1 /* sched.c - SPU scheduler.
3 * Copyright (C) IBM 2005
4 * Author: Mark Nutter <mnutter@us.ibm.com>
6 * 2006-03-31 NUMA domains added.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2, or (at your option)
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #include <linux/module.h>
26 #include <linux/errno.h>
27 #include <linux/sched.h>
28 #include <linux/kernel.h>
30 #include <linux/completion.h>
31 #include <linux/vmalloc.h>
32 #include <linux/smp.h>
33 #include <linux/smp_lock.h>
34 #include <linux/stddef.h>
35 #include <linux/unistd.h>
36 #include <linux/numa.h>
37 #include <linux/mutex.h>
38 #include <linux/notifier.h>
41 #include <asm/mmu_context.h>
43 #include <asm/spu_csa.h>
44 #include <asm/spu_priv1.h>
47 #define SPU_TIMESLICE (HZ)
49 struct spu_prio_array {
50 DECLARE_BITMAP(bitmap, MAX_PRIO);
51 struct list_head runq[MAX_PRIO];
53 struct list_head active_list[MAX_NUMNODES];
54 struct mutex active_mutex[MAX_NUMNODES];
57 static struct spu_prio_array *spu_prio;
58 static struct workqueue_struct *spu_sched_wq;
60 static inline int node_allowed(int node)
64 if (!nr_cpus_node(node))
66 mask = node_to_cpumask(node);
67 if (!cpus_intersects(mask, current->cpus_allowed))
72 void spu_start_tick(struct spu_context *ctx)
74 if (ctx->policy == SCHED_RR) {
76 * Make sure the exiting bit is cleared.
78 clear_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
80 queue_delayed_work(spu_sched_wq, &ctx->sched_work, SPU_TIMESLICE);
84 void spu_stop_tick(struct spu_context *ctx)
86 if (ctx->policy == SCHED_RR) {
88 * While the work can be rearming normally setting this flag
89 * makes sure it does not rearm itself anymore.
91 set_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
93 cancel_delayed_work(&ctx->sched_work);
97 void spu_sched_tick(struct work_struct *work)
99 struct spu_context *ctx =
100 container_of(work, struct spu_context, sched_work.work);
105 * If this context is being stopped avoid rescheduling from the
106 * scheduler tick because we would block on the state_mutex.
107 * The caller will yield the spu later on anyway.
109 if (test_bit(SPU_SCHED_EXITING, &ctx->sched_flags))
112 mutex_lock(&ctx->state_mutex);
115 int best = sched_find_first_bit(spu_prio->bitmap);
116 if (best <= ctx->prio) {
121 mutex_unlock(&ctx->state_mutex);
125 * We need to break out of the wait loop in spu_run manually
126 * to ensure this context gets put on the runqueue again
129 wake_up(&ctx->stop_wq);
135 * spu_add_to_active_list - add spu to active list
136 * @spu: spu to add to the active list
138 static void spu_add_to_active_list(struct spu *spu)
140 mutex_lock(&spu_prio->active_mutex[spu->node]);
141 list_add_tail(&spu->list, &spu_prio->active_list[spu->node]);
142 mutex_unlock(&spu_prio->active_mutex[spu->node]);
146 * spu_remove_from_active_list - remove spu from active list
147 * @spu: spu to remove from the active list
149 static void spu_remove_from_active_list(struct spu *spu)
151 int node = spu->node;
153 mutex_lock(&spu_prio->active_mutex[node]);
154 list_del_init(&spu->list);
155 mutex_unlock(&spu_prio->active_mutex[node]);
158 static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
160 static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
162 blocking_notifier_call_chain(&spu_switch_notifier,
163 ctx ? ctx->object_id : 0, spu);
166 int spu_switch_event_register(struct notifier_block * n)
168 return blocking_notifier_chain_register(&spu_switch_notifier, n);
171 int spu_switch_event_unregister(struct notifier_block * n)
173 return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
177 * spu_bind_context - bind spu context to physical spu
178 * @spu: physical spu to bind to
179 * @ctx: context to bind
181 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
183 pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
184 spu->number, spu->node);
188 ctx->ops = &spu_hw_ops;
189 spu->pid = current->pid;
190 spu_associate_mm(spu, ctx->owner);
191 spu->ibox_callback = spufs_ibox_callback;
192 spu->wbox_callback = spufs_wbox_callback;
193 spu->stop_callback = spufs_stop_callback;
194 spu->mfc_callback = spufs_mfc_callback;
195 spu->dma_callback = spufs_dma_callback;
197 spu_unmap_mappings(ctx);
198 spu_restore(&ctx->csa, spu);
199 spu->timestamp = jiffies;
200 spu_cpu_affinity_set(spu, raw_smp_processor_id());
201 spu_switch_notify(spu, ctx);
202 spu_add_to_active_list(spu);
203 ctx->state = SPU_STATE_RUNNABLE;
207 * spu_unbind_context - unbind spu context from physical spu
208 * @spu: physical spu to unbind from
209 * @ctx: context to unbind
211 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
213 pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
214 spu->pid, spu->number, spu->node);
216 spu_remove_from_active_list(spu);
217 spu_switch_notify(spu, NULL);
218 spu_unmap_mappings(ctx);
219 spu_save(&ctx->csa, spu);
220 spu->timestamp = jiffies;
221 ctx->state = SPU_STATE_SAVED;
222 spu->ibox_callback = NULL;
223 spu->wbox_callback = NULL;
224 spu->stop_callback = NULL;
225 spu->mfc_callback = NULL;
226 spu->dma_callback = NULL;
227 spu_associate_mm(spu, NULL);
229 ctx->ops = &spu_backing_ops;
236 * spu_add_to_rq - add a context to the runqueue
237 * @ctx: context to add
239 static void spu_add_to_rq(struct spu_context *ctx)
241 spin_lock(&spu_prio->runq_lock);
242 list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
243 set_bit(ctx->prio, spu_prio->bitmap);
245 spin_unlock(&spu_prio->runq_lock);
249 * spu_del_from_rq - remove a context from the runqueue
250 * @ctx: context to remove
252 static void spu_del_from_rq(struct spu_context *ctx)
254 spin_lock(&spu_prio->runq_lock);
255 list_del_init(&ctx->rq);
256 if (list_empty(&spu_prio->runq[ctx->prio]))
257 clear_bit(ctx->prio, spu_prio->bitmap);
258 spin_unlock(&spu_prio->runq_lock);
262 * spu_grab_context - remove one context from the runqueue
263 * @prio: priority of the context to be removed
265 * This function removes one context from the runqueue for priority @prio.
266 * If there is more than one context with the given priority the first
267 * task on the runqueue will be taken.
269 * Returns the spu_context it just removed.
271 * Must be called with spu_prio->runq_lock held.
273 static struct spu_context *spu_grab_context(int prio)
275 struct list_head *rq = &spu_prio->runq[prio];
279 return list_entry(rq->next, struct spu_context, rq);
282 static void spu_prio_wait(struct spu_context *ctx)
286 prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
287 if (!signal_pending(current)) {
288 mutex_unlock(&ctx->state_mutex);
290 mutex_lock(&ctx->state_mutex);
292 __set_current_state(TASK_RUNNING);
293 remove_wait_queue(&ctx->stop_wq, &wait);
297 * spu_reschedule - try to find a runnable context for a spu
298 * @spu: spu available
300 * This function is called whenever a spu becomes idle. It looks for the
301 * most suitable runnable spu context and schedules it for execution.
303 static void spu_reschedule(struct spu *spu)
309 spin_lock(&spu_prio->runq_lock);
310 best = sched_find_first_bit(spu_prio->bitmap);
311 if (best < MAX_PRIO) {
312 struct spu_context *ctx = spu_grab_context(best);
314 wake_up(&ctx->stop_wq);
316 spin_unlock(&spu_prio->runq_lock);
319 static struct spu *spu_get_idle(struct spu_context *ctx)
321 struct spu *spu = NULL;
322 int node = cpu_to_node(raw_smp_processor_id());
325 for (n = 0; n < MAX_NUMNODES; n++, node++) {
326 node = (node < MAX_NUMNODES) ? node : 0;
327 if (!node_allowed(node))
329 spu = spu_alloc_node(node);
337 * find_victim - find a lower priority context to preempt
338 * @ctx: canidate context for running
340 * Returns the freed physical spu to run the new context on.
342 static struct spu *find_victim(struct spu_context *ctx)
344 struct spu_context *victim = NULL;
349 * Look for a possible preemption candidate on the local node first.
350 * If there is no candidate look at the other nodes. This isn't
351 * exactly fair, but so far the whole spu schedule tries to keep
352 * a strong node affinity. We might want to fine-tune this in
356 node = cpu_to_node(raw_smp_processor_id());
357 for (n = 0; n < MAX_NUMNODES; n++, node++) {
358 node = (node < MAX_NUMNODES) ? node : 0;
359 if (!node_allowed(node))
362 mutex_lock(&spu_prio->active_mutex[node]);
363 list_for_each_entry(spu, &spu_prio->active_list[node], list) {
364 struct spu_context *tmp = spu->ctx;
366 if (tmp->rt_priority < ctx->rt_priority &&
367 (!victim || tmp->rt_priority < victim->rt_priority))
370 mutex_unlock(&spu_prio->active_mutex[node]);
374 * This nests ctx->state_mutex, but we always lock
375 * higher priority contexts before lower priority
376 * ones, so this is safe until we introduce
377 * priority inheritance schemes.
379 if (!mutex_trylock(&victim->state_mutex)) {
387 * This race can happen because we've dropped
388 * the active list mutex. No a problem, just
389 * restart the search.
391 mutex_unlock(&victim->state_mutex);
395 spu_unbind_context(spu, victim);
396 mutex_unlock(&victim->state_mutex);
398 * We need to break out of the wait loop in spu_run
399 * manually to ensure this context gets put on the
400 * runqueue again ASAP.
402 wake_up(&victim->stop_wq);
411 * spu_activate - find a free spu for a context and execute it
412 * @ctx: spu context to schedule
413 * @flags: flags (currently ignored)
415 * Tries to find a free spu to run @ctx. If no free spu is available
416 * add the context to the runqueue so it gets woken up once an spu
419 int spu_activate(struct spu_context *ctx, unsigned long flags)
428 spu = spu_get_idle(ctx);
430 * If this is a realtime thread we try to get it running by
431 * preempting a lower priority thread.
433 if (!spu && ctx->rt_priority)
434 spu = find_victim(ctx);
436 spu_bind_context(spu, ctx);
442 spu_del_from_rq(ctx);
443 } while (!signal_pending(current));
449 * spu_deactivate - unbind a context from it's physical spu
450 * @ctx: spu context to unbind
452 * Unbind @ctx from the physical spu it is running on and schedule
453 * the highest priority context to run on the freed physical spu.
455 void spu_deactivate(struct spu_context *ctx)
457 struct spu *spu = ctx->spu;
460 spu_unbind_context(spu, ctx);
466 * spu_yield - yield a physical spu if others are waiting
467 * @ctx: spu context to yield
469 * Check if there is a higher priority context waiting and if yes
470 * unbind @ctx from the physical spu and schedule the highest
471 * priority context to run on the freed physical spu instead.
473 void spu_yield(struct spu_context *ctx)
477 if (mutex_trylock(&ctx->state_mutex)) {
478 if ((spu = ctx->spu) != NULL) {
479 int best = sched_find_first_bit(spu_prio->bitmap);
480 if (best < MAX_PRIO) {
481 pr_debug("%s: yielding SPU %d NODE %d\n",
482 __FUNCTION__, spu->number, spu->node);
486 mutex_unlock(&ctx->state_mutex);
490 int __init spu_sched_init(void)
494 spu_sched_wq = create_singlethread_workqueue("spusched");
498 spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
500 printk(KERN_WARNING "%s: Unable to allocate priority queue.\n",
502 destroy_workqueue(spu_sched_wq);
505 for (i = 0; i < MAX_PRIO; i++) {
506 INIT_LIST_HEAD(&spu_prio->runq[i]);
507 __clear_bit(i, spu_prio->bitmap);
509 __set_bit(MAX_PRIO, spu_prio->bitmap);
510 for (i = 0; i < MAX_NUMNODES; i++) {
511 mutex_init(&spu_prio->active_mutex[i]);
512 INIT_LIST_HEAD(&spu_prio->active_list[i]);
514 spin_lock_init(&spu_prio->runq_lock);
518 void __exit spu_sched_exit(void)
520 struct spu *spu, *tmp;
523 for (node = 0; node < MAX_NUMNODES; node++) {
524 mutex_lock(&spu_prio->active_mutex[node]);
525 list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
527 list_del_init(&spu->list);
530 mutex_unlock(&spu_prio->active_mutex[node]);
533 destroy_workqueue(spu_sched_wq);