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];
54 struct list_head active_list[MAX_NUMNODES];
55 struct mutex active_mutex[MAX_NUMNODES];
56 int nr_active[MAX_NUMNODES];
60 static unsigned long spu_avenrun[3];
61 static struct spu_prio_array *spu_prio;
62 static struct task_struct *spusched_task;
63 static struct timer_list spusched_timer;
66 * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
68 #define NORMAL_PRIO 120
71 * Frequency of the spu scheduler tick. By default we do one SPU scheduler
72 * tick for every 10 CPU scheduler ticks.
74 #define SPUSCHED_TICK (10)
77 * These are the 'tuning knobs' of the scheduler:
79 * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is
80 * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs.
82 #define MIN_SPU_TIMESLICE max(5 * HZ / (1000 * SPUSCHED_TICK), 1)
83 #define DEF_SPU_TIMESLICE (100 * HZ / (1000 * SPUSCHED_TICK))
85 #define MAX_USER_PRIO (MAX_PRIO - MAX_RT_PRIO)
86 #define SCALE_PRIO(x, prio) \
87 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE)
90 * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values:
91 * [800ms ... 100ms ... 5ms]
93 * The higher a thread's priority, the bigger timeslices
94 * it gets during one round of execution. But even the lowest
95 * priority thread gets MIN_TIMESLICE worth of execution time.
97 void spu_set_timeslice(struct spu_context *ctx)
99 if (ctx->prio < NORMAL_PRIO)
100 ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio);
102 ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
106 * Update scheduling information from the owning thread.
108 void __spu_update_sched_info(struct spu_context *ctx)
111 * 32-Bit assignment are atomic on powerpc, and we don't care about
112 * memory ordering here because retriving the controlling thread is
113 * per defintion racy.
115 ctx->tid = current->pid;
118 * We do our own priority calculations, so we normally want
119 * ->static_prio to start with. Unfortunately thies field
120 * contains junk for threads with a realtime scheduling
121 * policy so we have to look at ->prio in this case.
123 if (rt_prio(current->prio))
124 ctx->prio = current->prio;
126 ctx->prio = current->static_prio;
127 ctx->policy = current->policy;
130 * A lot of places that don't hold active_mutex poke into
131 * cpus_allowed, including grab_runnable_context which
132 * already holds the runq_lock. So abuse runq_lock
133 * to protect this field aswell.
135 spin_lock(&spu_prio->runq_lock);
136 ctx->cpus_allowed = current->cpus_allowed;
137 spin_unlock(&spu_prio->runq_lock);
140 void spu_update_sched_info(struct spu_context *ctx)
142 int node = ctx->spu->node;
144 mutex_lock(&spu_prio->active_mutex[node]);
145 __spu_update_sched_info(ctx);
146 mutex_unlock(&spu_prio->active_mutex[node]);
149 static int __node_allowed(struct spu_context *ctx, int node)
151 if (nr_cpus_node(node)) {
152 cpumask_t mask = node_to_cpumask(node);
154 if (cpus_intersects(mask, ctx->cpus_allowed))
161 static int node_allowed(struct spu_context *ctx, int node)
165 spin_lock(&spu_prio->runq_lock);
166 rval = __node_allowed(ctx, node);
167 spin_unlock(&spu_prio->runq_lock);
173 * spu_add_to_active_list - add spu to active list
174 * @spu: spu to add to the active list
176 static void spu_add_to_active_list(struct spu *spu)
178 int node = spu->node;
180 mutex_lock(&spu_prio->active_mutex[node]);
181 spu_prio->nr_active[node]++;
182 list_add_tail(&spu->list, &spu_prio->active_list[node]);
183 mutex_unlock(&spu_prio->active_mutex[node]);
186 static void __spu_remove_from_active_list(struct spu *spu)
188 list_del_init(&spu->list);
189 spu_prio->nr_active[spu->node]--;
193 * spu_remove_from_active_list - remove spu from active list
194 * @spu: spu to remove from the active list
196 static void spu_remove_from_active_list(struct spu *spu)
198 int node = spu->node;
200 mutex_lock(&spu_prio->active_mutex[node]);
201 __spu_remove_from_active_list(spu);
202 mutex_unlock(&spu_prio->active_mutex[node]);
205 static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
207 static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
209 blocking_notifier_call_chain(&spu_switch_notifier,
210 ctx ? ctx->object_id : 0, spu);
213 int spu_switch_event_register(struct notifier_block * n)
215 return blocking_notifier_chain_register(&spu_switch_notifier, n);
218 int spu_switch_event_unregister(struct notifier_block * n)
220 return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
224 * spu_bind_context - bind spu context to physical spu
225 * @spu: physical spu to bind to
226 * @ctx: context to bind
228 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
230 pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
231 spu->number, spu->node);
232 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
234 if (ctx->flags & SPU_CREATE_NOSCHED)
235 atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
236 if (!list_empty(&ctx->aff_list))
237 atomic_inc(&ctx->gang->aff_sched_count);
239 ctx->stats.slb_flt_base = spu->stats.slb_flt;
240 ctx->stats.class2_intr_base = spu->stats.class2_intr;
245 ctx->ops = &spu_hw_ops;
246 spu->pid = current->pid;
247 spu_associate_mm(spu, ctx->owner);
248 spu->ibox_callback = spufs_ibox_callback;
249 spu->wbox_callback = spufs_wbox_callback;
250 spu->stop_callback = spufs_stop_callback;
251 spu->mfc_callback = spufs_mfc_callback;
252 spu->dma_callback = spufs_dma_callback;
254 spu_unmap_mappings(ctx);
255 spu_restore(&ctx->csa, spu);
256 spu->timestamp = jiffies;
257 spu_cpu_affinity_set(spu, raw_smp_processor_id());
258 spu_switch_notify(spu, ctx);
259 ctx->state = SPU_STATE_RUNNABLE;
261 spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
265 * XXX(hch): needs locking.
267 static inline int sched_spu(struct spu *spu)
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 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
322 if ((!mem_aff || spu->has_mem_affinity) &&
330 static void aff_set_ref_point_location(struct spu_gang *gang)
332 int mem_aff, gs, lowest_offset;
333 struct spu_context *ctx;
336 mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
340 list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
343 list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
345 if (&ctx->aff_list == &gang->aff_list_head)
347 lowest_offset = ctx->aff_offset;
350 gang->aff_ref_spu = aff_ref_location(ctx, mem_aff, gs, lowest_offset);
353 static struct spu *ctx_location(struct spu *ref, int offset)
359 list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
366 list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
377 * affinity_check is called each time a context is going to be scheduled.
378 * It returns the spu ptr on which the context must run.
380 struct spu *affinity_check(struct spu_context *ctx)
382 struct spu_gang *gang;
384 if (list_empty(&ctx->aff_list))
387 mutex_lock(&gang->aff_mutex);
388 if (!gang->aff_ref_spu) {
389 if (!(gang->aff_flags & AFF_MERGED))
390 aff_merge_remaining_ctxs(gang);
391 if (!(gang->aff_flags & AFF_OFFSETS_SET))
392 aff_set_offsets(gang);
393 aff_set_ref_point_location(gang);
395 mutex_unlock(&gang->aff_mutex);
396 if (!gang->aff_ref_spu)
398 return ctx_location(gang->aff_ref_spu, ctx->aff_offset);
402 * spu_unbind_context - unbind spu context from physical spu
403 * @spu: physical spu to unbind from
404 * @ctx: context to unbind
406 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
408 pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
409 spu->pid, spu->number, spu->node);
410 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
412 if (spu->ctx->flags & SPU_CREATE_NOSCHED)
413 atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
414 if (!list_empty(&ctx->aff_list))
415 if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
416 ctx->gang->aff_ref_spu = NULL;
417 spu_switch_notify(spu, NULL);
418 spu_unmap_mappings(ctx);
419 spu_save(&ctx->csa, spu);
420 spu->timestamp = jiffies;
421 ctx->state = SPU_STATE_SAVED;
422 spu->ibox_callback = NULL;
423 spu->wbox_callback = NULL;
424 spu->stop_callback = NULL;
425 spu->mfc_callback = NULL;
426 spu->dma_callback = NULL;
427 spu_associate_mm(spu, NULL);
429 ctx->ops = &spu_backing_ops;
433 ctx->stats.slb_flt +=
434 (spu->stats.slb_flt - ctx->stats.slb_flt_base);
435 ctx->stats.class2_intr +=
436 (spu->stats.class2_intr - ctx->stats.class2_intr_base);
438 /* This maps the underlying spu state to idle */
439 spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
444 * spu_add_to_rq - add a context to the runqueue
445 * @ctx: context to add
447 static void __spu_add_to_rq(struct spu_context *ctx)
450 * Unfortunately this code path can be called from multiple threads
451 * on behalf of a single context due to the way the problem state
452 * mmap support works.
454 * Fortunately we need to wake up all these threads at the same time
455 * and can simply skip the runqueue addition for every but the first
456 * thread getting into this codepath.
458 * It's still quite hacky, and long-term we should proxy all other
459 * threads through the owner thread so that spu_run is in control
460 * of all the scheduling activity for a given context.
462 if (list_empty(&ctx->rq)) {
463 list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
464 set_bit(ctx->prio, spu_prio->bitmap);
465 if (!spu_prio->nr_waiting++)
466 __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
470 static void __spu_del_from_rq(struct spu_context *ctx)
472 int prio = ctx->prio;
474 if (!list_empty(&ctx->rq)) {
475 if (!--spu_prio->nr_waiting)
476 del_timer(&spusched_timer);
477 list_del_init(&ctx->rq);
479 if (list_empty(&spu_prio->runq[prio]))
480 clear_bit(prio, spu_prio->bitmap);
484 static void spu_prio_wait(struct spu_context *ctx)
488 spin_lock(&spu_prio->runq_lock);
489 prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
490 if (!signal_pending(current)) {
491 __spu_add_to_rq(ctx);
492 spin_unlock(&spu_prio->runq_lock);
493 mutex_unlock(&ctx->state_mutex);
495 mutex_lock(&ctx->state_mutex);
496 spin_lock(&spu_prio->runq_lock);
497 __spu_del_from_rq(ctx);
499 spin_unlock(&spu_prio->runq_lock);
500 __set_current_state(TASK_RUNNING);
501 remove_wait_queue(&ctx->stop_wq, &wait);
504 static struct spu *spu_get_idle(struct spu_context *ctx)
506 struct spu *spu = NULL;
507 int node = cpu_to_node(raw_smp_processor_id());
510 for (n = 0; n < MAX_NUMNODES; n++, node++) {
511 node = (node < MAX_NUMNODES) ? node : 0;
512 if (!node_allowed(ctx, node))
514 spu = spu_alloc_node(node);
522 * find_victim - find a lower priority context to preempt
523 * @ctx: canidate context for running
525 * Returns the freed physical spu to run the new context on.
527 static struct spu *find_victim(struct spu_context *ctx)
529 struct spu_context *victim = NULL;
534 * Look for a possible preemption candidate on the local node first.
535 * If there is no candidate look at the other nodes. This isn't
536 * exactly fair, but so far the whole spu schedule tries to keep
537 * a strong node affinity. We might want to fine-tune this in
541 node = cpu_to_node(raw_smp_processor_id());
542 for (n = 0; n < MAX_NUMNODES; n++, node++) {
543 node = (node < MAX_NUMNODES) ? node : 0;
544 if (!node_allowed(ctx, node))
547 mutex_lock(&spu_prio->active_mutex[node]);
548 list_for_each_entry(spu, &spu_prio->active_list[node], list) {
549 struct spu_context *tmp = spu->ctx;
551 if (tmp->prio > ctx->prio &&
552 (!victim || tmp->prio > victim->prio))
555 mutex_unlock(&spu_prio->active_mutex[node]);
559 * This nests ctx->state_mutex, but we always lock
560 * higher priority contexts before lower priority
561 * ones, so this is safe until we introduce
562 * priority inheritance schemes.
564 if (!mutex_trylock(&victim->state_mutex)) {
572 * This race can happen because we've dropped
573 * the active list mutex. No a problem, just
574 * restart the search.
576 mutex_unlock(&victim->state_mutex);
580 spu_remove_from_active_list(spu);
581 spu_unbind_context(spu, victim);
582 victim->stats.invol_ctx_switch++;
583 spu->stats.invol_ctx_switch++;
584 mutex_unlock(&victim->state_mutex);
586 * We need to break out of the wait loop in spu_run
587 * manually to ensure this context gets put on the
588 * runqueue again ASAP.
590 wake_up(&victim->stop_wq);
599 * spu_activate - find a free spu for a context and execute it
600 * @ctx: spu context to schedule
601 * @flags: flags (currently ignored)
603 * Tries to find a free spu to run @ctx. If no free spu is available
604 * add the context to the runqueue so it gets woken up once an spu
607 int spu_activate(struct spu_context *ctx, unsigned long flags)
613 * If there are multiple threads waiting for a single context
614 * only one actually binds the context while the others will
615 * only be able to acquire the state_mutex once the context
616 * already is in runnable state.
621 spu = spu_get_idle(ctx);
623 * If this is a realtime thread we try to get it running by
624 * preempting a lower priority thread.
626 if (!spu && rt_prio(ctx->prio))
627 spu = find_victim(ctx);
629 spu_bind_context(spu, ctx);
630 spu_add_to_active_list(spu);
635 } while (!signal_pending(current));
641 * grab_runnable_context - try to find a runnable context
643 * Remove the highest priority context on the runqueue and return it
644 * to the caller. Returns %NULL if no runnable context was found.
646 static struct spu_context *grab_runnable_context(int prio, int node)
648 struct spu_context *ctx;
651 spin_lock(&spu_prio->runq_lock);
652 best = find_first_bit(spu_prio->bitmap, prio);
653 while (best < prio) {
654 struct list_head *rq = &spu_prio->runq[best];
656 list_for_each_entry(ctx, rq, rq) {
657 /* XXX(hch): check for affinity here aswell */
658 if (__node_allowed(ctx, node)) {
659 __spu_del_from_rq(ctx);
667 spin_unlock(&spu_prio->runq_lock);
671 static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
673 struct spu *spu = ctx->spu;
674 struct spu_context *new = NULL;
677 new = grab_runnable_context(max_prio, spu->node);
679 spu_remove_from_active_list(spu);
680 spu_unbind_context(spu, ctx);
681 ctx->stats.vol_ctx_switch++;
682 spu->stats.vol_ctx_switch++;
685 wake_up(&new->stop_wq);
694 * spu_deactivate - unbind a context from it's physical spu
695 * @ctx: spu context to unbind
697 * Unbind @ctx from the physical spu it is running on and schedule
698 * the highest priority context to run on the freed physical spu.
700 void spu_deactivate(struct spu_context *ctx)
702 __spu_deactivate(ctx, 1, MAX_PRIO);
706 * spu_yield - yield a physical spu if others are waiting
707 * @ctx: spu context to yield
709 * Check if there is a higher priority context waiting and if yes
710 * unbind @ctx from the physical spu and schedule the highest
711 * priority context to run on the freed physical spu instead.
713 void spu_yield(struct spu_context *ctx)
715 if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
716 mutex_lock(&ctx->state_mutex);
717 __spu_deactivate(ctx, 0, MAX_PRIO);
718 mutex_unlock(&ctx->state_mutex);
722 static void spusched_tick(struct spu_context *ctx)
724 if (ctx->flags & SPU_CREATE_NOSCHED)
726 if (ctx->policy == SCHED_FIFO)
729 if (--ctx->time_slice)
733 * Unfortunately active_mutex ranks outside of state_mutex, so
734 * we have to trylock here. If we fail give the context another
735 * tick and try again.
737 if (mutex_trylock(&ctx->state_mutex)) {
738 struct spu *spu = ctx->spu;
739 struct spu_context *new;
741 new = grab_runnable_context(ctx->prio + 1, spu->node);
744 __spu_remove_from_active_list(spu);
745 spu_unbind_context(spu, ctx);
746 ctx->stats.invol_ctx_switch++;
747 spu->stats.invol_ctx_switch++;
749 wake_up(&new->stop_wq);
751 * We need to break out of the wait loop in
752 * spu_run manually to ensure this context
753 * gets put on the runqueue again ASAP.
755 wake_up(&ctx->stop_wq);
757 spu_set_timeslice(ctx);
758 mutex_unlock(&ctx->state_mutex);
765 * count_active_contexts - count nr of active tasks
767 * Return the number of tasks currently running or waiting to run.
769 * Note that we don't take runq_lock / active_mutex here. Reading
770 * a single 32bit value is atomic on powerpc, and we don't care
771 * about memory ordering issues here.
773 static unsigned long count_active_contexts(void)
775 int nr_active = 0, node;
777 for (node = 0; node < MAX_NUMNODES; node++)
778 nr_active += spu_prio->nr_active[node];
779 nr_active += spu_prio->nr_waiting;
785 * spu_calc_load - given tick count, update the avenrun load estimates.
788 * No locking against reading these values from userspace, as for
789 * the CPU loadavg code.
791 static void spu_calc_load(unsigned long ticks)
793 unsigned long active_tasks; /* fixed-point */
794 static int count = LOAD_FREQ;
798 if (unlikely(count < 0)) {
799 active_tasks = count_active_contexts() * FIXED_1;
801 CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks);
802 CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks);
803 CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks);
809 static void spusched_wake(unsigned long data)
811 mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
812 wake_up_process(spusched_task);
813 spu_calc_load(SPUSCHED_TICK);
816 static int spusched_thread(void *unused)
818 struct spu *spu, *next;
821 while (!kthread_should_stop()) {
822 set_current_state(TASK_INTERRUPTIBLE);
824 for (node = 0; node < MAX_NUMNODES; node++) {
825 mutex_lock(&spu_prio->active_mutex[node]);
826 list_for_each_entry_safe(spu, next,
827 &spu_prio->active_list[node],
829 spusched_tick(spu->ctx);
830 mutex_unlock(&spu_prio->active_mutex[node]);
837 #define LOAD_INT(x) ((x) >> FSHIFT)
838 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
840 static int show_spu_loadavg(struct seq_file *s, void *private)
844 a = spu_avenrun[0] + (FIXED_1/200);
845 b = spu_avenrun[1] + (FIXED_1/200);
846 c = spu_avenrun[2] + (FIXED_1/200);
849 * Note that last_pid doesn't really make much sense for the
850 * SPU loadavg (it even seems very odd on the CPU side..),
851 * but we include it here to have a 100% compatible interface.
853 seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
854 LOAD_INT(a), LOAD_FRAC(a),
855 LOAD_INT(b), LOAD_FRAC(b),
856 LOAD_INT(c), LOAD_FRAC(c),
857 count_active_contexts(),
858 atomic_read(&nr_spu_contexts),
859 current->nsproxy->pid_ns->last_pid);
863 static int spu_loadavg_open(struct inode *inode, struct file *file)
865 return single_open(file, show_spu_loadavg, NULL);
868 static const struct file_operations spu_loadavg_fops = {
869 .open = spu_loadavg_open,
872 .release = single_release,
875 int __init spu_sched_init(void)
877 struct proc_dir_entry *entry;
878 int err = -ENOMEM, i;
880 spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
884 for (i = 0; i < MAX_PRIO; i++) {
885 INIT_LIST_HEAD(&spu_prio->runq[i]);
886 __clear_bit(i, spu_prio->bitmap);
888 for (i = 0; i < MAX_NUMNODES; i++) {
889 mutex_init(&spu_prio->active_mutex[i]);
890 INIT_LIST_HEAD(&spu_prio->active_list[i]);
892 spin_lock_init(&spu_prio->runq_lock);
894 setup_timer(&spusched_timer, spusched_wake, 0);
896 spusched_task = kthread_run(spusched_thread, NULL, "spusched");
897 if (IS_ERR(spusched_task)) {
898 err = PTR_ERR(spusched_task);
899 goto out_free_spu_prio;
902 entry = create_proc_entry("spu_loadavg", 0, NULL);
904 goto out_stop_kthread;
905 entry->proc_fops = &spu_loadavg_fops;
907 pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
908 SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE);
912 kthread_stop(spusched_task);
919 void spu_sched_exit(void)
921 struct spu *spu, *tmp;
924 remove_proc_entry("spu_loadavg", NULL);
926 del_timer_sync(&spusched_timer);
927 kthread_stop(spusched_task);
929 for (node = 0; node < MAX_NUMNODES; node++) {
930 mutex_lock(&spu_prio->active_mutex[node]);
931 list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
933 list_del_init(&spu->list);
936 mutex_unlock(&spu_prio->active_mutex[node]);