source: mainline/kernel/generic/src/proc/scheduler.c@ 111b9b9

ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since 111b9b9 was 111b9b9, checked in by Jiří Zárevúcky <zarevucky.jiri@…>, 2 years ago

Reimplement waitq using thread_wait/wakeup

This adds a few functions to the thread API which can be
summarized as "stop running until woken up by others".
The ordering and context-switching concerns are thus yeeted
to this abstraction and waitq only deals with maintaining
the queues. Overall, this makes the control flow in waitq
much easier to navigate.
  • Property mode set to 100644
File size: 16.7 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 kernel_generic_proc
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 <assert.h>
42#include <atomic.h>
43#include <proc/scheduler.h>
44#include <proc/thread.h>
45#include <proc/task.h>
46#include <mm/frame.h>
47#include <mm/page.h>
48#include <mm/as.h>
49#include <time/timeout.h>
50#include <time/delay.h>
51#include <arch/asm.h>
52#include <arch/faddr.h>
53#include <arch/cycle.h>
54#include <atomic.h>
55#include <synch/spinlock.h>
56#include <config.h>
57#include <context.h>
58#include <fpu_context.h>
59#include <halt.h>
60#include <arch.h>
61#include <adt/list.h>
62#include <panic.h>
63#include <cpu.h>
64#include <stdio.h>
65#include <log.h>
66#include <stacktrace.h>
67
68static void scheduler_separated_stack(void);
69
70atomic_size_t nrdy; /**< Number of ready threads in the system. */
71
72/** Carry out actions before new task runs. */
73static void 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 * thread is passed control.
83 *
84 * THREAD->lock is locked on entry
85 *
86 */
87static void before_thread_runs(void)
88{
89 before_thread_runs_arch();
90
91#ifdef CONFIG_FPU_LAZY
92 if (THREAD == CPU->fpu_owner)
93 fpu_enable();
94 else
95 fpu_disable();
96#elif defined CONFIG_FPU
97 fpu_enable();
98 if (THREAD->fpu_context_exists)
99 fpu_context_restore(&THREAD->fpu_context);
100 else {
101 fpu_init();
102 THREAD->fpu_context_exists = true;
103 }
104#endif
105
106#ifdef CONFIG_UDEBUG
107 if (THREAD->btrace) {
108 istate_t *istate = THREAD->udebug.uspace_state;
109 if (istate != NULL) {
110 printf("Thread %" PRIu64 " stack trace:\n", THREAD->tid);
111 stack_trace_istate(istate);
112 }
113
114 THREAD->btrace = false;
115 }
116#endif
117}
118
119/** Take actions after THREAD had run.
120 *
121 * Perform actions that need to be
122 * taken after the running thread
123 * had been preempted by the scheduler.
124 *
125 * THREAD->lock is locked on entry
126 *
127 */
128static void after_thread_ran(void)
129{
130 after_thread_ran_arch();
131}
132
133#ifdef CONFIG_FPU_LAZY
134void scheduler_fpu_lazy_request(void)
135{
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->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->fpu_context);
153 } else {
154 fpu_init();
155 THREAD->fpu_context_exists = true;
156 }
157
158 CPU->fpu_owner = THREAD;
159 THREAD->fpu_context_engaged = true;
160 irq_spinlock_unlock(&THREAD->lock, false);
161
162 irq_spinlock_unlock(&CPU->lock, false);
163}
164#endif /* CONFIG_FPU_LAZY */
165
166/** Initialize scheduler
167 *
168 * Initialize kernel scheduler.
169 *
170 */
171void scheduler_init(void)
172{
173}
174
175/** Get thread to be scheduled
176 *
177 * Get the optimal thread to be scheduled
178 * according to thread accounting and scheduler
179 * policy.
180 *
181 * @return Thread to be scheduled.
182 *
183 */
184static thread_t *find_best_thread(void)
185{
186 assert(CPU != NULL);
187
188loop:
189
190 if (atomic_load(&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 irq_spinlock_lock(&CPU->lock, false);
197 CPU->idle = true;
198 irq_spinlock_unlock(&CPU->lock, false);
199 interrupts_enable();
200
201 /*
202 * An interrupt might occur right now and wake up a thread.
203 * In such case, the CPU will continue to go to sleep
204 * even though there is a runnable thread.
205 */
206 cpu_sleep();
207 interrupts_disable();
208 goto loop;
209 }
210
211 assert(!CPU->idle);
212
213 unsigned int i;
214 for (i = 0; i < RQ_COUNT; i++) {
215 irq_spinlock_lock(&(CPU->rq[i].lock), false);
216 if (CPU->rq[i].n == 0) {
217 /*
218 * If this queue is empty, try a lower-priority queue.
219 */
220 irq_spinlock_unlock(&(CPU->rq[i].lock), false);
221 continue;
222 }
223
224 atomic_dec(&CPU->nrdy);
225 atomic_dec(&nrdy);
226 CPU->rq[i].n--;
227
228 /*
229 * Take the first thread from the queue.
230 */
231 thread_t *thread = list_get_instance(
232 list_first(&CPU->rq[i].rq), thread_t, rq_link);
233 list_remove(&thread->rq_link);
234
235 irq_spinlock_pass(&(CPU->rq[i].lock), &thread->lock);
236
237 thread->cpu = CPU;
238 thread->priority = i; /* Correct rq index */
239
240 /* Time allocation in microseconds. */
241 uint64_t time_to_run = (i + 1) * 10000;
242
243 /* This is safe because interrupts are disabled. */
244 CPU->preempt_deadline = CPU->current_clock_tick + us2ticks(time_to_run);
245
246 /*
247 * Clear the stolen flag so that it can be migrated
248 * when load balancing needs emerge.
249 */
250 thread->stolen = false;
251 irq_spinlock_unlock(&thread->lock, false);
252
253 return thread;
254 }
255
256 goto loop;
257}
258
259/** Prevent rq starvation
260 *
261 * Prevent low priority threads from starving in rq's.
262 *
263 * When the function decides to relink rq's, it reconnects
264 * respective pointers so that in result threads with 'pri'
265 * greater or equal start are moved to a higher-priority queue.
266 *
267 * @param start Threshold priority.
268 *
269 */
270static void relink_rq(int start)
271{
272 if (CPU->current_clock_tick < CPU->relink_deadline)
273 return;
274
275 CPU->relink_deadline = CPU->current_clock_tick + NEEDS_RELINK_MAX;
276
277 list_t list;
278 list_initialize(&list);
279
280 irq_spinlock_lock(&CPU->lock, false);
281
282 for (int i = start; i < RQ_COUNT - 1; i++) {
283 /* Remember and empty rq[i + 1] */
284
285 irq_spinlock_lock(&CPU->rq[i + 1].lock, false);
286 list_concat(&list, &CPU->rq[i + 1].rq);
287 size_t n = CPU->rq[i + 1].n;
288 CPU->rq[i + 1].n = 0;
289 irq_spinlock_unlock(&CPU->rq[i + 1].lock, false);
290
291 /* Append rq[i + 1] to rq[i] */
292
293 irq_spinlock_lock(&CPU->rq[i].lock, false);
294 list_concat(&CPU->rq[i].rq, &list);
295 CPU->rq[i].n += n;
296 irq_spinlock_unlock(&CPU->rq[i].lock, false);
297 }
298
299 irq_spinlock_unlock(&CPU->lock, false);
300}
301
302void scheduler(void)
303{
304 ipl_t ipl = interrupts_disable();
305
306 if (atomic_load(&haltstate))
307 halt();
308
309 if (THREAD) {
310 irq_spinlock_lock(&THREAD->lock, false);
311 }
312
313 scheduler_locked(ipl);
314}
315
316/** The scheduler
317 *
318 * The thread scheduling procedure.
319 * Passes control directly to
320 * scheduler_separated_stack().
321 *
322 */
323void scheduler_locked(ipl_t ipl)
324{
325 assert(CPU != NULL);
326
327 if (THREAD) {
328 /* Update thread kernel accounting */
329 THREAD->kcycles += get_cycle() - THREAD->last_cycle;
330
331#if (defined CONFIG_FPU) && (!defined CONFIG_FPU_LAZY)
332 fpu_context_save(&THREAD->fpu_context);
333#endif
334 if (!context_save(&THREAD->saved_context)) {
335 /*
336 * This is the place where threads leave scheduler();
337 */
338
339 /* Save current CPU cycle */
340 THREAD->last_cycle = get_cycle();
341
342 irq_spinlock_unlock(&THREAD->lock, false);
343 interrupts_restore(THREAD->saved_ipl);
344
345 return;
346 }
347
348 /*
349 * Interrupt priority level of preempted thread is recorded
350 * here to facilitate scheduler() invocations from
351 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
352 *
353 */
354 THREAD->saved_ipl = ipl;
355 }
356
357 /*
358 * Through the 'CURRENT' structure, we keep track of THREAD, TASK, CPU, AS
359 * and preemption counter. At this point CURRENT could be coming either
360 * from THREAD's or CPU's stack.
361 *
362 */
363 current_copy(CURRENT, (current_t *) CPU->stack);
364
365 /*
366 * We may not keep the old stack.
367 * Reason: If we kept the old stack and got blocked, for instance, in
368 * find_best_thread(), the old thread could get rescheduled by another
369 * CPU and overwrite the part of its own stack that was also used by
370 * the scheduler on this CPU.
371 *
372 * Moreover, we have to bypass the compiler-generated POP sequence
373 * which is fooled by SP being set to the very top of the stack.
374 * Therefore the scheduler() function continues in
375 * scheduler_separated_stack().
376 *
377 */
378 context_t ctx;
379 context_save(&ctx);
380 context_set(&ctx, FADDR(scheduler_separated_stack),
381 (uintptr_t) CPU->stack, STACK_SIZE);
382 context_restore(&ctx);
383
384 /* Not reached */
385}
386
387/** Scheduler stack switch wrapper
388 *
389 * Second part of the scheduler() function
390 * using new stack. Handling the actual context
391 * switch to a new thread.
392 *
393 */
394void scheduler_separated_stack(void)
395{
396 task_t *old_task = TASK;
397 as_t *old_as = AS;
398
399 assert((!THREAD) || (irq_spinlock_locked(&THREAD->lock)));
400 assert(CPU != NULL);
401 assert(interrupts_disabled());
402
403 /*
404 * Hold the current task and the address space to prevent their
405 * possible destruction should thread_destroy() be called on this or any
406 * other processor while the scheduler is still using them.
407 */
408 if (old_task)
409 task_hold(old_task);
410
411 if (old_as)
412 as_hold(old_as);
413
414 if (THREAD) {
415 /* Must be run after the switch to scheduler stack */
416 after_thread_ran();
417
418 switch (THREAD->state) {
419 case Running:
420 irq_spinlock_unlock(&THREAD->lock, false);
421 thread_ready(THREAD);
422 break;
423
424 case Exiting:
425 irq_spinlock_unlock(&THREAD->lock, false);
426 waitq_close(&THREAD->join_wq);
427
428 /*
429 * Release the reference CPU has for the thread.
430 * If there are no other references (e.g. threads calling join),
431 * the thread structure is deallocated.
432 */
433 thread_put(THREAD);
434 break;
435
436 case Sleeping:
437 /*
438 * Prefer the thread after it's woken up.
439 */
440 THREAD->priority = -1;
441 irq_spinlock_unlock(&THREAD->lock, false);
442 break;
443
444 default:
445 /*
446 * Entering state is unexpected.
447 */
448 panic("tid%" PRIu64 ": unexpected state %s.",
449 THREAD->tid, thread_states[THREAD->state]);
450 break;
451 }
452
453 THREAD = NULL;
454 }
455
456 THREAD = find_best_thread();
457
458 irq_spinlock_lock(&THREAD->lock, false);
459 int priority = THREAD->priority;
460 irq_spinlock_unlock(&THREAD->lock, false);
461
462 relink_rq(priority);
463
464 /*
465 * If both the old and the new task are the same,
466 * lots of work is avoided.
467 */
468 if (TASK != THREAD->task) {
469 as_t *new_as = THREAD->task->as;
470
471 /*
472 * Note that it is possible for two tasks
473 * to share one address space.
474 */
475 if (old_as != new_as) {
476 /*
477 * Both tasks and address spaces are different.
478 * Replace the old one with the new one.
479 */
480 as_switch(old_as, new_as);
481 }
482
483 TASK = THREAD->task;
484 before_task_runs();
485 }
486
487 if (old_task)
488 task_release(old_task);
489
490 if (old_as)
491 as_release(old_as);
492
493 irq_spinlock_lock(&THREAD->lock, false);
494 THREAD->state = Running;
495
496#ifdef SCHEDULER_VERBOSE
497 log(LF_OTHER, LVL_DEBUG,
498 "cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
499 ", nrdy=%zu)", CPU->id, THREAD->tid, THREAD->priority,
500 THREAD->ticks, atomic_load(&CPU->nrdy));
501#endif
502
503 /*
504 * Some architectures provide late kernel PA2KA(identity)
505 * mapping in a page fault handler. However, the page fault
506 * handler uses the kernel stack of the running thread and
507 * therefore cannot be used to map it. The kernel stack, if
508 * necessary, is to be mapped in before_thread_runs(). This
509 * function must be executed before the switch to the new stack.
510 */
511 before_thread_runs();
512
513 /*
514 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
515 * thread's stack.
516 */
517 current_copy(CURRENT, (current_t *) THREAD->kstack);
518
519 context_restore(&THREAD->saved_context);
520
521 /* Not reached */
522}
523
524#ifdef CONFIG_SMP
525/** Load balancing thread
526 *
527 * SMP load balancing thread, supervising thread supplies
528 * for the CPU it's wired to.
529 *
530 * @param arg Generic thread argument (unused).
531 *
532 */
533void kcpulb(void *arg)
534{
535 size_t average;
536 size_t rdy;
537
538loop:
539 /*
540 * Work in 1s intervals.
541 */
542 thread_sleep(1);
543
544not_satisfied:
545 /*
546 * Calculate the number of threads that will be migrated/stolen from
547 * other CPU's. Note that situation can have changed between two
548 * passes. Each time get the most up to date counts.
549 *
550 */
551 average = atomic_load(&nrdy) / config.cpu_active + 1;
552 rdy = atomic_load(&CPU->nrdy);
553
554 if (average <= rdy)
555 goto satisfied;
556
557 size_t count = average - rdy;
558
559 /*
560 * Searching least priority queues on all CPU's first and most priority
561 * queues on all CPU's last.
562 */
563 size_t acpu;
564 size_t acpu_bias = 0;
565 int rq;
566
567 for (rq = RQ_COUNT - 1; rq >= 0; rq--) {
568 for (acpu = 0; acpu < config.cpu_active; acpu++) {
569 cpu_t *cpu = &cpus[(acpu + acpu_bias) % config.cpu_active];
570
571 /*
572 * Not interested in ourselves.
573 * Doesn't require interrupt disabling for kcpulb has
574 * THREAD_FLAG_WIRED.
575 *
576 */
577 if (CPU == cpu)
578 continue;
579
580 if (atomic_load(&cpu->nrdy) <= average)
581 continue;
582
583 irq_spinlock_lock(&(cpu->rq[rq].lock), true);
584 if (cpu->rq[rq].n == 0) {
585 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
586 continue;
587 }
588
589 thread_t *thread = NULL;
590
591 /* Search rq from the back */
592 link_t *link = list_last(&cpu->rq[rq].rq);
593
594 while (link != NULL) {
595 thread = (thread_t *) list_get_instance(link,
596 thread_t, rq_link);
597
598 /*
599 * Do not steal CPU-wired threads, threads
600 * already stolen, threads for which migration
601 * was temporarily disabled or threads whose
602 * FPU context is still in the CPU.
603 */
604 irq_spinlock_lock(&thread->lock, false);
605
606 if ((!thread->wired) && (!thread->stolen) &&
607 (!thread->nomigrate) &&
608 (!thread->fpu_context_engaged)) {
609 /*
610 * Remove thread from ready queue.
611 */
612 irq_spinlock_unlock(&thread->lock,
613 false);
614
615 atomic_dec(&cpu->nrdy);
616 atomic_dec(&nrdy);
617
618 cpu->rq[rq].n--;
619 list_remove(&thread->rq_link);
620
621 break;
622 }
623
624 irq_spinlock_unlock(&thread->lock, false);
625
626 link = list_prev(link, &cpu->rq[rq].rq);
627 thread = NULL;
628 }
629
630 if (thread) {
631 /*
632 * Ready thread on local CPU
633 */
634
635 irq_spinlock_pass(&(cpu->rq[rq].lock),
636 &thread->lock);
637
638#ifdef KCPULB_VERBOSE
639 log(LF_OTHER, LVL_DEBUG,
640 "kcpulb%u: TID %" PRIu64 " -> cpu%u, "
641 "nrdy=%ld, avg=%ld", CPU->id, thread->tid,
642 CPU->id, atomic_load(&CPU->nrdy),
643 atomic_load(&nrdy) / config.cpu_active);
644#endif
645
646 thread->stolen = true;
647 thread->state = Entering;
648
649 irq_spinlock_unlock(&thread->lock, true);
650 thread_ready(thread);
651
652 if (--count == 0)
653 goto satisfied;
654
655 /*
656 * We are not satisfied yet, focus on another
657 * CPU next time.
658 *
659 */
660 acpu_bias++;
661
662 continue;
663 } else
664 irq_spinlock_unlock(&(cpu->rq[rq].lock), true);
665
666 }
667 }
668
669 if (atomic_load(&CPU->nrdy)) {
670 /*
671 * Be a little bit light-weight and let migrated threads run.
672 *
673 */
674 scheduler();
675 } else {
676 /*
677 * We failed to migrate a single thread.
678 * Give up this turn.
679 *
680 */
681 goto loop;
682 }
683
684 goto not_satisfied;
685
686satisfied:
687 goto loop;
688}
689#endif /* CONFIG_SMP */
690
691/** Print information about threads & scheduler queues
692 *
693 */
694void sched_print_list(void)
695{
696 size_t cpu;
697 for (cpu = 0; cpu < config.cpu_count; cpu++) {
698 if (!cpus[cpu].active)
699 continue;
700
701 irq_spinlock_lock(&cpus[cpu].lock, true);
702
703 /* Technically a data race, but we don't really care in this case. */
704 int needs_relink = cpus[cpu].relink_deadline - cpus[cpu].current_clock_tick;
705
706 printf("cpu%u: address=%p, nrdy=%zu, needs_relink=%d\n",
707 cpus[cpu].id, &cpus[cpu], atomic_load(&cpus[cpu].nrdy),
708 needs_relink);
709
710 unsigned int i;
711 for (i = 0; i < RQ_COUNT; i++) {
712 irq_spinlock_lock(&(cpus[cpu].rq[i].lock), false);
713 if (cpus[cpu].rq[i].n == 0) {
714 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
715 continue;
716 }
717
718 printf("\trq[%u]: ", i);
719 list_foreach(cpus[cpu].rq[i].rq, rq_link, thread_t,
720 thread) {
721 printf("%" PRIu64 "(%s) ", thread->tid,
722 thread_states[thread->state]);
723 }
724 printf("\n");
725
726 irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
727 }
728
729 irq_spinlock_unlock(&cpus[cpu].lock, true);
730 }
731}
732
733/** @}
734 */
Note: See TracBrowser for help on using the repository browser.