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>
38 #include <linux/kthread.h>
39 #include <linux/pid_namespace.h>
40 #include <linux/proc_fs.h>
41 #include <linux/seq_file.h>
44 #include <asm/mmu_context.h>
46 #include <asm/spu_csa.h>
47 #include <asm/spu_priv1.h>
50 struct spu_prio_array {
51 DECLARE_BITMAP(bitmap, MAX_PRIO);
52 struct list_head runq[MAX_PRIO];
57 static unsigned long spu_avenrun[3];
58 static struct spu_prio_array *spu_prio;
59 static struct task_struct *spusched_task;
60 static struct timer_list spusched_timer;
63 * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
65 #define NORMAL_PRIO 120
68 * Frequency of the spu scheduler tick. By default we do one SPU scheduler
69 * tick for every 10 CPU scheduler ticks.
71 #define SPUSCHED_TICK (10)
74 * These are the 'tuning knobs' of the scheduler:
76 * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is
77 * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs.
79 #define MIN_SPU_TIMESLICE max(5 * HZ / (1000 * SPUSCHED_TICK), 1)
80 #define DEF_SPU_TIMESLICE (100 * HZ / (1000 * SPUSCHED_TICK))
82 #define MAX_USER_PRIO (MAX_PRIO - MAX_RT_PRIO)
83 #define SCALE_PRIO(x, prio) \
84 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE)
87 * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values:
88 * [800ms ... 100ms ... 5ms]
90 * The higher a thread's priority, the bigger timeslices
91 * it gets during one round of execution. But even the lowest
92 * priority thread gets MIN_TIMESLICE worth of execution time.
94 void spu_set_timeslice(struct spu_context *ctx)
96 if (ctx->prio < NORMAL_PRIO)
97 ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio);
99 ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
103 * Update scheduling information from the owning thread.
105 void __spu_update_sched_info(struct spu_context *ctx)
108 * 32-Bit assignment are atomic on powerpc, and we don't care about
109 * memory ordering here because retriving the controlling thread is
110 * per defintion racy.
112 ctx->tid = current->pid;
115 * We do our own priority calculations, so we normally want
116 * ->static_prio to start with. Unfortunately thies field
117 * contains junk for threads with a realtime scheduling
118 * policy so we have to look at ->prio in this case.
120 if (rt_prio(current->prio))
121 ctx->prio = current->prio;
123 ctx->prio = current->static_prio;
124 ctx->policy = current->policy;
127 * A lot of places that don't hold list_mutex poke into
128 * cpus_allowed, including grab_runnable_context which
129 * already holds the runq_lock. So abuse runq_lock
130 * to protect this field aswell.
132 spin_lock(&spu_prio->runq_lock);
133 ctx->cpus_allowed = current->cpus_allowed;
134 spin_unlock(&spu_prio->runq_lock);
137 void spu_update_sched_info(struct spu_context *ctx)
139 int node = ctx->spu->node;
141 mutex_lock(&cbe_spu_info[node].list_mutex);
142 __spu_update_sched_info(ctx);
143 mutex_unlock(&cbe_spu_info[node].list_mutex);
146 static int __node_allowed(struct spu_context *ctx, int node)
148 if (nr_cpus_node(node)) {
149 cpumask_t mask = node_to_cpumask(node);
151 if (cpus_intersects(mask, ctx->cpus_allowed))
158 static int node_allowed(struct spu_context *ctx, int node)
162 spin_lock(&spu_prio->runq_lock);
163 rval = __node_allowed(ctx, node);
164 spin_unlock(&spu_prio->runq_lock);
169 static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
171 void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
173 blocking_notifier_call_chain(&spu_switch_notifier,
174 ctx ? ctx->object_id : 0, spu);
177 static void notify_spus_active(void)
182 * Wake up the active spu_contexts.
184 * When the awakened processes see their "notify_active" flag is set,
185 * they will call spu_switch_notify();
187 for_each_online_node(node) {
190 mutex_lock(&cbe_spu_info[node].list_mutex);
191 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
192 if (spu->alloc_state != SPU_FREE) {
193 struct spu_context *ctx = spu->ctx;
194 set_bit(SPU_SCHED_NOTIFY_ACTIVE,
197 wake_up_all(&ctx->stop_wq);
200 mutex_unlock(&cbe_spu_info[node].list_mutex);
204 int spu_switch_event_register(struct notifier_block * n)
207 ret = blocking_notifier_chain_register(&spu_switch_notifier, n);
209 notify_spus_active();
212 EXPORT_SYMBOL_GPL(spu_switch_event_register);
214 int spu_switch_event_unregister(struct notifier_block * n)
216 return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
218 EXPORT_SYMBOL_GPL(spu_switch_event_unregister);
221 * spu_bind_context - bind spu context to physical spu
222 * @spu: physical spu to bind to
223 * @ctx: context to bind
225 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
227 pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
228 spu->number, spu->node);
229 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
231 if (ctx->flags & SPU_CREATE_NOSCHED)
232 atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
233 if (!list_empty(&ctx->aff_list))
234 atomic_inc(&ctx->gang->aff_sched_count);
236 ctx->stats.slb_flt_base = spu->stats.slb_flt;
237 ctx->stats.class2_intr_base = spu->stats.class2_intr;
242 ctx->ops = &spu_hw_ops;
243 spu->pid = current->pid;
244 spu->tgid = current->tgid;
245 spu_associate_mm(spu, ctx->owner);
246 spu->ibox_callback = spufs_ibox_callback;
247 spu->wbox_callback = spufs_wbox_callback;
248 spu->stop_callback = spufs_stop_callback;
249 spu->mfc_callback = spufs_mfc_callback;
250 spu->dma_callback = spufs_dma_callback;
252 spu_unmap_mappings(ctx);
253 spu_restore(&ctx->csa, spu);
254 spu->timestamp = jiffies;
255 spu_cpu_affinity_set(spu, raw_smp_processor_id());
256 spu_switch_notify(spu, ctx);
257 ctx->state = SPU_STATE_RUNNABLE;
259 spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
263 * Must be used with the list_mutex held.
265 static inline int sched_spu(struct spu *spu)
267 BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex));
269 return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED));
272 static void aff_merge_remaining_ctxs(struct spu_gang *gang)
274 struct spu_context *ctx;
276 list_for_each_entry(ctx, &gang->aff_list_head, aff_list) {
277 if (list_empty(&ctx->aff_list))
278 list_add(&ctx->aff_list, &gang->aff_list_head);
280 gang->aff_flags |= AFF_MERGED;
283 static void aff_set_offsets(struct spu_gang *gang)
285 struct spu_context *ctx;
289 list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
291 if (&ctx->aff_list == &gang->aff_list_head)
293 ctx->aff_offset = offset--;
297 list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) {
298 if (&ctx->aff_list == &gang->aff_list_head)
300 ctx->aff_offset = offset++;
303 gang->aff_flags |= AFF_OFFSETS_SET;
306 static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff,
307 int group_size, int lowest_offset)
313 * TODO: A better algorithm could be used to find a good spu to be
314 * used as reference location for the ctxs chain.
316 node = cpu_to_node(raw_smp_processor_id());
317 for (n = 0; n < MAX_NUMNODES; n++, node++) {
318 node = (node < MAX_NUMNODES) ? node : 0;
319 if (!node_allowed(ctx, node))
321 mutex_lock(&cbe_spu_info[node].list_mutex);
322 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
323 if ((!mem_aff || spu->has_mem_affinity) &&
325 mutex_unlock(&cbe_spu_info[node].list_mutex);
329 mutex_unlock(&cbe_spu_info[node].list_mutex);
334 static void aff_set_ref_point_location(struct spu_gang *gang)
336 int mem_aff, gs, lowest_offset;
337 struct spu_context *ctx;
340 mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
344 list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
347 list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
349 if (&ctx->aff_list == &gang->aff_list_head)
351 lowest_offset = ctx->aff_offset;
354 gang->aff_ref_spu = aff_ref_location(ctx, mem_aff, gs, lowest_offset);
357 static struct spu *ctx_location(struct spu *ref, int offset, int node)
363 list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
364 BUG_ON(spu->node != node);
371 list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
372 BUG_ON(spu->node != node);
384 * affinity_check is called each time a context is going to be scheduled.
385 * It returns the spu ptr on which the context must run.
387 static int has_affinity(struct spu_context *ctx)
389 struct spu_gang *gang = ctx->gang;
391 if (list_empty(&ctx->aff_list))
394 mutex_lock(&gang->aff_mutex);
395 if (!gang->aff_ref_spu) {
396 if (!(gang->aff_flags & AFF_MERGED))
397 aff_merge_remaining_ctxs(gang);
398 if (!(gang->aff_flags & AFF_OFFSETS_SET))
399 aff_set_offsets(gang);
400 aff_set_ref_point_location(gang);
402 mutex_unlock(&gang->aff_mutex);
404 return gang->aff_ref_spu != NULL;
408 * spu_unbind_context - unbind spu context from physical spu
409 * @spu: physical spu to unbind from
410 * @ctx: context to unbind
412 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
414 pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
415 spu->pid, spu->number, spu->node);
416 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
418 if (spu->ctx->flags & SPU_CREATE_NOSCHED)
419 atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
420 if (!list_empty(&ctx->aff_list))
421 if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
422 ctx->gang->aff_ref_spu = NULL;
423 spu_switch_notify(spu, NULL);
424 spu_unmap_mappings(ctx);
425 spu_save(&ctx->csa, spu);
426 spu->timestamp = jiffies;
427 ctx->state = SPU_STATE_SAVED;
428 spu->ibox_callback = NULL;
429 spu->wbox_callback = NULL;
430 spu->stop_callback = NULL;
431 spu->mfc_callback = NULL;
432 spu->dma_callback = NULL;
433 spu_associate_mm(spu, NULL);
436 ctx->ops = &spu_backing_ops;
440 ctx->stats.slb_flt +=
441 (spu->stats.slb_flt - ctx->stats.slb_flt_base);
442 ctx->stats.class2_intr +=
443 (spu->stats.class2_intr - ctx->stats.class2_intr_base);
445 /* This maps the underlying spu state to idle */
446 spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
451 * spu_add_to_rq - add a context to the runqueue
452 * @ctx: context to add
454 static void __spu_add_to_rq(struct spu_context *ctx)
457 * Unfortunately this code path can be called from multiple threads
458 * on behalf of a single context due to the way the problem state
459 * mmap support works.
461 * Fortunately we need to wake up all these threads at the same time
462 * and can simply skip the runqueue addition for every but the first
463 * thread getting into this codepath.
465 * It's still quite hacky, and long-term we should proxy all other
466 * threads through the owner thread so that spu_run is in control
467 * of all the scheduling activity for a given context.
469 if (list_empty(&ctx->rq)) {
470 list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
471 set_bit(ctx->prio, spu_prio->bitmap);
472 if (!spu_prio->nr_waiting++)
473 __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
477 static void __spu_del_from_rq(struct spu_context *ctx)
479 int prio = ctx->prio;
481 if (!list_empty(&ctx->rq)) {
482 if (!--spu_prio->nr_waiting)
483 del_timer(&spusched_timer);
484 list_del_init(&ctx->rq);
486 if (list_empty(&spu_prio->runq[prio]))
487 clear_bit(prio, spu_prio->bitmap);
491 static void spu_prio_wait(struct spu_context *ctx)
495 spin_lock(&spu_prio->runq_lock);
496 prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
497 if (!signal_pending(current)) {
498 __spu_add_to_rq(ctx);
499 spin_unlock(&spu_prio->runq_lock);
500 mutex_unlock(&ctx->state_mutex);
502 mutex_lock(&ctx->state_mutex);
503 spin_lock(&spu_prio->runq_lock);
504 __spu_del_from_rq(ctx);
506 spin_unlock(&spu_prio->runq_lock);
507 __set_current_state(TASK_RUNNING);
508 remove_wait_queue(&ctx->stop_wq, &wait);
511 static struct spu *spu_get_idle(struct spu_context *ctx)
516 if (has_affinity(ctx)) {
517 node = ctx->gang->aff_ref_spu->node;
519 mutex_lock(&cbe_spu_info[node].list_mutex);
520 spu = ctx_location(ctx->gang->aff_ref_spu, ctx->aff_offset, node);
521 if (spu && spu->alloc_state == SPU_FREE)
523 mutex_unlock(&cbe_spu_info[node].list_mutex);
527 node = cpu_to_node(raw_smp_processor_id());
528 for (n = 0; n < MAX_NUMNODES; n++, node++) {
529 node = (node < MAX_NUMNODES) ? node : 0;
530 if (!node_allowed(ctx, node))
533 mutex_lock(&cbe_spu_info[node].list_mutex);
534 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
535 if (spu->alloc_state == SPU_FREE)
538 mutex_unlock(&cbe_spu_info[node].list_mutex);
544 spu->alloc_state = SPU_USED;
545 mutex_unlock(&cbe_spu_info[node].list_mutex);
546 pr_debug("Got SPU %d %d\n", spu->number, spu->node);
547 spu_init_channels(spu);
552 * find_victim - find a lower priority context to preempt
553 * @ctx: canidate context for running
555 * Returns the freed physical spu to run the new context on.
557 static struct spu *find_victim(struct spu_context *ctx)
559 struct spu_context *victim = NULL;
564 * Look for a possible preemption candidate on the local node first.
565 * If there is no candidate look at the other nodes. This isn't
566 * exactly fair, but so far the whole spu schedule tries to keep
567 * a strong node affinity. We might want to fine-tune this in
571 node = cpu_to_node(raw_smp_processor_id());
572 for (n = 0; n < MAX_NUMNODES; n++, node++) {
573 node = (node < MAX_NUMNODES) ? node : 0;
574 if (!node_allowed(ctx, node))
577 mutex_lock(&cbe_spu_info[node].list_mutex);
578 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
579 struct spu_context *tmp = spu->ctx;
581 if (tmp->prio > ctx->prio &&
582 (!victim || tmp->prio > victim->prio))
585 mutex_unlock(&cbe_spu_info[node].list_mutex);
589 * This nests ctx->state_mutex, but we always lock
590 * higher priority contexts before lower priority
591 * ones, so this is safe until we introduce
592 * priority inheritance schemes.
594 if (!mutex_trylock(&victim->state_mutex)) {
602 * This race can happen because we've dropped
603 * the active list mutex. No a problem, just
604 * restart the search.
606 mutex_unlock(&victim->state_mutex);
611 mutex_lock(&cbe_spu_info[node].list_mutex);
612 cbe_spu_info[node].nr_active--;
613 mutex_unlock(&cbe_spu_info[node].list_mutex);
615 spu_unbind_context(spu, victim);
616 victim->stats.invol_ctx_switch++;
617 spu->stats.invol_ctx_switch++;
618 mutex_unlock(&victim->state_mutex);
620 * We need to break out of the wait loop in spu_run
621 * manually to ensure this context gets put on the
622 * runqueue again ASAP.
624 wake_up(&victim->stop_wq);
633 * spu_activate - find a free spu for a context and execute it
634 * @ctx: spu context to schedule
635 * @flags: flags (currently ignored)
637 * Tries to find a free spu to run @ctx. If no free spu is available
638 * add the context to the runqueue so it gets woken up once an spu
641 int spu_activate(struct spu_context *ctx, unsigned long flags)
647 * If there are multiple threads waiting for a single context
648 * only one actually binds the context while the others will
649 * only be able to acquire the state_mutex once the context
650 * already is in runnable state.
655 spu = spu_get_idle(ctx);
657 * If this is a realtime thread we try to get it running by
658 * preempting a lower priority thread.
660 if (!spu && rt_prio(ctx->prio))
661 spu = find_victim(ctx);
663 int node = spu->node;
665 mutex_lock(&cbe_spu_info[node].list_mutex);
666 spu_bind_context(spu, ctx);
667 cbe_spu_info[node].nr_active++;
668 mutex_unlock(&cbe_spu_info[node].list_mutex);
673 } while (!signal_pending(current));
679 * grab_runnable_context - try to find a runnable context
681 * Remove the highest priority context on the runqueue and return it
682 * to the caller. Returns %NULL if no runnable context was found.
684 static struct spu_context *grab_runnable_context(int prio, int node)
686 struct spu_context *ctx;
689 spin_lock(&spu_prio->runq_lock);
690 best = find_first_bit(spu_prio->bitmap, prio);
691 while (best < prio) {
692 struct list_head *rq = &spu_prio->runq[best];
694 list_for_each_entry(ctx, rq, rq) {
695 /* XXX(hch): check for affinity here aswell */
696 if (__node_allowed(ctx, node)) {
697 __spu_del_from_rq(ctx);
705 spin_unlock(&spu_prio->runq_lock);
709 static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
711 struct spu *spu = ctx->spu;
712 struct spu_context *new = NULL;
715 new = grab_runnable_context(max_prio, spu->node);
717 int node = spu->node;
719 mutex_lock(&cbe_spu_info[node].list_mutex);
720 spu_unbind_context(spu, ctx);
721 spu->alloc_state = SPU_FREE;
722 cbe_spu_info[node].nr_active--;
723 mutex_unlock(&cbe_spu_info[node].list_mutex);
725 ctx->stats.vol_ctx_switch++;
726 spu->stats.vol_ctx_switch++;
729 wake_up(&new->stop_wq);
738 * spu_deactivate - unbind a context from it's physical spu
739 * @ctx: spu context to unbind
741 * Unbind @ctx from the physical spu it is running on and schedule
742 * the highest priority context to run on the freed physical spu.
744 void spu_deactivate(struct spu_context *ctx)
746 __spu_deactivate(ctx, 1, MAX_PRIO);
750 * spu_yield - yield a physical spu if others are waiting
751 * @ctx: spu context to yield
753 * Check if there is a higher priority context waiting and if yes
754 * unbind @ctx from the physical spu and schedule the highest
755 * priority context to run on the freed physical spu instead.
757 void spu_yield(struct spu_context *ctx)
759 if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
760 mutex_lock(&ctx->state_mutex);
761 __spu_deactivate(ctx, 0, MAX_PRIO);
762 mutex_unlock(&ctx->state_mutex);
766 static noinline void spusched_tick(struct spu_context *ctx)
768 if (ctx->flags & SPU_CREATE_NOSCHED)
770 if (ctx->policy == SCHED_FIFO)
773 if (--ctx->time_slice)
777 * Unfortunately list_mutex ranks outside of state_mutex, so
778 * we have to trylock here. If we fail give the context another
779 * tick and try again.
781 if (mutex_trylock(&ctx->state_mutex)) {
782 struct spu *spu = ctx->spu;
783 struct spu_context *new;
785 new = grab_runnable_context(ctx->prio + 1, spu->node);
787 spu_unbind_context(spu, ctx);
788 ctx->stats.invol_ctx_switch++;
789 spu->stats.invol_ctx_switch++;
790 spu->alloc_state = SPU_FREE;
791 cbe_spu_info[spu->node].nr_active--;
792 wake_up(&new->stop_wq);
794 * We need to break out of the wait loop in
795 * spu_run manually to ensure this context
796 * gets put on the runqueue again ASAP.
798 wake_up(&ctx->stop_wq);
800 spu_set_timeslice(ctx);
801 mutex_unlock(&ctx->state_mutex);
808 * count_active_contexts - count nr of active tasks
810 * Return the number of tasks currently running or waiting to run.
812 * Note that we don't take runq_lock / list_mutex here. Reading
813 * a single 32bit value is atomic on powerpc, and we don't care
814 * about memory ordering issues here.
816 static unsigned long count_active_contexts(void)
818 int nr_active = 0, node;
820 for (node = 0; node < MAX_NUMNODES; node++)
821 nr_active += cbe_spu_info[node].nr_active;
822 nr_active += spu_prio->nr_waiting;
828 * spu_calc_load - given tick count, update the avenrun load estimates.
831 * No locking against reading these values from userspace, as for
832 * the CPU loadavg code.
834 static void spu_calc_load(unsigned long ticks)
836 unsigned long active_tasks; /* fixed-point */
837 static int count = LOAD_FREQ;
841 if (unlikely(count < 0)) {
842 active_tasks = count_active_contexts() * FIXED_1;
844 CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks);
845 CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks);
846 CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks);
852 static void spusched_wake(unsigned long data)
854 mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
855 wake_up_process(spusched_task);
856 spu_calc_load(SPUSCHED_TICK);
859 static int spusched_thread(void *unused)
864 while (!kthread_should_stop()) {
865 set_current_state(TASK_INTERRUPTIBLE);
867 for (node = 0; node < MAX_NUMNODES; node++) {
868 mutex_lock(&cbe_spu_info[node].list_mutex);
869 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
871 spusched_tick(spu->ctx);
872 mutex_unlock(&cbe_spu_info[node].list_mutex);
879 #define LOAD_INT(x) ((x) >> FSHIFT)
880 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
882 static int show_spu_loadavg(struct seq_file *s, void *private)
886 a = spu_avenrun[0] + (FIXED_1/200);
887 b = spu_avenrun[1] + (FIXED_1/200);
888 c = spu_avenrun[2] + (FIXED_1/200);
891 * Note that last_pid doesn't really make much sense for the
892 * SPU loadavg (it even seems very odd on the CPU side..),
893 * but we include it here to have a 100% compatible interface.
895 seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
896 LOAD_INT(a), LOAD_FRAC(a),
897 LOAD_INT(b), LOAD_FRAC(b),
898 LOAD_INT(c), LOAD_FRAC(c),
899 count_active_contexts(),
900 atomic_read(&nr_spu_contexts),
901 current->nsproxy->pid_ns->last_pid);
905 static int spu_loadavg_open(struct inode *inode, struct file *file)
907 return single_open(file, show_spu_loadavg, NULL);
910 static const struct file_operations spu_loadavg_fops = {
911 .open = spu_loadavg_open,
914 .release = single_release,
917 int __init spu_sched_init(void)
919 struct proc_dir_entry *entry;
920 int err = -ENOMEM, i;
922 spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
926 for (i = 0; i < MAX_PRIO; i++) {
927 INIT_LIST_HEAD(&spu_prio->runq[i]);
928 __clear_bit(i, spu_prio->bitmap);
930 for (i = 0; i < MAX_NUMNODES; i++) {
931 mutex_init(&cbe_spu_info[i].list_mutex);
932 INIT_LIST_HEAD(&cbe_spu_info[i].spus);
934 spin_lock_init(&spu_prio->runq_lock);
936 setup_timer(&spusched_timer, spusched_wake, 0);
938 spusched_task = kthread_run(spusched_thread, NULL, "spusched");
939 if (IS_ERR(spusched_task)) {
940 err = PTR_ERR(spusched_task);
941 goto out_free_spu_prio;
944 entry = create_proc_entry("spu_loadavg", 0, NULL);
946 goto out_stop_kthread;
947 entry->proc_fops = &spu_loadavg_fops;
949 pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
950 SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE);
954 kthread_stop(spusched_task);
961 void spu_sched_exit(void)
966 remove_proc_entry("spu_loadavg", NULL);
968 del_timer_sync(&spusched_timer);
969 kthread_stop(spusched_task);
971 for (node = 0; node < MAX_NUMNODES; node++) {
972 mutex_lock(&cbe_spu_info[node].list_mutex);
973 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
974 if (spu->alloc_state != SPU_FREE)
975 spu->alloc_state = SPU_FREE;
976 mutex_unlock(&cbe_spu_info[node].list_mutex);