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 void do_notify_spus_active(void)
174 * Wake up the active spu_contexts.
176 * When the awakened processes see their "notify_active" flag is set,
177 * they will call spu_switch_notify();
179 for_each_online_node(node) {
182 mutex_lock(&cbe_spu_info[node].list_mutex);
183 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
184 if (spu->alloc_state != SPU_FREE) {
185 struct spu_context *ctx = spu->ctx;
186 set_bit(SPU_SCHED_NOTIFY_ACTIVE,
189 wake_up_all(&ctx->stop_wq);
192 mutex_unlock(&cbe_spu_info[node].list_mutex);
197 * spu_bind_context - bind spu context to physical spu
198 * @spu: physical spu to bind to
199 * @ctx: context to bind
201 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
203 pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
204 spu->number, spu->node);
205 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
207 if (ctx->flags & SPU_CREATE_NOSCHED)
208 atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
210 ctx->stats.slb_flt_base = spu->stats.slb_flt;
211 ctx->stats.class2_intr_base = spu->stats.class2_intr;
216 ctx->ops = &spu_hw_ops;
217 spu->pid = current->pid;
218 spu->tgid = current->tgid;
219 spu_associate_mm(spu, ctx->owner);
220 spu->ibox_callback = spufs_ibox_callback;
221 spu->wbox_callback = spufs_wbox_callback;
222 spu->stop_callback = spufs_stop_callback;
223 spu->mfc_callback = spufs_mfc_callback;
224 spu->dma_callback = spufs_dma_callback;
226 spu_unmap_mappings(ctx);
227 spu_restore(&ctx->csa, spu);
228 spu->timestamp = jiffies;
229 spu_cpu_affinity_set(spu, raw_smp_processor_id());
230 spu_switch_notify(spu, ctx);
231 ctx->state = SPU_STATE_RUNNABLE;
233 spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
237 * Must be used with the list_mutex held.
239 static inline int sched_spu(struct spu *spu)
241 BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex));
243 return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED));
246 static void aff_merge_remaining_ctxs(struct spu_gang *gang)
248 struct spu_context *ctx;
250 list_for_each_entry(ctx, &gang->aff_list_head, aff_list) {
251 if (list_empty(&ctx->aff_list))
252 list_add(&ctx->aff_list, &gang->aff_list_head);
254 gang->aff_flags |= AFF_MERGED;
257 static void aff_set_offsets(struct spu_gang *gang)
259 struct spu_context *ctx;
263 list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
265 if (&ctx->aff_list == &gang->aff_list_head)
267 ctx->aff_offset = offset--;
271 list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) {
272 if (&ctx->aff_list == &gang->aff_list_head)
274 ctx->aff_offset = offset++;
277 gang->aff_flags |= AFF_OFFSETS_SET;
280 static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff,
281 int group_size, int lowest_offset)
287 * TODO: A better algorithm could be used to find a good spu to be
288 * used as reference location for the ctxs chain.
290 node = cpu_to_node(raw_smp_processor_id());
291 for (n = 0; n < MAX_NUMNODES; n++, node++) {
292 node = (node < MAX_NUMNODES) ? node : 0;
293 if (!node_allowed(ctx, node))
295 mutex_lock(&cbe_spu_info[node].list_mutex);
296 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
297 if ((!mem_aff || spu->has_mem_affinity) &&
299 mutex_unlock(&cbe_spu_info[node].list_mutex);
303 mutex_unlock(&cbe_spu_info[node].list_mutex);
308 static void aff_set_ref_point_location(struct spu_gang *gang)
310 int mem_aff, gs, lowest_offset;
311 struct spu_context *ctx;
314 mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
318 list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
321 list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
323 if (&ctx->aff_list == &gang->aff_list_head)
325 lowest_offset = ctx->aff_offset;
328 gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs,
332 static struct spu *ctx_location(struct spu *ref, int offset, int node)
338 list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
339 BUG_ON(spu->node != node);
346 list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
347 BUG_ON(spu->node != node);
359 * affinity_check is called each time a context is going to be scheduled.
360 * It returns the spu ptr on which the context must run.
362 static int has_affinity(struct spu_context *ctx)
364 struct spu_gang *gang = ctx->gang;
366 if (list_empty(&ctx->aff_list))
369 if (!gang->aff_ref_spu) {
370 if (!(gang->aff_flags & AFF_MERGED))
371 aff_merge_remaining_ctxs(gang);
372 if (!(gang->aff_flags & AFF_OFFSETS_SET))
373 aff_set_offsets(gang);
374 aff_set_ref_point_location(gang);
377 return gang->aff_ref_spu != NULL;
381 * spu_unbind_context - unbind spu context from physical spu
382 * @spu: physical spu to unbind from
383 * @ctx: context to unbind
385 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
387 pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
388 spu->pid, spu->number, spu->node);
389 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
391 if (spu->ctx->flags & SPU_CREATE_NOSCHED)
392 atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
395 mutex_lock(&ctx->gang->aff_mutex);
396 if (has_affinity(ctx)) {
397 if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
398 ctx->gang->aff_ref_spu = NULL;
400 mutex_unlock(&ctx->gang->aff_mutex);
403 spu_switch_notify(spu, NULL);
404 spu_unmap_mappings(ctx);
405 spu_save(&ctx->csa, spu);
406 spu->timestamp = jiffies;
407 ctx->state = SPU_STATE_SAVED;
408 spu->ibox_callback = NULL;
409 spu->wbox_callback = NULL;
410 spu->stop_callback = NULL;
411 spu->mfc_callback = NULL;
412 spu->dma_callback = NULL;
413 spu_associate_mm(spu, NULL);
416 ctx->ops = &spu_backing_ops;
420 ctx->stats.slb_flt +=
421 (spu->stats.slb_flt - ctx->stats.slb_flt_base);
422 ctx->stats.class2_intr +=
423 (spu->stats.class2_intr - ctx->stats.class2_intr_base);
425 /* This maps the underlying spu state to idle */
426 spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
431 * spu_add_to_rq - add a context to the runqueue
432 * @ctx: context to add
434 static void __spu_add_to_rq(struct spu_context *ctx)
437 * Unfortunately this code path can be called from multiple threads
438 * on behalf of a single context due to the way the problem state
439 * mmap support works.
441 * Fortunately we need to wake up all these threads at the same time
442 * and can simply skip the runqueue addition for every but the first
443 * thread getting into this codepath.
445 * It's still quite hacky, and long-term we should proxy all other
446 * threads through the owner thread so that spu_run is in control
447 * of all the scheduling activity for a given context.
449 if (list_empty(&ctx->rq)) {
450 list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
451 set_bit(ctx->prio, spu_prio->bitmap);
452 if (!spu_prio->nr_waiting++)
453 __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
457 static void __spu_del_from_rq(struct spu_context *ctx)
459 int prio = ctx->prio;
461 if (!list_empty(&ctx->rq)) {
462 if (!--spu_prio->nr_waiting)
463 del_timer(&spusched_timer);
464 list_del_init(&ctx->rq);
466 if (list_empty(&spu_prio->runq[prio]))
467 clear_bit(prio, spu_prio->bitmap);
471 static void spu_prio_wait(struct spu_context *ctx)
475 spin_lock(&spu_prio->runq_lock);
476 prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
477 if (!signal_pending(current)) {
478 __spu_add_to_rq(ctx);
479 spin_unlock(&spu_prio->runq_lock);
480 mutex_unlock(&ctx->state_mutex);
482 mutex_lock(&ctx->state_mutex);
483 spin_lock(&spu_prio->runq_lock);
484 __spu_del_from_rq(ctx);
486 spin_unlock(&spu_prio->runq_lock);
487 __set_current_state(TASK_RUNNING);
488 remove_wait_queue(&ctx->stop_wq, &wait);
491 static struct spu *spu_get_idle(struct spu_context *ctx)
493 struct spu *spu, *aff_ref_spu;
497 mutex_lock(&ctx->gang->aff_mutex);
498 if (has_affinity(ctx)) {
499 aff_ref_spu = ctx->gang->aff_ref_spu;
500 atomic_inc(&ctx->gang->aff_sched_count);
501 mutex_unlock(&ctx->gang->aff_mutex);
502 node = aff_ref_spu->node;
504 mutex_lock(&cbe_spu_info[node].list_mutex);
505 spu = ctx_location(aff_ref_spu, ctx->aff_offset, node);
506 if (spu && spu->alloc_state == SPU_FREE)
508 mutex_unlock(&cbe_spu_info[node].list_mutex);
510 mutex_lock(&ctx->gang->aff_mutex);
511 if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
512 ctx->gang->aff_ref_spu = NULL;
513 mutex_unlock(&ctx->gang->aff_mutex);
517 mutex_unlock(&ctx->gang->aff_mutex);
519 node = cpu_to_node(raw_smp_processor_id());
520 for (n = 0; n < MAX_NUMNODES; n++, node++) {
521 node = (node < MAX_NUMNODES) ? node : 0;
522 if (!node_allowed(ctx, node))
525 mutex_lock(&cbe_spu_info[node].list_mutex);
526 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
527 if (spu->alloc_state == SPU_FREE)
530 mutex_unlock(&cbe_spu_info[node].list_mutex);
536 spu->alloc_state = SPU_USED;
537 mutex_unlock(&cbe_spu_info[node].list_mutex);
538 pr_debug("Got SPU %d %d\n", spu->number, spu->node);
539 spu_init_channels(spu);
544 * find_victim - find a lower priority context to preempt
545 * @ctx: canidate context for running
547 * Returns the freed physical spu to run the new context on.
549 static struct spu *find_victim(struct spu_context *ctx)
551 struct spu_context *victim = NULL;
556 * Look for a possible preemption candidate on the local node first.
557 * If there is no candidate look at the other nodes. This isn't
558 * exactly fair, but so far the whole spu schedule tries to keep
559 * a strong node affinity. We might want to fine-tune this in
563 node = cpu_to_node(raw_smp_processor_id());
564 for (n = 0; n < MAX_NUMNODES; n++, node++) {
565 node = (node < MAX_NUMNODES) ? node : 0;
566 if (!node_allowed(ctx, node))
569 mutex_lock(&cbe_spu_info[node].list_mutex);
570 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
571 struct spu_context *tmp = spu->ctx;
573 if (tmp && tmp->prio > ctx->prio &&
574 (!victim || tmp->prio > victim->prio))
577 mutex_unlock(&cbe_spu_info[node].list_mutex);
581 * This nests ctx->state_mutex, but we always lock
582 * higher priority contexts before lower priority
583 * ones, so this is safe until we introduce
584 * priority inheritance schemes.
586 if (!mutex_trylock(&victim->state_mutex)) {
594 * This race can happen because we've dropped
595 * the active list mutex. No a problem, just
596 * restart the search.
598 mutex_unlock(&victim->state_mutex);
603 mutex_lock(&cbe_spu_info[node].list_mutex);
604 cbe_spu_info[node].nr_active--;
605 spu_unbind_context(spu, victim);
606 mutex_unlock(&cbe_spu_info[node].list_mutex);
608 victim->stats.invol_ctx_switch++;
609 spu->stats.invol_ctx_switch++;
610 mutex_unlock(&victim->state_mutex);
612 * We need to break out of the wait loop in spu_run
613 * manually to ensure this context gets put on the
614 * runqueue again ASAP.
616 wake_up(&victim->stop_wq);
625 * spu_activate - find a free spu for a context and execute it
626 * @ctx: spu context to schedule
627 * @flags: flags (currently ignored)
629 * Tries to find a free spu to run @ctx. If no free spu is available
630 * add the context to the runqueue so it gets woken up once an spu
633 int spu_activate(struct spu_context *ctx, unsigned long flags)
639 * If there are multiple threads waiting for a single context
640 * only one actually binds the context while the others will
641 * only be able to acquire the state_mutex once the context
642 * already is in runnable state.
647 spu = spu_get_idle(ctx);
649 * If this is a realtime thread we try to get it running by
650 * preempting a lower priority thread.
652 if (!spu && rt_prio(ctx->prio))
653 spu = find_victim(ctx);
655 int node = spu->node;
657 mutex_lock(&cbe_spu_info[node].list_mutex);
658 spu_bind_context(spu, ctx);
659 cbe_spu_info[node].nr_active++;
660 mutex_unlock(&cbe_spu_info[node].list_mutex);
665 } while (!signal_pending(current));
671 * grab_runnable_context - try to find a runnable context
673 * Remove the highest priority context on the runqueue and return it
674 * to the caller. Returns %NULL if no runnable context was found.
676 static struct spu_context *grab_runnable_context(int prio, int node)
678 struct spu_context *ctx;
681 spin_lock(&spu_prio->runq_lock);
682 best = find_first_bit(spu_prio->bitmap, prio);
683 while (best < prio) {
684 struct list_head *rq = &spu_prio->runq[best];
686 list_for_each_entry(ctx, rq, rq) {
687 /* XXX(hch): check for affinity here aswell */
688 if (__node_allowed(ctx, node)) {
689 __spu_del_from_rq(ctx);
697 spin_unlock(&spu_prio->runq_lock);
701 static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
703 struct spu *spu = ctx->spu;
704 struct spu_context *new = NULL;
707 new = grab_runnable_context(max_prio, spu->node);
709 int node = spu->node;
711 mutex_lock(&cbe_spu_info[node].list_mutex);
712 spu_unbind_context(spu, ctx);
713 spu->alloc_state = SPU_FREE;
714 cbe_spu_info[node].nr_active--;
715 mutex_unlock(&cbe_spu_info[node].list_mutex);
717 ctx->stats.vol_ctx_switch++;
718 spu->stats.vol_ctx_switch++;
721 wake_up(&new->stop_wq);
730 * spu_deactivate - unbind a context from it's physical spu
731 * @ctx: spu context to unbind
733 * Unbind @ctx from the physical spu it is running on and schedule
734 * the highest priority context to run on the freed physical spu.
736 void spu_deactivate(struct spu_context *ctx)
738 __spu_deactivate(ctx, 1, MAX_PRIO);
742 * spu_yield - yield a physical spu if others are waiting
743 * @ctx: spu context to yield
745 * Check if there is a higher priority context waiting and if yes
746 * unbind @ctx from the physical spu and schedule the highest
747 * priority context to run on the freed physical spu instead.
749 void spu_yield(struct spu_context *ctx)
751 if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
752 mutex_lock(&ctx->state_mutex);
753 __spu_deactivate(ctx, 0, MAX_PRIO);
754 mutex_unlock(&ctx->state_mutex);
758 static noinline void spusched_tick(struct spu_context *ctx)
760 if (ctx->flags & SPU_CREATE_NOSCHED)
762 if (ctx->policy == SCHED_FIFO)
765 if (--ctx->time_slice)
769 * Unfortunately list_mutex ranks outside of state_mutex, so
770 * we have to trylock here. If we fail give the context another
771 * tick and try again.
773 if (mutex_trylock(&ctx->state_mutex)) {
774 struct spu *spu = ctx->spu;
775 struct spu_context *new;
777 new = grab_runnable_context(ctx->prio + 1, spu->node);
779 spu_unbind_context(spu, ctx);
780 ctx->stats.invol_ctx_switch++;
781 spu->stats.invol_ctx_switch++;
782 spu->alloc_state = SPU_FREE;
783 cbe_spu_info[spu->node].nr_active--;
784 wake_up(&new->stop_wq);
786 * We need to break out of the wait loop in
787 * spu_run manually to ensure this context
788 * gets put on the runqueue again ASAP.
790 wake_up(&ctx->stop_wq);
792 spu_set_timeslice(ctx);
793 mutex_unlock(&ctx->state_mutex);
800 * count_active_contexts - count nr of active tasks
802 * Return the number of tasks currently running or waiting to run.
804 * Note that we don't take runq_lock / list_mutex here. Reading
805 * a single 32bit value is atomic on powerpc, and we don't care
806 * about memory ordering issues here.
808 static unsigned long count_active_contexts(void)
810 int nr_active = 0, node;
812 for (node = 0; node < MAX_NUMNODES; node++)
813 nr_active += cbe_spu_info[node].nr_active;
814 nr_active += spu_prio->nr_waiting;
820 * spu_calc_load - given tick count, update the avenrun load estimates.
823 * No locking against reading these values from userspace, as for
824 * the CPU loadavg code.
826 static void spu_calc_load(unsigned long ticks)
828 unsigned long active_tasks; /* fixed-point */
829 static int count = LOAD_FREQ;
833 if (unlikely(count < 0)) {
834 active_tasks = count_active_contexts() * FIXED_1;
836 CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks);
837 CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks);
838 CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks);
844 static void spusched_wake(unsigned long data)
846 mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
847 wake_up_process(spusched_task);
848 spu_calc_load(SPUSCHED_TICK);
851 static int spusched_thread(void *unused)
856 while (!kthread_should_stop()) {
857 set_current_state(TASK_INTERRUPTIBLE);
859 for (node = 0; node < MAX_NUMNODES; node++) {
860 mutex_lock(&cbe_spu_info[node].list_mutex);
861 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
863 spusched_tick(spu->ctx);
864 mutex_unlock(&cbe_spu_info[node].list_mutex);
871 #define LOAD_INT(x) ((x) >> FSHIFT)
872 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
874 static int show_spu_loadavg(struct seq_file *s, void *private)
878 a = spu_avenrun[0] + (FIXED_1/200);
879 b = spu_avenrun[1] + (FIXED_1/200);
880 c = spu_avenrun[2] + (FIXED_1/200);
883 * Note that last_pid doesn't really make much sense for the
884 * SPU loadavg (it even seems very odd on the CPU side..),
885 * but we include it here to have a 100% compatible interface.
887 seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
888 LOAD_INT(a), LOAD_FRAC(a),
889 LOAD_INT(b), LOAD_FRAC(b),
890 LOAD_INT(c), LOAD_FRAC(c),
891 count_active_contexts(),
892 atomic_read(&nr_spu_contexts),
893 current->nsproxy->pid_ns->last_pid);
897 static int spu_loadavg_open(struct inode *inode, struct file *file)
899 return single_open(file, show_spu_loadavg, NULL);
902 static const struct file_operations spu_loadavg_fops = {
903 .open = spu_loadavg_open,
906 .release = single_release,
909 int __init spu_sched_init(void)
911 struct proc_dir_entry *entry;
912 int err = -ENOMEM, i;
914 spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
918 for (i = 0; i < MAX_PRIO; i++) {
919 INIT_LIST_HEAD(&spu_prio->runq[i]);
920 __clear_bit(i, spu_prio->bitmap);
922 spin_lock_init(&spu_prio->runq_lock);
924 setup_timer(&spusched_timer, spusched_wake, 0);
926 spusched_task = kthread_run(spusched_thread, NULL, "spusched");
927 if (IS_ERR(spusched_task)) {
928 err = PTR_ERR(spusched_task);
929 goto out_free_spu_prio;
932 entry = create_proc_entry("spu_loadavg", 0, NULL);
934 goto out_stop_kthread;
935 entry->proc_fops = &spu_loadavg_fops;
937 pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
938 SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE);
942 kthread_stop(spusched_task);
949 void spu_sched_exit(void)
954 remove_proc_entry("spu_loadavg", NULL);
956 del_timer_sync(&spusched_timer);
957 kthread_stop(spusched_task);
959 for (node = 0; node < MAX_NUMNODES; node++) {
960 mutex_lock(&cbe_spu_info[node].list_mutex);
961 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
962 if (spu->alloc_state != SPU_FREE)
963 spu->alloc_state = SPU_FREE;
964 mutex_unlock(&cbe_spu_info[node].list_mutex);