drm/i915: Do not unmap object unless no other VMAs reference it
Subsystems (2)
PATCH 1/1 drm/i915: Do not unmap object unless no other VMAs reference it
From: Armin Reese <[email protected]>
--- drivers/gpu/drm/i915/i915_gem.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +3 -3
@@ -2922,8 +2922,6 @@ int i915_vma_unbind(struct i915_vma *vma) vma->unbind_vma (vma); -i915_gem_gtt_finish_object (obj); -list_del_init (&vma->mm_list);/* Avoid an unnecessary call to unbind on rebind. */ if (i915_is_ggtt (vma->vm))
@@ -2934,8 +2932,10 @@ int i915_vma_unbind(struct i915_vma *vma)/* Since the unbound list is global, only move to that list if * no more VMAs exist. */ -if (list_empty (&obj->vma_list)) +if (list_empty (&obj->vma_list)) { +i915_gem_gtt_finish_object (obj);list_move_tail (&obj->global_list, &dev_priv->mm.unbound_list); + }/* And finally now the object is completely decoupled from this vma, * we can drop its hold on the backing storage and allow it to be
Signature · 1 line
1.9.1
PATCH 1/1 drm/i915: Optimize the i915_gem_gtt_finish_object function
From: Armin Reese <[email protected]>
--- drivers/gpu/drm/i915/i915_gem_gtt.c | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +6 -7
@@ -1624,18 +1624,17 @@ static void ggtt_unbind_vma(struct i915_vma *vma)void i915_gem_gtt_finish_object (struct drm_i915_gem_object *obj) { -struct drm_device *dev = obj->base.dev; -struct drm_i915_private *dev_priv = dev->dev_private; -bool interruptible; - - interruptible =do_idling (dev_priv); +if (!obj->has_dma_mapping) { +struct drm_device *dev = obj->base.dev; +struct drm_i915_private *dev_priv = dev->dev_private; +bool interruptible =do_idling (dev_priv); -if (!obj->has_dma_mapping)dma_unmap_sg (&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,PCI_DMA_BIDIRECTIONAL ); -undo_idling (dev_priv, interruptible); +undo_idling (dev_priv, interruptible); + } } static void i915_gtt_color_adjust(struct drm_mm_node *node,
Signature · 1 line
1.9.1
Re: [PATCH] drm/i915: Optimize the i915_gem_gtt_finish_object function
Quoted 26 lines · On Fri, Jul 11, 2014 at 10:20:08AM -0700, [email protected] wrote:
From: Armin Reese <[email protected]>
Signed-off-by Armin Reese <[email protected]>--- drivers/gpu/drm/i915/i915_gem_gtt.c | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-)drivers/gpu/drm/i915/i915_gem_gtt.c +4
-6@@ -1624,18 +1624,17 @@ static void ggtt_unbind_vma(struct i915_vma *vma)void i915_gem_gtt_finish_object (struct drm_i915_gem_object *obj) { -struct drm_device *dev = obj->base.dev; -struct drm_i915_private *dev_priv = dev->dev_private; -bool interruptible; - - interruptible =do_idling (dev_priv); +if (!obj->has_dma_mapping) { +struct drm_device *dev = obj->base.dev; +struct drm_i915_private *dev_priv = dev->dev_private; +bool interruptible =do_idling (dev_priv); -if (!obj->has_dma_mapping)
drivers/gpu/drm/i915/i915_gem_gtt.c static void ggtt_unbind_vma(struct i915_vma *vma) · line 1640
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +2
-1dma_unmap_sg (&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,PCI_DMA_BIDIRECTIONAL ); -undo_idling (dev_priv, interruptible); +undo_idling (dev_priv, interruptible); + } } static void i915_gtt_color_adjust(struct drm_mm_node *node,
Note that this doesn't do much on platforms you care about.
Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
Re: [PATCH] drm/i915: Optimize the i915_gem_gtt_finish_object function
Quoted 31 lines · Ben Widawsky
Quoted 30 lines · On Fri, Jul 11, 2014 at 10:20:08AM -0700, [email protected] wrote:
From: Armin Reese <[email protected]>
Signed-off-by Armin Reese <[email protected]>--- drivers/gpu/drm/i915/i915_gem_gtt.c | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-)drivers/gpu/drm/i915/i915_gem_gtt.c +4
-6@@ -1624,18 +1624,17 @@ static void ggtt_unbind_vma(struct i915_vma *vma)void i915_gem_gtt_finish_object (struct drm_i915_gem_object *obj) { -struct drm_device *dev = obj->base.dev; -struct drm_i915_private *dev_priv = dev->dev_private; -bool interruptible; - - interruptible =do_idling (dev_priv); +if (!obj->has_dma_mapping) { +struct drm_device *dev = obj->base.dev; +struct drm_i915_private *dev_priv = dev->dev_private; +bool interruptible =do_idling (dev_priv); -if (!obj->has_dma_mapping)dma_unmap_sg (&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,PCI_DMA_BIDIRECTIONAL );
Quoted 8 lines
drivers/gpu/drm/i915/i915_gem_gtt.c static void ggtt_unbind_vma(struct i915_vma *vma) · line 1640
Quoted 6 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +2
-1-undo_idling (dev_priv, interruptible); +undo_idling (dev_priv, interruptible); + } } static void i915_gtt_color_adjust(struct drm_mm_node *node,Note that this doesn't do much on platforms you care about.
To elaborate: This w/a is _only_ used on ironlake when the iommu is enabled. Also the performance impact of it is so disastrous that any kind of optimization is pointless.
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
Quoted 22 lines · On Fri, Jul 11, 2014 at 10:20:07AM -0700, [email protected] wrote:
From: Armin Reese <[email protected]>
Signed-off-by Armin Reese <[email protected]>--- drivers/gpu/drm/i915/i915_gem.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/gpu/drm/i915/i915_gem.c +0
-2@@ -2922,8 +2922,6 @@ int i915_vma_unbind(struct i915_vma *vma) vma->unbind_vma (vma); -i915_gem_gtt_finish_object (obj); -list_del_init (&vma->mm_list);/* Avoid an unnecessary call to unbind on rebind. */ if (i915_is_ggtt (vma->vm))@@ -2934,8 +2932,10 @@ int i915_vma_unbind(struct i915_vma *vma)
drivers/gpu/drm/i915/i915_gem.c int i915_vma_unbind(struct i915_vma *vma) · line 2941
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +3
-1/* Since the unbound list is global, only move to that list if * no more VMAs exist. */ -if (list_empty (&obj->vma_list)) +if (list_empty (&obj->vma_list)) { +i915_gem_gtt_finish_object (obj);list_move_tail (&obj->global_list, &dev_priv->mm.unbound_list); + }/* And finally now the object is completely decoupled from this vma, * we can drop its hold on the backing storage and allow it to be
This patch is a fix, and therefore needs a commit message to explain the bug that you're trying to fix. I will help.
"
When using an IOMMU, GEM objects are mapped by their DMA address as the physical address is unknown. This depends on the underlying IOMMU driver to map and unmap the physical pages properly as defined in intel_iommu.c.
The current code will tell the IOMMU to unmap the GEM BO's pages on the destruction of the first VMA that "maps" that BO. This is clearly wrong as there may be other VMAs "mapping" that BO (using flink). The scanout is one such example.
The patch fixes this issue by only unmapping the DMA maps when there are no more VMAs mapping that object. This is equivalent to when an object is considered unbound as can be seen by the code. On the first VMA that again because bound, we will remap.
An alternate solution would be to move the dma mapping to object creation and destrubtion. I am not sure if this is considered an unfriendly thing to do.
Some notes to backporters:
The bug can never be hit without enabling the IOMMU. The existing code will also do the right thing when the object is shared via dmabuf. The failure should be demonstrable with flink. In cases when not using intel_iommu_strict it is likely (likely, as defined by: off the top of my head) on current workloads to *not* hit this bug since we often teardown all VMAs for an object shared across multiple VMs. We also finish access to that object before the first dma_unmapping. intel_iommu_strict with flinked buffers is likely to hit this issue. "
As a side note:
I was expecting this to be fixed as part of Daniels "lifetime tracking" plans. I had imagined that would have a handler run when refcount hits 0 that can do the dma_unmap. Therefore I assume everything in that list_empty() condition is a temporary hack.
This patch is:
Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
Quoted 58 lines · Ben Widawsky
Quoted 22 lines · On Fri, Jul 11, 2014 at 10:20:07AM -0700, [email protected] wrote:
From: Armin Reese <[email protected]>
Signed-off-by Armin Reese <[email protected]>--- drivers/gpu/drm/i915/i915_gem.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/gpu/drm/i915/i915_gem.c +0
-2@@ -2922,8 +2922,6 @@ int i915_vma_unbind(struct i915_vma *vma) vma->unbind_vma (vma); -i915_gem_gtt_finish_object (obj); -list_del_init (&vma->mm_list);/* Avoid an unnecessary call to unbind on rebind. */ if (i915_is_ggtt (vma->vm))@@ -2934,8 +2932,10 @@ int i915_vma_unbind(struct i915_vma *vma)drivers/gpu/drm/i915/i915_gem.c int i915_vma_unbind(struct i915_vma *vma) · line 2941
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +3
-1/* Since the unbound list is global, only move to that list if * no more VMAs exist. */ -if (list_empty (&obj->vma_list)) +if (list_empty (&obj->vma_list)) { +i915_gem_gtt_finish_object (obj);list_move_tail (&obj->global_list, &dev_priv->mm.unbound_list); + }/* And finally now the object is completely decoupled from this vma, * we can drop its hold on the backing storage and allow it to be This patch is a fix, and therefore needs a commit message to explain the bug that you're trying to fix. I will help.
"
When using an IOMMU, GEM objects are mapped by their DMA address as the physical address is unknown. This depends on the underlying IOMMU driver to map and unmap the physical pages properly as defined inintel_iommu.c.The current code will tell the IOMMU to unmap the GEM BO's pages on the destruction of the first VMA that "maps" that BO. This is clearly wrong as there may be other VMAs "mapping" that BO (using flink). The scanout is one such example.
The patch fixes this issue by only unmapping the DMA maps when there are no more VMAs mapping that object. This is equivalent to when an object is considered unbound as can be seen by the code. On the first VMA that again because bound, we will remap.
An alternate solution would be to move the dma mapping to object creation and destrubtion. I am not sure if this is considered an unfriendly thing to do.
Quoted 10 lines
Some notes to backporters:
The bug can never be hit without enabling the IOMMU. The existing code will also do the right thing when the object is shared via dmabuf. The failure should be demonstrable with flink. In cases when not usingintel_iommu_strictit is likely (likely, as defined by: off the top of my head) on current workloads to *not* hit this bug since we often teardown all VMAs for an object shared across multiple VMs. We also finish access to that object before the firstdma_unmapping.intel_iommu_strictwith flinked buffers is likely to hit this issue. "
Crap. I left out the important part. Note to backporters trying to backport full PPGTT, which is disabled by default on all kernels to date.
Quoted 1 line
Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
Quoted 31 lines · Ben Widawsky
Quoted 30 lines · On Fri, Jul 11, 2014 at 10:20:07AM -0700, [email protected] wrote:
From: Armin Reese <[email protected]>
Signed-off-by Armin Reese <[email protected]>--- drivers/gpu/drm/i915/i915_gem.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/gpu/drm/i915/i915_gem.c +3
-3@@ -2922,8 +2922,6 @@ int i915_vma_unbind(struct i915_vma *vma) vma->unbind_vma (vma); -i915_gem_gtt_finish_object (obj); -list_del_init (&vma->mm_list);/* Avoid an unnecessary call to unbind on rebind. */ if (i915_is_ggtt (vma->vm))@@ -2934,8 +2932,10 @@ int i915_vma_unbind(struct i915_vma *vma)/* Since the unbound list is global, only move to that list if * no more VMAs exist. */ -if (list_empty (&obj->vma_list)) +if (list_empty (&obj->vma_list)) { +i915_gem_gtt_finish_object (obj);list_move_tail (&obj->global_list, &dev_priv->mm.unbound_list); + }
Quoted 7 lines
drivers/gpu/drm/i915/i915_gem.c int i915_vma_unbind(struct i915_vma *vma) · line 2941
Quoted 2 lines
drivers/gpu/drm/i915/i915_gem.c +0
-0/* And finally now the object is completely decoupled from this vma, * we can drop its hold on the backing storage and allow it to be This patch is a fix, and therefore needs a commit message to explain the bug that you're trying to fix. I will help.
Indeed, and thanks a lot for providing the commit message. I've added "for full ppgtt" as requested and merged the patch to dinq.
Thanks, Daniel
Quoted 47 lines
"
When using an IOMMU, GEM objects are mapped by their DMA address as the physical address is unknown. This depends on the underlying IOMMU driver to map and unmap the physical pages properly as defined inintel_iommu.c.The current code will tell the IOMMU to unmap the GEM BO's pages on the destruction of the first VMA that "maps" that BO. This is clearly wrong as there may be other VMAs "mapping" that BO (using flink). The scanout is one such example.
The patch fixes this issue by only unmapping the DMA maps when there are no more VMAs mapping that object. This is equivalent to when an object is considered unbound as can be seen by the code. On the first VMA that again because bound, we will remap.
An alternate solution would be to move the dma mapping to object creation and destrubtion. I am not sure if this is considered an unfriendly thing to do.
Some notes to backporters:
The bug can never be hit without enabling the IOMMU. The existing code will also do the right thing when the object is shared via dmabuf. The failure should be demonstrable with flink. In cases when not usingintel_iommu_strictit is likely (likely, as defined by: off the top of my head) on current workloads to *not* hit this bug since we often teardown all VMAs for an object shared across multiple VMs. We also finish access to that object before the firstdma_unmapping.intel_iommu_strictwith flinked buffers is likely to hit this issue. "As a side note:
I was expecting this to be fixed as part of Daniels "lifetime tracking" plans. I had imagined that would have a handler run when refcount hits 0 that can do thedma_unmap. Therefore I assume everything in thatlist_empty()condition is a temporary hack.This patch is:
Reviewed-by Ben Widawsky <[email protected]>Signature · 5 lines
Ben Widawsky, Intel Open Source Technology Center _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch