source: mainline/generic/src/proc/scheduler.c@ 444ec64

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since 444ec64 was 444ec64, checked in by Ondrej Palkovsky <ondrap@…>, 19 years ago

Cleanup of scheduler code
  • Property mode set to 100644
File size: 13.9 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#include <proc/scheduler.h>
30#include <proc/thread.h>
31#include <proc/task.h>
32#include <mm/heap.h>
33#include <mm/frame.h>
34#include <mm/page.h>
35#include <mm/as.h>
36#include <arch/asm.h>
37#include <arch/faddr.h>
38#include <arch/atomic.h>
39#include <synch/spinlock.h>
40#include <config.h>
41#include <context.h>
42#include <func.h>
43#include <arch.h>
44#include <list.h>
45#include <panic.h>
46#include <typedefs.h>
47#include <cpu.h>
48#include <print.h>
49#include <debug.h>
50
51atomic_t nrdy;
52
53/** Take actions before new thread runs
54 *
55 * Perform actions that need to be
56 * taken before the newly selected
57 * tread is passed control.
58 *
59 */
60void before_thread_runs(void)
61{
62 before_thread_runs_arch();
63#ifdef CONFIG_FPU_LAZY
64 if(THREAD==CPU->fpu_owner)
65 fpu_enable();
66 else
67 fpu_disable();
68#else
69 fpu_enable();
70 if (THREAD->fpu_context_exists)
71 fpu_context_restore(&(THREAD->saved_fpu_context));
72 else {
73 fpu_init();
74 THREAD->fpu_context_exists=1;
75 }
76#endif
77}
78
79#ifdef CONFIG_FPU_LAZY
80void scheduler_fpu_lazy_request(void)
81{
82 fpu_enable();
83 if (CPU->fpu_owner != NULL) {
84 fpu_context_save(&CPU->fpu_owner->saved_fpu_context);
85 /* don't prevent migration */
86 CPU->fpu_owner->fpu_context_engaged=0;
87 }
88 if (THREAD->fpu_context_exists)
89 fpu_context_restore(&THREAD->saved_fpu_context);
90 else {
91 fpu_init();
92 THREAD->fpu_context_exists=1;
93 }
94 CPU->fpu_owner=THREAD;
95 THREAD->fpu_context_engaged = 1;
96}
97#endif
98
99/** Initialize scheduler
100 *
101 * Initialize kernel scheduler.
102 *
103 */
104void scheduler_init(void)
105{
106}
107
108
109/** Get thread to be scheduled
110 *
111 * Get the optimal thread to be scheduled
112 * according to thread accounting and scheduler
113 * policy.
114 *
115 * @return Thread to be scheduled.
116 *
117 */
118static thread_t *find_best_thread(void)
119{
120 thread_t *t;
121 runq_t *r;
122 int i;
123
124 ASSERT(CPU != NULL);
125
126loop:
127 interrupts_enable();
128
129 if (atomic_get(&CPU->nrdy) == 0) {
130 /*
131 * For there was nothing to run, the CPU goes to sleep
132 * until a hardware interrupt or an IPI comes.
133 * This improves energy saving and hyperthreading.
134 *
135 * - we might get an interrupt here that makes some thread runnable,
136 * in such a case we must wait for the next quantum to come
137 */
138 cpu_sleep();
139 goto loop;
140 }
141
142 interrupts_disable();
143
144 i = 0;
145 for (; i<RQ_COUNT; i++) {
146 r = &CPU->rq[i];
147 spinlock_lock(&r->lock);
148 if (r->n == 0) {
149 /*
150 * If this queue is empty, try a lower-priority queue.
151 */
152 spinlock_unlock(&r->lock);
153 continue;
154 }
155
156 atomic_dec(&CPU->nrdy);
157 atomic_dec(&nrdy);
158 r->n--;
159
160 /*
161 * Take the first thread from the queue.
162 */
163 t = list_get_instance(r->rq_head.next, thread_t, rq_link);
164 list_remove(&t->rq_link);
165
166 spinlock_unlock(&r->lock);
167
168 spinlock_lock(&t->lock);
169 t->cpu = CPU;
170
171 t->ticks = us2ticks((i+1)*10000);
172 t->priority = i; /* eventually correct rq index */
173
174 /*
175 * Clear the X_STOLEN flag so that t can be migrated when load balancing needs emerge.
176 */
177 t->flags &= ~X_STOLEN;
178 spinlock_unlock(&t->lock);
179
180 return t;
181 }
182 goto loop;
183
184}
185
186
187/** Prevent rq starvation
188 *
189 * Prevent low priority threads from starving in rq's.
190 *
191 * When the function decides to relink rq's, it reconnects
192 * respective pointers so that in result threads with 'pri'
193 * greater or equal 'start' are moved to a higher-priority queue.
194 *
195 * @param start Threshold priority.
196 *
197 */
198static void relink_rq(int start)
199{
200 link_t head;
201 runq_t *r;
202 int i, n;
203
204 list_initialize(&head);
205 spinlock_lock(&CPU->lock);
206 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
207 for (i = start; i<RQ_COUNT-1; i++) {
208 /* remember and empty rq[i + 1] */
209 r = &CPU->rq[i + 1];
210 spinlock_lock(&r->lock);
211 list_concat(&head, &r->rq_head);
212 n = r->n;
213 r->n = 0;
214 spinlock_unlock(&r->lock);
215
216 /* append rq[i + 1] to rq[i] */
217 r = &CPU->rq[i];
218 spinlock_lock(&r->lock);
219 list_concat(&r->rq_head, &head);
220 r->n += n;
221 spinlock_unlock(&r->lock);
222 }
223 CPU->needs_relink = 0;
224 }
225 spinlock_unlock(&CPU->lock);
226
227}
228
229
230/** Scheduler stack switch wrapper
231 *
232 * Second part of the scheduler() function
233 * using new stack. Handling the actual context
234 * switch to a new thread.
235 *
236 */
237static void scheduler_separated_stack(void)
238{
239 int priority;
240
241 ASSERT(CPU != NULL);
242
243 if (THREAD) {
244 switch (THREAD->state) {
245 case Running:
246 THREAD->state = Ready;
247 spinlock_unlock(&THREAD->lock);
248 thread_ready(THREAD);
249 break;
250
251 case Exiting:
252 frame_free((__address) THREAD->kstack);
253 if (THREAD->ustack) {
254 frame_free((__address) THREAD->ustack);
255 }
256
257 /*
258 * Detach from the containing task.
259 */
260 spinlock_lock(&TASK->lock);
261 list_remove(&THREAD->th_link);
262 spinlock_unlock(&TASK->lock);
263
264 spinlock_unlock(&THREAD->lock);
265
266 spinlock_lock(&threads_lock);
267 list_remove(&THREAD->threads_link);
268 spinlock_unlock(&threads_lock);
269
270 spinlock_lock(&CPU->lock);
271 if(CPU->fpu_owner==THREAD)
272 CPU->fpu_owner=NULL;
273 spinlock_unlock(&CPU->lock);
274
275 free(THREAD);
276
277 break;
278
279 case Sleeping:
280 /*
281 * Prefer the thread after it's woken up.
282 */
283 THREAD->priority = -1;
284
285 /*
286 * We need to release wq->lock which we locked in waitq_sleep().
287 * Address of wq->lock is kept in THREAD->sleep_queue.
288 */
289 spinlock_unlock(&THREAD->sleep_queue->lock);
290
291 /*
292 * Check for possible requests for out-of-context invocation.
293 */
294 if (THREAD->call_me) {
295 THREAD->call_me(THREAD->call_me_with);
296 THREAD->call_me = NULL;
297 THREAD->call_me_with = NULL;
298 }
299
300 spinlock_unlock(&THREAD->lock);
301
302 break;
303
304 default:
305 /*
306 * Entering state is unexpected.
307 */
308 panic("tid%d: unexpected state %s\n", THREAD->tid, thread_states[THREAD->state]);
309 break;
310 }
311 THREAD = NULL;
312 }
313
314
315 THREAD = find_best_thread();
316
317 spinlock_lock(&THREAD->lock);
318 priority = THREAD->priority;
319 spinlock_unlock(&THREAD->lock);
320
321 relink_rq(priority);
322
323 spinlock_lock(&THREAD->lock);
324
325 /*
326 * If both the old and the new task are the same, lots of work is avoided.
327 */
328 if (TASK != THREAD->task) {
329 as_t *as1 = NULL;
330 as_t *as2;
331
332 if (TASK) {
333 spinlock_lock(&TASK->lock);
334 as1 = TASK->as;
335 spinlock_unlock(&TASK->lock);
336 }
337
338 spinlock_lock(&THREAD->task->lock);
339 as2 = THREAD->task->as;
340 spinlock_unlock(&THREAD->task->lock);
341
342 /*
343 * Note that it is possible for two tasks to share one address space.
344 */
345 if (as1 != as2) {
346 /*
347 * Both tasks and address spaces are different.
348 * Replace the old one with the new one.
349 */
350 as_install(as2);
351 }
352 TASK = THREAD->task;
353 }
354
355 THREAD->state = Running;
356
357 #ifdef SCHEDULER_VERBOSE
358 printf("cpu%d: tid %d (priority=%d,ticks=%d,nrdy=%d)\n", CPU->id, THREAD->tid, THREAD->priority, THREAD->ticks, CPU->nrdy);
359 #endif
360
361 /*
362 * Copy the knowledge of CPU, TASK, THREAD and preemption counter to thread's stack.
363 */
364 the_copy(THE, (the_t *) THREAD->kstack);
365
366 context_restore(&THREAD->saved_context);
367 /* not reached */
368}
369
370
371/** The scheduler
372 *
373 * The thread scheduling procedure.
374 * Passes control directly to
375 * scheduler_separated_stack().
376 *
377 */
378void scheduler(void)
379{
380 volatile ipl_t ipl;
381
382 ASSERT(CPU != NULL);
383
384 ipl = interrupts_disable();
385
386 if (atomic_get(&haltstate))
387 halt();
388
389 if (THREAD) {
390 spinlock_lock(&THREAD->lock);
391#ifndef CONFIG_FPU_LAZY
392 fpu_context_save(&(THREAD->saved_fpu_context));
393#endif
394 if (!context_save(&THREAD->saved_context)) {
395 /*
396 * This is the place where threads leave scheduler();
397 */
398 before_thread_runs();
399 spinlock_unlock(&THREAD->lock);
400 interrupts_restore(THREAD->saved_context.ipl);
401 return;
402 }
403
404 /*
405 * Interrupt priority level of preempted thread is recorded here
406 * to facilitate scheduler() invocations from interrupts_disable()'d
407 * code (e.g. waitq_sleep_timeout()).
408 */
409 THREAD->saved_context.ipl = ipl;
410 }
411
412 /*
413 * Through the 'THE' structure, we keep track of THREAD, TASK, CPU, VM
414 * and preemption counter. At this point THE could be coming either
415 * from THREAD's or CPU's stack.
416 */
417 the_copy(THE, (the_t *) CPU->stack);
418
419 /*
420 * We may not keep the old stack.
421 * Reason: If we kept the old stack and got blocked, for instance, in
422 * find_best_thread(), the old thread could get rescheduled by another
423 * CPU and overwrite the part of its own stack that was also used by
424 * the scheduler on this CPU.
425 *
426 * Moreover, we have to bypass the compiler-generated POP sequence
427 * which is fooled by SP being set to the very top of the stack.
428 * Therefore the scheduler() function continues in
429 * scheduler_separated_stack().
430 */
431 context_save(&CPU->saved_context);
432 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack), (__address) CPU->stack, CPU_STACK_SIZE);
433 context_restore(&CPU->saved_context);
434 /* not reached */
435}
436
437
438
439
440
441#ifdef CONFIG_SMP
442/** Load balancing thread
443 *
444 * SMP load balancing thread, supervising thread supplies
445 * for the CPU it's wired to.
446 *
447 * @param arg Generic thread argument (unused).
448 *
449 */
450void kcpulb(void *arg)
451{
452 thread_t *t;
453 int count, average, i, j, k = 0;
454 ipl_t ipl;
455
456loop:
457 /*
458 * Work in 1s intervals.
459 */
460 thread_sleep(1);
461
462not_satisfied:
463 /*
464 * Calculate the number of threads that will be migrated/stolen from
465 * other CPU's. Note that situation can have changed between two
466 * passes. Each time get the most up to date counts.
467 */
468 average = atomic_get(&nrdy) / config.cpu_active + 1;
469 count = average - atomic_get(&CPU->nrdy);
470
471 if (count <= 0)
472 goto satisfied;
473
474 /*
475 * Searching least priority queues on all CPU's first and most priority queues on all CPU's last.
476 */
477 for (j=RQ_COUNT-1; j >= 0; j--) {
478 for (i=0; i < config.cpu_active; i++) {
479 link_t *l;
480 runq_t *r;
481 cpu_t *cpu;
482
483 cpu = &cpus[(i + k) % config.cpu_active];
484
485 /*
486 * Not interested in ourselves.
487 * Doesn't require interrupt disabling for kcpulb is X_WIRED.
488 */
489 if (CPU == cpu)
490 continue;
491 if (atomic_get(&cpu->nrdy) <= average)
492 continue;
493
494 ipl = interrupts_disable();
495 r = &cpu->rq[j];
496 spinlock_lock(&r->lock);
497 if (r->n == 0) {
498 spinlock_unlock(&r->lock);
499 interrupts_restore(ipl);
500 continue;
501 }
502
503 t = NULL;
504 l = r->rq_head.prev; /* search rq from the back */
505 while (l != &r->rq_head) {
506 t = list_get_instance(l, thread_t, rq_link);
507 /*
508 * We don't want to steal CPU-wired threads neither threads already stolen.
509 * The latter prevents threads from migrating between CPU's without ever being run.
510 * We don't want to steal threads whose FPU context is still in CPU.
511 */
512 spinlock_lock(&t->lock);
513 if ( (!(t->flags & (X_WIRED | X_STOLEN))) && (!(t->fpu_context_engaged)) ) {
514 /*
515 * Remove t from r.
516 */
517 spinlock_unlock(&t->lock);
518
519 atomic_dec(&cpu->nrdy);
520 atomic_dec(&nrdy);
521
522 r->n--;
523 list_remove(&t->rq_link);
524
525 break;
526 }
527 spinlock_unlock(&t->lock);
528 l = l->prev;
529 t = NULL;
530 }
531 spinlock_unlock(&r->lock);
532
533 if (t) {
534 /*
535 * Ready t on local CPU
536 */
537 spinlock_lock(&t->lock);
538 #ifdef KCPULB_VERBOSE
539 printf("kcpulb%d: TID %d -> cpu%d, nrdy=%d, avg=%d\n", CPU->id, t->tid, CPU->id, atomic_get(&CPU->nrdy), atomic_get(&nrdy) / config.cpu_active);
540 #endif
541 t->flags |= X_STOLEN;
542 spinlock_unlock(&t->lock);
543
544 thread_ready(t);
545
546 interrupts_restore(ipl);
547
548 if (--count == 0)
549 goto satisfied;
550
551 /*
552 * We are not satisfied yet, focus on another CPU next time.
553 */
554 k++;
555
556 continue;
557 }
558 interrupts_restore(ipl);
559 }
560 }
561
562 if (atomic_get(&CPU->nrdy)) {
563 /*
564 * Be a little bit light-weight and let migrated threads run.
565 */
566 scheduler();
567 } else {
568 /*
569 * We failed to migrate a single thread.
570 * Give up this turn.
571 */
572 goto loop;
573 }
574
575 goto not_satisfied;
576
577satisfied:
578 goto loop;
579}
580
581#endif /* CONFIG_SMP */
582
583
584/** Print information about threads & scheduler queues */
585void sched_print_list(void)
586{
587 ipl_t ipl;
588 int cpu,i;
589 runq_t *r;
590 thread_t *t;
591 link_t *cur;
592
593 /* We are going to mess with scheduler structures,
594 * let's not be interrupted */
595 ipl = interrupts_disable();
596 printf("*********** Scheduler dump ***********\n");
597 for (cpu=0;cpu < config.cpu_count; cpu++) {
598 if (!cpus[cpu].active)
599 continue;
600 spinlock_lock(&cpus[cpu].lock);
601 printf("cpu%d: nrdy: %d needs_relink: %d\n",
602 cpus[cpu].id, atomic_get(&cpus[cpu].nrdy), cpus[cpu].needs_relink);
603
604 for (i=0; i<RQ_COUNT; i++) {
605 r = &cpus[cpu].rq[i];
606 spinlock_lock(&r->lock);
607 if (!r->n) {
608 spinlock_unlock(&r->lock);
609 continue;
610 }
611 printf("\tRq %d: ", i);
612 for (cur=r->rq_head.next; cur!=&r->rq_head; cur=cur->next) {
613 t = list_get_instance(cur, thread_t, rq_link);
614 printf("%d(%s) ", t->tid,
615 thread_states[t->state]);
616 }
617 printf("\n");
618 spinlock_unlock(&r->lock);
619 }
620 spinlock_unlock(&cpus[cpu].lock);
621 }
622
623 interrupts_restore(ipl);
624}
Note: See TracBrowser for help on using the repository browser.