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
;
236 ctx
->ops
= &spu_hw_ops
;
237 spu
->pid
= current
->pid
;
238 spu
->tgid
= current
->tgid
;
239 spu_associate_mm(spu
, ctx
->owner
);
240 spu
->ibox_callback
= spufs_ibox_callback
;
241 spu
->wbox_callback
= spufs_wbox_callback
;
242 spu
->stop_callback
= spufs_stop_callback
;
243 spu
->mfc_callback
= spufs_mfc_callback
;
245 spu_unmap_mappings(ctx
);
246 spu_switch_log_notify(spu
, ctx
, SWITCH_LOG_START
, 0);
247 spu_restore(&ctx
->csa
, spu
);
248 spu
->timestamp
= jiffies
;
249 spu_switch_notify(spu
, ctx
);
250 ctx
->state
= SPU_STATE_RUNNABLE
;
252 spuctx_switch_state(ctx
, SPU_UTIL_USER
);
256 * Must be used with the list_mutex held.
258 static inline int sched_spu(struct spu
*spu
)
260 BUG_ON(!mutex_is_locked(&cbe_spu_info
[spu
->node
].list_mutex
));
262 return (!spu
->ctx
|| !(spu
->ctx
->flags
& SPU_CREATE_NOSCHED
));
265 static void aff_merge_remaining_ctxs(struct spu_gang
*gang
)
267 struct spu_context
*ctx
;
269 list_for_each_entry(ctx
, &gang
->aff_list_head
, aff_list
) {
270 if (list_empty(&ctx
->aff_list
))
271 list_add(&ctx
->aff_list
, &gang
->aff_list_head
);
273 gang
->aff_flags
|= AFF_MERGED
;
276 static void aff_set_offsets(struct spu_gang
*gang
)
278 struct spu_context
*ctx
;
282 list_for_each_entry_reverse(ctx
, &gang
->aff_ref_ctx
->aff_list
,
284 if (&ctx
->aff_list
== &gang
->aff_list_head
)
286 ctx
->aff_offset
= offset
--;
290 list_for_each_entry(ctx
, gang
->aff_ref_ctx
->aff_list
.prev
, aff_list
) {
291 if (&ctx
->aff_list
== &gang
->aff_list_head
)
293 ctx
->aff_offset
= offset
++;
296 gang
->aff_flags
|= AFF_OFFSETS_SET
;
299 static struct spu
*aff_ref_location(struct spu_context
*ctx
, int mem_aff
,
300 int group_size
, int lowest_offset
)
306 * TODO: A better algorithm could be used to find a good spu to be
307 * used as reference location for the ctxs chain.
309 node
= cpu_to_node(raw_smp_processor_id());
310 for (n
= 0; n
< MAX_NUMNODES
; n
++, node
++) {
311 node
= (node
< MAX_NUMNODES
) ? node
: 0;
312 if (!node_allowed(ctx
, node
))
314 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
315 list_for_each_entry(spu
, &cbe_spu_info
[node
].spus
, cbe_list
) {
316 if ((!mem_aff
|| spu
->has_mem_affinity
) &&
318 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
322 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
327 static void aff_set_ref_point_location(struct spu_gang
*gang
)
329 int mem_aff
, gs
, lowest_offset
;
330 struct spu_context
*ctx
;
333 mem_aff
= gang
->aff_ref_ctx
->flags
& SPU_CREATE_AFFINITY_MEM
;
337 list_for_each_entry(tmp
, &gang
->aff_list_head
, aff_list
)
340 list_for_each_entry_reverse(ctx
, &gang
->aff_ref_ctx
->aff_list
,
342 if (&ctx
->aff_list
== &gang
->aff_list_head
)
344 lowest_offset
= ctx
->aff_offset
;
347 gang
->aff_ref_spu
= aff_ref_location(gang
->aff_ref_ctx
, mem_aff
, gs
,
351 static struct spu
*ctx_location(struct spu
*ref
, int offset
, int node
)
357 list_for_each_entry(spu
, ref
->aff_list
.prev
, aff_list
) {
358 BUG_ON(spu
->node
!= node
);
365 list_for_each_entry_reverse(spu
, ref
->aff_list
.next
, aff_list
) {
366 BUG_ON(spu
->node
!= node
);
378 * affinity_check is called each time a context is going to be scheduled.
379 * It returns the spu ptr on which the context must run.
381 static int has_affinity(struct spu_context
*ctx
)
383 struct spu_gang
*gang
= ctx
->gang
;
385 if (list_empty(&ctx
->aff_list
))
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
);
396 return gang
->aff_ref_spu
!= NULL
;
400 * spu_unbind_context - unbind spu context from physical spu
401 * @spu: physical spu to unbind from
402 * @ctx: context to unbind
404 static void spu_unbind_context(struct spu
*spu
, struct spu_context
*ctx
)
406 spu_context_trace(spu_unbind_context__enter
, ctx
, spu
);
408 spuctx_switch_state(ctx
, SPU_UTIL_SYSTEM
);
410 if (spu
->ctx
->flags
& SPU_CREATE_NOSCHED
)
411 atomic_dec(&cbe_spu_info
[spu
->node
].reserved_spus
);
414 mutex_lock(&ctx
->gang
->aff_mutex
);
415 if (has_affinity(ctx
)) {
416 if (atomic_dec_and_test(&ctx
->gang
->aff_sched_count
))
417 ctx
->gang
->aff_ref_spu
= NULL
;
419 mutex_unlock(&ctx
->gang
->aff_mutex
);
422 spu_switch_notify(spu
, NULL
);
423 spu_unmap_mappings(ctx
);
424 spu_save(&ctx
->csa
, spu
);
425 spu_switch_log_notify(spu
, ctx
, SWITCH_LOG_STOP
, 0);
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_associate_mm(spu
, NULL
);
435 ctx
->ops
= &spu_backing_ops
;
439 ctx
->stats
.slb_flt
+=
440 (spu
->stats
.slb_flt
- ctx
->stats
.slb_flt_base
);
441 ctx
->stats
.class2_intr
+=
442 (spu
->stats
.class2_intr
- ctx
->stats
.class2_intr_base
);
444 /* This maps the underlying spu state to idle */
445 spuctx_switch_state(ctx
, SPU_UTIL_IDLE_LOADED
);
450 * spu_add_to_rq - add a context to the runqueue
451 * @ctx: context to add
453 static void __spu_add_to_rq(struct spu_context
*ctx
)
456 * Unfortunately this code path can be called from multiple threads
457 * on behalf of a single context due to the way the problem state
458 * mmap support works.
460 * Fortunately we need to wake up all these threads at the same time
461 * and can simply skip the runqueue addition for every but the first
462 * thread getting into this codepath.
464 * It's still quite hacky, and long-term we should proxy all other
465 * threads through the owner thread so that spu_run is in control
466 * of all the scheduling activity for a given context.
468 if (list_empty(&ctx
->rq
)) {
469 list_add_tail(&ctx
->rq
, &spu_prio
->runq
[ctx
->prio
]);
470 set_bit(ctx
->prio
, spu_prio
->bitmap
);
471 if (!spu_prio
->nr_waiting
++)
472 __mod_timer(&spusched_timer
, jiffies
+ SPUSCHED_TICK
);
476 static void spu_add_to_rq(struct spu_context
*ctx
)
478 spin_lock(&spu_prio
->runq_lock
);
479 __spu_add_to_rq(ctx
);
480 spin_unlock(&spu_prio
->runq_lock
);
483 static void __spu_del_from_rq(struct spu_context
*ctx
)
485 int prio
= ctx
->prio
;
487 if (!list_empty(&ctx
->rq
)) {
488 if (!--spu_prio
->nr_waiting
)
489 del_timer(&spusched_timer
);
490 list_del_init(&ctx
->rq
);
492 if (list_empty(&spu_prio
->runq
[prio
]))
493 clear_bit(prio
, spu_prio
->bitmap
);
497 void spu_del_from_rq(struct spu_context
*ctx
)
499 spin_lock(&spu_prio
->runq_lock
);
500 __spu_del_from_rq(ctx
);
501 spin_unlock(&spu_prio
->runq_lock
);
504 static void spu_prio_wait(struct spu_context
*ctx
)
509 * The caller must explicitly wait for a context to be loaded
510 * if the nosched flag is set. If NOSCHED is not set, the caller
511 * queues the context and waits for an spu event or error.
513 BUG_ON(!(ctx
->flags
& SPU_CREATE_NOSCHED
));
515 spin_lock(&spu_prio
->runq_lock
);
516 prepare_to_wait_exclusive(&ctx
->stop_wq
, &wait
, TASK_INTERRUPTIBLE
);
517 if (!signal_pending(current
)) {
518 __spu_add_to_rq(ctx
);
519 spin_unlock(&spu_prio
->runq_lock
);
520 mutex_unlock(&ctx
->state_mutex
);
522 mutex_lock(&ctx
->state_mutex
);
523 spin_lock(&spu_prio
->runq_lock
);
524 __spu_del_from_rq(ctx
);
526 spin_unlock(&spu_prio
->runq_lock
);
527 __set_current_state(TASK_RUNNING
);
528 remove_wait_queue(&ctx
->stop_wq
, &wait
);
531 static struct spu
*spu_get_idle(struct spu_context
*ctx
)
533 struct spu
*spu
, *aff_ref_spu
;
536 spu_context_nospu_trace(spu_get_idle__enter
, ctx
);
539 mutex_lock(&ctx
->gang
->aff_mutex
);
540 if (has_affinity(ctx
)) {
541 aff_ref_spu
= ctx
->gang
->aff_ref_spu
;
542 atomic_inc(&ctx
->gang
->aff_sched_count
);
543 mutex_unlock(&ctx
->gang
->aff_mutex
);
544 node
= aff_ref_spu
->node
;
546 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
547 spu
= ctx_location(aff_ref_spu
, ctx
->aff_offset
, node
);
548 if (spu
&& spu
->alloc_state
== SPU_FREE
)
550 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
552 mutex_lock(&ctx
->gang
->aff_mutex
);
553 if (atomic_dec_and_test(&ctx
->gang
->aff_sched_count
))
554 ctx
->gang
->aff_ref_spu
= NULL
;
555 mutex_unlock(&ctx
->gang
->aff_mutex
);
558 mutex_unlock(&ctx
->gang
->aff_mutex
);
560 node
= cpu_to_node(raw_smp_processor_id());
561 for (n
= 0; n
< MAX_NUMNODES
; n
++, node
++) {
562 node
= (node
< MAX_NUMNODES
) ? node
: 0;
563 if (!node_allowed(ctx
, node
))
566 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
567 list_for_each_entry(spu
, &cbe_spu_info
[node
].spus
, cbe_list
) {
568 if (spu
->alloc_state
== SPU_FREE
)
571 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
575 spu_context_nospu_trace(spu_get_idle__not_found
, ctx
);
579 spu
->alloc_state
= SPU_USED
;
580 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
581 spu_context_trace(spu_get_idle__found
, ctx
, spu
);
582 spu_init_channels(spu
);
587 * find_victim - find a lower priority context to preempt
588 * @ctx: canidate context for running
590 * Returns the freed physical spu to run the new context on.
592 static struct spu
*find_victim(struct spu_context
*ctx
)
594 struct spu_context
*victim
= NULL
;
598 spu_context_nospu_trace(spu_find_victim__enter
, ctx
);
601 * Look for a possible preemption candidate on the local node first.
602 * If there is no candidate look at the other nodes. This isn't
603 * exactly fair, but so far the whole spu scheduler tries to keep
604 * a strong node affinity. We might want to fine-tune this in
608 node
= cpu_to_node(raw_smp_processor_id());
609 for (n
= 0; n
< MAX_NUMNODES
; n
++, node
++) {
610 node
= (node
< MAX_NUMNODES
) ? node
: 0;
611 if (!node_allowed(ctx
, node
))
614 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
615 list_for_each_entry(spu
, &cbe_spu_info
[node
].spus
, cbe_list
) {
616 struct spu_context
*tmp
= spu
->ctx
;
618 if (tmp
&& tmp
->prio
> ctx
->prio
&&
619 !(tmp
->flags
& SPU_CREATE_NOSCHED
) &&
620 (!victim
|| tmp
->prio
> victim
->prio
))
623 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
627 * This nests ctx->state_mutex, but we always lock
628 * higher priority contexts before lower priority
629 * ones, so this is safe until we introduce
630 * priority inheritance schemes.
632 * XXX if the highest priority context is locked,
633 * this can loop a long time. Might be better to
634 * look at another context or give up after X retries.
636 if (!mutex_trylock(&victim
->state_mutex
)) {
642 if (!spu
|| victim
->prio
<= ctx
->prio
) {
644 * This race can happen because we've dropped
645 * the active list mutex. Not a problem, just
646 * restart the search.
648 mutex_unlock(&victim
->state_mutex
);
653 spu_context_trace(__spu_deactivate__unload
, ctx
, spu
);
655 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
656 cbe_spu_info
[node
].nr_active
--;
657 spu_unbind_context(spu
, victim
);
658 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
660 victim
->stats
.invol_ctx_switch
++;
661 spu
->stats
.invol_ctx_switch
++;
662 if (test_bit(SPU_SCHED_SPU_RUN
, &victim
->sched_flags
))
663 spu_add_to_rq(victim
);
665 mutex_unlock(&victim
->state_mutex
);
674 static void __spu_schedule(struct spu
*spu
, struct spu_context
*ctx
)
676 int node
= spu
->node
;
679 spu_set_timeslice(ctx
);
681 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
682 if (spu
->ctx
== NULL
) {
683 spu_bind_context(spu
, ctx
);
684 cbe_spu_info
[node
].nr_active
++;
685 spu
->alloc_state
= SPU_USED
;
688 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
691 wake_up_all(&ctx
->run_wq
);
696 static void spu_schedule(struct spu
*spu
, struct spu_context
*ctx
)
698 /* not a candidate for interruptible because it's called either
699 from the scheduler thread or from spu_deactivate */
700 mutex_lock(&ctx
->state_mutex
);
701 __spu_schedule(spu
, ctx
);
705 static void spu_unschedule(struct spu
*spu
, struct spu_context
*ctx
)
707 int node
= spu
->node
;
709 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
710 cbe_spu_info
[node
].nr_active
--;
711 spu
->alloc_state
= SPU_FREE
;
712 spu_unbind_context(spu
, ctx
);
713 ctx
->stats
.invol_ctx_switch
++;
714 spu
->stats
.invol_ctx_switch
++;
715 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);
719 * spu_activate - find a free spu for a context and execute it
720 * @ctx: spu context to schedule
721 * @flags: flags (currently ignored)
723 * Tries to find a free spu to run @ctx. If no free spu is available
724 * add the context to the runqueue so it gets woken up once an spu
727 int spu_activate(struct spu_context
*ctx
, unsigned long flags
)
732 * If there are multiple threads waiting for a single context
733 * only one actually binds the context while the others will
734 * only be able to acquire the state_mutex once the context
735 * already is in runnable state.
741 if (signal_pending(current
))
744 spu
= spu_get_idle(ctx
);
746 * If this is a realtime thread we try to get it running by
747 * preempting a lower priority thread.
749 if (!spu
&& rt_prio(ctx
->prio
))
750 spu
= find_victim(ctx
);
752 unsigned long runcntl
;
754 runcntl
= ctx
->ops
->runcntl_read(ctx
);
755 __spu_schedule(spu
, ctx
);
756 if (runcntl
& SPU_RUNCNTL_RUNNABLE
)
757 spuctx_switch_state(ctx
, SPU_UTIL_USER
);
762 if (ctx
->flags
& SPU_CREATE_NOSCHED
) {
764 goto spu_activate_top
;
773 * grab_runnable_context - try to find a runnable context
775 * Remove the highest priority context on the runqueue and return it
776 * to the caller. Returns %NULL if no runnable context was found.
778 static struct spu_context
*grab_runnable_context(int prio
, int node
)
780 struct spu_context
*ctx
;
783 spin_lock(&spu_prio
->runq_lock
);
784 best
= find_first_bit(spu_prio
->bitmap
, prio
);
785 while (best
< prio
) {
786 struct list_head
*rq
= &spu_prio
->runq
[best
];
788 list_for_each_entry(ctx
, rq
, rq
) {
789 /* XXX(hch): check for affinity here aswell */
790 if (__node_allowed(ctx
, node
)) {
791 __spu_del_from_rq(ctx
);
799 spin_unlock(&spu_prio
->runq_lock
);
803 static int __spu_deactivate(struct spu_context
*ctx
, int force
, int max_prio
)
805 struct spu
*spu
= ctx
->spu
;
806 struct spu_context
*new = NULL
;
809 new = grab_runnable_context(max_prio
, spu
->node
);
811 spu_unschedule(spu
, ctx
);
813 if (new->flags
& SPU_CREATE_NOSCHED
)
814 wake_up(&new->stop_wq
);
817 spu_schedule(spu
, new);
818 /* this one can't easily be made
820 mutex_lock(&ctx
->state_mutex
);
830 * spu_deactivate - unbind a context from it's physical spu
831 * @ctx: spu context to unbind
833 * Unbind @ctx from the physical spu it is running on and schedule
834 * the highest priority context to run on the freed physical spu.
836 void spu_deactivate(struct spu_context
*ctx
)
838 spu_context_nospu_trace(spu_deactivate__enter
, ctx
);
839 __spu_deactivate(ctx
, 1, MAX_PRIO
);
843 * spu_yield - yield a physical spu if others are waiting
844 * @ctx: spu context to yield
846 * Check if there is a higher priority context waiting and if yes
847 * unbind @ctx from the physical spu and schedule the highest
848 * priority context to run on the freed physical spu instead.
850 void spu_yield(struct spu_context
*ctx
)
852 spu_context_nospu_trace(spu_yield__enter
, ctx
);
853 if (!(ctx
->flags
& SPU_CREATE_NOSCHED
)) {
854 mutex_lock(&ctx
->state_mutex
);
855 __spu_deactivate(ctx
, 0, MAX_PRIO
);
856 mutex_unlock(&ctx
->state_mutex
);
860 static noinline
void spusched_tick(struct spu_context
*ctx
)
862 struct spu_context
*new = NULL
;
863 struct spu
*spu
= NULL
;
865 if (spu_acquire(ctx
))
866 BUG(); /* a kernel thread never has signals pending */
868 if (ctx
->state
!= SPU_STATE_RUNNABLE
)
870 if (ctx
->flags
& SPU_CREATE_NOSCHED
)
872 if (ctx
->policy
== SCHED_FIFO
)
875 if (--ctx
->time_slice
&& test_bit(SPU_SCHED_SPU_RUN
, &ctx
->sched_flags
))
880 spu_context_trace(spusched_tick__preempt
, ctx
, spu
);
882 new = grab_runnable_context(ctx
->prio
+ 1, spu
->node
);
884 spu_unschedule(spu
, ctx
);
885 if (test_bit(SPU_SCHED_SPU_RUN
, &ctx
->sched_flags
))
888 spu_context_nospu_trace(spusched_tick__newslice
, ctx
);
895 spu_schedule(spu
, new);
899 * count_active_contexts - count nr of active tasks
901 * Return the number of tasks currently running or waiting to run.
903 * Note that we don't take runq_lock / list_mutex here. Reading
904 * a single 32bit value is atomic on powerpc, and we don't care
905 * about memory ordering issues here.
907 static unsigned long count_active_contexts(void)
909 int nr_active
= 0, node
;
911 for (node
= 0; node
< MAX_NUMNODES
; node
++)
912 nr_active
+= cbe_spu_info
[node
].nr_active
;
913 nr_active
+= spu_prio
->nr_waiting
;
919 * spu_calc_load - update the avenrun load estimates.
921 * No locking against reading these values from userspace, as for
922 * the CPU loadavg code.
924 static void spu_calc_load(void)
926 unsigned long active_tasks
; /* fixed-point */
928 active_tasks
= count_active_contexts() * FIXED_1
;
929 CALC_LOAD(spu_avenrun
[0], EXP_1
, active_tasks
);
930 CALC_LOAD(spu_avenrun
[1], EXP_5
, active_tasks
);
931 CALC_LOAD(spu_avenrun
[2], EXP_15
, active_tasks
);
934 static void spusched_wake(unsigned long data
)
936 mod_timer(&spusched_timer
, jiffies
+ SPUSCHED_TICK
);
937 wake_up_process(spusched_task
);
940 static void spuloadavg_wake(unsigned long data
)
942 mod_timer(&spuloadavg_timer
, jiffies
+ LOAD_FREQ
);
946 static int spusched_thread(void *unused
)
951 while (!kthread_should_stop()) {
952 set_current_state(TASK_INTERRUPTIBLE
);
954 for (node
= 0; node
< MAX_NUMNODES
; node
++) {
955 struct mutex
*mtx
= &cbe_spu_info
[node
].list_mutex
;
958 list_for_each_entry(spu
, &cbe_spu_info
[node
].spus
,
960 struct spu_context
*ctx
= spu
->ctx
;
975 void spuctx_switch_state(struct spu_context
*ctx
,
976 enum spu_utilization_state new_state
)
978 unsigned long long curtime
;
979 signed long long delta
;
982 enum spu_utilization_state old_state
;
985 curtime
= timespec_to_ns(&ts
);
986 delta
= curtime
- ctx
->stats
.tstamp
;
988 WARN_ON(!mutex_is_locked(&ctx
->state_mutex
));
992 old_state
= ctx
->stats
.util_state
;
993 ctx
->stats
.util_state
= new_state
;
994 ctx
->stats
.tstamp
= curtime
;
997 * Update the physical SPU utilization statistics.
1000 ctx
->stats
.times
[old_state
] += delta
;
1001 spu
->stats
.times
[old_state
] += delta
;
1002 spu
->stats
.util_state
= new_state
;
1003 spu
->stats
.tstamp
= curtime
;
1007 #define LOAD_INT(x) ((x) >> FSHIFT)
1008 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
1010 static int show_spu_loadavg(struct seq_file
*s
, void *private)
1014 a
= spu_avenrun
[0] + (FIXED_1
/200);
1015 b
= spu_avenrun
[1] + (FIXED_1
/200);
1016 c
= spu_avenrun
[2] + (FIXED_1
/200);
1019 * Note that last_pid doesn't really make much sense for the
1020 * SPU loadavg (it even seems very odd on the CPU side...),
1021 * but we include it here to have a 100% compatible interface.
1023 seq_printf(s
, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
1024 LOAD_INT(a
), LOAD_FRAC(a
),
1025 LOAD_INT(b
), LOAD_FRAC(b
),
1026 LOAD_INT(c
), LOAD_FRAC(c
),
1027 count_active_contexts(),
1028 atomic_read(&nr_spu_contexts
),
1029 current
->nsproxy
->pid_ns
->last_pid
);
1033 static int spu_loadavg_open(struct inode
*inode
, struct file
*file
)
1035 return single_open(file
, show_spu_loadavg
, NULL
);
1038 static const struct file_operations spu_loadavg_fops
= {
1039 .open
= spu_loadavg_open
,
1041 .llseek
= seq_lseek
,
1042 .release
= single_release
,
1045 int __init
spu_sched_init(void)
1047 struct proc_dir_entry
*entry
;
1048 int err
= -ENOMEM
, i
;
1050 spu_prio
= kzalloc(sizeof(struct spu_prio_array
), GFP_KERNEL
);
1054 for (i
= 0; i
< MAX_PRIO
; i
++) {
1055 INIT_LIST_HEAD(&spu_prio
->runq
[i
]);
1056 __clear_bit(i
, spu_prio
->bitmap
);
1058 spin_lock_init(&spu_prio
->runq_lock
);
1060 setup_timer(&spusched_timer
, spusched_wake
, 0);
1061 setup_timer(&spuloadavg_timer
, spuloadavg_wake
, 0);
1063 spusched_task
= kthread_run(spusched_thread
, NULL
, "spusched");
1064 if (IS_ERR(spusched_task
)) {
1065 err
= PTR_ERR(spusched_task
);
1066 goto out_free_spu_prio
;
1069 mod_timer(&spuloadavg_timer
, 0);
1071 entry
= proc_create("spu_loadavg", 0, NULL
, &spu_loadavg_fops
);
1073 goto out_stop_kthread
;
1075 pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
1076 SPUSCHED_TICK
, MIN_SPU_TIMESLICE
, DEF_SPU_TIMESLICE
);
1080 kthread_stop(spusched_task
);
1087 void spu_sched_exit(void)
1092 remove_proc_entry("spu_loadavg", NULL
);
1094 del_timer_sync(&spusched_timer
);
1095 del_timer_sync(&spuloadavg_timer
);
1096 kthread_stop(spusched_task
);
1098 for (node
= 0; node
< MAX_NUMNODES
; node
++) {
1099 mutex_lock(&cbe_spu_info
[node
].list_mutex
);
1100 list_for_each_entry(spu
, &cbe_spu_info
[node
].spus
, cbe_list
)
1101 if (spu
->alloc_state
!= SPU_FREE
)
1102 spu
->alloc_state
= SPU_FREE
;
1103 mutex_unlock(&cbe_spu_info
[node
].list_mutex
);