source: mainline/src/proc/scheduler.c@ b07769b6

lfn serial ticket/834-toolchain-update topic/msim-upgrade topic/simplify-dev-export
Last change on this file since b07769b6 was b07769b6, checked in by Jakub Jermar <jakub@…>, 20 years ago

Switch over to per vm_t page tables. (breaks ia32 userspace)

Define dummy memcopy() for amd64 and ppc to satisfy compiler.

Get rid of problematic #include <arch/mm/page.h> in mm/page.h.

Indentation fixes and small changes here and there.
  • Property mode set to 100644
File size: 12.3 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 <cpu.h>
33#include <mm/vm.h>
34#include <config.h>
35#include <context.h>
36#include <func.h>
37#include <arch.h>
38#include <arch/asm.h>
39#include <list.h>
40#include <panic.h>
41#include <typedefs.h>
42#include <mm/page.h>
43#include <synch/spinlock.h>
44#include <arch/faddr.h>
45#include <arch/atomic.h>
46
47volatile int nrdy;
48
49
50/** Take actions before new thread runs
51 *
52 * Perform actions that need to be
53 * taken before the newly selected
54 * tread is passed control.
55 *
56 */
57void before_thread_runs(void)
58{
59 before_thread_runs_arch();
60 fpu_context_restore(&(THREAD->saved_fpu_context));
61}
62
63
64/** Initialize scheduler
65 *
66 * Initialize kernel scheduler.
67 *
68 */
69void scheduler_init(void)
70{
71}
72
73
74/** Get thread to be scheduled
75 *
76 * Get the optimal thread to be scheduled
77 * according to thread accounting and scheduler
78 * policy.
79 *
80 * @return Thread to be scheduled.
81 *
82 */
83struct thread *find_best_thread(void)
84{
85 thread_t *t;
86 runq_t *r;
87 int i, n;
88
89loop:
90 cpu_priority_high();
91
92 spinlock_lock(&CPU->lock);
93 n = CPU->nrdy;
94 spinlock_unlock(&CPU->lock);
95
96 cpu_priority_low();
97
98 if (n == 0) {
99 #ifdef __SMP__
100 /*
101 * If the load balancing thread is not running, wake it up and
102 * set CPU-private flag that the kcpulb has been started.
103 */
104 if (test_and_set(&CPU->kcpulbstarted) == 0) {
105 waitq_wakeup(&CPU->kcpulb_wq, 0);
106 goto loop;
107 }
108 #endif /* __SMP__ */
109
110 /*
111 * For there was nothing to run, the CPU goes to sleep
112 * until a hardware interrupt or an IPI comes.
113 * This improves energy saving and hyperthreading.
114 * On the other hand, several hardware interrupts can be ignored.
115 */
116 cpu_sleep();
117 goto loop;
118 }
119
120 cpu_priority_high();
121
122 i = 0;
123retry:
124 for (; i<RQ_COUNT; i++) {
125 r = &CPU->rq[i];
126 spinlock_lock(&r->lock);
127 if (r->n == 0) {
128 /*
129 * If this queue is empty, try a lower-priority queue.
130 */
131 spinlock_unlock(&r->lock);
132 continue;
133 }
134
135 /* avoid deadlock with relink_rq() */
136 if (!spinlock_trylock(&CPU->lock)) {
137 /*
138 * Unlock r and try again.
139 */
140 spinlock_unlock(&r->lock);
141 goto retry;
142 }
143 CPU->nrdy--;
144 spinlock_unlock(&CPU->lock);
145
146 atomic_dec(&nrdy);
147 r->n--;
148
149 /*
150 * Take the first thread from the queue.
151 */
152 t = list_get_instance(r->rq_head.next, thread_t, rq_link);
153 list_remove(&t->rq_link);
154
155 spinlock_unlock(&r->lock);
156
157 spinlock_lock(&t->lock);
158 t->cpu = CPU;
159
160 t->ticks = us2ticks((i+1)*10000);
161 t->pri = i; /* eventually correct rq index */
162
163 /*
164 * Clear the X_STOLEN flag so that t can be migrated when load balancing needs emerge.
165 */
166 t->flags &= ~X_STOLEN;
167 spinlock_unlock(&t->lock);
168
169 return t;
170 }
171 goto loop;
172
173}
174
175
176/** Prevent rq starvation
177 *
178 * Prevent low priority threads from starving in rq's.
179 *
180 * When the function decides to relink rq's, it reconnects
181 * respective pointers so that in result threads with 'pri'
182 * greater or equal 'start' are moved to a higher-priority queue.
183 *
184 * @param start Threshold priority.
185 *
186 */
187void relink_rq(int start)
188{
189 link_t head;
190 runq_t *r;
191 int i, n;
192
193 list_initialize(&head);
194 spinlock_lock(&CPU->lock);
195 if (CPU->needs_relink > NEEDS_RELINK_MAX) {
196 for (i = start; i<RQ_COUNT-1; i++) {
197 /* remember and empty rq[i + 1] */
198 r = &CPU->rq[i + 1];
199 spinlock_lock(&r->lock);
200 list_concat(&head, &r->rq_head);
201 n = r->n;
202 r->n = 0;
203 spinlock_unlock(&r->lock);
204
205 /* append rq[i + 1] to rq[i] */
206 r = &CPU->rq[i];
207 spinlock_lock(&r->lock);
208 list_concat(&r->rq_head, &head);
209 r->n += n;
210 spinlock_unlock(&r->lock);
211 }
212 CPU->needs_relink = 0;
213 }
214 spinlock_unlock(&CPU->lock);
215
216}
217
218
219/** The scheduler
220 *
221 * The thread scheduling procedure.
222 *
223 */
224void scheduler(void)
225{
226 volatile pri_t pri;
227
228 pri = cpu_priority_high();
229
230 if (haltstate)
231 halt();
232
233 if (THREAD) {
234 spinlock_lock(&THREAD->lock);
235 fpu_context_save(&(THREAD->saved_fpu_context));
236 if (!context_save(&THREAD->saved_context)) {
237 /*
238 * This is the place where threads leave scheduler();
239 */
240 before_thread_runs();
241 spinlock_unlock(&THREAD->lock);
242 cpu_priority_restore(THREAD->saved_context.pri);
243 return;
244 }
245 THREAD->saved_context.pri = pri;
246 }
247
248 /*
249 * We may not keep the old stack.
250 * Reason: If we kept the old stack and got blocked, for instance, in
251 * find_best_thread(), the old thread could get rescheduled by another
252 * CPU and overwrite the part of its own stack that was also used by
253 * the scheduler on this CPU.
254 *
255 * Moreover, we have to bypass the compiler-generated POP sequence
256 * which is fooled by SP being set to the very top of the stack.
257 * Therefore the scheduler() function continues in
258 * scheduler_separated_stack().
259 */
260 context_save(&CPU->saved_context);
261 context_set(&CPU->saved_context, FADDR(scheduler_separated_stack), CPU->stack, CPU_STACK_SIZE);
262 context_restore(&CPU->saved_context);
263 /* not reached */
264}
265
266
267/** Scheduler stack switch wrapper
268 *
269 * Second part of the scheduler() function
270 * using new stack. Handling the actual context
271 * switch to a new thread.
272 *
273 */
274void scheduler_separated_stack(void)
275{
276 int priority;
277
278 if (THREAD) {
279 switch (THREAD->state) {
280 case Running:
281 THREAD->state = Ready;
282 spinlock_unlock(&THREAD->lock);
283 thread_ready(THREAD);
284 break;
285
286 case Exiting:
287 frame_free((__address) THREAD->kstack);
288 if (THREAD->ustack) {
289 frame_free((__address) THREAD->ustack);
290 }
291
292 /*
293 * Detach from the containing task.
294 */
295 spinlock_lock(&TASK->lock);
296 list_remove(&THREAD->th_link);
297 spinlock_unlock(&TASK->lock);
298
299 spinlock_unlock(&THREAD->lock);
300
301 spinlock_lock(&threads_lock);
302 list_remove(&THREAD->threads_link);
303 spinlock_unlock(&threads_lock);
304
305 spinlock_lock(&CPU->lock);
306 if(CPU->fpu_owner==THREAD) CPU->fpu_owner=NULL;
307 spinlock_unlock(&CPU->lock);
308
309 free(THREAD);
310
311 break;
312
313 case Sleeping:
314 /*
315 * Prefer the thread after it's woken up.
316 */
317 THREAD->pri = -1;
318
319 /*
320 * We need to release wq->lock which we locked in waitq_sleep().
321 * Address of wq->lock is kept in THREAD->sleep_queue.
322 */
323 spinlock_unlock(&THREAD->sleep_queue->lock);
324
325 /*
326 * Check for possible requests for out-of-context invocation.
327 */
328 if (THREAD->call_me) {
329 THREAD->call_me(THREAD->call_me_with);
330 THREAD->call_me = NULL;
331 THREAD->call_me_with = NULL;
332 }
333
334 spinlock_unlock(&THREAD->lock);
335
336 break;
337
338 default:
339 /*
340 * Entering state is unexpected.
341 */
342 panic("tid%d: unexpected state %s\n", THREAD->tid, thread_states[THREAD->state]);
343 break;
344 }
345 THREAD = NULL;
346 }
347
348 THREAD = find_best_thread();
349
350 spinlock_lock(&THREAD->lock);
351 priority = THREAD->pri;
352 spinlock_unlock(&THREAD->lock);
353
354 relink_rq(priority);
355
356 spinlock_lock(&THREAD->lock);
357
358 /*
359 * If both the old and the new task are the same, lots of work is avoided.
360 */
361 if (TASK != THREAD->task) {
362 vm_t *m1 = NULL;
363 vm_t *m2;
364
365 if (TASK) {
366 spinlock_lock(&TASK->lock);
367 m1 = TASK->vm;
368 spinlock_unlock(&TASK->lock);
369 }
370
371 spinlock_lock(&THREAD->task->lock);
372 m2 = THREAD->task->vm;
373 spinlock_unlock(&THREAD->task->lock);
374
375 /*
376 * Note that it is possible for two tasks to share one vm mapping.
377 */
378 if (m1 != m2) {
379 /*
380 * Both tasks and vm mappings are different.
381 * Replace the old one with the new one.
382 */
383 vm_install(m2);
384 }
385 TASK = THREAD->task;
386 }
387
388 THREAD->state = Running;
389
390 #ifdef SCHEDULER_VERBOSE
391 printf("cpu%d: tid %d (pri=%d,ticks=%d,nrdy=%d)\n", CPU->id, THREAD->tid, THREAD->pri, THREAD->ticks, CPU->nrdy);
392 #endif
393
394 context_restore(&THREAD->saved_context);
395 /* not reached */
396}
397
398
399#ifdef __SMP__
400/** Load balancing thread
401 *
402 * SMP load balancing thread, supervising thread supplies
403 * for the CPU it's wired to.
404 *
405 * @param arg Generic thread argument (unused).
406 *
407 */
408void kcpulb(void *arg)
409{
410 thread_t *t;
411 int count, i, j, k = 0;
412 pri_t pri;
413
414loop:
415 /*
416 * Sleep until there's some work to do.
417 */
418 waitq_sleep(&CPU->kcpulb_wq);
419
420not_satisfied:
421 /*
422 * Calculate the number of threads that will be migrated/stolen from
423 * other CPU's. Note that situation can have changed between two
424 * passes. Each time get the most up to date counts.
425 */
426 pri = cpu_priority_high();
427 spinlock_lock(&CPU->lock);
428 count = nrdy / config.cpu_active;
429 count -= CPU->nrdy;
430 spinlock_unlock(&CPU->lock);
431 cpu_priority_restore(pri);
432
433 if (count <= 0)
434 goto satisfied;
435
436 /*
437 * Searching least priority queues on all CPU's first and most priority queues on all CPU's last.
438 */
439 for (j=RQ_COUNT-1; j >= 0; j--) {
440 for (i=0; i < config.cpu_active; i++) {
441 link_t *l;
442 runq_t *r;
443 cpu_t *cpu;
444
445 cpu = &cpus[(i + k) % config.cpu_active];
446
447 /*
448 * Not interested in ourselves.
449 * Doesn't require interrupt disabling for kcpulb is X_WIRED.
450 */
451 if (CPU == cpu)
452 continue;
453
454restart: pri = cpu_priority_high();
455 r = &cpu->rq[j];
456 spinlock_lock(&r->lock);
457 if (r->n == 0) {
458 spinlock_unlock(&r->lock);
459 cpu_priority_restore(pri);
460 continue;
461 }
462
463 t = NULL;
464 l = r->rq_head.prev; /* search rq from the back */
465 while (l != &r->rq_head) {
466 t = list_get_instance(l, thread_t, rq_link);
467 /*
468 * We don't want to steal CPU-wired threads neither threads already stolen.
469 * The latter prevents threads from migrating between CPU's without ever being run.
470 * We don't want to steal threads whose FPU context is still in CPU.
471 */
472 spinlock_lock(&t->lock);
473 if ( (!(t->flags & (X_WIRED | X_STOLEN))) && (!(t->fpu_context_engaged)) ) {
474
475 /*
476 * Remove t from r.
477 */
478
479 spinlock_unlock(&t->lock);
480
481 /*
482 * Here we have to avoid deadlock with relink_rq(),
483 * because it locks cpu and r in a different order than we do.
484 */
485 if (!spinlock_trylock(&cpu->lock)) {
486 /* Release all locks and try again. */
487 spinlock_unlock(&r->lock);
488 cpu_priority_restore(pri);
489 goto restart;
490 }
491 cpu->nrdy--;
492 spinlock_unlock(&cpu->lock);
493
494 atomic_dec(&nrdy);
495
496 r->n--;
497 list_remove(&t->rq_link);
498
499 break;
500 }
501 spinlock_unlock(&t->lock);
502 l = l->prev;
503 t = NULL;
504 }
505 spinlock_unlock(&r->lock);
506
507 if (t) {
508 /*
509 * Ready t on local CPU
510 */
511 spinlock_lock(&t->lock);
512 #ifdef KCPULB_VERBOSE
513 printf("kcpulb%d: TID %d -> cpu%d, nrdy=%d, avg=%d\n", CPU->id, t->tid, CPU->id, CPU->nrdy, nrdy / config.cpu_active);
514 #endif
515 t->flags |= X_STOLEN;
516 spinlock_unlock(&t->lock);
517
518 thread_ready(t);
519
520 cpu_priority_restore(pri);
521
522 if (--count == 0)
523 goto satisfied;
524
525 /*
526 * We are not satisfied yet, focus on another CPU next time.
527 */
528 k++;
529
530 continue;
531 }
532 cpu_priority_restore(pri);
533 }
534 }
535
536 if (CPU->nrdy) {
537 /*
538 * Be a little bit light-weight and let migrated threads run.
539 */
540 scheduler();
541 }
542 else {
543 /*
544 * We failed to migrate a single thread.
545 * Something more sophisticated should be done.
546 */
547 scheduler();
548 }
549
550 goto not_satisfied;
551
552satisfied:
553 /*
554 * Tell find_best_thread() to wake us up later again.
555 */
556 CPU->kcpulbstarted = 0;
557 goto loop;
558}
559
560#endif /* __SMP__ */
Note: See TracBrowser for help on using the repository browser.