[Cbe-oss-dev] [PATCH 11/11 v2] powerpc/spufs: Implement SPU affinity on top of gang scheduling
Andre Detsch
adetsch at br.ibm.com
Fri Sep 12 22:22:02 EST 2008
SPU affinity, originally implemented before we had gang
scheduling, was disabled after gang scheduling was
introduced.
This patch re-enables SPU affinity, making it fit the new scheduling
algorithm.
Signed-off-by: Andre Detsch <adetsch at br.ibm.com>
---
arch/powerpc/platforms/cell/spufs/sched.c | 169
+++++++++++++----------------
arch/powerpc/platforms/cell/spufs/spufs.h | 1 -
2 files changed, 78 insertions(+), 92 deletions(-)
diff --git a/arch/powerpc/platforms/cell/spufs/sched.c
b/arch/powerpc/platforms/cell/spufs/sched.c
index 8326034..d4ef6e3 100644
--- a/arch/powerpc/platforms/cell/spufs/sched.c
+++ b/arch/powerpc/platforms/cell/spufs/sched.c
@@ -348,126 +348,97 @@ static void aff_set_offsets(struct spu_gang *gang)
gang->aff_flags |= AFF_OFFSETS_SET;
}
-static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff,
- int group_size, int lowest_offset)
+static struct spu *ctx_location(struct spu *ref, int offset,
+ struct spu_gang *gang)
+{
+ struct spu *spu;
+
+ spu = NULL;
+ if (offset >= 0) {
+ list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
+ if ((!gang && spu->alloc_state == SPU_FREE) ||
+ spu->gang == gang) {
+ if (offset == 0)
+ break;
+ offset--;
+ }
+ }
+ } else {
+ list_for_each_entry_reverse(spu, ref->aff_list.next,
+ aff_list) {
+ if ((!gang && spu->alloc_state == SPU_FREE) ||
+ spu->gang == gang) {
+ if (offset == 0)
+ break;
+ offset++;
+ }
+ }
+ }
+
+ return spu;
+}
+
+static int aff_place_gang(struct spu_gang *gang)
{
struct spu *spu;
int node, n;
+ struct spu_context *ctx;
+ int mem_aff;
+
+
+ BUG_ON(list_empty(&gang->aff_list_head));
+
+ if (!(gang->aff_flags & AFF_MERGED))
+ aff_merge_remaining_ctxs(gang);
+ if (!(gang->aff_flags & AFF_OFFSETS_SET))
+ aff_set_offsets(gang);
+
+ mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
- /*
- * TODO: A better algorithm could be used to find a good spu to be
- * used as reference location for the ctxs chain.
- */
node = cpu_to_node(raw_smp_processor_id());
for (n = 0; n < MAX_NUMNODES; n++, node++) {
int available_spus;
node = (node < MAX_NUMNODES) ? node : 0;
- if (!node_allowed(ctx->gang, node))
+ if (!node_allowed(gang, node))
continue;
available_spus = 0;
spin_lock(&cbe_spu_info[node].list_lock);
list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
- if (spu->ctx && spu->ctx->gang
- && spu->ctx->aff_offset == 0)
- available_spus -=
- (spu->ctx->gang->contexts - 1);
- else
+ if (spu->alloc_state == SPU_FREE)
available_spus++;
}
- if (available_spus < ctx->gang->contexts) {
+ if (available_spus < gang->contexts) {
spin_unlock(&cbe_spu_info[node].list_lock);
continue;
}
list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
if ((!mem_aff || spu->has_mem_affinity) &&
- sched_spu(spu)) {
+ spu->alloc_state == SPU_FREE) {
+ gang->aff_ref_spu = spu;
spin_unlock(&cbe_spu_info[node].list_lock);
- return spu;
}
}
- spin_unlock(&cbe_spu_info[node].list_lock);
- }
- return NULL;
-}
-static void aff_set_ref_point_location(struct spu_gang *gang)
-{
- int mem_aff, gs, lowest_offset;
- struct spu_context *ctx;
- struct spu *tmp;
-
- mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
- lowest_offset = 0;
- gs = 0;
-
- list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
- gs++;
+ list_for_each_entry(ctx, &gang->list, gang_list) {
+ spu = ctx_location(gang->aff_ref_spu, ctx->aff_offset,
+ NULL);
+ BUG_ON(!spu);
+ spu->gang = gang;
+ }
- list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
- aff_list) {
- if (&ctx->aff_list == &gang->aff_list_head)
- break;
- lowest_offset = ctx->aff_offset;
+ spin_unlock(&cbe_spu_info[node].list_lock);
+ return 1;
}
-
- gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs,
- lowest_offset);
+ return 0;
}
-static struct spu *ctx_location(struct spu *ref, int offset, int node)
-{
- struct spu *spu;
-
- spu = NULL;
- if (offset >= 0) {
- list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
- BUG_ON(spu->node != node);
- if (offset == 0)
- break;
- if (sched_spu(spu))
- offset--;
- }
- } else {
- list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
- BUG_ON(spu->node != node);
- if (offset == 0)
- break;
- if (sched_spu(spu))
- offset++;
- }
- }
-
- return spu;
-}
-/*
- * affinity_check is called each time a context is going to be scheduled.
- * It returns the spu ptr on which the context must run.
- */
static int has_affinity(struct spu_gang *gang)
{
- if (list_empty(&gang->aff_list_head))
- return 0;
-
- /*
- * TODO: fix SPU Affinity to work with gang scheduling.
- */
-
- if (atomic_read(&gang->aff_sched_count) == 0)
- gang->aff_ref_spu = NULL;
-
- if (!gang->aff_ref_spu) {
- if (!(gang->aff_flags & AFF_MERGED))
- aff_merge_remaining_ctxs(gang);
- if (!(gang->aff_flags & AFF_OFFSETS_SET))
- aff_set_offsets(gang);
- aff_set_ref_point_location(gang);
- }
-
- return gang->aff_ref_spu != NULL;
+ return !list_empty(&gang->aff_list_head);
}
/**
@@ -486,9 +457,6 @@ static void spu_unbind_context(struct spu *spu, struct
spu_context *ctx)
if (spu->ctx->flags & SPU_CREATE_NOSCHED)
atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
- if (ctx->gang)
- atomic_dec_if_positive(&ctx->gang->aff_sched_count);
-
spu_switch_notify(spu, NULL);
spu_unmap_mappings(ctx);
spu_save(&ctx->csa, spu);
@@ -582,6 +550,21 @@ static struct spu *spu_bind(struct spu_gang *gang,
if (!node_allowed(gang, node))
continue;
+ if (has_affinity(gang)) {
+ spin_lock(&cbe_spu_info[node].list_lock);
+ spu = ctx_location(gang->aff_ref_spu, ctx->aff_offset,
+ gang);
+ if (spu && (spu->alloc_state == SPU_FREE) &&
+ (spu->gang == gang)) {
+ goto found;
+ }
+
+ /* If ctx_location returns an spu, it should be valid */
+ WARN_ON(spu);
+
+ spin_unlock(&cbe_spu_info[node].list_lock);
+ }
+
spin_lock(&cbe_spu_info[node].list_lock);
list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
if ((spu->alloc_state == SPU_FREE) &&
@@ -646,6 +629,7 @@ static void spu_unschedule(struct spu_gang *gang)
spin_unlock(&cbe_spu_info[node].list_lock);
mutex_unlock(&ctx->state_mutex);
}
+ gang->aff_ref_spu = NULL;
}
static int spu_get_idle(struct spu_gang *gang, int node)
@@ -656,7 +640,9 @@ static int spu_get_idle(struct spu_gang *gang, int node)
spu_context_nospu_trace(spu_get_idle__enter, gang);
- /* TO DO: SPU affinity scheduling. */
+ if (has_affinity(gang) && !gang->aff_ref_spu)
+ ret = aff_place_gang(gang);
+ goto spu_get_idle_out;
mode = SPU_RESERVE;
@@ -676,6 +662,7 @@ spu_get_idle_top:
continue;
spin_lock(&cbe_spu_info[lnode].list_lock);
+
list_for_each_entry(spu, &cbe_spu_info[lnode].spus, cbe_list) {
switch (mode) {
case SPU_RESERVE :
diff --git a/arch/powerpc/platforms/cell/spufs/spufs.h
b/arch/powerpc/platforms/cell/spufs/spufs.h
index 6afc514..907baf9 100644
--- a/arch/powerpc/platforms/cell/spufs/spufs.h
+++ b/arch/powerpc/platforms/cell/spufs/spufs.h
@@ -178,7 +178,6 @@ struct spu_gang {
struct mutex aff_mutex;
int aff_flags;
struct spu *aff_ref_spu;
- atomic_t aff_sched_count;
/* spu scheduler statistics for zombie ctxts */
struct {
--
1.5.4.1
More information about the cbe-oss-dev
mailing list