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/stddef.h>
34 #include <linux/unistd.h>
35 #include <linux/numa.h>
36 #include <linux/mutex.h>
37 #include <linux/notifier.h>
40 #include <asm/mmu_context.h>
42 #include <asm/spu_csa.h>
43 #include <asm/spu_priv1.h>
46 #define SPU_TIMESLICE (HZ)
48 struct spu_prio_array {
49 DECLARE_BITMAP(bitmap, MAX_PRIO);
50 struct list_head runq[MAX_PRIO];
52 struct list_head active_list[MAX_NUMNODES];
53 struct mutex active_mutex[MAX_NUMNODES];
56 static struct spu_prio_array *spu_prio;
57 static struct workqueue_struct *spu_sched_wq;
59 static inline int node_allowed(int node)
63 if (!nr_cpus_node(node))
65 mask = node_to_cpumask(node);
66 if (!cpus_intersects(mask, current->cpus_allowed))
71 void spu_start_tick(struct spu_context *ctx)
73 if (ctx->policy == SCHED_RR) {
75 * Make sure the exiting bit is cleared.
77 clear_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
79 queue_delayed_work(spu_sched_wq, &ctx->sched_work, SPU_TIMESLICE);
83 void spu_stop_tick(struct spu_context *ctx)
85 if (ctx->policy == SCHED_RR) {
87 * While the work can be rearming normally setting this flag
88 * makes sure it does not rearm itself anymore.
90 set_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
92 cancel_delayed_work(&ctx->sched_work);
97 * spu_add_to_active_list - add spu to active list
98 * @spu: spu to add to the active list
100 static void spu_add_to_active_list(struct spu *spu)
102 mutex_lock(&spu_prio->active_mutex[spu->node]);
103 list_add_tail(&spu->list, &spu_prio->active_list[spu->node]);
104 mutex_unlock(&spu_prio->active_mutex[spu->node]);
108 * spu_remove_from_active_list - remove spu from active list
109 * @spu: spu to remove from the active list
111 static void spu_remove_from_active_list(struct spu *spu)
113 int node = spu->node;
115 mutex_lock(&spu_prio->active_mutex[node]);
116 list_del_init(&spu->list);
117 mutex_unlock(&spu_prio->active_mutex[node]);
120 static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
122 static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
124 blocking_notifier_call_chain(&spu_switch_notifier,
125 ctx ? ctx->object_id : 0, spu);
128 int spu_switch_event_register(struct notifier_block * n)
130 return blocking_notifier_chain_register(&spu_switch_notifier, n);
133 int spu_switch_event_unregister(struct notifier_block * n)
135 return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
139 * spu_bind_context - bind spu context to physical spu
140 * @spu: physical spu to bind to
141 * @ctx: context to bind
143 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
145 pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
146 spu->number, spu->node);
150 ctx->ops = &spu_hw_ops;
151 spu->pid = current->pid;
152 spu_associate_mm(spu, ctx->owner);
153 spu->ibox_callback = spufs_ibox_callback;
154 spu->wbox_callback = spufs_wbox_callback;
155 spu->stop_callback = spufs_stop_callback;
156 spu->mfc_callback = spufs_mfc_callback;
157 spu->dma_callback = spufs_dma_callback;
159 spu_unmap_mappings(ctx);
160 spu_restore(&ctx->csa, spu);
161 spu->timestamp = jiffies;
162 spu_cpu_affinity_set(spu, raw_smp_processor_id());
163 spu_switch_notify(spu, ctx);
164 spu_add_to_active_list(spu);
165 ctx->state = SPU_STATE_RUNNABLE;
169 * spu_unbind_context - unbind spu context from physical spu
170 * @spu: physical spu to unbind from
171 * @ctx: context to unbind
173 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
175 pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
176 spu->pid, spu->number, spu->node);
178 spu_remove_from_active_list(spu);
179 spu_switch_notify(spu, NULL);
180 spu_unmap_mappings(ctx);
181 spu_save(&ctx->csa, spu);
182 spu->timestamp = jiffies;
183 ctx->state = SPU_STATE_SAVED;
184 spu->ibox_callback = NULL;
185 spu->wbox_callback = NULL;
186 spu->stop_callback = NULL;
187 spu->mfc_callback = NULL;
188 spu->dma_callback = NULL;
189 spu_associate_mm(spu, NULL);
191 ctx->ops = &spu_backing_ops;
198 * spu_add_to_rq - add a context to the runqueue
199 * @ctx: context to add
201 static void __spu_add_to_rq(struct spu_context *ctx)
203 int prio = ctx->prio;
205 list_add_tail(&ctx->rq, &spu_prio->runq[prio]);
206 set_bit(prio, spu_prio->bitmap);
209 static void __spu_del_from_rq(struct spu_context *ctx)
211 int prio = ctx->prio;
213 if (!list_empty(&ctx->rq))
214 list_del_init(&ctx->rq);
215 if (list_empty(&spu_prio->runq[prio]))
216 clear_bit(prio, spu_prio->bitmap);
219 static void spu_prio_wait(struct spu_context *ctx)
223 spin_lock(&spu_prio->runq_lock);
224 prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
225 if (!signal_pending(current)) {
226 __spu_add_to_rq(ctx);
227 spin_unlock(&spu_prio->runq_lock);
228 mutex_unlock(&ctx->state_mutex);
230 mutex_lock(&ctx->state_mutex);
231 spin_lock(&spu_prio->runq_lock);
232 __spu_del_from_rq(ctx);
234 spin_unlock(&spu_prio->runq_lock);
235 __set_current_state(TASK_RUNNING);
236 remove_wait_queue(&ctx->stop_wq, &wait);
239 static struct spu *spu_get_idle(struct spu_context *ctx)
241 struct spu *spu = NULL;
242 int node = cpu_to_node(raw_smp_processor_id());
245 for (n = 0; n < MAX_NUMNODES; n++, node++) {
246 node = (node < MAX_NUMNODES) ? node : 0;
247 if (!node_allowed(node))
249 spu = spu_alloc_node(node);
257 * find_victim - find a lower priority context to preempt
258 * @ctx: canidate context for running
260 * Returns the freed physical spu to run the new context on.
262 static struct spu *find_victim(struct spu_context *ctx)
264 struct spu_context *victim = NULL;
269 * Look for a possible preemption candidate on the local node first.
270 * If there is no candidate look at the other nodes. This isn't
271 * exactly fair, but so far the whole spu schedule tries to keep
272 * a strong node affinity. We might want to fine-tune this in
276 node = cpu_to_node(raw_smp_processor_id());
277 for (n = 0; n < MAX_NUMNODES; n++, node++) {
278 node = (node < MAX_NUMNODES) ? node : 0;
279 if (!node_allowed(node))
282 mutex_lock(&spu_prio->active_mutex[node]);
283 list_for_each_entry(spu, &spu_prio->active_list[node], list) {
284 struct spu_context *tmp = spu->ctx;
286 if (tmp->rt_priority < ctx->rt_priority &&
287 (!victim || tmp->rt_priority < victim->rt_priority))
290 mutex_unlock(&spu_prio->active_mutex[node]);
294 * This nests ctx->state_mutex, but we always lock
295 * higher priority contexts before lower priority
296 * ones, so this is safe until we introduce
297 * priority inheritance schemes.
299 if (!mutex_trylock(&victim->state_mutex)) {
307 * This race can happen because we've dropped
308 * the active list mutex. No a problem, just
309 * restart the search.
311 mutex_unlock(&victim->state_mutex);
315 spu_unbind_context(spu, victim);
316 mutex_unlock(&victim->state_mutex);
318 * We need to break out of the wait loop in spu_run
319 * manually to ensure this context gets put on the
320 * runqueue again ASAP.
322 wake_up(&victim->stop_wq);
331 * spu_activate - find a free spu for a context and execute it
332 * @ctx: spu context to schedule
333 * @flags: flags (currently ignored)
335 * Tries to find a free spu to run @ctx. If no free spu is available
336 * add the context to the runqueue so it gets woken up once an spu
339 int spu_activate(struct spu_context *ctx, unsigned long flags)
348 spu = spu_get_idle(ctx);
350 * If this is a realtime thread we try to get it running by
351 * preempting a lower priority thread.
353 if (!spu && ctx->rt_priority)
354 spu = find_victim(ctx);
356 spu_bind_context(spu, ctx);
361 } while (!signal_pending(current));
367 * grab_runnable_context - try to find a runnable context
369 * Remove the highest priority context on the runqueue and return it
370 * to the caller. Returns %NULL if no runnable context was found.
372 static struct spu_context *grab_runnable_context(int prio)
374 struct spu_context *ctx = NULL;
377 spin_lock(&spu_prio->runq_lock);
378 best = sched_find_first_bit(spu_prio->bitmap);
380 struct list_head *rq = &spu_prio->runq[best];
382 BUG_ON(list_empty(rq));
384 ctx = list_entry(rq->next, struct spu_context, rq);
385 __spu_del_from_rq(ctx);
387 spin_unlock(&spu_prio->runq_lock);
392 static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
394 struct spu *spu = ctx->spu;
395 struct spu_context *new = NULL;
398 new = grab_runnable_context(max_prio);
400 spu_unbind_context(spu, ctx);
403 wake_up(&new->stop_wq);
412 * spu_deactivate - unbind a context from it's physical spu
413 * @ctx: spu context to unbind
415 * Unbind @ctx from the physical spu it is running on and schedule
416 * the highest priority context to run on the freed physical spu.
418 void spu_deactivate(struct spu_context *ctx)
420 __spu_deactivate(ctx, 1, MAX_PRIO);
424 * spu_yield - yield a physical spu if others are waiting
425 * @ctx: spu context to yield
427 * Check if there is a higher priority context waiting and if yes
428 * unbind @ctx from the physical spu and schedule the highest
429 * priority context to run on the freed physical spu instead.
431 void spu_yield(struct spu_context *ctx)
433 if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
434 mutex_lock(&ctx->state_mutex);
435 __spu_deactivate(ctx, 0, MAX_PRIO);
436 mutex_unlock(&ctx->state_mutex);
440 void spu_sched_tick(struct work_struct *work)
442 struct spu_context *ctx =
443 container_of(work, struct spu_context, sched_work.work);
447 * If this context is being stopped avoid rescheduling from the
448 * scheduler tick because we would block on the state_mutex.
449 * The caller will yield the spu later on anyway.
451 if (test_bit(SPU_SCHED_EXITING, &ctx->sched_flags))
454 mutex_lock(&ctx->state_mutex);
455 preempted = __spu_deactivate(ctx, 0, ctx->prio + 1);
456 mutex_unlock(&ctx->state_mutex);
460 * We need to break out of the wait loop in spu_run manually
461 * to ensure this context gets put on the runqueue again
464 wake_up(&ctx->stop_wq);
470 int __init spu_sched_init(void)
474 spu_sched_wq = create_singlethread_workqueue("spusched");
478 spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
480 printk(KERN_WARNING "%s: Unable to allocate priority queue.\n",
482 destroy_workqueue(spu_sched_wq);
485 for (i = 0; i < MAX_PRIO; i++) {
486 INIT_LIST_HEAD(&spu_prio->runq[i]);
487 __clear_bit(i, spu_prio->bitmap);
489 __set_bit(MAX_PRIO, spu_prio->bitmap);
490 for (i = 0; i < MAX_NUMNODES; i++) {
491 mutex_init(&spu_prio->active_mutex[i]);
492 INIT_LIST_HEAD(&spu_prio->active_list[i]);
494 spin_lock_init(&spu_prio->runq_lock);
498 void __exit spu_sched_exit(void)
500 struct spu *spu, *tmp;
503 for (node = 0; node < MAX_NUMNODES; node++) {
504 mutex_lock(&spu_prio->active_mutex[node]);
505 list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
507 list_del_init(&spu->list);
510 mutex_unlock(&spu_prio->active_mutex[node]);
513 destroy_workqueue(spu_sched_wq);