Skip to contentSkip to conversation

Fix DMA ops layering violations in vmwgfx v2

PATCH In review Thomas Hellstrom 2 participants 8 messages Jan 25, 2019 – Feb 5, 2019 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Fix DMA ops layering violations in vmwgfx v2

This is a patch set from Christoph Hellwig. Patch 2 has been updated by me to not remove checks for TTM coherent pool presence. Ideally, we should query TTM for this instead and I'll have a patch set intended for 5.1 so that we can do that. But I don't want to introduce cross-module API additions outside of a merge window.

Original patch set description:
vmwgfx has been doing some odd checks based on DMA ops which rely on deep DMA mapping layer internals, and I think the changes in Linux 4.21 finally broke most of these implicit assumptions.

The real fix is in patch 3, but I think the others are important to make it clear what is actually going on.

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#1 lore raw

PATCH 1/4 drm/vmwgfx: remove CONFIG_X86 ifdefs

From: Christoph Hellwig <[email protected]>

The driver depends on CONFIG_X86 so these are dead code.

---
 drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 6 ------
 1 file changed, 6 deletions(-)
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +0 -6
index 25afb1d594e3..69e325b2d954 100644
@@ -565,7 +565,6 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 		[vmw_dma_alloc_coherent] = "Using coherent TTM pages.",
 		[vmw_dma_map_populate] = "Keeping DMA mappings.",
 		[vmw_dma_map_bind] = "Giving up DMA mappings early."};
-#ifdef CONFIG_X86
 	const struct dma_map_ops *dma_ops = get_dma_ops(dev_priv->dev->dev);
 
 #ifdef CONFIG_INTEL_IOMMU
@@ -607,11 +606,6 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 	if (dev_priv->map_mode == vmw_dma_alloc_coherent)
 		return -EINVAL;
 #endif
-
-#else /* CONFIG_X86 */
-	dev_priv->map_mode = vmw_dma_map_populate;
-#endif /* CONFIG_X86 */
-
 	DRM_INFO("DMA map mode: %s\n", names[dev_priv->map_mode]);
 
 	return 0;
Signature · 6 lines
2.19.0.rc1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#2 lore raw

PATCH 2/4 drm/vmwgfx: remove CONFIG_INTEL_IOMMU ifdefs v2

In reply to Thomas Hellstrom

From: Christoph Hellwig <[email protected]>

intel_iommu_enabled is defined as always false for !CONFIG_INTEL_IOMMU, so remove the ifdefs around it.

---
v2: Retain the check for TTM dma page pool presence.

---
 drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 21 ++++-----------------
 1 file changed, 4 insertions(+), 17 deletions(-)
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +4 -17
index 69e325b2d954..b7777b5b4a81 100644
@@ -567,12 +567,10 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 		[vmw_dma_map_bind] = "Giving up DMA mappings early."};
 	const struct dma_map_ops *dma_ops = get_dma_ops(dev_priv->dev->dev);
 
-#ifdef CONFIG_INTEL_IOMMU
 	if (intel_iommu_enabled) {
 		dev_priv->map_mode = vmw_dma_map_populate;
 		goto out_fixup;
 	}
-#endif
 
 	if (!(vmw_force_iommu || vmw_force_coherent)) {
 		dev_priv->map_mode = vmw_dma_phys;
@@ -589,9 +587,7 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 		dev_priv->map_mode = vmw_dma_map_populate;
 #endif
 
-#ifdef CONFIG_INTEL_IOMMU
 out_fixup:
-#endif
 	if (dev_priv->map_mode == vmw_dma_map_populate &&
 	    vmw_restrict_iommu)
 		dev_priv->map_mode = vmw_dma_map_bind;
@@ -599,13 +595,11 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 	if (vmw_force_coherent)
 		dev_priv->map_mode = vmw_dma_alloc_coherent;
 
-#if !defined(CONFIG_SWIOTLB) && !defined(CONFIG_INTEL_IOMMU)
-	/*
-	 * No coherent page pool
-	 */
-	if (dev_priv->map_mode == vmw_dma_alloc_coherent)
+	/* No TTM coherent page pool? FIXME: Ask TTM instead! */
+	if (!(IS_ENABLED(CONFIG_SWIOTLB) || IS_ENABLED(CONFIG_INTEL_IOMMU)) &&
+	    (dev_priv->map_mode == vmw_dma_alloc_coherent))
 		return -EINVAL;
-#endif
+
 	DRM_INFO("DMA map mode: %s\n", names[dev_priv->map_mode]);
 
 	return 0;
@@ -619,7 +613,6 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
  * With 32-bit we can only handle 32 bit PFNs. Optionally set that
  * restriction also for 64-bit systems.
  */
-#ifdef CONFIG_INTEL_IOMMU
 static int vmw_dma_masks(struct vmw_private *dev_priv)
 {
 	struct drm_device *dev = dev_priv->dev;
@@ -631,12 +624,6 @@ static int vmw_dma_masks(struct vmw_private *dev_priv)
 	}
 	return 0;
 }
-#else
-static int vmw_dma_masks(struct vmw_private *dev_priv)
-{
-	return 0;
-}
-#endif
 
 static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 {
Signature · 6 lines
2.19.0.rc1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): [email protected]

#4 lore raw

Re: [PATCH 2/4] drm/vmwgfx: remove CONFIG_INTEL_IOMMU ifdefs v2

Quoted 17 lines · Christoph Hellwig

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c static int vmw_dma_select_mode(struct vmw_private *dev_priv) · line 602

Quoted 12 lines · Thomas Hellstrom
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +2 -5
-#if !defined(CONFIG_SWIOTLB) && !defined(CONFIG_INTEL_IOMMU)
-	/*
-	 * No coherent page pool
-	 */
-	if (dev_priv->map_mode == vmw_dma_alloc_coherent)
+	/* No TTM coherent page pool? FIXME: Ask TTM instead! */
+	if (!(IS_ENABLED(CONFIG_SWIOTLB) ||

IS_ENABLED(CONFIG_INTEL_IOMMU)) &&

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +2 -1
+	    (dev_priv->map_mode == vmw_dma_alloc_coherent))
 		return -EINVAL;
-#endif
+

I don't think this edited in change makes any sense. The swiotlb vs dma-direct versions of dma_alloc_coherent are the same, so this check seems very obsfucating.

So this part of code is identical in functionality to the previous version. It checks whether the TTM module has the coherent page pool enabled. (an identical test is present in TTM). What we *really* need to do here instead is to ask TTM whether it has enabled its coherent page pool instead of trying to mimic TTM's test, and I have a changeset under review for that. But as mentioned previously, I don't want to change the TTM interface outside of a merge window, so we either have to live with the above for 5.0 or keep the old defines. I'd prefer the former so I don't have to respin the patch series once more.

Thanks,
Thoams

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#5 lore raw

Re: [PATCH 2/4] drm/vmwgfx: remove CONFIG_INTEL_IOMMU ifdefs v2

Quoted 33 lines · Thomas Hellström
Quoted 11 lines · Christoph Hellwig
Quoted 10 lines · Thomas Hellstrom
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +2 -5
-#if !defined(CONFIG_SWIOTLB) && !defined(CONFIG_INTEL_IOMMU)
-	/*
-	 * No coherent page pool
-	 */
-	if (dev_priv->map_mode == vmw_dma_alloc_coherent)
+	/* No TTM coherent page pool? FIXME: Ask TTM instead! */
+	if (!(IS_ENABLED(CONFIG_SWIOTLB) ||

IS_ENABLED(CONFIG_INTEL_IOMMU)) &&

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +1 -0
+	    (dev_priv->map_mode == vmw_dma_alloc_coherent))
 		return -EINVAL;
Quoted 8 lines

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c static int vmw_dma_select_mode(struct vmw_private *dev_priv) · line 602

Quoted 2 lines
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +1 -1
-#endif
+

I don't think this edited in change makes any sense. The swiotlb vs
dma-direct versions of dma_alloc_coherent are the same, so this check
seems very obsfucating.

So this part of code is identical in functionality to the previous version. It checks whether the TTM module has the coherent page pool enabled. (an identical test is present in TTM). What we *really* need to do here instead is to ask TTM whether it has enabled its coherent page pool instead of trying to mimic TTM's test, and I have a changeset
under review for that. But as mentioned previously, I don't want to change the TTM interface outside of a merge window, so we either have to live with the above for 5.0 or keep the old defines. I'd prefer the
former so I don't have to respin the patch series once more.

Quoted 3 lines

Thanks,
Thoams

Hi, Christoph,

I need to get this merged this week. Could you please ack or ack removing this hunk + updating the following patches for merge errors?

If no response, I'll add a Cc: tag on the patch and a #v1 to your s-o- b.

Thanks,
Thomas

#6 lore raw

Re: [PATCH 2/4] drm/vmwgfx: remove CONFIG_INTEL_IOMMU ifdefs v2

Quoted 38 lines · Thomas Hellstrom
Quoted 33 lines · Thomas Hellström
Quoted 11 lines · Christoph Hellwig
Quoted 10 lines · Thomas Hellstrom
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +2 -5
-#if !defined(CONFIG_SWIOTLB) && !defined(CONFIG_INTEL_IOMMU)
-	/*
-	 * No coherent page pool
-	 */
-	if (dev_priv->map_mode == vmw_dma_alloc_coherent)
+	/* No TTM coherent page pool? FIXME: Ask TTM instead! */
+	if (!(IS_ENABLED(CONFIG_SWIOTLB) ||

IS_ENABLED(CONFIG_INTEL_IOMMU)) &&

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +1 -0
+	    (dev_priv->map_mode == vmw_dma_alloc_coherent))
 		return -EINVAL;
Quoted 8 lines

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c static int vmw_dma_select_mode(struct vmw_private *dev_priv) · line 602

Quoted 2 lines
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +1 -1
-#endif
+

I don't think this edited in change makes any sense. The swiotlb vs
dma-direct versions of dma_alloc_coherent are the same, so this check
seems very obsfucating.

So this part of code is identical in functionality to the previous version. It checks whether the TTM module has the coherent page pool enabled. (an identical test is present in TTM). What we *really* need to do here instead is to ask TTM whether it has enabled its coherent page pool instead of trying to mimic TTM's test, and I have a changeset
under review for that. But as mentioned previously, I don't want to change the TTM interface outside of a merge window, so we either have to live with the above for 5.0 or keep the old defines. I'd prefer the
former so I don't have to respin the patch series once more.

Quoted 3 lines

Thanks,
Thoams

Quoted 7 lines

Hi, Christoph,

I need to get this merged this week. Could you please ack or ack removing this hunk + updating the following patches for merge errors?

If no response, I'll add a Cc: tag on the patch and a #v1 to your s-o- b.

Please go ahead. I'll look into the fallout later.

#7 lore raw

PATCH 3/4 drm/vmwgfx: fix the check when to use dma_alloc_coherent

In reply to Thomas Hellstrom

From: Christoph Hellwig <[email protected]>

Since Linux 4.21 we merged the swiotlb ops into the DMA direct ops, so they would always have a the sync_single methods. But late in the cicle we also removed the direct ops entirely, so we'd see NULL DMA ops. Switch vmw_dma_select_mode to only detect swiotlb presence using swiotlb_nr_tbl() instead.

Fixes 55897af630 ("dma-direct: merge swiotlb_dma_ops into the dma_direct code")
Fixes 356da6d0cd ("dma-mapping: bypass indirect calls for dma-direct")
---
 drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +4 -7
index b7777b5b4a81..1456101e67a9 100644
@@ -565,7 +565,6 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 		[vmw_dma_alloc_coherent] = "Using coherent TTM pages.",
 		[vmw_dma_map_populate] = "Keeping DMA mappings.",
 		[vmw_dma_map_bind] = "Giving up DMA mappings early."};
-	const struct dma_map_ops *dma_ops = get_dma_ops(dev_priv->dev->dev);
 
 	if (intel_iommu_enabled) {
 		dev_priv->map_mode = vmw_dma_map_populate;
@@ -578,14 +577,12 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 		return 0;
 	}
 
-	dev_priv->map_mode = vmw_dma_map_populate;
-
-	if (dma_ops && dma_ops->sync_single_for_cpu)
-		dev_priv->map_mode = vmw_dma_alloc_coherent;
 #ifdef CONFIG_SWIOTLB
-	if (swiotlb_nr_tbl() == 0)
-		dev_priv->map_mode = vmw_dma_map_populate;
+	if (swiotlb_nr_tbl())
+		dev_priv->map_mode = vmw_dma_alloc_coherent;
+	else
 #endif
+		dev_priv->map_mode = vmw_dma_map_populate;
 
 out_fixup:
 	if (dev_priv->map_mode == vmw_dma_map_populate &&
Signature · 6 lines
2.19.0.rc1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#8 lore raw

PATCH 4/4 drm/vmwgfx: unwind spaghetti code in vmw_dma_select_mode

In reply to Thomas Hellstrom

From: Christoph Hellwig <[email protected]>

Just use a simple if/else chain to select the DMA mode.

---
 drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 25 ++++++-------------------
 1 file changed, 6 insertions(+), 19 deletions(-)
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +6 -19
index 1456101e67a9..3e2bcff34032 100644
@@ -566,39 +566,26 @@ static int vmw_dma_select_mode(struct vmw_private *dev_priv)
 		[vmw_dma_map_populate] = "Keeping DMA mappings.",
 		[vmw_dma_map_bind] = "Giving up DMA mappings early."};
 
-	if (intel_iommu_enabled) {
+	if (vmw_force_coherent)
+		dev_priv->map_mode = vmw_dma_alloc_coherent;
+	else if (intel_iommu_enabled)
 		dev_priv->map_mode = vmw_dma_map_populate;
-		goto out_fixup;
-	}
-
-	if (!(vmw_force_iommu || vmw_force_coherent)) {
+	else if (!vmw_force_iommu)
 		dev_priv->map_mode = vmw_dma_phys;
-		DRM_INFO("DMA map mode: %s\n", names[dev_priv->map_mode]);
-		return 0;
-	}
-
-#ifdef CONFIG_SWIOTLB
-	if (swiotlb_nr_tbl())
+	else if (IS_ENABLED(CONFIG_SWIOTLB) && swiotlb_nr_tbl())
 		dev_priv->map_mode = vmw_dma_alloc_coherent;
 	else
-#endif
 		dev_priv->map_mode = vmw_dma_map_populate;
 
-out_fixup:
-	if (dev_priv->map_mode == vmw_dma_map_populate &&
-	    vmw_restrict_iommu)
+	if (dev_priv->map_mode == vmw_dma_map_populate && vmw_restrict_iommu)
 		dev_priv->map_mode = vmw_dma_map_bind;
 
-	if (vmw_force_coherent)
-		dev_priv->map_mode = vmw_dma_alloc_coherent;
-
 	/* No TTM coherent page pool? FIXME: Ask TTM instead! */
         if (!(IS_ENABLED(CONFIG_SWIOTLB) || IS_ENABLED(CONFIG_INTEL_IOMMU)) &&
 	    (dev_priv->map_mode == vmw_dma_alloc_coherent))
 		return -EINVAL;
 
 	DRM_INFO("DMA map mode: %s\n", names[dev_priv->map_mode]);
-
 	return 0;
 }
 
Signature · 6 lines
2.19.0.rc1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel