source: mainline/kernel/generic/src/proc/scheduler.c@ da1bafb

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since da1bafb was da1bafb, checked in by Martin Decky <martin@…>, 15 years ago

major code revision

  • replace spinlocks taken with interrupts disabled with irq_spinlocks
  • change spacing (not indendation) to be tab-size independent
  • use unsigned integer types where appropriate (especially bit flags)
  • visual separation
  • remove argument names in function prototypes
  • string changes
  • correct some formating directives
  • replace various cryptic single-character variables (t, a, m, c, b, etc.) with proper identifiers (thread, task, timeout, as, itm, itc, etc.)
  • unify some assembler constructs
  • unused page table levels are now optimized out in compile time
  • replace several ints (with boolean semantics) with bools
  • use specifically sized types instead of generic types where appropriate (size_t, uint32_t, btree_key_t)
  • improve comments
  • split asserts with conjuction into multiple independent asserts
  • Property mode set to 100644
File size: 17.8 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
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 irq_spinlock_lock(&CPU->lock, false);
126
127 /* Save old context */
128 if (CPU->fpu_owner != NULL) {
129 irq_spinlock_lock(&CPU->fpu_owner->lock, false);
130 fpu_context_save(CPU->fpu_owner->saved_fpu_context);
131
132 /* Don't prevent migration */
133 CPU->fpu_owner->fpu_context_engaged = 0;
134 irq_spinlock_unlock(&CPU->fpu_owner->lock, false);
135 CPU->fpu_owner = NULL;
136 }
137
138 irq_spinlock_lock(&THREAD->lock, false);
139 if (THREAD->fpu_context_exists) {
140 fpu_context_restore(THREAD->saved_fpu_context);
141 } else {
142 /* Allocate FPU context */
143 if (!THREAD->saved_fpu_context) {
144 /* Might sleep */
145 irq_spinlock_unlock(&THREAD->lock, false);
146 irq_spinlock_unlock(&CPU->lock, false);
147 THREAD->saved_fpu_context =
148 (fpu_context_t *) slab_alloc(fpu_context_slab, 0);
149
150 /* We may have switched CPUs during slab_alloc */
151 goto restart;
152 }
153 fpu_init();
154 THREAD->fpu_context_exists = 1;
155 }
156
157 CPU->fpu_owner = THREAD;
158 THREAD->fpu_context_engaged = 1;
159 irq_spinlock_unlock(&THREAD->lock, false);
160
161 irq_spinlock_unlock(&CPU->lock, false);
162}
163#endif /* CONFIG_FPU_LAZY */
164
165/** Initialize scheduler
166 *
167 * Initialize kernel scheduler.
168 *
169 */
170void scheduler_init(void)
171{
172}
173
174/** Get thread to be scheduled
175 *
176 * Get the optimal thread to be scheduled
177 * according to thread accounting and scheduler
178 * policy.
179 *
180 * @return Thread to be scheduled.
181 *
182 */
183static thread_t *find_best_thread(void)
184{
185 ASSERT(CPU != NULL);
186
187loop:
188
189 if (atomic_get(&CPU->nrdy) == 0) {
190 /*
191 * For there was nothing to run, the CPU goes to sleep
192 * until a hardware interrupt or an IPI comes.
193 * This improves energy saving and hyperthreading.
194 */
195
196 /* Mark CPU as it was idle this clock tick */
197 irq_spinlock_lock(&CPU->lock, false);
198 CPU->idle = true;
199 irq_spinlock_unlock(&CPU->lock, false);
200
201 interrupts_enable();
202 /*
203 * An interrupt might occur right now and wake up a thread.
204 * In such case, the CPU will continue to go to sleep
205 * even though there is a runnable thread.
206 */
207 cpu_sleep();
208 interrupts_disable();
209 goto loop;
210 }
211
212 unsigned int i;
213 for (i = 0; i < RQ_COUNT; i++) {
214 irq_spinlock_lock(&(CPU->rq[i].lock), false);
215 if (CPU->rq[i].n == 0) {
216 /*
217 * If this queue is empty, try a lower-priority queue.
218 */
219 irq_spinlock_unlock(&(CPU->rq[i].lock), false);
220 continue;
221 }
222
223 atomic_dec(&CPU->nrdy);
224 atomic_dec(&nrdy);
225 CPU->rq[i].n--;
226
227 /*
228 * Take the first thread from the queue.
229 */
230 thread_t *thread =
231 list_get_instance(CPU->rq[i].rq_head.next, thread_t, rq_link);
232 list_remove(&thread->rq_link);
233
234 irq_spinlock_pass(&(CPU->rq[i].lock), &thread->lock);
235
236 thread->cpu = CPU;
237 thread->ticks = us2ticks((i + 1) * 10000);
238 thread->priority = i; /* Correct rq index */
239
240 /*
241 * Clear the THREAD_FLAG_STOLEN flag so that t can be migrated
242 * when load balancing needs emerge.
243 */
244 thread->flags &= ~THREAD_FLAG_STOLEN;
245 irq_spinlock_unlock(&thread->lock, false);
246
247 return thread;
248 }
249
250 goto loop;
251}
252
253/** Prevent rq starvation
254 *
255 * Prevent low priority threads from starving in rq's.
256 *
257 * When the function decides to relink rq's, it reconnects
258 * respective pointers so that in result threads with 'pri'
259 * greater or equal start are moved to a higher-priority queue.
260 *
261 * @param start Threshold priority.
262 *
263 */
264static void relink_rq(int start)
265{
266 link_t head;
267
268 list_initialize(&head);
269 irq_spinlock_lock(&CPU->lock, false);
270
271 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
272 int i;
273 for (i = start; i < RQ_COUNT - 1; i++) {
274 /* Remember and empty rq[i + 1] */
275
276 irq_spinlock_lock(&CPU->rq[i + 1].lock, false);
277 list_concat(&head, &CPU->rq[i + 1].rq_head);
278 size_t n = CPU->rq[i + 1].n;
279 CPU->rq[i + 1].n = 0;
280 irq_spinlock_unlock(&CPU->rq[i + 1].lock, false);
281
282 /* Append rq[i + 1] to rq[i] */
283
284 irq_spinlock_lock(&CPU->rq[i].lock, false);
285 list_concat(&CPU->rq[i].rq_head, &head);
286 CPU->rq[i].n += n;
287 irq_spinlock_unlock(&CPU->rq[i].lock, false);
288 }
289
290 CPU->needs_relink = 0;
291 }
292
293 irq_spinlock_unlock(&CPU->lock, false);
294}
295
296/** The scheduler
297 *
298 * The thread scheduling procedure.
299 * Passes control directly to
300 * scheduler_separated_stack().
301 *
302 */
303void scheduler(void)
304{
305 volatile ipl_t ipl;
306
307 ASSERT(CPU != NULL);
308
309 ipl = interrupts_disable();
310
311 if (atomic_get(&haltstate))
312 halt();
313
314 if (THREAD) {
315 irq_spinlock_lock(&THREAD->lock, false);
316
317 /* Update thread kernel accounting */
318 THREAD->kcycles += get_cycle() - THREAD->last_cycle;
319
320#ifndef CONFIG_FPU_LAZY
321 fpu_context_save(THREAD->saved_fpu_context);
322#endif
323 if (!context_save(&THREAD->saved_context)) {
324 /*
325 * This is the place where threads leave scheduler();
326 */
327
328 /* Save current CPU cycle */
329 THREAD->last_cycle = get_cycle();
330
331 irq_spinlock_unlock(&THREAD->lock, false);
332 interrupts_restore(THREAD->saved_context.ipl);
333
334 return;
335 }
336
337 /*
338 * Interrupt priority level of preempted thread is recorded
339 * here to facilitate scheduler() invocations from
340 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
341 *
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 */
352 the_copy(THE, (the_t *) CPU->stack);
353
354 /*
355 * We may not keep the old stack.
356 * Reason: If we kept the old stack and got blocked, for instance, in
357 * find_best_thread(), the old thread could get rescheduled by another
358 * CPU and overwrite the part of its own stack that was also used by
359 * the scheduler on this CPU.
360 *
361 * Moreover, we have to bypass the compiler-generated POP sequence
362 * which is fooled by SP being set to the very top of the stack.
363 * Therefore the scheduler() function continues in
364 * scheduler_separated_stack().
365 *
366 */
367 context_save(&CPU->saved_context);
368 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack),
369 (uintptr_t) CPU->stack, CPU_STACK_SIZE);
370 context_restore(&CPU->saved_context);
371
372 /* Not reached */
373}
374
375/** Scheduler stack switch wrapper
376 *
377 * Second part of the scheduler() function
378 * using new stack. Handling the actual context
379 * switch to a new thread.
380 *
381 * Assume THREAD->lock is held.
382 *
383 */
384void scheduler_separated_stack(void)
385{
386 DEADLOCK_PROBE_INIT(p_joinwq);
387 task_t *old_task = TASK;
388 as_t *old_as = AS;
389
390 ASSERT(CPU != NULL);
391
392 /*
393 * Hold the current task and the address space to prevent their
394 * possible destruction should thread_destroy() be called on this or any
395 * other processor while the scheduler is still using them.
396 *
397 */
398 if (old_task)
399 task_hold(old_task);
400
401 if (old_as)
402 as_hold(old_as);
403
404 if (THREAD) {
405 /* Must be run after the switch to scheduler stack */
406 after_thread_ran();
407
408 switch (THREAD->state) {
409 case Running:
410 irq_spinlock_unlock(&THREAD->lock, false);
411 thread_ready(THREAD);
412 break;
413
414 case Exiting:
415repeat:
416 if (THREAD->detached) {
417 thread_destroy(THREAD, false);
418 } else {
419 /*
420 * The thread structure is kept allocated until
421 * somebody calls thread_detach() on it.
422 *
423 */
424 if (!irq_spinlock_trylock(&THREAD->join_wq.lock)) {
425 /*
426 * Avoid deadlock.
427 *
428 */
429 irq_spinlock_unlock(&THREAD->lock, false);
430 delay(HZ);
431 irq_spinlock_lock(&THREAD->lock, false);
432 DEADLOCK_PROBE(p_joinwq,
433 DEADLOCK_THRESHOLD);
434 goto repeat;
435 }
436 _waitq_wakeup_unsafe(&THREAD->join_wq,
437 WAKEUP_FIRST);
438 irq_spinlock_unlock(&THREAD->join_wq.lock, false);
439
440 THREAD->state = Lingering;
441 irq_spinlock_unlock(&THREAD->lock, false);
442 }
443 break;
444
445 case Sleeping:
446 /*
447 * Prefer the thread after it's woken up.
448 *
449 */
450 THREAD->priority = -1;
451
452 /*
453 * We need to release wq->lock which we locked in
454 * waitq_sleep(). Address of wq->lock is kept in
455 * THREAD->sleep_queue.
456 *
457 */
458 irq_spinlock_unlock(&THREAD->sleep_queue->lock, false);
459
460 /*
461 * Check for possible requests for out-of-context
462 * invocation.
463 *
464 */
465 if (THREAD->call_me) {
466 THREAD->call_me(THREAD->call_me_with);
467 THREAD->call_me = NULL;
468 THREAD->call_me_with = NULL;
469 }
470
471 irq_spinlock_unlock(&THREAD->lock, false);
472
473 break;
474
475 default:
476 /*
477 * Entering state is unexpected.
478 *
479 */
480 panic("tid%" PRIu64 ": unexpected state %s.",
481 THREAD->tid, thread_states[THREAD->state]);
482 break;
483 }
484
485 THREAD = NULL;
486 }
487
488 THREAD = find_best_thread();
489
490 irq_spinlock_lock(&THREAD->lock, false);
491 int priority = THREAD->priority;
492 irq_spinlock_unlock(&THREAD->lock, false);
493
494 relink_rq(priority);
495
496 /*
497 * If both the old and the new task are the same, lots of work is
498 * avoided.
499 *
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 to share one address
506 * space.
507 (
508 */
509 if (old_as != new_as) {
510 /*
511 * Both tasks and address spaces are different.
512 * Replace the old one with the new one.
513 *
514 */
515 as_switch(old_as, new_as);
516 }
517
518 TASK = THREAD->task;
519 before_task_runs();
520 }
521
522 if (old_task)
523 task_release(old_task);
524
525 if (old_as)
526 as_release(old_as);
527
528 irq_spinlock_lock(&THREAD->lock, false);
529 THREAD->state = Running;
530
531#ifdef SCHEDULER_VERBOSE
532 printf("cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
533 ", nrdy=%ld)\n", CPU->id, THREAD->tid, THREAD->priority,
534 THREAD->ticks, atomic_get(&CPU->nrdy));
535#endif
536
537 /*
538 * Some architectures provide late kernel PA2KA(identity)
539 * mapping in a page fault handler. However, the page fault
540 * handler uses the kernel stack of the running thread and
541 * therefore cannot be used to map it. The kernel stack, if
542 * necessary, is to be mapped in before_thread_runs(). This
543 * function must be executed before the switch to the new stack.
544 *
545 */
546 before_thread_runs();
547
548 /*
549 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
550 * thread's stack.
551 *
552 */
553 the_copy(THE, (the_t *) THREAD->kstack);
554
555 context_restore(&THREAD->saved_context);
556
557 /* Not reached */
558}
559
560#ifdef CONFIG_SMP
561/** Load balancing thread
562 *
563 * SMP load balancing thread, supervising thread supplies
564 * for the CPU it's wired to.
565 *
566 * @param arg Generic thread argument (unused).
567 *
568 */
569void kcpulb(void *arg)
570{
571 atomic_count_t average;
572 atomic_count_t rdy;
573
574 /*
575 * Detach kcpulb as nobody will call thread_join_timeout() on it.
576 */
577 thread_detach(THREAD);
578
579loop:
580 /*
581 * Work in 1s intervals.
582 */
583 thread_sleep(1);
584
585not_satisfied:
586 /*
587 * Calculate the number of threads that will be migrated/stolen from
588 * other CPU's. Note that situation can have changed between two
589 * passes. Each time get the most up to date counts.
590 *
591 */
592 average = atomic_get(&nrdy) / config.cpu_active + 1;
593 rdy = atomic_get(&CPU->nrdy);
594
595 if (average <= rdy)
596 goto satisfied;
597
598 atomic_count_t count = average - rdy;
599
600 /*
601 * Searching least priority queues on all CPU's first and most priority
602 * queues on all CPU's last.
603 *
604 */
605 size_t acpu;
606 size_t acpu_bias = 0;
607 int rq;
608
609 for (rq = RQ_COUNT - 1; rq >= 0; rq--) {
610 for (acpu = 0; acpu < config.cpu_active; acpu++) {
611 cpu_t *cpu = &cpus[(acpu + acpu_bias) % config.cpu_active];
612
613 /*
614 * Not interested in ourselves.
615 * Doesn't require interrupt disabling for kcpulb has
616 * THREAD_FLAG_WIRED.
617 *
618 */
619 if (CPU == cpu)
620 continue;
621
622 if (atomic_get(&cpu->nrdy) <= average)
623 continue;
624
625 irq_spinlock_lock(&(cpu->rq[rq].lock), true);
626 if (cpu->rq[rq].n == 0) {
627 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
628 continue;
629 }
630
631 thread_t *thread = NULL;
632
633 /* Search rq from the back */
634 link_t *link = cpu->rq[rq].rq_head.prev;
635
636 while (link != &(cpu->rq[rq].rq_head)) {
637 thread = (thread_t *) list_get_instance(link, thread_t, rq_link);
638
639 /*
640 * We don't want to steal CPU-wired threads
641 * neither threads already stolen. The latter
642 * prevents threads from migrating between CPU's
643 * without ever being run. We don't want to
644 * steal threads whose FPU context is still in
645 * CPU.
646 *
647 */
648 irq_spinlock_lock(&thread->lock, false);
649
650 if ((!(thread->flags & (THREAD_FLAG_WIRED | THREAD_FLAG_STOLEN)))
651 && (!(thread->fpu_context_engaged))) {
652 /*
653 * Remove thread from ready queue.
654 */
655 irq_spinlock_unlock(&thread->lock, false);
656
657 atomic_dec(&cpu->nrdy);
658 atomic_dec(&nrdy);
659
660 cpu->rq[rq].n--;
661 list_remove(&thread->rq_link);
662
663 break;
664 }
665
666 irq_spinlock_unlock(&thread->lock, false);
667
668 link = link->prev;
669 thread = NULL;
670 }
671
672 if (thread) {
673 /*
674 * Ready thread on local CPU
675 *
676 */
677
678 irq_spinlock_pass(&(cpu->rq[rq].lock), &thread->lock);
679
680#ifdef KCPULB_VERBOSE
681 printf("kcpulb%u: TID %" PRIu64 " -> cpu%u, "
682 "nrdy=%ld, avg=%ld\n", CPU->id, t->tid,
683 CPU->id, atomic_get(&CPU->nrdy),
684 atomic_get(&nrdy) / config.cpu_active);
685#endif
686
687 thread->flags |= THREAD_FLAG_STOLEN;
688 thread->state = Entering;
689
690 irq_spinlock_unlock(&thread->lock, true);
691 thread_ready(thread);
692
693 if (--count == 0)
694 goto satisfied;
695
696 /*
697 * We are not satisfied yet, focus on another
698 * CPU next time.
699 *
700 */
701 acpu_bias++;
702
703 continue;
704 } else
705 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
706
707 }
708 }
709
710 if (atomic_get(&CPU->nrdy)) {
711 /*
712 * Be a little bit light-weight and let migrated threads run.
713 *
714 */
715 scheduler();
716 } else {
717 /*
718 * We failed to migrate a single thread.
719 * Give up this turn.
720 *
721 */
722 goto loop;
723 }
724
725 goto not_satisfied;
726
727satisfied:
728 goto loop;
729}
730#endif /* CONFIG_SMP */
731
732/** Print information about threads & scheduler queues
733 *
734 */
735void sched_print_list(void)
736{
737 size_t cpu;
738 for (cpu = 0; cpu < config.cpu_count; cpu++) {
739 if (!cpus[cpu].active)
740 continue;
741
742 irq_spinlock_lock(&cpus[cpu].lock, true);
743
744 printf("cpu%u: address=%p, nrdy=%ld, needs_relink=%" PRIs "\n",
745 cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy),
746 cpus[cpu].needs_relink);
747
748 unsigned int i;
749 for (i = 0; i < RQ_COUNT; i++) {
750 irq_spinlock_lock(&(cpus[cpu].rq[i].lock), false);
751 if (cpus[cpu].rq[i].n == 0) {
752 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
753 continue;
754 }
755
756 printf("\trq[%u]: ", i);
757 link_t *cur;
758 for (cur = cpus[cpu].rq[i].rq_head.next;
759 cur != &(cpus[cpu].rq[i].rq_head);
760 cur = cur->next) {
761 thread_t *thread = list_get_instance(cur, thread_t, rq_link);
762 printf("%" PRIu64 "(%s) ", thread->tid,
763 thread_states[thread->state]);
764 }
765 printf("\n");
766
767 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
768 }
769
770 irq_spinlock_unlock(&cpus[cpu].lock, true);
771 }
772}
773
774/** @}
775 */
Note: See TracBrowser for help on using the repository browser.