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>
42 #include <linux/marker.h>
45 #include <asm/mmu_context.h>
47 #include <asm/spu_csa.h>
48 #include <asm/spu_priv1.h>
51 struct spu_prio_array {
52 DECLARE_BITMAP(bitmap, MAX_PRIO);
53 struct list_head runq[MAX_PRIO];
58 static unsigned long spu_avenrun[3];
59 static struct spu_prio_array *spu_prio;
60 static struct task_struct *spusched_task;
61 static struct timer_list spusched_timer;
62 static struct timer_list spuloadavg_timer;
65 * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
67 #define NORMAL_PRIO 120
70 * Frequency of the spu scheduler tick. By default we do one SPU scheduler
71 * tick for every 10 CPU scheduler ticks.
73 #define SPUSCHED_TICK (10)
76 * These are the 'tuning knobs' of the scheduler:
78 * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is
79 * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs.
81 #define MIN_SPU_TIMESLICE max(5 * HZ / (1000 * SPUSCHED_TICK), 1)
82 #define DEF_SPU_TIMESLICE (100 * HZ / (1000 * SPUSCHED_TICK))
84 #define MAX_USER_PRIO (MAX_PRIO - MAX_RT_PRIO)
85 #define SCALE_PRIO(x, prio) \
86 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE)
89 * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values:
90 * [800ms ... 100ms ... 5ms]
92 * The higher a thread's priority, the bigger timeslices
93 * it gets during one round of execution. But even the lowest
94 * priority thread gets MIN_TIMESLICE worth of execution time.
96 void spu_set_timeslice(struct spu_context *ctx)
98 if (ctx->prio < NORMAL_PRIO)
99 ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio);
101 ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
105 * Update scheduling information from the owning thread.
107 void __spu_update_sched_info(struct spu_context *ctx)
110 * assert that the context is not on the runqueue, so it is safe
111 * to change its scheduling parameters.
113 BUG_ON(!list_empty(&ctx->rq));
116 * 32-Bit assignments are atomic on powerpc, and we don't care about
117 * memory ordering here because retrieving the controlling thread is
118 * per definition racy.
120 ctx->tid = current->pid;
123 * We do our own priority calculations, so we normally want
124 * ->static_prio to start with. Unfortunately this field
125 * contains junk for threads with a realtime scheduling
126 * policy so we have to look at ->prio in this case.
128 if (rt_prio(current->prio))
129 ctx->prio = current->prio;
131 ctx->prio = current->static_prio;
132 ctx->policy = current->policy;
135 * TO DO: the context may be loaded, so we may need to activate
136 * it again on a different node. But it shouldn't hurt anything
137 * to update its parameters, because we know that the scheduler
138 * is not actively looking at this field, since it is not on the
139 * runqueue. The context will be rescheduled on the proper node
140 * if it is timesliced or preempted.
142 ctx->cpus_allowed = current->cpus_allowed;
144 /* Save the current cpu id for spu interrupt routing. */
145 ctx->last_ran = raw_smp_processor_id();
148 void spu_update_sched_info(struct spu_context *ctx)
152 if (ctx->state == SPU_STATE_RUNNABLE) {
153 node = ctx->spu->node;
156 * Take list_mutex to sync with find_victim().
158 mutex_lock(&cbe_spu_info[node].list_mutex);
159 __spu_update_sched_info(ctx);
160 mutex_unlock(&cbe_spu_info[node].list_mutex);
162 __spu_update_sched_info(ctx);
166 static int __node_allowed(struct spu_context *ctx, int node)
168 if (nr_cpus_node(node)) {
169 cpumask_t mask = node_to_cpumask(node);
171 if (cpus_intersects(mask, ctx->cpus_allowed))
178 static int node_allowed(struct spu_context *ctx, int node)
182 spin_lock(&spu_prio->runq_lock);
183 rval = __node_allowed(ctx, node);
184 spin_unlock(&spu_prio->runq_lock);
189 void do_notify_spus_active(void)
194 * Wake up the active spu_contexts.
196 * When the awakened processes see their "notify_active" flag is set,
197 * they will call spu_switch_notify().
199 for_each_online_node(node) {
202 mutex_lock(&cbe_spu_info[node].list_mutex);
203 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
204 if (spu->alloc_state != SPU_FREE) {
205 struct spu_context *ctx = spu->ctx;
206 set_bit(SPU_SCHED_NOTIFY_ACTIVE,
209 wake_up_all(&ctx->stop_wq);
212 mutex_unlock(&cbe_spu_info[node].list_mutex);
217 * spu_bind_context - bind spu context to physical spu
218 * @spu: physical spu to bind to
219 * @ctx: context to bind
221 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
223 spu_context_trace(spu_bind_context__enter, ctx, spu);
225 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
227 if (ctx->flags & SPU_CREATE_NOSCHED)
228 atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
230 ctx->stats.slb_flt_base = spu->stats.slb_flt;
231 ctx->stats.class2_intr_base = spu->stats.class2_intr;
233 spu_associate_mm(spu, ctx->owner);
235 spin_lock_irq(&spu->register_lock);
239 ctx->ops = &spu_hw_ops;
240 spu->pid = current->pid;
241 spu->tgid = current->tgid;
242 spu->ibox_callback = spufs_ibox_callback;
243 spu->wbox_callback = spufs_wbox_callback;
244 spu->stop_callback = spufs_stop_callback;
245 spu->mfc_callback = spufs_mfc_callback;
246 spin_unlock_irq(&spu->register_lock);
248 spu_unmap_mappings(ctx);
250 spu_switch_log_notify(spu, ctx, SWITCH_LOG_START, 0);
251 spu_restore(&ctx->csa, spu);
252 spu->timestamp = jiffies;
253 spu_switch_notify(spu, ctx);
254 ctx->state = SPU_STATE_RUNNABLE;
256 spuctx_switch_state(ctx, SPU_UTIL_USER);
260 * Must be used with the list_mutex held.
262 static inline int sched_spu(struct spu *spu)
264 BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex));
266 return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED));
269 static void aff_merge_remaining_ctxs(struct spu_gang *gang)
271 struct spu_context *ctx;
273 list_for_each_entry(ctx, &gang->aff_list_head, aff_list) {
274 if (list_empty(&ctx->aff_list))
275 list_add(&ctx->aff_list, &gang->aff_list_head);
277 gang->aff_flags |= AFF_MERGED;
280 static void aff_set_offsets(struct spu_gang *gang)
282 struct spu_context *ctx;
286 list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
288 if (&ctx->aff_list == &gang->aff_list_head)
290 ctx->aff_offset = offset--;
294 list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) {
295 if (&ctx->aff_list == &gang->aff_list_head)
297 ctx->aff_offset = offset++;
300 gang->aff_flags |= AFF_OFFSETS_SET;
303 static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff,
304 int group_size, int lowest_offset)
310 * TODO: A better algorithm could be used to find a good spu to be
311 * used as reference location for the ctxs chain.
313 node = cpu_to_node(raw_smp_processor_id());
314 for (n = 0; n < MAX_NUMNODES; n++, node++) {
316 * "available_spus" counts how many spus are not potentially
317 * going to be used by other affinity gangs whose reference
318 * context is already in place. Although this code seeks to
319 * avoid having affinity gangs with a summed amount of
320 * contexts bigger than the amount of spus in the node,
321 * this may happen sporadically. In this case, available_spus
322 * becomes negative, which is harmless.
326 node = (node < MAX_NUMNODES) ? node : 0;
327 if (!node_allowed(ctx, node))
331 mutex_lock(&cbe_spu_info[node].list_mutex);
332 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
333 if (spu->ctx && spu->ctx->gang && !spu->ctx->aff_offset
334 && spu->ctx->gang->aff_ref_spu)
335 available_spus -= spu->ctx->gang->contexts;
338 if (available_spus < ctx->gang->contexts) {
339 mutex_unlock(&cbe_spu_info[node].list_mutex);
343 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
344 if ((!mem_aff || spu->has_mem_affinity) &&
346 mutex_unlock(&cbe_spu_info[node].list_mutex);
350 mutex_unlock(&cbe_spu_info[node].list_mutex);
355 static void aff_set_ref_point_location(struct spu_gang *gang)
357 int mem_aff, gs, lowest_offset;
358 struct spu_context *ctx;
361 mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
365 list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
368 list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
370 if (&ctx->aff_list == &gang->aff_list_head)
372 lowest_offset = ctx->aff_offset;
375 gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs,
379 static struct spu *ctx_location(struct spu *ref, int offset, int node)
385 list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
386 BUG_ON(spu->node != node);
393 list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
394 BUG_ON(spu->node != node);
406 * affinity_check is called each time a context is going to be scheduled.
407 * It returns the spu ptr on which the context must run.
409 static int has_affinity(struct spu_context *ctx)
411 struct spu_gang *gang = ctx->gang;
413 if (list_empty(&ctx->aff_list))
416 if (atomic_read(&ctx->gang->aff_sched_count) == 0)
417 ctx->gang->aff_ref_spu = NULL;
419 if (!gang->aff_ref_spu) {
420 if (!(gang->aff_flags & AFF_MERGED))
421 aff_merge_remaining_ctxs(gang);
422 if (!(gang->aff_flags & AFF_OFFSETS_SET))
423 aff_set_offsets(gang);
424 aff_set_ref_point_location(gang);
427 return gang->aff_ref_spu != NULL;
431 * spu_unbind_context - unbind spu context from physical spu
432 * @spu: physical spu to unbind from
433 * @ctx: context to unbind
435 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
439 spu_context_trace(spu_unbind_context__enter, ctx, spu);
441 spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
443 if (spu->ctx->flags & SPU_CREATE_NOSCHED)
444 atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
447 atomic_dec_if_positive(&ctx->gang->aff_sched_count);
449 spu_switch_notify(spu, NULL);
450 spu_unmap_mappings(ctx);
451 spu_save(&ctx->csa, spu);
452 spu_switch_log_notify(spu, ctx, SWITCH_LOG_STOP, 0);
454 spin_lock_irq(&spu->register_lock);
455 spu->timestamp = jiffies;
456 ctx->state = SPU_STATE_SAVED;
457 spu->ibox_callback = NULL;
458 spu->wbox_callback = NULL;
459 spu->stop_callback = NULL;
460 spu->mfc_callback = NULL;
463 ctx->ops = &spu_backing_ops;
466 spin_unlock_irq(&spu->register_lock);
468 spu_associate_mm(spu, NULL);
470 ctx->stats.slb_flt +=
471 (spu->stats.slb_flt - ctx->stats.slb_flt_base);
472 ctx->stats.class2_intr +=
473 (spu->stats.class2_intr - ctx->stats.class2_intr_base);
475 /* This maps the underlying spu state to idle */
476 spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
479 if (spu_stopped(ctx, &status))
480 wake_up_all(&ctx->stop_wq);
484 * spu_add_to_rq - add a context to the runqueue
485 * @ctx: context to add
487 static void __spu_add_to_rq(struct spu_context *ctx)
490 * Unfortunately this code path can be called from multiple threads
491 * on behalf of a single context due to the way the problem state
492 * mmap support works.
494 * Fortunately we need to wake up all these threads at the same time
495 * and can simply skip the runqueue addition for every but the first
496 * thread getting into this codepath.
498 * It's still quite hacky, and long-term we should proxy all other
499 * threads through the owner thread so that spu_run is in control
500 * of all the scheduling activity for a given context.
502 if (list_empty(&ctx->rq)) {
503 list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
504 set_bit(ctx->prio, spu_prio->bitmap);
505 if (!spu_prio->nr_waiting++)
506 __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
510 static void spu_add_to_rq(struct spu_context *ctx)
512 spin_lock(&spu_prio->runq_lock);
513 __spu_add_to_rq(ctx);
514 spin_unlock(&spu_prio->runq_lock);
517 static void __spu_del_from_rq(struct spu_context *ctx)
519 int prio = ctx->prio;
521 if (!list_empty(&ctx->rq)) {
522 if (!--spu_prio->nr_waiting)
523 del_timer(&spusched_timer);
524 list_del_init(&ctx->rq);
526 if (list_empty(&spu_prio->runq[prio]))
527 clear_bit(prio, spu_prio->bitmap);
531 void spu_del_from_rq(struct spu_context *ctx)
533 spin_lock(&spu_prio->runq_lock);
534 __spu_del_from_rq(ctx);
535 spin_unlock(&spu_prio->runq_lock);
538 static void spu_prio_wait(struct spu_context *ctx)
543 * The caller must explicitly wait for a context to be loaded
544 * if the nosched flag is set. If NOSCHED is not set, the caller
545 * queues the context and waits for an spu event or error.
547 BUG_ON(!(ctx->flags & SPU_CREATE_NOSCHED));
549 spin_lock(&spu_prio->runq_lock);
550 prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
551 if (!signal_pending(current)) {
552 __spu_add_to_rq(ctx);
553 spin_unlock(&spu_prio->runq_lock);
554 mutex_unlock(&ctx->state_mutex);
556 mutex_lock(&ctx->state_mutex);
557 spin_lock(&spu_prio->runq_lock);
558 __spu_del_from_rq(ctx);
560 spin_unlock(&spu_prio->runq_lock);
561 __set_current_state(TASK_RUNNING);
562 remove_wait_queue(&ctx->stop_wq, &wait);
565 static struct spu *spu_get_idle(struct spu_context *ctx)
567 struct spu *spu, *aff_ref_spu;
570 spu_context_nospu_trace(spu_get_idle__enter, ctx);
573 mutex_lock(&ctx->gang->aff_mutex);
574 if (has_affinity(ctx)) {
575 aff_ref_spu = ctx->gang->aff_ref_spu;
576 atomic_inc(&ctx->gang->aff_sched_count);
577 mutex_unlock(&ctx->gang->aff_mutex);
578 node = aff_ref_spu->node;
580 mutex_lock(&cbe_spu_info[node].list_mutex);
581 spu = ctx_location(aff_ref_spu, ctx->aff_offset, node);
582 if (spu && spu->alloc_state == SPU_FREE)
584 mutex_unlock(&cbe_spu_info[node].list_mutex);
586 atomic_dec(&ctx->gang->aff_sched_count);
589 mutex_unlock(&ctx->gang->aff_mutex);
591 node = cpu_to_node(raw_smp_processor_id());
592 for (n = 0; n < MAX_NUMNODES; n++, node++) {
593 node = (node < MAX_NUMNODES) ? node : 0;
594 if (!node_allowed(ctx, node))
597 mutex_lock(&cbe_spu_info[node].list_mutex);
598 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
599 if (spu->alloc_state == SPU_FREE)
602 mutex_unlock(&cbe_spu_info[node].list_mutex);
606 spu_context_nospu_trace(spu_get_idle__not_found, ctx);
610 spu->alloc_state = SPU_USED;
611 mutex_unlock(&cbe_spu_info[node].list_mutex);
612 spu_context_trace(spu_get_idle__found, ctx, spu);
613 spu_init_channels(spu);
618 * find_victim - find a lower priority context to preempt
619 * @ctx: canidate context for running
621 * Returns the freed physical spu to run the new context on.
623 static struct spu *find_victim(struct spu_context *ctx)
625 struct spu_context *victim = NULL;
629 spu_context_nospu_trace(spu_find_victim__enter, ctx);
632 * Look for a possible preemption candidate on the local node first.
633 * If there is no candidate look at the other nodes. This isn't
634 * exactly fair, but so far the whole spu scheduler tries to keep
635 * a strong node affinity. We might want to fine-tune this in
639 node = cpu_to_node(raw_smp_processor_id());
640 for (n = 0; n < MAX_NUMNODES; n++, node++) {
641 node = (node < MAX_NUMNODES) ? node : 0;
642 if (!node_allowed(ctx, node))
645 mutex_lock(&cbe_spu_info[node].list_mutex);
646 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
647 struct spu_context *tmp = spu->ctx;
649 if (tmp && tmp->prio > ctx->prio &&
650 !(tmp->flags & SPU_CREATE_NOSCHED) &&
651 (!victim || tmp->prio > victim->prio)) {
656 get_spu_context(victim);
657 mutex_unlock(&cbe_spu_info[node].list_mutex);
661 * This nests ctx->state_mutex, but we always lock
662 * higher priority contexts before lower priority
663 * ones, so this is safe until we introduce
664 * priority inheritance schemes.
666 * XXX if the highest priority context is locked,
667 * this can loop a long time. Might be better to
668 * look at another context or give up after X retries.
670 if (!mutex_trylock(&victim->state_mutex)) {
671 put_spu_context(victim);
677 if (!spu || victim->prio <= ctx->prio) {
679 * This race can happen because we've dropped
680 * the active list mutex. Not a problem, just
681 * restart the search.
683 mutex_unlock(&victim->state_mutex);
684 put_spu_context(victim);
689 spu_context_trace(__spu_deactivate__unload, ctx, spu);
691 mutex_lock(&cbe_spu_info[node].list_mutex);
692 cbe_spu_info[node].nr_active--;
693 spu_unbind_context(spu, victim);
694 mutex_unlock(&cbe_spu_info[node].list_mutex);
696 victim->stats.invol_ctx_switch++;
697 spu->stats.invol_ctx_switch++;
698 if (test_bit(SPU_SCHED_SPU_RUN, &victim->sched_flags))
699 spu_add_to_rq(victim);
701 mutex_unlock(&victim->state_mutex);
702 put_spu_context(victim);
711 static void __spu_schedule(struct spu *spu, struct spu_context *ctx)
713 int node = spu->node;
716 spu_set_timeslice(ctx);
718 mutex_lock(&cbe_spu_info[node].list_mutex);
719 if (spu->ctx == NULL) {
720 spu_bind_context(spu, ctx);
721 cbe_spu_info[node].nr_active++;
722 spu->alloc_state = SPU_USED;
725 mutex_unlock(&cbe_spu_info[node].list_mutex);
728 wake_up_all(&ctx->run_wq);
733 static void spu_schedule(struct spu *spu, struct spu_context *ctx)
735 /* not a candidate for interruptible because it's called either
736 from the scheduler thread or from spu_deactivate */
737 mutex_lock(&ctx->state_mutex);
738 if (ctx->state == SPU_STATE_SAVED)
739 __spu_schedule(spu, ctx);
744 * spu_unschedule - remove a context from a spu, and possibly release it.
745 * @spu: The SPU to unschedule from
746 * @ctx: The context currently scheduled on the SPU
747 * @free_spu Whether to free the SPU for other contexts
749 * Unbinds the context @ctx from the SPU @spu. If @free_spu is non-zero, the
750 * SPU is made available for other contexts (ie, may be returned by
751 * spu_get_idle). If this is zero, the caller is expected to schedule another
752 * context to this spu.
754 * Should be called with ctx->state_mutex held.
756 static void spu_unschedule(struct spu *spu, struct spu_context *ctx,
759 int node = spu->node;
761 mutex_lock(&cbe_spu_info[node].list_mutex);
762 cbe_spu_info[node].nr_active--;
764 spu->alloc_state = SPU_FREE;
765 spu_unbind_context(spu, ctx);
766 ctx->stats.invol_ctx_switch++;
767 spu->stats.invol_ctx_switch++;
768 mutex_unlock(&cbe_spu_info[node].list_mutex);
772 * spu_activate - find a free spu for a context and execute it
773 * @ctx: spu context to schedule
774 * @flags: flags (currently ignored)
776 * Tries to find a free spu to run @ctx. If no free spu is available
777 * add the context to the runqueue so it gets woken up once an spu
780 int spu_activate(struct spu_context *ctx, unsigned long flags)
785 * If there are multiple threads waiting for a single context
786 * only one actually binds the context while the others will
787 * only be able to acquire the state_mutex once the context
788 * already is in runnable state.
794 if (signal_pending(current))
797 spu = spu_get_idle(ctx);
799 * If this is a realtime thread we try to get it running by
800 * preempting a lower priority thread.
802 if (!spu && rt_prio(ctx->prio))
803 spu = find_victim(ctx);
805 unsigned long runcntl;
807 runcntl = ctx->ops->runcntl_read(ctx);
808 __spu_schedule(spu, ctx);
809 if (runcntl & SPU_RUNCNTL_RUNNABLE)
810 spuctx_switch_state(ctx, SPU_UTIL_USER);
815 if (ctx->flags & SPU_CREATE_NOSCHED) {
817 goto spu_activate_top;
826 * grab_runnable_context - try to find a runnable context
828 * Remove the highest priority context on the runqueue and return it
829 * to the caller. Returns %NULL if no runnable context was found.
831 static struct spu_context *grab_runnable_context(int prio, int node)
833 struct spu_context *ctx;
836 spin_lock(&spu_prio->runq_lock);
837 best = find_first_bit(spu_prio->bitmap, prio);
838 while (best < prio) {
839 struct list_head *rq = &spu_prio->runq[best];
841 list_for_each_entry(ctx, rq, rq) {
842 /* XXX(hch): check for affinity here aswell */
843 if (__node_allowed(ctx, node)) {
844 __spu_del_from_rq(ctx);
852 spin_unlock(&spu_prio->runq_lock);
856 static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
858 struct spu *spu = ctx->spu;
859 struct spu_context *new = NULL;
862 new = grab_runnable_context(max_prio, spu->node);
864 spu_unschedule(spu, ctx, new == NULL);
866 if (new->flags & SPU_CREATE_NOSCHED)
867 wake_up(&new->stop_wq);
870 spu_schedule(spu, new);
871 /* this one can't easily be made
873 mutex_lock(&ctx->state_mutex);
883 * spu_deactivate - unbind a context from it's physical spu
884 * @ctx: spu context to unbind
886 * Unbind @ctx from the physical spu it is running on and schedule
887 * the highest priority context to run on the freed physical spu.
889 void spu_deactivate(struct spu_context *ctx)
891 spu_context_nospu_trace(spu_deactivate__enter, ctx);
892 __spu_deactivate(ctx, 1, MAX_PRIO);
896 * spu_yield - yield a physical spu if others are waiting
897 * @ctx: spu context to yield
899 * Check if there is a higher priority context waiting and if yes
900 * unbind @ctx from the physical spu and schedule the highest
901 * priority context to run on the freed physical spu instead.
903 void spu_yield(struct spu_context *ctx)
905 spu_context_nospu_trace(spu_yield__enter, ctx);
906 if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
907 mutex_lock(&ctx->state_mutex);
908 __spu_deactivate(ctx, 0, MAX_PRIO);
909 mutex_unlock(&ctx->state_mutex);
913 static noinline void spusched_tick(struct spu_context *ctx)
915 struct spu_context *new = NULL;
916 struct spu *spu = NULL;
918 if (spu_acquire(ctx))
919 BUG(); /* a kernel thread never has signals pending */
921 if (ctx->state != SPU_STATE_RUNNABLE)
923 if (ctx->flags & SPU_CREATE_NOSCHED)
925 if (ctx->policy == SCHED_FIFO)
928 if (--ctx->time_slice && test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
933 spu_context_trace(spusched_tick__preempt, ctx, spu);
935 new = grab_runnable_context(ctx->prio + 1, spu->node);
937 spu_unschedule(spu, ctx, 0);
938 if (test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
941 spu_context_nospu_trace(spusched_tick__newslice, ctx);
942 if (!ctx->time_slice)
949 spu_schedule(spu, new);
953 * count_active_contexts - count nr of active tasks
955 * Return the number of tasks currently running or waiting to run.
957 * Note that we don't take runq_lock / list_mutex here. Reading
958 * a single 32bit value is atomic on powerpc, and we don't care
959 * about memory ordering issues here.
961 static unsigned long count_active_contexts(void)
963 int nr_active = 0, node;
965 for (node = 0; node < MAX_NUMNODES; node++)
966 nr_active += cbe_spu_info[node].nr_active;
967 nr_active += spu_prio->nr_waiting;
973 * spu_calc_load - update the avenrun load estimates.
975 * No locking against reading these values from userspace, as for
976 * the CPU loadavg code.
978 static void spu_calc_load(void)
980 unsigned long active_tasks; /* fixed-point */
982 active_tasks = count_active_contexts() * FIXED_1;
983 CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks);
984 CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks);
985 CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks);
988 static void spusched_wake(unsigned long data)
990 mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
991 wake_up_process(spusched_task);
994 static void spuloadavg_wake(unsigned long data)
996 mod_timer(&spuloadavg_timer, jiffies + LOAD_FREQ);
1000 static int spusched_thread(void *unused)
1005 while (!kthread_should_stop()) {
1006 set_current_state(TASK_INTERRUPTIBLE);
1008 for (node = 0; node < MAX_NUMNODES; node++) {
1009 struct mutex *mtx = &cbe_spu_info[node].list_mutex;
1012 list_for_each_entry(spu, &cbe_spu_info[node].spus,
1014 struct spu_context *ctx = spu->ctx;
1017 get_spu_context(ctx);
1021 put_spu_context(ctx);
1031 void spuctx_switch_state(struct spu_context *ctx,
1032 enum spu_utilization_state new_state)
1034 unsigned long long curtime;
1035 signed long long delta;
1038 enum spu_utilization_state old_state;
1042 curtime = timespec_to_ns(&ts);
1043 delta = curtime - ctx->stats.tstamp;
1045 WARN_ON(!mutex_is_locked(&ctx->state_mutex));
1049 old_state = ctx->stats.util_state;
1050 ctx->stats.util_state = new_state;
1051 ctx->stats.tstamp = curtime;
1054 * Update the physical SPU utilization statistics.
1057 ctx->stats.times[old_state] += delta;
1058 spu->stats.times[old_state] += delta;
1059 spu->stats.util_state = new_state;
1060 spu->stats.tstamp = curtime;
1062 if (old_state == SPU_UTIL_USER)
1063 atomic_dec(&cbe_spu_info[node].busy_spus);
1064 if (new_state == SPU_UTIL_USER)
1065 atomic_inc(&cbe_spu_info[node].busy_spus);
1069 #define LOAD_INT(x) ((x) >> FSHIFT)
1070 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
1072 static int show_spu_loadavg(struct seq_file *s, void *private)
1076 a = spu_avenrun[0] + (FIXED_1/200);
1077 b = spu_avenrun[1] + (FIXED_1/200);
1078 c = spu_avenrun[2] + (FIXED_1/200);
1081 * Note that last_pid doesn't really make much sense for the
1082 * SPU loadavg (it even seems very odd on the CPU side...),
1083 * but we include it here to have a 100% compatible interface.
1085 seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
1086 LOAD_INT(a), LOAD_FRAC(a),
1087 LOAD_INT(b), LOAD_FRAC(b),
1088 LOAD_INT(c), LOAD_FRAC(c),
1089 count_active_contexts(),
1090 atomic_read(&nr_spu_contexts),
1091 current->nsproxy->pid_ns->last_pid);
1095 static int spu_loadavg_open(struct inode *inode, struct file *file)
1097 return single_open(file, show_spu_loadavg, NULL);
1100 static const struct file_operations spu_loadavg_fops = {
1101 .open = spu_loadavg_open,
1103 .llseek = seq_lseek,
1104 .release = single_release,
1107 int __init spu_sched_init(void)
1109 struct proc_dir_entry *entry;
1110 int err = -ENOMEM, i;
1112 spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
1116 for (i = 0; i < MAX_PRIO; i++) {
1117 INIT_LIST_HEAD(&spu_prio->runq[i]);
1118 __clear_bit(i, spu_prio->bitmap);
1120 spin_lock_init(&spu_prio->runq_lock);
1122 setup_timer(&spusched_timer, spusched_wake, 0);
1123 setup_timer(&spuloadavg_timer, spuloadavg_wake, 0);
1125 spusched_task = kthread_run(spusched_thread, NULL, "spusched");
1126 if (IS_ERR(spusched_task)) {
1127 err = PTR_ERR(spusched_task);
1128 goto out_free_spu_prio;
1131 mod_timer(&spuloadavg_timer, 0);
1133 entry = proc_create("spu_loadavg", 0, NULL, &spu_loadavg_fops);
1135 goto out_stop_kthread;
1137 pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
1138 SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE);
1142 kthread_stop(spusched_task);
1149 void spu_sched_exit(void)
1154 remove_proc_entry("spu_loadavg", NULL);
1156 del_timer_sync(&spusched_timer);
1157 del_timer_sync(&spuloadavg_timer);
1158 kthread_stop(spusched_task);
1160 for (node = 0; node < MAX_NUMNODES; node++) {
1161 mutex_lock(&cbe_spu_info[node].list_mutex);
1162 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
1163 if (spu->alloc_state != SPU_FREE)
1164 spu->alloc_state = SPU_FREE;
1165 mutex_unlock(&cbe_spu_info[node].list_mutex);