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)
75 queue_delayed_work(spu_sched_wq, &ctx->sched_work, SPU_TIMESLICE);
78 void spu_stop_tick(struct spu_context *ctx)
80 if (ctx->policy == SCHED_RR)
81 cancel_delayed_work(&ctx->sched_work);
84 void spu_sched_tick(struct work_struct *work)
86 struct spu_context *ctx =
87 container_of(work, struct spu_context, sched_work.work);
91 mutex_lock(&ctx->state_mutex);
94 int best = sched_find_first_bit(spu_prio->bitmap);
95 if (best <= ctx->prio) {
100 mutex_unlock(&ctx->state_mutex);
107 * spu_add_to_active_list - add spu to active list
108 * @spu: spu to add to the active list
110 static void spu_add_to_active_list(struct spu *spu)
112 mutex_lock(&spu_prio->active_mutex[spu->node]);
113 list_add_tail(&spu->list, &spu_prio->active_list[spu->node]);
114 mutex_unlock(&spu_prio->active_mutex[spu->node]);
118 * spu_remove_from_active_list - remove spu from active list
119 * @spu: spu to remove from the active list
121 static void spu_remove_from_active_list(struct spu *spu)
123 int node = spu->node;
125 mutex_lock(&spu_prio->active_mutex[node]);
126 list_del_init(&spu->list);
127 mutex_unlock(&spu_prio->active_mutex[node]);
130 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
132 int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
134 /* Global TLBIE broadcast required with SPEs. */
135 __cpus_setall(&mm->cpu_vm_mask, nr);
138 static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
140 static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
142 blocking_notifier_call_chain(&spu_switch_notifier,
143 ctx ? ctx->object_id : 0, spu);
146 int spu_switch_event_register(struct notifier_block * n)
148 return blocking_notifier_chain_register(&spu_switch_notifier, n);
151 int spu_switch_event_unregister(struct notifier_block * n)
153 return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
157 * spu_bind_context - bind spu context to physical spu
158 * @spu: physical spu to bind to
159 * @ctx: context to bind
161 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
163 pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
164 spu->number, spu->node);
168 ctx->ops = &spu_hw_ops;
169 spu->pid = current->pid;
170 spu->mm = ctx->owner;
171 mm_needs_global_tlbie(spu->mm);
172 spu->ibox_callback = spufs_ibox_callback;
173 spu->wbox_callback = spufs_wbox_callback;
174 spu->stop_callback = spufs_stop_callback;
175 spu->mfc_callback = spufs_mfc_callback;
176 spu->dma_callback = spufs_dma_callback;
178 spu_unmap_mappings(ctx);
179 spu_restore(&ctx->csa, spu);
180 spu->timestamp = jiffies;
181 spu_cpu_affinity_set(spu, raw_smp_processor_id());
182 spu_switch_notify(spu, ctx);
183 spu_add_to_active_list(spu);
184 ctx->state = SPU_STATE_RUNNABLE;
188 * spu_unbind_context - unbind spu context from physical spu
189 * @spu: physical spu to unbind from
190 * @ctx: context to unbind
192 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
194 pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
195 spu->pid, spu->number, spu->node);
197 spu_remove_from_active_list(spu);
198 spu_switch_notify(spu, NULL);
199 spu_unmap_mappings(ctx);
200 spu_save(&ctx->csa, spu);
201 spu->timestamp = jiffies;
202 ctx->state = SPU_STATE_SAVED;
203 spu->ibox_callback = NULL;
204 spu->wbox_callback = NULL;
205 spu->stop_callback = NULL;
206 spu->mfc_callback = NULL;
207 spu->dma_callback = NULL;
210 ctx->ops = &spu_backing_ops;
217 * spu_add_to_rq - add a context to the runqueue
218 * @ctx: context to add
220 static void spu_add_to_rq(struct spu_context *ctx)
222 spin_lock(&spu_prio->runq_lock);
223 list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
224 set_bit(ctx->prio, spu_prio->bitmap);
225 spin_unlock(&spu_prio->runq_lock);
229 * spu_del_from_rq - remove a context from the runqueue
230 * @ctx: context to remove
232 static void spu_del_from_rq(struct spu_context *ctx)
234 spin_lock(&spu_prio->runq_lock);
235 list_del_init(&ctx->rq);
236 if (list_empty(&spu_prio->runq[ctx->prio]))
237 clear_bit(ctx->prio, spu_prio->bitmap);
238 spin_unlock(&spu_prio->runq_lock);
242 * spu_grab_context - remove one context from the runqueue
243 * @prio: priority of the context to be removed
245 * This function removes one context from the runqueue for priority @prio.
246 * If there is more than one context with the given priority the first
247 * task on the runqueue will be taken.
249 * Returns the spu_context it just removed.
251 * Must be called with spu_prio->runq_lock held.
253 static struct spu_context *spu_grab_context(int prio)
255 struct list_head *rq = &spu_prio->runq[prio];
259 return list_entry(rq->next, struct spu_context, rq);
262 static void spu_prio_wait(struct spu_context *ctx)
266 set_bit(SPU_SCHED_WAKE, &ctx->sched_flags);
267 prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
268 if (!signal_pending(current)) {
269 mutex_unlock(&ctx->state_mutex);
271 mutex_lock(&ctx->state_mutex);
273 __set_current_state(TASK_RUNNING);
274 remove_wait_queue(&ctx->stop_wq, &wait);
275 clear_bit(SPU_SCHED_WAKE, &ctx->sched_flags);
279 * spu_reschedule - try to find a runnable context for a spu
280 * @spu: spu available
282 * This function is called whenever a spu becomes idle. It looks for the
283 * most suitable runnable spu context and schedules it for execution.
285 static void spu_reschedule(struct spu *spu)
291 spin_lock(&spu_prio->runq_lock);
292 best = sched_find_first_bit(spu_prio->bitmap);
293 if (best < MAX_PRIO) {
294 struct spu_context *ctx = spu_grab_context(best);
295 if (ctx && test_bit(SPU_SCHED_WAKE, &ctx->sched_flags))
296 wake_up(&ctx->stop_wq);
298 spin_unlock(&spu_prio->runq_lock);
301 static struct spu *spu_get_idle(struct spu_context *ctx)
303 struct spu *spu = NULL;
304 int node = cpu_to_node(raw_smp_processor_id());
307 for (n = 0; n < MAX_NUMNODES; n++, node++) {
308 node = (node < MAX_NUMNODES) ? node : 0;
309 if (!node_allowed(node))
311 spu = spu_alloc_node(node);
319 * find_victim - find a lower priority context to preempt
320 * @ctx: canidate context for running
322 * Returns the freed physical spu to run the new context on.
324 static struct spu *find_victim(struct spu_context *ctx)
326 struct spu_context *victim = NULL;
331 * Look for a possible preemption candidate on the local node first.
332 * If there is no candidate look at the other nodes. This isn't
333 * exactly fair, but so far the whole spu schedule tries to keep
334 * a strong node affinity. We might want to fine-tune this in
338 node = cpu_to_node(raw_smp_processor_id());
339 for (n = 0; n < MAX_NUMNODES; n++, node++) {
340 node = (node < MAX_NUMNODES) ? node : 0;
341 if (!node_allowed(node))
344 mutex_lock(&spu_prio->active_mutex[node]);
345 list_for_each_entry(spu, &spu_prio->active_list[node], list) {
346 struct spu_context *tmp = spu->ctx;
348 if (tmp->rt_priority < ctx->rt_priority &&
349 (!victim || tmp->rt_priority < victim->rt_priority))
352 mutex_unlock(&spu_prio->active_mutex[node]);
356 * This nests ctx->state_mutex, but we always lock
357 * higher priority contexts before lower priority
358 * ones, so this is safe until we introduce
359 * priority inheritance schemes.
361 if (!mutex_trylock(&victim->state_mutex)) {
369 * This race can happen because we've dropped
370 * the active list mutex. No a problem, just
371 * restart the search.
373 mutex_unlock(&victim->state_mutex);
377 spu_unbind_context(spu, victim);
378 mutex_unlock(&victim->state_mutex);
387 * spu_activate - find a free spu for a context and execute it
388 * @ctx: spu context to schedule
389 * @flags: flags (currently ignored)
391 * Tries to find a free spu to run @ctx. If no free spu is availble
392 * add the context to the runqueue so it gets woken up once an spu
395 int spu_activate(struct spu_context *ctx, unsigned long flags)
404 spu = spu_get_idle(ctx);
406 * If this is a realtime thread we try to get it running by
407 * preempting a lower priority thread.
409 if (!spu && ctx->rt_priority)
410 spu = find_victim(ctx);
412 spu_bind_context(spu, ctx);
417 if (!(flags & SPU_ACTIVATE_NOWAKE))
419 spu_del_from_rq(ctx);
420 } while (!signal_pending(current));
426 * spu_deactivate - unbind a context from it's physical spu
427 * @ctx: spu context to unbind
429 * Unbind @ctx from the physical spu it is running on and schedule
430 * the highest priority context to run on the freed physical spu.
432 void spu_deactivate(struct spu_context *ctx)
434 struct spu *spu = ctx->spu;
437 spu_unbind_context(spu, ctx);
443 * spu_yield - yield a physical spu if others are waiting
444 * @ctx: spu context to yield
446 * Check if there is a higher priority context waiting and if yes
447 * unbind @ctx from the physical spu and schedule the highest
448 * priority context to run on the freed physical spu instead.
450 void spu_yield(struct spu_context *ctx)
455 if (mutex_trylock(&ctx->state_mutex)) {
456 if ((spu = ctx->spu) != NULL) {
457 int best = sched_find_first_bit(spu_prio->bitmap);
458 if (best < MAX_PRIO) {
459 pr_debug("%s: yielding SPU %d NODE %d\n",
460 __FUNCTION__, spu->number, spu->node);
465 mutex_unlock(&ctx->state_mutex);
467 if (unlikely(need_yield))
471 int __init spu_sched_init(void)
475 spu_sched_wq = create_singlethread_workqueue("spusched");
479 spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
481 printk(KERN_WARNING "%s: Unable to allocate priority queue.\n",
483 destroy_workqueue(spu_sched_wq);
486 for (i = 0; i < MAX_PRIO; i++) {
487 INIT_LIST_HEAD(&spu_prio->runq[i]);
488 __clear_bit(i, spu_prio->bitmap);
490 __set_bit(MAX_PRIO, spu_prio->bitmap);
491 for (i = 0; i < MAX_NUMNODES; i++) {
492 mutex_init(&spu_prio->active_mutex[i]);
493 INIT_LIST_HEAD(&spu_prio->active_list[i]);
495 spin_lock_init(&spu_prio->runq_lock);
499 void __exit spu_sched_exit(void)
501 struct spu *spu, *tmp;
504 for (node = 0; node < MAX_NUMNODES; node++) {
505 mutex_lock(&spu_prio->active_mutex[node]);
506 list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
508 list_del_init(&spu->list);
511 mutex_unlock(&spu_prio->active_mutex[node]);
514 destroy_workqueue(spu_sched_wq);