iommu: Avoid DMA ops domain refcount contention

John raised the issue[1] that we have some unnecessary refcount contention in the DMA ops path which shows scalability problems now that we have more real high-performance hardware using iommu-dma. The x86 IOMMU drivers are sidestepping this by stashing domain references in archdata, but since that's not very nice for architecture-agnostic code, I think it's time to look at a generic API-level solution.

These are a couple of quick patches based on the idea I had back when first implementing iommu-dma, but didn't have any way to justify at the time. However, the reports of 10-25% better networking performance on v1 suggest that it's very worthwhile (and far more significant than I ever would have guessed).

As far as merging goes, I don't mind at all whether this goes via IOMMU, or via dma-mapping provided Joerg's happy to ack it.

Hi Robin,

Quoted 7 lines

John raised the issue[1] that we have some unnecessary refcount contention in the DMA ops path which shows scalability problems now that we have more real high-performance hardware using iommu-dma. The x86 IOMMU drivers are sidestepping this by stashing domain references in archdata, but since that's not very nice for architecture-agnostic code, I think it's time to look at a generic API-level solution.

Quoted 8 lines

These are a couple of quick patches based on the idea I had back when first implementing iommu-dma, but didn't have any way to justify at the time. However, the reports of 10-25% better networking performance on v1 suggest that it's very worthwhile (and far more significant than I ever would have guessed).

As far as merging goes, I don't mind at all whether this goes via IOMMU, or via dma-mapping provided Joerg's happy to ack it.

I think it makes most sense for Joerg to take this series via his tree.

Anyway, I've been running them on my TX2 box and things are happy enough, so:

Tested-by Will Deacon <[email protected]>

Will

.

Quoted 2 lines

As far as merging goes, I don't mind at all whether this goes via IOMMU, or via dma-mapping provided Joerg's happy to ack it.

I think it makes most sense for Joerg to take this series via his tree.

Not sure how fast Tom needs the common dma-iommu code for the x86 AMD iommu conversion.

Quoted 4 lines · Will Deacon

Quoted 2 lines

As far as merging goes, I don't mind at all whether this goes via IOMMU, or via dma-mapping provided Joerg's happy to ack it.

I think it makes most sense for Joerg to take this series via his tree.

FYI, I have WIP patches to move the arm dma-iommu wrappers to common code:

http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/dma-maybe-coherent

which will require some synchronization of the involved trees. In the end it is iommu code, so the actual iommu patches should probably go into the iommu tree after all, but it might have to pull in the dma-mapping branch for that. Or we just punt it until the next merge window. Not sure how fast Tom needs the common dma-iommu code for the x86 AMD iommu conversion.

Quoted 2 lines · Will Deacon

Hi Robin,

Hi Robin,

I just spoke with Dongdong and we will test this version also so that we may provide a "Tested-by" tag.

Thanks,
John

Quoted 23 lines · Robin Murphy

Quoted 11 lines

John raised the issue[1] that we have some unnecessary refcount contention
in the DMA ops path which shows scalability problems now that we have more
real high-performance hardware using iommu-dma. The x86 IOMMU drivers are
sidestepping this by stashing domain references in archdata, but since that's not very nice for architecture-agnostic code, I think it's time to
look at a generic API-level solution.

Quoted 8 lines

These are a couple of quick patches based on the idea I had back when first implementing iommu-dma, but didn't have any way to justify at the time. However, the reports of 10-25% better networking performance on v1 suggest that it's very worthwhile (and far more significant than I ever would have guessed).

As far as merging goes, I don't mind at all whether this goes via IOMMU, or via dma-mapping provided Joerg's happy to ack it.

I think it makes most sense for Joerg to take this series via his tree.

Quoted 9 lines

Anyway, I've been running them on my TX2 box and things are happy enough, so:

Tested-by Will Deacon <[email protected]>

Will

.

_______________________________________________
iommu mailing list
iommu at lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu

.

Robin Murphy (3):
iommu: Add fast hook for getting DMA domains iommu/dma: Use fast DMA domain lookup arm64/dma-mapping: Mildly optimise non-coherent IOMMU ops

Most parts of iommu-dma already assume they are operating on a default domain set up by iommu_dma_init_domain(), and can be converted straight over to avoid the refcounting bottleneck. MSI page mappings may be in an unmanaged domain with an explicit MSI-only cookie, so retain the non-specific lookup, but that's OK since they're far from a contended fast path either way.

Signed-off-by Robin Murphy <[email protected]>

---
  drivers/iommu/dma-iommu.c | 23 ++++++++++++-----------
  1 file changed, 12 insertions(+), 11 deletions(-)

drivers/iommu/dma-iommu.c +11 -10

index 511ff9a1d6d9..320f9ea82f3f 100644
@@ -491,7 +491,7 @@ static struct page **__iommu_dma_alloc_pages(unsigned int count,
  void iommu_dma_free(struct device *dev, struct page **pages, size_t size,
  		dma_addr_t *handle)
  {
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), *handle, size);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), *handle, size);
  	__iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
  	*handle = IOMMU_MAPPING_ERROR;
  }

@@ -518,7 +518,7 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
  		unsigned long attrs, int prot, dma_addr_t *handle,
  		void (*flush_page)(struct device *, const void *, phys_addr_t))
  {
-	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
+	struct iommu_domain *domain = iommu_get_dma_domain(dev);
  	struct iommu_dma_cookie *cookie = domain->iova_cookie;
  	struct iova_domain *iovad = &cookie->iovad;
  	struct page **pages;

@@ -606,9 +606,8 @@ int iommu_dma_mmap(struct page **pages, size_t size, struct vm_area_struct *vma)
  }
 
  static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
-		size_t size, int prot)
+		size_t size, int prot, struct iommu_domain *domain)
  {
-	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
  	struct iommu_dma_cookie *cookie = domain->iova_cookie;
  	size_t iova_off = 0;
  	dma_addr_t iova;

@@ -632,13 +631,14 @@ static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
  dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
  		unsigned long offset, size_t size, int prot)
  {
-	return __iommu_dma_map(dev, page_to_phys(page) + offset, size, prot);
+	return __iommu_dma_map(dev, page_to_phys(page) + offset, size, prot,
+			iommu_get_dma_domain(dev));
  }
 
  void iommu_dma_unmap_page(struct device *dev, dma_addr_t handle, size_t size,
  		enum dma_data_direction dir, unsigned long attrs)
  {
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), handle, size);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), handle, size);
  }
 
  /*

@@ -726,7 +726,7 @@ static void __invalidate_sg(struct scatterlist *sg, int nents)
  int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
  		int nents, int prot)
  {
-	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
+	struct iommu_domain *domain = iommu_get_dma_domain(dev);
  	struct iommu_dma_cookie *cookie = domain->iova_cookie;
  	struct iova_domain *iovad = &cookie->iovad;
  	struct scatterlist *s, *prev = NULL;

@@ -811,20 +811,21 @@ void iommu_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
  		sg = tmp;
  	}
  	end = sg_dma_address(sg) + sg_dma_len(sg);
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), start, end - start);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), start, end - start);
  }
 
  dma_addr_t iommu_dma_map_resource(struct device *dev, phys_addr_t phys,
  		size_t size, enum dma_data_direction dir, unsigned long attrs)
  {
  	return __iommu_dma_map(dev, phys, size,
-			dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO);
+			dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO,
+			iommu_get_dma_domain(dev));
  }
 
  void iommu_dma_unmap_resource(struct device *dev, dma_addr_t handle,
  		size_t size, enum dma_data_direction dir, unsigned long attrs)
  {
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), handle, size);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), handle, size);
  }
 
  int iommu_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)

Hi All,

Quoted 8 lines · Robin Murphy

Most parts of iommu-dma already assume they are operating on a default domain set up by iommu_dma_init_domain(), and can be converted straight over to avoid the refcounting bottleneck. MSI page mappings may be in an unmanaged domain with an explicit MSI-only cookie, so retain the non-specific lookup, but that's OK since they're far from a contended fast path either way.

Signed-off-by Robin Murphy <[email protected]>

This breaks Exynos DRM driver with Exynos IOMMU on ARM64 (Exynos5433). Exynos DRM creates it's own domain, attach all devices which performs DMA access (typically CRTC devices) to it and uses standard DMA-mapping calls to allocate/map buffers. This way it can use the same one DMA address for each buffer regardless of the CRTC/display/processing device.

This no longer works with this patch. The simplest solution to fix this is add API to change default_domain to the one allocated by the Exynos DRM driver.

Quoted 89 lines

---
   drivers/iommu/dma-iommu.c | 23 ++++++++++++-----------
   1 file changed, 12 insertions(+), 11 deletions(-)

drivers/iommu/dma-iommu.c +11 -10

index 511ff9a1d6d9..320f9ea82f3f 100644
@@ -491,7 +491,7 @@ static struct page **__iommu_dma_alloc_pages(unsigned int count,
   void iommu_dma_free(struct device *dev, struct page **pages, size_t size,
   		dma_addr_t *handle)
   {
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), *handle, size);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), *handle, size);
   	__iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
   	*handle = IOMMU_MAPPING_ERROR;
   }

@@ -518,7 +518,7 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
   		unsigned long attrs, int prot, dma_addr_t *handle,
   		void (*flush_page)(struct device *, const void *, phys_addr_t))
   {
-	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
+	struct iommu_domain *domain = iommu_get_dma_domain(dev);
   	struct iommu_dma_cookie *cookie = domain->iova_cookie;
   	struct iova_domain *iovad = &cookie->iovad;
   	struct page **pages;

@@ -606,9 +606,8 @@ int iommu_dma_mmap(struct page **pages, size_t size, struct vm_area_struct *vma)
   }
 
   static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
-		size_t size, int prot)
+		size_t size, int prot, struct iommu_domain *domain)
   {
-	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
   	struct iommu_dma_cookie *cookie = domain->iova_cookie;
   	size_t iova_off = 0;
   	dma_addr_t iova;

@@ -632,13 +631,14 @@ static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
   dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
   		unsigned long offset, size_t size, int prot)
   {
-	return __iommu_dma_map(dev, page_to_phys(page) + offset, size, prot);
+	return __iommu_dma_map(dev, page_to_phys(page) + offset, size, prot,
+			iommu_get_dma_domain(dev));
   }
 
   void iommu_dma_unmap_page(struct device *dev, dma_addr_t handle, size_t size,
   		enum dma_data_direction dir, unsigned long attrs)
   {
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), handle, size);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), handle, size);
   }
 
   /*

@@ -726,7 +726,7 @@ static void __invalidate_sg(struct scatterlist *sg, int nents)
   int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
   		int nents, int prot)
   {
-	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
+	struct iommu_domain *domain = iommu_get_dma_domain(dev);
   	struct iommu_dma_cookie *cookie = domain->iova_cookie;
   	struct iova_domain *iovad = &cookie->iovad;
   	struct scatterlist *s, *prev = NULL;

@@ -811,20 +811,21 @@ void iommu_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
   		sg = tmp;
   	}
   	end = sg_dma_address(sg) + sg_dma_len(sg);
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), start, end - start);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), start, end - start);
   }
 
   dma_addr_t iommu_dma_map_resource(struct device *dev, phys_addr_t phys,
   		size_t size, enum dma_data_direction dir, unsigned long attrs)
   {
   	return __iommu_dma_map(dev, phys, size,
-			dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO);
+			dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO,
+			iommu_get_dma_domain(dev));
   }
 
   void iommu_dma_unmap_resource(struct device *dev, dma_addr_t handle,
   		size_t size, enum dma_data_direction dir, unsigned long attrs)
   {
-	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), handle, size);
+	__iommu_dma_unmap(iommu_get_dma_domain(dev), handle, size);
   }
 
   int iommu_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)

drivers/iommu/dma-iommu.c static struct iommu_dma_msi_page *iommu_dma_get_msi_page(struct device *dev, · line 857

Quoted 8 lines

drivers/iommu/dma-iommu.c +1 -1

@@ -850,7 +851,7 @@ static struct iommu_dma_msi_page *iommu_dma_get_msi_page(struct device *dev,
   	if (!msi_page)
   		return NULL;
 
-	iova = __iommu_dma_map(dev, msi_addr, size, prot);
+	iova = __iommu_dma_map(dev, msi_addr, size, prot, domain);
   	if (iommu_dma_mapping_error(dev, iova))
   		goto out_free_page;
 

Best regards

Quoted 17 lines · Marek Szyprowski

Quoted 8 lines · Robin Murphy

Most parts of iommu-dma already assume they are operating on a default domain set up by iommu_dma_init_domain(), and can be converted straight over to avoid the refcounting bottleneck. MSI page mappings may be in an unmanaged domain with an explicit MSI-only cookie, so retain the non-specific lookup, but that's OK since they're far from a contended fast path either way.

Signed-off-by Robin Murphy <[email protected]>

This breaks Exynos DRM driver with Exynos IOMMU on ARM64 (Exynos5433). Exynos DRM creates it's own domain, attach all devices which performs DMA access (typically CRTC devices) to it and uses standard DMA-mapping calls to allocate/map buffers. This way it can use the same one DMA address for each buffer regardless of the CRTC/display/processing device.

Quoted 3 lines

This no longer works with this patch. The simplest solution to fix this is add API to change default_domain to the one allocated by the Exynos DRM driver.

Ugh, I hadn't realised there were any drivers trying to fake up their own default domains - that's always going to be fragile. In fact, one of my proposals for un-breaking Tegra and other DRM drivers involves preventing iommu-dma from trying to operate on non-IOMMU_DOMAIN_DMA domains at all, which would stop that from working permanently.

Can Exynos not put all the DRM components into a single group, or at least just pick one of the real default domains to attach everyone to instead of allocating a fake one?

...

Hi Robin,

Quoted 22 lines · Robin Murphy

Quoted 17 lines · Marek Szyprowski

Quoted 8 lines · Robin Murphy

Most parts of iommu-dma already assume they are operating on a default domain set up by iommu_dma_init_domain(), and can be converted straight over to avoid the refcounting bottleneck. MSI page mappings may be in an unmanaged domain with an explicit MSI-only cookie, so retain the non-specific lookup, but that's OK since they're far from a contended fast path either way.

Signed-off-by Robin Murphy <[email protected]>

This breaks Exynos DRM driver with Exynos IOMMU on ARM64 (Exynos5433). Exynos DRM creates it's own domain, attach all devices which performs DMA access (typically CRTC devices) to it and uses standard DMA-mapping calls to allocate/map buffers. This way it can use the same one DMA address for each buffer regardless of the CRTC/display/processing device.

Quoted 3 lines

This no longer works with this patch. The simplest solution to fix this is add API to change default_domain to the one allocated by the Exynos DRM driver.

Quoted 5 lines

Ugh, I hadn't realised there were any drivers trying to fake up their own default domains - that's always going to be fragile. In fact, one of my proposals for un-breaking Tegra and other DRM drivers involves preventing iommu-dma from trying to operate on non-IOMMU_DOMAIN_DMA domains at all, which would stop that from working permanently.

IMHO there should be a way to let drivers like DRM to use DMA-mapping infrastructure on their own iommu domain. Better provide such API to avoid hacking in the DRM drivers to get this done without help from IOMMU core.

Quoted 3 lines

Can Exynos not put all the DRM components into a single group, or at least just pick one of the real default domains to attach everyone to instead of allocating a fake one?

Exynos DRM components are not in the single group. Currently iommu-groups have no meaning on Exynos5433 and Exynos IOMMU simply allocate one group per each iommu-master device in the system.

Exynos DRM already selects one of its component devices as the 'main DMA device'. It is being used for managing buffers if no IOMMU is available, so it can also use its iommu domain instead of allocating a fake one. I've checked and it fixes Exynos DRM now. The question is how to merge this fix to keep bisectability.

From: Marek Szyprowski <[email protected]>
Date: Fri, 28 Sep 2018 16:17:27 +0200
Subject: [PATCH] drm/exynos: Use selected dma_dev default iommu domain instead
?of a fake one

Instead of allocating a fake IOMMU domain for all Exynos DRM components, simply reuse the default IOMMU domain of the already selected DMA device. This allows some design changes in IOMMU framework without breaking IOMMU support in Exynos DRM.

Signed-off-by Marek Szyprowski <[email protected]>

---
  ?drivers/gpu/drm/exynos/exynos_drm_iommu.h | 34 ++++-------------------

?1 file changed, 6 insertions(+), 28 deletions(-)

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -7

index 87f6b5672e11..51d3b7dcd529 100644
@@ -55,37 +55,12 @@ static inline void __exynos_iommu_detach(struct
 exynos_drm_private *priv,
  ?static inline int __exynos_iommu_create_mapping(struct
 exynos_drm_private *priv,
  ???? ??? ??? ??? ??? unsigned long start, unsigned long size)
  ?{
-??? struct iommu_domain *domain;
-??? int ret;
-
-??? domain = iommu_domain_alloc(priv->dma_dev->bus);
-??? if (!domain)
-??? ??? return -ENOMEM;
-

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -10

  ???? return 0;
-
-put_cookie:
-??? iommu_put_dma_cookie(domain);
-free_domain:
-??? iommu_domain_free(domain);
-??? return ret;
  ?}
 
  ?static inline void __exynos_iommu_release_mapping(struct
 exynos_drm_private *priv)
  ?{
-??? struct iommu_domain *domain = priv->mapping;
-
-??? iommu_put_dma_cookie(domain);
-??? iommu_domain_free(domain);

  ???? priv->mapping = NULL;
  ?}

drivers/gpu/drm/exynos/exynos_drm_iommu.h +3 -1

@@ -94,7 +69,9 @@ static inline int __exynos_iommu_attach(struct
 exynos_drm_private *priv,
  ?{
  ???? struct iommu_domain *domain = priv->mapping;
 
-??? return iommu_attach_device(domain, dev);
+??? if (dev != priv->dma_dev)
+??? ??? return iommu_attach_device(domain, dev);
+??? return 0;
  ?}
 

?static inline void __exynos_iommu_detach(struct exynos_drm_private *priv,

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -0

  ?}
  ?#else

?#error Unsupported architecture and IOMMU/DMA-mapping glue code

Quoted 1 line

...

Best regards

The tricky part is how to reconcile that with those other drivers which want to do explicit IOMMU management with their own domain but still use the DMA API for coherency of the underlying memory. I do have a couple of half-formed ideas, but they're not quite there yet.

Quoted 4 lines · Robin Murphy

The tricky part is how to reconcile that with those other drivers which want to do explicit IOMMU management with their own domain but still use the DMA API for coherency of the underlying memory. I do have a couple of half-formed ideas, but they're not quite there yet.

It think the only sensible answer is that they can't, and we'll need coherency API in the iommu API (or augmenting it). Which might be a real possibility now that I'm almost done lifting the coherency management to arch hooks instead of dma op methods.

Instead of allocating a fake IOMMU domain for all Exynos DRM components, simply reuse the default IOMMU domain of the already selected DMA device. This allows some design changes in IOMMU framework without breaking IOMMU support in Exynos DRM.

Signed-off-by Marek Szyprowski <[email protected]>

---
Inki:
If possible, please consider this patch as a fix for v4.19-rc, this will help doing a rework in IOMMU and DMA-IOMMU frameworks in v4.20 without breaking Exynos DRM. It worked for current IOMMU code, but such usage is considered as a hack.

---
  drivers/gpu/drm/exynos/exynos_drm_iommu.h | 34 ++++-------------------
  1 file changed, 6 insertions(+), 28 deletions(-)

drivers/gpu/drm/exynos/exynos_drm_iommu.h +6 -28

index 87f6b5672e11..797d9ee5f15a 100644
@@ -55,37 +55,12 @@ static inline void __exynos_iommu_detach(struct exynos_drm_private *priv,
  static inline int __exynos_iommu_create_mapping(struct exynos_drm_private *priv,
  					unsigned long start, unsigned long size)
  {
-	struct iommu_domain *domain;
-	int ret;
-
-	domain = iommu_domain_alloc(priv->dma_dev->bus);
-	if (!domain)
-		return -ENOMEM;
-
-	ret = iommu_get_dma_cookie(domain);
-	if (ret)
-		goto free_domain;
-
-	ret = iommu_dma_init_domain(domain, start, size, NULL);
-	if (ret)
-		goto put_cookie;
-
-	priv->mapping = domain;
+	priv->mapping = iommu_get_domain_for_dev(priv->dma_dev);
  	return 0;
-
-put_cookie:
-	iommu_put_dma_cookie(domain);
-free_domain:
-	iommu_domain_free(domain);
-	return ret;
  }
 
  static inline void __exynos_iommu_release_mapping(struct exynos_drm_private *priv)
  {
-	struct iommu_domain *domain = priv->mapping;
-
-	iommu_put_dma_cookie(domain);
-	iommu_domain_free(domain);
  	priv->mapping = NULL;
  }
 

@@ -94,7 +69,9 @@ static inline int __exynos_iommu_attach(struct exynos_drm_private *priv,
  {
  	struct iommu_domain *domain = priv->mapping;
 
-	return iommu_attach_device(domain, dev);
+	if (dev != priv->dma_dev)
+		return iommu_attach_device(domain, dev);
+	return 0;
  }
 
  static inline void __exynos_iommu_detach(struct exynos_drm_private *priv,

@@ -102,7 +79,8 @@ static inline void __exynos_iommu_detach(struct exynos_drm_private *priv,
  {
  	struct iommu_domain *domain = priv->mapping;
 
-	iommu_detach_device(domain, dev);
+	if (dev != priv->dma_dev)
+		iommu_detach_device(domain, dev);
  }
  #else
  #error Unsupported architecture and IOMMU/DMA-mapping glue code

Quoted 33 lines · Marek Szyprowski

Quoted 26 lines · Robin Murphy

Quoted 20 lines · Marek Szyprowski

Quoted 10 lines · Robin Murphy

Most parts of iommu-dma already assume they are operating on a default
domain set up by iommu_dma_init_domain(), and can be converted straight
over to avoid the refcounting bottleneck. MSI page mappings may be in an unmanaged domain with an explicit MSI-only cookie, so retain the non-specific lookup, but that's OK since they're far from a contended fast path either way.

Signed-off-by Robin Murphy <[email protected]>

This breaks Exynos DRM driver with Exynos IOMMU on ARM64 (Exynos5433). Exynos DRM creates it's own domain, attach all devices which performs DMA access (typically CRTC devices) to it and uses standard DMA-mapping
calls to allocate/map buffers. This way it can use the same one DMA address for each buffer regardless of the CRTC/display/processing device.

Quoted 4 lines

This no longer works with this patch. The simplest solution to fix this
is add API to change default_domain to the one allocated by the Exynos DRM driver.

Quoted 5 lines

Ugh, I hadn't realised there were any drivers trying to fake up their own default domains - that's always going to be fragile. In fact, one of my proposals for un-breaking Tegra and other DRM drivers involves preventing iommu-dma from trying to operate on non-IOMMU_DOMAIN_DMA domains at all, which would stop that from working permanently.

Quoted 4 lines

IMHO there should be a way to let drivers like DRM to use DMA-mapping infrastructure on their own iommu domain. Better provide such API to avoid hacking in the DRM drivers to get this done without help from IOMMU core.

The tricky part is how to reconcile that with those other drivers which want to do explicit IOMMU management with their own domain but still use the DMA API for coherency of the underlying memory. I do have a couple of half-formed ideas, but they're not quite there yet.

Quoted 8 lines

Quoted 4 lines

Can Exynos not put all the DRM components into a single group, or at least just pick one of the real default domains to attach everyone to instead of allocating a fake one?

Exynos DRM components are not in the single group. Currently iommu-groups have no meaning on Exynos5433 and Exynos IOMMU simply allocate one group per each iommu-master device in the system.

Yeah, a group spanning multiple IOMMU devices would have been a bit of a hack for your topology, but still *technically* possible ;)

Quoted 43 lines

Exynos DRM already selects one of its component devices as the 'main DMA device'. It is being used for managing buffers if no IOMMU is available, so it can also use its iommu domain instead of allocating a fake one. I've checked and it fixes Exynos DRM now. The question is how to merge this fix to keep bisectability.

From: Marek Szyprowski <[email protected]>
Date: Fri, 28 Sep 2018 16:17:27 +0200
Subject: [PATCH] drm/exynos: Use selected dma_dev default iommu domain instead
? ?of a fake one

Instead of allocating a fake IOMMU domain for all Exynos DRM components, simply reuse the default IOMMU domain of the already selected DMA device.
This allows some design changes in IOMMU framework without breaking IOMMU
support in Exynos DRM.

Signed-off-by Marek Szyprowski <[email protected]>

---
? ?drivers/gpu/drm/exynos/exynos_drm_iommu.h | 34
++++-------------------
? ?1 file changed, 6 insertions(+), 28 deletions(-)

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -0

index 87f6b5672e11..51d3b7dcd529 100644
@@ -55,37 +55,12 @@ static inline void __exynos_iommu_detach(struct
 exynos_drm_private *priv,

? ?static inline int __exynos_iommu_create_mapping(struct exynos_drm_private *priv,
? ???? ??? ??? ??? ??? unsigned long start, unsigned long size) ? ?{
-??? struct iommu_domain *domain;
-??? int ret;

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -5

-
-??? domain = iommu_domain_alloc(priv->dma_dev->bus);
-??? if (!domain)
-??? ??? return -ENOMEM;
-

drivers/gpu/drm/exynos/exynos_drm_iommu.h

Quoted 10 lines

drivers/gpu/drm/exynos/exynos_drm_iommu.h +1 -9

-??? ret = iommu_get_dma_cookie(domain);
-??? if (ret)
-??? ??? goto free_domain;
-
-??? ret = iommu_dma_init_domain(domain, start, size, NULL);
-??? if (ret)
-??? ??? goto put_cookie;
-
-??? priv->mapping = domain;
+??? priv->mapping = iommu_get_dma_domain(priv->dma_dev);

iommu_get_domain_for_dev(), please - this isn't a performance-critical path inside a DMA API implementation, plus without an actual dependency on this series maybe there's a chance of sneaking it into 4.19?

(you could always still check domain->type == IOMMU_DOMAIN_DMA if you want to be really really paranoid.)

Quoted 37 lines

? ???? return 0;

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -6

-
-put_cookie:
-??? iommu_put_dma_cookie(domain);
-free_domain:
-??? iommu_domain_free(domain);
-??? return ret;

? ?}

? ?static inline void __exynos_iommu_release_mapping(struct exynos_drm_private *priv)
? ?{
-??? struct iommu_domain *domain = priv->mapping; -
-??? iommu_put_dma_cookie(domain);
-??? iommu_domain_free(domain);
? ???? priv->mapping = NULL;
? ?}

drivers/gpu/drm/exynos/exynos_drm_iommu.h +3 -1

@@ -94,7 +69,9 @@ static inline int __exynos_iommu_attach(struct
 exynos_drm_private *priv,
 ? ?{
 ? ???? struct iommu_domain *domain = priv->mapping;
 
-??? return iommu_attach_device(domain, dev);
+??? if (dev != priv->dma_dev)
+??? ??? return iommu_attach_device(domain, dev);
+??? return 0;
 ? ?}
 

? ?static inline void __exynos_iommu_detach(struct exynos_drm_private *priv,

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -0

@@ -102,7 +79,8 @@ static inline void __exynos_iommu_detach(struct
 exynos_drm_private *priv,
 ? ?{
 ? ???? struct iommu_domain *domain = priv->mapping;
 

drivers/gpu/drm/exynos/exynos_drm_iommu.h static inline void __exynos_iommu_detach(struct · line 84

Quoted 3 lines

drivers/gpu/drm/exynos/exynos_drm_iommu.h +2 -1

-??? iommu_detach_device(domain, dev);
+??? if (dev != priv->dma_dev)
+??? ??? iommu_detach_device(domain, dev);

Strictly you could skip that check, as detaching the master device from its real default domain is just a no-op, but I guess maintaining the symmetry is probably more intuitive.

...

Hi Robin,

Quoted 39 lines · Robin Murphy

Quoted 33 lines · Marek Szyprowski

Quoted 26 lines · Robin Murphy

Quoted 20 lines · Marek Szyprowski

Quoted 10 lines · Robin Murphy

Most parts of iommu-dma already assume they are operating on a default
domain set up by iommu_dma_init_domain(), and can be converted straight
over to avoid the refcounting bottleneck. MSI page mappings may be in an unmanaged domain with an explicit MSI-only cookie, so retain the non-specific lookup, but that's OK since they're far from a contended fast path either way.

Signed-off-by Robin Murphy <[email protected]>

This breaks Exynos DRM driver with Exynos IOMMU on ARM64 (Exynos5433). Exynos DRM creates it's own domain, attach all devices which performs DMA access (typically CRTC devices) to it and uses standard DMA-mapping
calls to allocate/map buffers. This way it can use the same one DMA address for each buffer regardless of the CRTC/display/processing device.

Quoted 4 lines

This no longer works with this patch. The simplest solution to fix this
is add API to change default_domain to the one allocated by the Exynos DRM driver.

Quoted 5 lines

Ugh, I hadn't realised there were any drivers trying to fake up their own default domains - that's always going to be fragile. In fact, one of my proposals for un-breaking Tegra and other DRM drivers involves preventing iommu-dma from trying to operate on non-IOMMU_DOMAIN_DMA domains at all, which would stop that from working permanently.

Quoted 4 lines

IMHO there should be a way to let drivers like DRM to use DMA-mapping infrastructure on their own iommu domain. Better provide such API to avoid hacking in the DRM drivers to get this done without help from IOMMU core.

Quoted 4 lines

The tricky part is how to reconcile that with those other drivers which want to do explicit IOMMU management with their own domain but still use the DMA API for coherency of the underlying memory. I do have a couple of half-formed ideas, but they're not quite there yet.

The only idea I have here is to add some flags to struct driver to let device core and frameworks note that this driver wants to manage everything on their own. Then such drivers, once they settle their IOMMU domain, should call some simple API to bless it for DMA API.

Quoted 11 lines

Quoted 8 lines

Quoted 4 lines

Can Exynos not put all the DRM components into a single group, or at least just pick one of the real default domains to attach everyone to instead of allocating a fake one?

Exynos DRM components are not in the single group. Currently iommu-groups have no meaning on Exynos5433 and Exynos IOMMU simply allocate one group per each iommu-master device in the system.

Yeah, a group spanning multiple IOMMU devices would have been a bit of a hack for your topology, but still *technically* possible ;)

The question if I want to put all DRM components into single IOMMU group, how to propagate such knowledge from Exynos DRM driver to Exynos IOMMU driver (groups are created by IOMMU driver)? Anyway, I don't see any benefits from using IOMMU groups as for now, especially when each Exynos DRM component has its own, separate IOMMU (or even more than one, but that a different story).

Quoted 59 lines

Quoted 43 lines

Exynos DRM already selects one of its component devices as the 'main DMA device'. It is being used for managing buffers if no IOMMU is available, so it can also use its iommu domain instead of allocating a fake one. I've checked and it fixes Exynos DRM now. The question is how to merge this fix to keep bisectability.

From: Marek Szyprowski <[email protected]>
Date: Fri, 28 Sep 2018 16:17:27 +0200
Subject: [PATCH] drm/exynos: Use selected dma_dev default iommu domain instead
of a fake one

Instead of allocating a fake IOMMU domain for all Exynos DRM components, simply reuse the default IOMMU domain of the already selected DMA device.
This allows some design changes in IOMMU framework without breaking IOMMU
support in Exynos DRM.

Signed-off-by Marek Szyprowski <[email protected]>

---
   drivers/gpu/drm/exynos/exynos_drm_iommu.h | 34
++++-------------------
   1 file changed, 6 insertions(+), 28 deletions(-)

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -7

index 87f6b5672e11..51d3b7dcd529 100644
@@ -55,37 +55,12 @@ static inline void __exynos_iommu_detach(struct
 exynos_drm_private *priv,
   static inline int __exynos_iommu_create_mapping(struct
 exynos_drm_private *priv,
                       unsigned long start, unsigned long size)
   {
-    struct iommu_domain *domain;
-    int ret;
-
-    domain = iommu_domain_alloc(priv->dma_dev->bus);
-    if (!domain)
-        return -ENOMEM;
-

drivers/gpu/drm/exynos/exynos_drm_iommu.h static inline void __exynos_iommu_detach(struct · line 60

Quoted 10 lines

drivers/gpu/drm/exynos/exynos_drm_iommu.h +1 -9

-    ret = iommu_get_dma_cookie(domain);
-    if (ret)
-        goto free_domain;
-
-    ret = iommu_dma_init_domain(domain, start, size, NULL);
-    if (ret)
-        goto put_cookie;
-
-    priv->mapping = domain;
+    priv->mapping = iommu_get_dma_domain(priv->dma_dev);

iommu_get_domain_for_dev(), please - this isn't a performance-critical path inside a DMA API implementation, plus without an actual dependency on this series maybe there's a chance of sneaking it into 4.19?

Quoted 3 lines

(you could always still check domain->type == IOMMU_DOMAIN_DMA if you want to be really really paranoid.)

I've sent a patch. I hope we will manage to get it as a fix into v4.19, so it won't block your changes in v4.20.

Quoted 37 lines

Quoted 37 lines

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -10

       return 0;
-
-put_cookie:
-    iommu_put_dma_cookie(domain);
-free_domain:
-    iommu_domain_free(domain);
-    return ret;
   }
 
   static inline void __exynos_iommu_release_mapping(struct
 exynos_drm_private *priv)
   {
-    struct iommu_domain *domain = priv->mapping;
-
-    iommu_put_dma_cookie(domain);
-    iommu_domain_free(domain);

       priv->mapping = NULL;

}

drivers/gpu/drm/exynos/exynos_drm_iommu.h +3 -1

@@ -94,7 +69,9 @@ static inline int __exynos_iommu_attach(struct
 exynos_drm_private *priv,
   {
       struct iommu_domain *domain = priv->mapping;
 
-    return iommu_attach_device(domain, dev);
+    if (dev != priv->dma_dev)
+        return iommu_attach_device(domain, dev);
+    return 0;
   }
 

static inline void __exynos_iommu_detach(struct exynos_drm_private
*priv,

drivers/gpu/drm/exynos/exynos_drm_iommu.h +0 -0

@@ -102,7 +79,8 @@ static inline void __exynos_iommu_detach(struct
 exynos_drm_private *priv,
   {
       struct iommu_domain *domain = priv->mapping;
 

Quoted 7 lines

drivers/gpu/drm/exynos/exynos_drm_iommu.h static inline void __exynos_iommu_detach(struct · line 84

Quoted 3 lines

drivers/gpu/drm/exynos/exynos_drm_iommu.h +2 -1

-    iommu_detach_device(domain, dev);
+    if (dev != priv->dma_dev)
+        iommu_detach_device(domain, dev);

Strictly you could skip that check, as detaching the master device from its real default domain is just a no-op, but I guess maintaining the symmetry is probably more intuitive.

I would like to keep the symmetry.

Quoted 1 line

...

Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland