switch to the new stack.
504 */
505 before_thread_runs();
506
507 /*
508 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to thread's stack.
509 */
510 the_copy(THE, (the_t *) THREAD->kstack);
511
512 context_restore(&THREAD->saved_context);
513 /* not reached */
514}
515
516#ifdef CONFIG_SMP
517/** Load balancing thread
518 *
519 * SMP load balancing thread, supervising thread supplies
520 * for the CPU it's wired to.
521 *
522 * @param arg Generic thread argument (unused).
523 *
524 */
525void kcpulb(void *arg)
526{
527 thread_t *t;
528 int count, average, i, j, k = 0;
529 ipl_t ipl;
530
531 /*
532 * Detach kcpulb as nobody will call thread_join_timeout() on it.
533 */
534 thread_detach(THREAD);
535
536loop:
537 /*
538 * Work in 1s intervals.
539 */
540 thread_sleep(1);
541
542not_satisfied:
543 /*
544 * Calculate the number of threads that will be migrated/stolen from
545 * other CPU's. Note that situation can have changed between two
546 * passes. Each time get the most up to date counts.
547 */
548 average = atomic_get(&nrdy) / config.cpu_active + 1;
549 count = average - atomic_get(&CPU->nrdy);
550
551 if (count <= 0)
552 goto satisfied;
553
554 /*
555 * Searching least priority queues on all CPU's first and most priority queues on all CPU's last.
556 */
557 for (j=RQ_COUNT-1; j >= 0; j--) {
558 for (i=0; i < config.cpu_active; i++) {
559 link_t *l;
560 runq_t *r;
561 cpu_t *cpu;
562
563 cpu = &cpus[(i + k) % config.cpu_active];
564
565 /*
566 * Not interested in ourselves.
567 * Doesn't require interrupt disabling for kcpulb is X_WIRED.
568 */
569 if (CPU == cpu)
570 continue;
571 if (atomic_get(&cpu->nrdy) <= average)
572 continue;
573
574 ipl = interrupts_disable();
575 r = &cpu->rq[j];
576 spinlock_lock(&r->lock);
577 if (r->n == 0) {
578 spinlock_unlock(&r->lock);
579 interrupts_restore(ipl);
580 continue;
581 }
582
583 t = NULL;
584 l = r->rq_head.prev; /* search rq from the back */
585 while (l != &r->rq_head) {
586 t = list_get_instance(l, thread_t, rq_link);
587 /*
588 * We don't want to steal CPU-wired threads neither threads already stolen.
589 * The latter prevents threads from migrating between CPU's without ever being run.
590 * We don't want to steal threads whose FPU context is still in CPU.
591 */
592 spinlock_lock(&t->lock);
593 if ( (!(t->flags & (X_WIRED | X_STOLEN))) && (!(t->fpu_context_engaged)) ) {
594 /*
595 * Remove t from r.
596 */
597 spinlock_unlock(&t->lock);
598
599 atomic_dec(&cpu->nrdy);
600 atomic_dec(&nrdy);
601
602 r->n--;
603 list_remove(&t->rq_link);
604
605 break;
606 }
607 spinlock_unlock(&t->lock);
608 l = l->prev;
609 t = NULL;
610 }
611 spinlock_unlock(&r->lock);
612
613 if (t) {
614 /*
615 * Ready t on local CPU
616 */
617 spinlock_lock(&t->lock);
618#ifdef KCPULB_VERBOSE
619 printf("kcpulb%d: TID %d -> cpu%d, nrdy=%ld, avg=%nd\n", CPU->id, t->tid, CPU->id, atomic_get(&CPU->nrdy), atomic_get(&nrdy) / config.cpu_active);
620#endif
621 t->flags |= X_STOLEN;
622 t->state = Entering;
623 spinlock_unlock(&t->lock);
624
625 thread_ready(t);
626
627 interrupts_restore(ipl);
628
629 if (--count == 0)
630 goto satisfied;
631
632 /*
633 * We are not satisfied yet, focus on another CPU next time.
634 */
635 k++;
636
637 continue;
638 }
639 interrupts_restore(ipl);
640 }
641 }
642
643 if (atomic_get(&CPU->nrdy)) {
644 /*
645 * Be a little bit light-weight and let migrated threads run.
646 */
647 scheduler();
648 } else {
649 /*
650 * We failed to migrate a single thread.
651 * Give up this turn.
652 */
653 goto loop;
654 }
655
656 goto not_satisfied;
657
658satisfied:
659 goto loop;
660}
661
662#endif /* CONFIG_SMP */
663
664
665/** Print information about threads & scheduler queues */
666void sched_print_list(void)
667{
668 ipl_t ipl;
669 int cpu,i;
670 runq_t *r;
671 thread_t *t;
672 link_t *cur;
673
674 /* We are going to mess with scheduler structures,
675 * let's not be interrupted */
676 ipl = interrupts_disable();
677 for (cpu=0;cpu < config.cpu_count; cpu++) {
678
679 if (!cpus[cpu].active)
680 continue;
681
682 spinlock_lock(&cpus[cpu].lock);
683 printf("cpu%d: address=%p, nrdy=%ld, needs_relink=%ld\n",
684 cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy), cpus[cpu].needs_relink);
685
686 for (i=0; i<RQ_COUNT; i++) {
687 r = &cpus[cpu].rq[i];
688 spinlock_lock(&r->lock);
689 if (!r->n) {
690 spinlock_unlock(&r->lock);
691 continue;
692 }
693 printf("\trq[%d]: ", i);
694 for (cur=r->rq_head.next; cur!=&r->rq_head; cur=cur->next) {
695 t = list_get_instance(cur, thread_t, rq_link);
696 printf("%d(%s) ", t->tid,
697 thread_states[t->state]);
698 }
699 printf("\n");
700 spinlock_unlock(&r->lock);
701 }
702 spinlock_unlock(&cpus[cpu].lock);
703 }
704
705 interrupts_restore(ipl);
706}
707
708 /** @}
709 */
710
Note: See TracBrowser for help on using the repository browser.