source: mainline/kernel/generic/src/proc/scheduler.c@ b2fa1204

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since b2fa1204 was b2fa1204, checked in by Martin Sucha <sucha14@…>, 11 years ago

Cherrypick usage of kernel logger
  • 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 <log.h>
64#include <debug.h>
65#include <stacktrace.h>
66
67static void scheduler_separated_stack(void);
68
69atomic_t nrdy; /**< Number of ready threads in the system. */
70
71/** Carry out actions before new task runs. */
72static void before_task_runs(void)
73{
74 before_task_runs_arch();
75}
76
77/** Take actions before new thread runs.
78 *
79 * Perform actions that need to be
80 * taken before the newly selected
81 * thread is passed control.
82 *
83 * THREAD->lock is locked on entry
84 *
85 */
86static void before_thread_runs(void)
87{
88 before_thread_runs_arch();
89
90#ifdef CONFIG_FPU_LAZY
91 if (THREAD == CPU->fpu_owner)
92 fpu_enable();
93 else
94 fpu_disable();
95#elif defined CONFIG_FPU
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 = true;
102 }
103#endif
104
105#ifdef CONFIG_UDEBUG
106 if (THREAD->btrace) {
107 istate_t *istate = THREAD->udebug.uspace_state;
108 if (istate != NULL) {
109 printf("Thread %" PRIu64 " stack trace:\n", THREAD->tid);
110 stack_trace_istate(istate);
111 }
112
113 THREAD->btrace = false;
114 }
115#endif
116}
117
118/** Take actions after THREAD had run.
119 *
120 * Perform actions that need to be
121 * taken after the running thread
122 * had been preempted by the scheduler.
123 *
124 * THREAD->lock is locked on entry
125 *
126 */
127static void after_thread_ran(void)
128{
129 after_thread_ran_arch();
130}
131
132#ifdef CONFIG_FPU_LAZY
133void scheduler_fpu_lazy_request(void)
134{
135restart:
136 fpu_enable();
137 irq_spinlock_lock(&CPU->lock, false);
138
139 /* Save old context */
140 if (CPU->fpu_owner != NULL) {
141 irq_spinlock_lock(&CPU->fpu_owner->lock, false);
142 fpu_context_save(CPU->fpu_owner->saved_fpu_context);
143
144 /* Don't prevent migration */
145 CPU->fpu_owner->fpu_context_engaged = false;
146 irq_spinlock_unlock(&CPU->fpu_owner->lock, false);
147 CPU->fpu_owner = NULL;
148 }
149
150 irq_spinlock_lock(&THREAD->lock, false);
151 if (THREAD->fpu_context_exists) {
152 fpu_context_restore(THREAD->saved_fpu_context);
153 } else {
154 /* Allocate FPU context */
155 if (!THREAD->saved_fpu_context) {
156 /* Might sleep */
157 irq_spinlock_unlock(&THREAD->lock, false);
158 irq_spinlock_unlock(&CPU->lock, false);
159 THREAD->saved_fpu_context =
160 (fpu_context_t *) slab_alloc(fpu_context_slab, 0);
161
162 /* We may have switched CPUs during slab_alloc */
163 goto restart;
164 }
165 fpu_init();
166 THREAD->fpu_context_exists = true;
167 }
168
169 CPU->fpu_owner = THREAD;
170 THREAD->fpu_context_engaged = true;
171 irq_spinlock_unlock(&THREAD->lock, false);
172
173 irq_spinlock_unlock(&CPU->lock, false);
174}
175#endif /* CONFIG_FPU_LAZY */
176
177/** Initialize scheduler
178 *
179 * Initialize kernel scheduler.
180 *
181 */
182void scheduler_init(void)
183{
184}
185
186/** Get thread to be scheduled
187 *
188 * Get the optimal thread to be scheduled
189 * according to thread accounting and scheduler
190 * policy.
191 *
192 * @return Thread to be scheduled.
193 *
194 */
195static thread_t *find_best_thread(void)
196{
197 ASSERT(CPU != NULL);
198
199loop:
200
201 if (atomic_get(&CPU->nrdy) == 0) {
202 /*
203 * For there was nothing to run, the CPU goes to sleep
204 * until a hardware interrupt or an IPI comes.
205 * This improves energy saving and hyperthreading.
206 */
207 irq_spinlock_lock(&CPU->lock, false);
208 CPU->idle = true;
209 irq_spinlock_unlock(&CPU->lock, false);
210 interrupts_enable();
211
212 /*
213 * An interrupt might occur right now and wake up a thread.
214 * In such case, the CPU will continue to go to sleep
215 * even though there is a runnable thread.
216 */
217 cpu_sleep();
218 interrupts_disable();
219 goto loop;
220 }
221
222 unsigned int i;
223 for (i = 0; i < RQ_COUNT; i++) {
224 irq_spinlock_lock(&(CPU->rq[i].lock), false);
225 if (CPU->rq[i].n == 0) {
226 /*
227 * If this queue is empty, try a lower-priority queue.
228 */
229 irq_spinlock_unlock(&(CPU->rq[i].lock), false);
230 continue;
231 }
232
233 atomic_dec(&CPU->nrdy);
234 atomic_dec(&nrdy);
235 CPU->rq[i].n--;
236
237 /*
238 * Take the first thread from the queue.
239 */
240 thread_t *thread = list_get_instance(
241 list_first(&CPU->rq[i].rq), thread_t, rq_link);
242 list_remove(&thread->rq_link);
243
244 irq_spinlock_pass(&(CPU->rq[i].lock), &thread->lock);
245
246 thread->cpu = CPU;
247 thread->ticks = us2ticks((i + 1) * 10000);
248 thread->priority = i; /* Correct rq index */
249
250 /*
251 * Clear the stolen flag so that it can be migrated
252 * when load balancing needs emerge.
253 */
254 thread->stolen = false;
255 irq_spinlock_unlock(&thread->lock, false);
256
257 return thread;
258 }
259
260 goto loop;
261}
262
263/** Prevent rq starvation
264 *
265 * Prevent low priority threads from starving in rq's.
266 *
267 * When the function decides to relink rq's, it reconnects
268 * respective pointers so that in result threads with 'pri'
269 * greater or equal start are moved to a higher-priority queue.
270 *
271 * @param start Threshold priority.
272 *
273 */
274static void relink_rq(int start)
275{
276 list_t list;
277
278 list_initialize(&list);
279 irq_spinlock_lock(&CPU->lock, false);
280
281 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
282 int i;
283 for (i = start; i < RQ_COUNT - 1; i++) {
284 /* Remember and empty rq[i + 1] */
285
286 irq_spinlock_lock(&CPU->rq[i + 1].lock, false);
287 list_concat(&list, &CPU->rq[i + 1].rq);
288 size_t n = CPU->rq[i + 1].n;
289 CPU->rq[i + 1].n = 0;
290 irq_spinlock_unlock(&CPU->rq[i + 1].lock, false);
291
292 /* Append rq[i + 1] to rq[i] */
293
294 irq_spinlock_lock(&CPU->rq[i].lock, false);
295 list_concat(&CPU->rq[i].rq, &list);
296 CPU->rq[i].n += n;
297 irq_spinlock_unlock(&CPU->rq[i].lock, false);
298 }
299
300 CPU->needs_relink = 0;
301 }
302
303 irq_spinlock_unlock(&CPU->lock, false);
304}
305
306/** The scheduler
307 *
308 * The thread scheduling procedure.
309 * Passes control directly to
310 * scheduler_separated_stack().
311 *
312 */
313void scheduler(void)
314{
315 volatile ipl_t ipl;
316
317 ASSERT(CPU != NULL);
318
319 ipl = interrupts_disable();
320
321 if (atomic_get(&haltstate))
322 halt();
323
324 if (THREAD) {
325 irq_spinlock_lock(&THREAD->lock, false);
326
327 /* Update thread kernel accounting */
328 THREAD->kcycles += get_cycle() - THREAD->last_cycle;
329
330#if (defined CONFIG_FPU) && (!defined CONFIG_FPU_LAZY)
331 fpu_context_save(THREAD->saved_fpu_context);
332#endif
333 if (!context_save(&THREAD->saved_context)) {
334 /*
335 * This is the place where threads leave scheduler();
336 */
337
338 /* Save current CPU cycle */
339 THREAD->last_cycle = get_cycle();
340
341 irq_spinlock_unlock(&THREAD->lock, false);
342 interrupts_restore(THREAD->saved_context.ipl);
343
344 return;
345 }
346
347 /*
348 * Interrupt priority level of preempted thread is recorded
349 * here to facilitate scheduler() invocations from
350 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
351 *
352 */
353 THREAD->saved_context.ipl = ipl;
354 }
355
356 /*
357 * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, AS
358 * and preemption counter. At this point THE could be coming either
359 * from THREAD's or CPU's stack.
360 *
361 */
362 the_copy(THE, (the_t *) CPU->stack);
363
364 /*
365 * We may not keep the old stack.
366 * Reason: If we kept the old stack and got blocked, for instance, in
367 * find_best_thread(), the old thread could get rescheduled by another
368 * CPU and overwrite the part of its own stack that was also used by
369 * the scheduler on this CPU.
370 *
371 * Moreover, we have to bypass the compiler-generated POP sequence
372 * which is fooled by SP being set to the very top of the stack.
373 * Therefore the scheduler() function continues in
374 * scheduler_separated_stack().
375 *
376 */
377 context_save(&CPU->saved_context);
378 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack),
379 (uintptr_t) CPU->stack, STACK_SIZE);
380 context_restore(&CPU->saved_context);
381
382 /* Not reached */
383}
384
385/** Scheduler stack switch wrapper
386 *
387 * Second part of the scheduler() function
388 * using new stack. Handling the actual context
389 * switch to a new thread.
390 *
391 */
392void scheduler_separated_stack(void)
393{
394 DEADLOCK_PROBE_INIT(p_joinwq);
395 task_t *old_task = TASK;
396 as_t *old_as = AS;
397
398 ASSERT((!THREAD) || (irq_spinlock_locked(&THREAD->lock)));
399 ASSERT(CPU != NULL);
400
401 /*
402 * Hold the current task and the address space to prevent their
403 * possible destruction should thread_destroy() be called on this or any
404 * other processor while the scheduler is still using them.
405 */
406 if (old_task)
407 task_hold(old_task);
408
409 if (old_as)
410 as_hold(old_as);
411
412 if (THREAD) {
413 /* Must be run after the switch to scheduler stack */
414 after_thread_ran();
415
416 switch (THREAD->state) {
417 case Running:
418 irq_spinlock_unlock(&THREAD->lock, false);
419 thread_ready(THREAD);
420 break;
421
422 case Exiting:
423repeat:
424 if (THREAD->detached) {
425 thread_destroy(THREAD, false);
426 } else {
427 /*
428 * The thread structure is kept allocated until
429 * somebody calls thread_detach() on it.
430 */
431 if (!irq_spinlock_trylock(&THREAD->join_wq.lock)) {
432 /*
433 * Avoid deadlock.
434 */
435 irq_spinlock_unlock(&THREAD->lock, false);
436 delay(HZ);
437 irq_spinlock_lock(&THREAD->lock, false);
438 DEADLOCK_PROBE(p_joinwq,
439 DEADLOCK_THRESHOLD);
440 goto repeat;
441 }
442 _waitq_wakeup_unsafe(&THREAD->join_wq,
443 WAKEUP_FIRST);
444 irq_spinlock_unlock(&THREAD->join_wq.lock, false);
445
446 THREAD->state = Lingering;
447 irq_spinlock_unlock(&THREAD->lock, false);
448 }
449 break;
450
451 case Sleeping:
452 /*
453 * Prefer the thread after it's woken up.
454 */
455 THREAD->priority = -1;
456
457 /*
458 * We need to release wq->lock which we locked in
459 * waitq_sleep(). Address of wq->lock is kept in
460 * THREAD->sleep_queue.
461 */
462 irq_spinlock_unlock(&THREAD->sleep_queue->lock, false);
463
464 irq_spinlock_unlock(&THREAD->lock, false);
465 break;
466
467 default:
468 /*
469 * Entering state is unexpected.
470 */
471 panic("tid%" PRIu64 ": unexpected state %s.",
472 THREAD->tid, thread_states[THREAD->state]);
473 break;
474 }
475
476 THREAD = NULL;
477 }
478
479 THREAD = find_best_thread();
480
481 irq_spinlock_lock(&THREAD->lock, false);
482 int priority = THREAD->priority;
483 irq_spinlock_unlock(&THREAD->lock, false);
484
485 relink_rq(priority);
486
487 /*
488 * If both the old and the new task are the same,
489 * lots of work is avoided.
490 */
491 if (TASK != THREAD->task) {
492 as_t *new_as = THREAD->task->as;
493
494 /*
495 * Note that it is possible for two tasks
496 * to share one address space.
497 */
498 if (old_as != new_as) {
499 /*
500 * Both tasks and address spaces are different.
501 * Replace the old one with the new one.
502 */
503 as_switch(old_as, new_as);
504 }
505
506 TASK = THREAD->task;
507 before_task_runs();
508 }
509
510 if (old_task)
511 task_release(old_task);
512
513 if (old_as)
514 as_release(old_as);
515
516 irq_spinlock_lock(&THREAD->lock, false);
517 THREAD->state = Running;
518
519#ifdef SCHEDULER_VERBOSE
520 log(LF_OTHER, LVL_DEBUG,
521 "cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
522 ", nrdy=%" PRIua ")", CPU->id, THREAD->tid, THREAD->priority,
523 THREAD->ticks, atomic_get(&CPU->nrdy));
524#endif
525
526 /*
527 * Some architectures provide late kernel PA2KA(identity)
528 * mapping in a page fault handler. However, the page fault
529 * handler uses the kernel stack of the running thread and
530 * therefore cannot be used to map it. The kernel stack, if
531 * necessary, is to be mapped in before_thread_runs(). This
532 * function must be executed before the switch to the new stack.
533 */
534 before_thread_runs();
535
536 /*
537 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
538 * thread's stack.
539 */
540 the_copy(THE, (the_t *) THREAD->kstack);
541
542 context_restore(&THREAD->saved_context);
543
544 /* Not reached */
545}
546
547#ifdef CONFIG_SMP
548/** Load balancing thread
549 *
550 * SMP load balancing thread, supervising thread supplies
551 * for the CPU it's wired to.
552 *
553 * @param arg Generic thread argument (unused).
554 *
555 */
556void kcpulb(void *arg)
557{
558 atomic_count_t average;
559 atomic_count_t rdy;
560
561 /*
562 * Detach kcpulb as nobody will call thread_join_timeout() on it.
563 */
564 thread_detach(THREAD);
565
566loop:
567 /*
568 * Work in 1s intervals.
569 */
570 thread_sleep(1);
571
572not_satisfied:
573 /*
574 * Calculate the number of threads that will be migrated/stolen from
575 * other CPU's. Note that situation can have changed between two
576 * passes. Each time get the most up to date counts.
577 *
578 */
579 average = atomic_get(&nrdy) / config.cpu_active + 1;
580 rdy = atomic_get(&CPU->nrdy);
581
582 if (average <= rdy)
583 goto satisfied;
584
585 atomic_count_t count = average - rdy;
586
587 /*
588 * Searching least priority queues on all CPU's first and most priority
589 * queues on all CPU's last.
590 */
591 size_t acpu;
592 size_t acpu_bias = 0;
593 int rq;
594
595 for (rq = RQ_COUNT - 1; rq >= 0; rq--) {
596 for (acpu = 0; acpu < config.cpu_active; acpu++) {
597 cpu_t *cpu = &cpus[(acpu + acpu_bias) % config.cpu_active];
598
599 /*
600 * Not interested in ourselves.
601 * Doesn't require interrupt disabling for kcpulb has
602 * THREAD_FLAG_WIRED.
603 *
604 */
605 if (CPU == cpu)
606 continue;
607
608 if (atomic_get(&cpu->nrdy) <= average)
609 continue;
610
611 irq_spinlock_lock(&(cpu->rq[rq].lock), true);
612 if (cpu->rq[rq].n == 0) {
613 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
614 continue;
615 }
616
617 thread_t *thread = NULL;
618
619 /* Search rq from the back */
620 link_t *link = cpu->rq[rq].rq.head.prev;
621
622 while (link != &(cpu->rq[rq].rq.head)) {
623 thread = (thread_t *) list_get_instance(link,
624 thread_t, rq_link);
625
626 /*
627 * Do not steal CPU-wired threads, threads
628 * already stolen, threads for which migration
629 * was temporarily disabled or threads whose
630 * FPU context is still in the CPU.
631 */
632 irq_spinlock_lock(&thread->lock, false);
633
634 if ((!thread->wired) && (!thread->stolen) &&
635 (!thread->nomigrate) &&
636 (!thread->fpu_context_engaged)) {
637 /*
638 * Remove thread from ready queue.
639 */
640 irq_spinlock_unlock(&thread->lock,
641 false);
642
643 atomic_dec(&cpu->nrdy);
644 atomic_dec(&nrdy);
645
646 cpu->rq[rq].n--;
647 list_remove(&thread->rq_link);
648
649 break;
650 }
651
652 irq_spinlock_unlock(&thread->lock, false);
653
654 link = link->prev;
655 thread = NULL;
656 }
657
658 if (thread) {
659 /*
660 * Ready thread on local CPU
661 */
662
663 irq_spinlock_pass(&(cpu->rq[rq].lock),
664 &thread->lock);
665
666#ifdef KCPULB_VERBOSE
667 log(LF_OTHER, LVL_DEBUG,
668 "kcpulb%u: TID %" PRIu64 " -> cpu%u, "
669 "nrdy=%ld, avg=%ld", CPU->id, t->tid,
670 CPU->id, atomic_get(&CPU->nrdy),
671 atomic_get(&nrdy) / config.cpu_active);
672#endif
673
674 thread->stolen = true;
675 thread->state = Entering;
676
677 irq_spinlock_unlock(&thread->lock, true);
678 thread_ready(thread);
679
680 if (--count == 0)
681 goto satisfied;
682
683 /*
684 * We are not satisfied yet, focus on another
685 * CPU next time.
686 *
687 */
688 acpu_bias++;
689
690 continue;
691 } else
692 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
693
694 }
695 }
696
697 if (atomic_get(&CPU->nrdy)) {
698 /*
699 * Be a little bit light-weight and let migrated threads run.
700 *
701 */
702 scheduler();
703 } else {
704 /*
705 * We failed to migrate a single thread.
706 * Give up this turn.
707 *
708 */
709 goto loop;
710 }
711
712 goto not_satisfied;
713
714satisfied:
715 goto loop;
716}
717#endif /* CONFIG_SMP */
718
719/** Print information about threads & scheduler queues
720 *
721 */
722void sched_print_list(void)
723{
724 size_t cpu;
725 for (cpu = 0; cpu < config.cpu_count; cpu++) {
726 if (!cpus[cpu].active)
727 continue;
728
729 irq_spinlock_lock(&cpus[cpu].lock, true);
730
731 printf("cpu%u: address=%p, nrdy=%" PRIua ", needs_relink=%zu\n",
732 cpus[cpu].id, &cpus[cpu], atomic_get(&cpus[cpu].nrdy),
733 cpus[cpu].needs_relink);
734
735 unsigned int i;
736 for (i = 0; i < RQ_COUNT; i++) {
737 irq_spinlock_lock(&(cpus[cpu].rq[i].lock), false);
738 if (cpus[cpu].rq[i].n == 0) {
739 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
740 continue;
741 }
742
743 printf("\trq[%u]: ", i);
744 list_foreach(cpus[cpu].rq[i].rq, rq_link, thread_t,
745 thread) {
746 printf("%" PRIu64 "(%s) ", thread->tid,
747 thread_states[thread->state]);
748 }
749 printf("\n");
750
751 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
752 }
753
754 irq_spinlock_unlock(&cpus[cpu].lock, true);
755 }
756}
757
758/** @}
759 */
Note: See TracBrowser for help on using the repository browser.