Ben Skeggs
2010-May-31 02:01 UTC
[Nouveau] [PATCH] drm/nouveau: reduce usage of fence spinlock to when absolutely necessary
From: Ben Skeggs <bskeggs at redhat.com>
Signed-off-by: Ben Skeggs <bskeggs at redhat.com>
---
drivers/gpu/drm/nouveau/nouveau_channel.c | 2 --
drivers/gpu/drm/nouveau/nouveau_drv.h | 1 -
drivers/gpu/drm/nouveau/nouveau_fence.c | 28 ++++------------------------
drivers/gpu/drm/nouveau/nv04_graph.c | 1 -
4 files changed, 4 insertions(+), 28 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_channel.c
b/drivers/gpu/drm/nouveau/nouveau_channel.c
index 1fc57ef..f9b2acf 100644
--- a/drivers/gpu/drm/nouveau/nouveau_channel.c
+++ b/drivers/gpu/drm/nouveau/nouveau_channel.c
@@ -257,9 +257,7 @@ nouveau_channel_free(struct nouveau_channel *chan)
nouveau_debugfs_channel_fini(chan);
/* Give outstanding push buffers a chance to complete */
- spin_lock_irqsave(&chan->fence.lock, flags);
nouveau_fence_update(chan);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
if (chan->fence.sequence != chan->fence.sequence_ack) {
struct nouveau_fence *fence = NULL;
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h
b/drivers/gpu/drm/nouveau/nouveau_drv.h
index 9eff502..cd87378 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -1141,7 +1141,6 @@ extern int nouveau_fence_wait(void *obj, void *arg, bool
lazy, bool intr);
extern int nouveau_fence_flush(void *obj, void *arg);
extern void nouveau_fence_unref(void **obj);
extern void *nouveau_fence_ref(void *obj);
-extern void nouveau_fence_handler(struct drm_device *dev, int channel);
/* nouveau_gem.c */
extern int nouveau_gem_new(struct drm_device *, struct nouveau_channel *,
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c
b/drivers/gpu/drm/nouveau/nouveau_fence.c
index faddf53..d0775fc 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -73,6 +73,7 @@ nouveau_fence_update(struct nouveau_channel *chan)
return;
chan->fence.sequence_ack = sequence;
+ spin_lock(&chan->fence.lock);
list_for_each_safe(entry, tmp, &chan->fence.pending) {
fence = list_entry(entry, struct nouveau_fence, entry);
@@ -84,6 +85,7 @@ nouveau_fence_update(struct nouveau_channel *chan)
if (sequence == chan->fence.sequence_ack)
break;
}
+ spin_unlock(&chan->fence.lock);
}
int
@@ -119,7 +121,6 @@ nouveau_fence_emit(struct nouveau_fence *fence)
{
struct drm_nouveau_private *dev_priv =
fence->channel->dev->dev_private;
struct nouveau_channel *chan = fence->channel;
- unsigned long flags;
int ret;
ret = RING_SPACE(chan, 2);
@@ -127,9 +128,7 @@ nouveau_fence_emit(struct nouveau_fence *fence)
return ret;
if (unlikely(chan->fence.sequence == chan->fence.sequence_ack - 1)) {
- spin_lock_irqsave(&chan->fence.lock, flags);
nouveau_fence_update(chan);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
BUG_ON(chan->fence.sequence = chan->fence.sequence_ack - 1);
@@ -138,9 +137,9 @@ nouveau_fence_emit(struct nouveau_fence *fence)
fence->sequence = ++chan->fence.sequence;
kref_get(&fence->refcount);
- spin_lock_irqsave(&chan->fence.lock, flags);
+ spin_lock(&chan->fence.lock);
list_add_tail(&fence->entry, &chan->fence.pending);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
+ spin_unlock(&chan->fence.lock);
BEGIN_RING(chan, NvSubSw, USE_REFCNT ? 0x0050 : 0x0150, 1);
OUT_RING(chan, fence->sequence);
@@ -173,14 +172,11 @@ nouveau_fence_signalled(void *sync_obj, void *sync_arg)
{
struct nouveau_fence *fence = nouveau_fence(sync_obj);
struct nouveau_channel *chan = fence->channel;
- unsigned long flags;
if (fence->signalled)
return true;
- spin_lock_irqsave(&chan->fence.lock, flags);
nouveau_fence_update(chan);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
return fence->signalled;
}
@@ -221,22 +217,6 @@ nouveau_fence_flush(void *sync_obj, void *sync_arg)
return 0;
}
-void
-nouveau_fence_handler(struct drm_device *dev, int channel)
-{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
- struct nouveau_channel *chan = NULL;
-
- if (channel >= 0 && channel < dev_priv->engine.fifo.channels)
- chan = dev_priv->fifos[channel];
-
- if (chan) {
- spin_lock_irq(&chan->fence.lock);
- nouveau_fence_update(chan);
- spin_unlock_irq(&chan->fence.lock);
- }
-}
-
int
nouveau_fence_init(struct nouveau_channel *chan)
{
diff --git a/drivers/gpu/drm/nouveau/nv04_graph.c
b/drivers/gpu/drm/nouveau/nv04_graph.c
index 618355e..7226e15 100644
--- a/drivers/gpu/drm/nouveau/nv04_graph.c
+++ b/drivers/gpu/drm/nouveau/nv04_graph.c
@@ -528,7 +528,6 @@ nv04_graph_mthd_set_ref(struct nouveau_channel *chan, int
grclass,
int mthd, uint32_t data)
{
chan->fence.last_sequence_irq = data;
- nouveau_fence_handler(chan->dev, chan->id);
return 0;
}
--
1.7.0.1
Stephane Marchesin
2010-May-31 02:14 UTC
[Nouveau] [PATCH] drm/nouveau: reduce usage of fence spinlock to when absolutely necessary
The update of last_sequence_irq is still racy with this patch... So let me explain: - non-interrupt context starts reading last_sequence_irq - irq happens in the middle (because it is not an atomic read, it's possible, albeit it doesn't happen on x86 so just saying "it works on my x86" isn't enough) - last_sequence_irq is modified in interrupt context - second half of last_sequence_irq happens in normal context and results in a bogus value Stephane -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.freedesktop.org/archives/nouveau/attachments/20100530/39fadf63/attachment-0001.html>
Ben Skeggs
2010-May-31 03:45 UTC
[Nouveau] [PATCH] drm/nouveau: reduce usage of fence spinlock to when absolutely necessary
From: Ben Skeggs <bskeggs at redhat.com>
Signed-off-by: Ben Skeggs <bskeggs at redhat.com>
---
drivers/gpu/drm/nouveau/nouveau_channel.c | 2 -
drivers/gpu/drm/nouveau/nouveau_drv.h | 3 +-
drivers/gpu/drm/nouveau/nouveau_fence.c | 31 +++++-----------------------
drivers/gpu/drm/nouveau/nv04_graph.c | 3 +-
4 files changed, 8 insertions(+), 31 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_channel.c
b/drivers/gpu/drm/nouveau/nouveau_channel.c
index 1fc57ef..f9b2acf 100644
--- a/drivers/gpu/drm/nouveau/nouveau_channel.c
+++ b/drivers/gpu/drm/nouveau/nouveau_channel.c
@@ -257,9 +257,7 @@ nouveau_channel_free(struct nouveau_channel *chan)
nouveau_debugfs_channel_fini(chan);
/* Give outstanding push buffers a chance to complete */
- spin_lock_irqsave(&chan->fence.lock, flags);
nouveau_fence_update(chan);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
if (chan->fence.sequence != chan->fence.sequence_ack) {
struct nouveau_fence *fence = NULL;
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h
b/drivers/gpu/drm/nouveau/nouveau_drv.h
index 9eff502..75b1ea7 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -196,7 +196,7 @@ struct nouveau_channel {
struct list_head pending;
uint32_t sequence;
uint32_t sequence_ack;
- uint32_t last_sequence_irq;
+ atomic_t last_sequence_irq;
} fence;
/* DMA push buffer */
@@ -1141,7 +1141,6 @@ extern int nouveau_fence_wait(void *obj, void *arg, bool
lazy, bool intr);
extern int nouveau_fence_flush(void *obj, void *arg);
extern void nouveau_fence_unref(void **obj);
extern void *nouveau_fence_ref(void *obj);
-extern void nouveau_fence_handler(struct drm_device *dev, int channel);
/* nouveau_gem.c */
extern int nouveau_gem_new(struct drm_device *, struct nouveau_channel *,
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c
b/drivers/gpu/drm/nouveau/nouveau_fence.c
index faddf53..813d853 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -67,12 +67,13 @@ nouveau_fence_update(struct nouveau_channel *chan)
if (USE_REFCNT)
sequence = nvchan_rd32(chan, 0x48);
else
- sequence = chan->fence.last_sequence_irq;
+ sequence = atomic_read(&chan->fence.last_sequence_irq);
if (chan->fence.sequence_ack == sequence)
return;
chan->fence.sequence_ack = sequence;
+ spin_lock(&chan->fence.lock);
list_for_each_safe(entry, tmp, &chan->fence.pending) {
fence = list_entry(entry, struct nouveau_fence, entry);
@@ -84,6 +85,7 @@ nouveau_fence_update(struct nouveau_channel *chan)
if (sequence == chan->fence.sequence_ack)
break;
}
+ spin_unlock(&chan->fence.lock);
}
int
@@ -119,7 +121,6 @@ nouveau_fence_emit(struct nouveau_fence *fence)
{
struct drm_nouveau_private *dev_priv =
fence->channel->dev->dev_private;
struct nouveau_channel *chan = fence->channel;
- unsigned long flags;
int ret;
ret = RING_SPACE(chan, 2);
@@ -127,9 +128,7 @@ nouveau_fence_emit(struct nouveau_fence *fence)
return ret;
if (unlikely(chan->fence.sequence == chan->fence.sequence_ack - 1)) {
- spin_lock_irqsave(&chan->fence.lock, flags);
nouveau_fence_update(chan);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
BUG_ON(chan->fence.sequence = chan->fence.sequence_ack - 1);
@@ -138,9 +137,9 @@ nouveau_fence_emit(struct nouveau_fence *fence)
fence->sequence = ++chan->fence.sequence;
kref_get(&fence->refcount);
- spin_lock_irqsave(&chan->fence.lock, flags);
+ spin_lock(&chan->fence.lock);
list_add_tail(&fence->entry, &chan->fence.pending);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
+ spin_unlock(&chan->fence.lock);
BEGIN_RING(chan, NvSubSw, USE_REFCNT ? 0x0050 : 0x0150, 1);
OUT_RING(chan, fence->sequence);
@@ -173,14 +172,11 @@ nouveau_fence_signalled(void *sync_obj, void *sync_arg)
{
struct nouveau_fence *fence = nouveau_fence(sync_obj);
struct nouveau_channel *chan = fence->channel;
- unsigned long flags;
if (fence->signalled)
return true;
- spin_lock_irqsave(&chan->fence.lock, flags);
nouveau_fence_update(chan);
- spin_unlock_irqrestore(&chan->fence.lock, flags);
return fence->signalled;
}
@@ -221,27 +217,12 @@ nouveau_fence_flush(void *sync_obj, void *sync_arg)
return 0;
}
-void
-nouveau_fence_handler(struct drm_device *dev, int channel)
-{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
- struct nouveau_channel *chan = NULL;
-
- if (channel >= 0 && channel < dev_priv->engine.fifo.channels)
- chan = dev_priv->fifos[channel];
-
- if (chan) {
- spin_lock_irq(&chan->fence.lock);
- nouveau_fence_update(chan);
- spin_unlock_irq(&chan->fence.lock);
- }
-}
-
int
nouveau_fence_init(struct nouveau_channel *chan)
{
INIT_LIST_HEAD(&chan->fence.pending);
spin_lock_init(&chan->fence.lock);
+ atomic_set(&chan->fence.last_sequence_irq, 0);
return 0;
}
diff --git a/drivers/gpu/drm/nouveau/nv04_graph.c
b/drivers/gpu/drm/nouveau/nv04_graph.c
index 618355e..2097b3c 100644
--- a/drivers/gpu/drm/nouveau/nv04_graph.c
+++ b/drivers/gpu/drm/nouveau/nv04_graph.c
@@ -527,8 +527,7 @@ static int
nv04_graph_mthd_set_ref(struct nouveau_channel *chan, int grclass,
int mthd, uint32_t data)
{
- chan->fence.last_sequence_irq = data;
- nouveau_fence_handler(chan->dev, chan->id);
+ atomic_set(&chan->fence.last_sequence_irq, data);
return 0;
}
--
1.7.0.1
Possibly Parallel Threads
- [PATCH 1/3] Introduce nouveau_bo_wait for waiting on a BO with a GPU channel (v2)
- [PATCH 1/3] Introduce nouveau_bo_wait for waiting on a BO with a GPU channel
- [PATCH 1/2] gallium/nouveau: decouple nouveau_fence implementation from screen
- [PATCH] drm/nouveau: initialize chan->fence.lock before use
- [PATCH 1/2] drm/nv50: align size of buffer object to the right boundaries.