Skip to contentSkip to conversation

dma_alloc_pages / dma_alloc_noncoherent fixups

PATCH In review Christoph Hellwig 5 participants 29 messages Sep 30, 2020 – Nov 24, 2020

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

Hi all,

this series has a bunch of fixups for the noncoherent DMA allocator rework that recently landed in linux-next.

I think the most important part is that the idea of vmap()ing non-contiguous allocations in dma_alloc_noncoherent doesn't work very well after all. It means we can't just rely on virt_to_page to get the page and just use remap_pfn_range or stuff it into other APIs, but on the other hand it also isn't really generic enought for what the media APIs seems to want.

So the first patch reverts that change, and the last patch suggests a different lower level API which should allow the media code to do all it wants.

I'd suggest all but the last patch for the current merge window, and we should have a discussion on how well the last one suits the media subsystem, and probably merge it together with any media changes to use the required API.

#1 lore raw

PATCH 1/8 dma-mapping: remove the {alloc,free}_noncoherent methods

It turns out allowing non-contigous allocations here was a rather bad idea, as we'll now need to define ways to get the pages for mmaping or dma_buf sharing. Revert this change and stick to the original concept. A different API for the use case of non-contigous allocations will be added back later.

---
 drivers/iommu/dma-iommu.c   | 30 ------------------------------
 include/linux/dma-mapping.h |  5 -----
 kernel/dma/mapping.c        | 33 ++++++---------------------------
 3 files changed, 6 insertions(+), 62 deletions(-)
drivers/iommu/dma-iommu.c +0 -30
index c12c1dc43d312e..b363b20a9f41ce 100644
@@ -1055,34 +1055,6 @@ static void *iommu_dma_alloc(struct device *dev, size_t size,
 	return cpu_addr;
 }
 
-#ifdef CONFIG_DMA_REMAP
-static void *iommu_dma_alloc_noncoherent(struct device *dev, size_t size,
-		dma_addr_t *handle, enum dma_data_direction dir, gfp_t gfp)
-{
-	if (!gfpflags_allow_blocking(gfp)) {
-		struct page *page;
-
-		page = dma_common_alloc_pages(dev, size, handle, dir, gfp);
-		if (!page)
-			return NULL;
-		return page_address(page);
-	}
-
-	return iommu_dma_alloc_remap(dev, size, handle, gfp | __GFP_ZERO,
-				     PAGE_KERNEL, 0);
-}
-
-static void iommu_dma_free_noncoherent(struct device *dev, size_t size,
-		void *cpu_addr, dma_addr_t handle, enum dma_data_direction dir)
-{
-	__iommu_dma_unmap(dev, handle, size);
-	__iommu_dma_free(dev, size, cpu_addr);
-}
-#else
-#define iommu_dma_alloc_noncoherent		NULL
-#define iommu_dma_free_noncoherent		NULL
-#endif /* CONFIG_DMA_REMAP */
-
 static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,
 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
 		unsigned long attrs)
@@ -1153,8 +1125,6 @@ static const struct dma_map_ops iommu_dma_ops = {
 	.free			= iommu_dma_free,
 	.alloc_pages		= dma_common_alloc_pages,
 	.free_pages		= dma_common_free_pages,
-	.alloc_noncoherent	= iommu_dma_alloc_noncoherent,
-	.free_noncoherent	= iommu_dma_free_noncoherent,
 	.mmap			= iommu_dma_mmap,
 	.get_sgtable		= iommu_dma_get_sgtable,
 	.map_page		= iommu_dma_map_page,
include/linux/dma-mapping.h +0 -5
index 7c77cd6f3604a7..4b9b1d64f5ec9e 100644
@@ -74,11 +74,6 @@ struct dma_map_ops {
 			gfp_t gfp);
 	void (*free_pages)(struct device *dev, size_t size, struct page *vaddr,
 			dma_addr_t dma_handle, enum dma_data_direction dir);
-	void* (*alloc_noncoherent)(struct device *dev, size_t size,
-			dma_addr_t *dma_handle, enum dma_data_direction dir,
-			gfp_t gfp);
-	void (*free_noncoherent)(struct device *dev, size_t size, void *vaddr,
-			dma_addr_t dma_handle, enum dma_data_direction dir);
 	int (*mmap)(struct device *, struct vm_area_struct *,
 			  void *, dma_addr_t, size_t,
 			  unsigned long attrs);
kernel/dma/mapping.c +6 -27
index 9669550656a0b4..06115f59f4ffbf 100644
@@ -513,40 +513,19 @@ EXPORT_SYMBOL_GPL(dma_free_pages);
 void *dma_alloc_noncoherent(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, enum dma_data_direction dir, gfp_t gfp)
 {
-	const struct dma_map_ops *ops = get_dma_ops(dev);
-	void *vaddr;
-
-	if (!ops || !ops->alloc_noncoherent) {
-		struct page *page;
-
-		page = dma_alloc_pages(dev, size, dma_handle, dir, gfp);
-		if (!page)
-			return NULL;
-		return page_address(page);
-	}
+	struct page *page;
 
-	size = PAGE_ALIGN(size);
-	vaddr = ops->alloc_noncoherent(dev, size, dma_handle, dir, gfp);
-	if (vaddr)
-		debug_dma_map_page(dev, virt_to_page(vaddr), 0, size, dir,
-				   *dma_handle);
-	return vaddr;
+	page = dma_alloc_pages(dev, size, dma_handle, dir, gfp);
+	if (!page)
+		return NULL;
+	return page_address(page);
 }
 EXPORT_SYMBOL_GPL(dma_alloc_noncoherent);
 
 void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
 		dma_addr_t dma_handle, enum dma_data_direction dir)
 {
-	const struct dma_map_ops *ops = get_dma_ops(dev);
-
-	if (!ops || !ops->free_noncoherent) {
-		dma_free_pages(dev, size, virt_to_page(vaddr), dma_handle, dir);
-		return;
-	}
-
-	size = PAGE_ALIGN(size);
-	debug_dma_unmap_page(dev, dma_handle, size, dir);
-	ops->free_noncoherent(dev, size, vaddr, dma_handle, dir);
+	dma_free_pages(dev, size, virt_to_page(vaddr), dma_handle, dir);
 }
 EXPORT_SYMBOL_GPL(dma_free_noncoherent);
 
Signature · 1 line
2.28.0
#2 lore raw

PATCH 2/8 dma-mapping: document dma_{alloc,free}_pages

In reply to Christoph Hellwig

Document the new dma_alloc_pages and dma_free_pages APIs, and fix up the documentation for dma_alloc_noncoherent and dma_free_noncoherent.

---
 Documentation/core-api/dma-api.rst | 45 ++++++++++++++++++++++++++----
 1 file changed, 40 insertions(+), 5 deletions(-)
Documentation/core-api/dma-api.rst +40 -5
index ea0413276ddb70..a75c469dbcaa7c 100644
@@ -534,11 +534,9 @@ an I/O device, you should not be using this part of the API.
 			dma_addr_t *dma_handle, enum dma_data_direction dir,
 			gfp_t gfp)
 
-This routine allocates a region of <size> bytes of consistent memory.  It
+This routine allocates a region of <size> bytes of non-coherent memory.  It
 returns a pointer to the allocated region (in the processor's virtual address
-space) or NULL if the allocation failed.  The returned memory may or may not
-be in the kernels direct mapping.  Drivers must not call virt_to_page on
-the returned memory region.
+space) or NULL if the allocation failed.
 
 It also returns a <dma_handle> which may be cast to an unsigned integer the
 same width as the bus and given to the device as the DMA address base of
@@ -565,7 +563,44 @@ reused.
 Free a region of memory previously allocated using dma_alloc_noncoherent().
 dev, size and dma_handle and dir must all be the same as those passed into
 dma_alloc_noncoherent().  cpu_addr must be the virtual address returned by
-the dma_alloc_noncoherent().
+dma_alloc_noncoherent().
+
+::
+
+	struct page *
+	dma_alloc_pages(struct device *dev, size_t size, dma_addr_t *dma_handle,
+			enum dma_data_direction dir, gfp_t gfp)
+
+This routine allocates a region of <size> bytes of non-coherent memory.  It
+returns a pointer to first struct page for the region, or NULL if the
+allocation failed.
+
+It also returns a <dma_handle> which may be cast to an unsigned integer the
+same width as the bus and given to the device as the DMA address base of
+the region.
+
+The dir parameter specified if data is read and/or written by the device,
+see dma_map_single() for details.
+
+The gfp parameter allows the caller to specify the ``GFP_`` flags (see
+kmalloc()) for the allocation, but rejects flags used to specify a memory
+zone such as GFP_DMA or GFP_HIGHMEM.
+
+Before giving the memory to the device, dma_sync_single_for_device() needs
+to be called, and before reading memory written by the device,
+dma_sync_single_for_cpu(), just like for streaming DMA mappings that are
+reused.
+
+::
+
+	void
+	dma_free_pages(struct device *dev, size_t size, struct page *page,
+			dma_addr_t dma_handle, enum dma_data_direction dir)
+
+Free a region of memory previously allocated using dma_alloc_pages().
+dev, size and dma_handle and dir must all be the same as those passed into
+dma_alloc_noncoherent().  page must be the pointer returned by
+dma_alloc_pages().
 
 ::
 
Signature · 6 lines
2.28.0

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#3 lore raw

PATCH 3/8 dma-direct check for highmem pages in dma_direct_alloc_pages

In reply to Christoph Hellwig

Check for highmem pages from CMA, just like in the dma_direct_alloc path.

---
 kernel/dma/direct.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
kernel/dma/direct.c +11 -0
index 121a9c1969dd3a..b5f20781d3a96f 100644
@@ -309,6 +309,17 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 	page = __dma_direct_alloc_pages(dev, size, gfp);
 	if (!page)
 		return NULL;
+	if (PageHighMem(page)) {
+		/*
+		 * Depending on the cma= arguments and per-arch setup
+		 * dma_alloc_contiguous could return highmem pages.
+		 * Without remapping there is no way to return them here,
+		 * so log an error and fail.
+		 */
+		dev_info(dev, "Rejecting highmem page from CMA.\n");
+		goto out_free_pages;
+	}
+
 	ret = page_address(page);
 	if (force_dma_unencrypted(dev)) {
 		if (set_memory_decrypted((unsigned long)ret,
Signature · 6 lines
2.28.0

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#4 lore raw

PATCH 4/8 dma-direct: use __GFP_ZERO in dma_direct_alloc_pages

In reply to Christoph Hellwig

Prepare for supporting the DMA_ATTR_NO_KERNEL_MAPPING flag in dma_alloc_pages.

---
 kernel/dma/direct.c | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)
kernel/dma/direct.c +4 -5
index b5f20781d3a96f..b5d56810130b22 100644
@@ -296,9 +296,10 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, enum dma_data_direction dir, gfp_t gfp)
 {
 	struct page *page;
-	void *ret;
 
 	if (dma_should_alloc_from_pool(dev, gfp, 0)) {
+		void *ret;
+
 		page = dma_alloc_from_pool(dev, size, &ret, gfp,
 				dma_coherent_ok);
 		if (!page)
@@ -306,7 +307,7 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 		goto done;
 	}
 
-	page = __dma_direct_alloc_pages(dev, size, gfp);
+	page = __dma_direct_alloc_pages(dev, size, gfp | __GFP_ZERO);
 	if (!page)
 		return NULL;
 	if (PageHighMem(page)) {
@@ -320,13 +321,11 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 		goto out_free_pages;
 	}
 
-	ret = page_address(page);
 	if (force_dma_unencrypted(dev)) {
-		if (set_memory_decrypted((unsigned long)ret,
+		if (set_memory_decrypted((unsigned long)page_address(page),
 				1 << get_order(size)))
 			goto out_free_pages;
 	}
-	memset(ret, 0, size);
 done:
 	*dma_handle = phys_to_dma_direct(dev, page_to_phys(page));
 	return page;
Signature · 6 lines
2.28.0

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#5 lore raw

Re: [PATCH 4/8] dma-direct: use __GFP_ZERO in dma_direct_alloc_pages

Quoted 36 lines · On Wed, 30 Sep 2020 18:09:13 Christoph Hellwig wrote:

Prepare for supporting the DMA_ATTR_NO_KERNEL_MAPPING flag in dma_alloc_pages.

---
 kernel/dma/direct.c | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)
kernel/dma/direct.c +3 -2
index b5f20781d3a96f..b5d56810130b22 100644
@@ -296,9 +296,10 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, enum dma_data_direction dir, gfp_t gfp)
 {
 	struct page *page;
-	void *ret;
 
 	if (dma_should_alloc_from_pool(dev, gfp, 0)) {
+		void *ret;
+
 		page = dma_alloc_from_pool(dev, size, &ret, gfp,
 				dma_coherent_ok);
 		if (!page)
@@ -306,7 +307,7 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 		goto done;
 	}
 
-	page = __dma_direct_alloc_pages(dev, size, gfp);
+	page = __dma_direct_alloc_pages(dev, size, gfp | __GFP_ZERO);
 	if (!page)
 		return NULL;
 	if (PageHighMem(page)) {
@@ -320,13 +321,11 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 		goto out_free_pages;

kernel/dma/direct.c struct page *dma_direct_alloc_pages(struct device *dev, size_t size, · line 329

Quoted 10 lines
kernel/dma/direct.c +1 -3
 	}
 
-	ret = page_address(page);
 	if (force_dma_unencrypted(dev)) {
-		if (set_memory_decrypted((unsigned long)ret,
+		if (set_memory_decrypted((unsigned long)page_address(page),
 				1 << get_order(size)))
 			goto out_free_pages;
 	}
-	memset(ret, 0, size);

Not sure this works without changes in cma_alloc().

Quoted 5 lines
kernel/dma/direct.c +0 -0
 done:
 	*dma_handle = phys_to_dma_direct(dev, page_to_phys(page));
 	return page;
Signature · 1 line
2.28.0

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#6 lore raw

PATCH 5/8 dma-direct: factor out a dma_direct_alloc_from_pool helper

In reply to Christoph Hellwig

This ensures dma_direct_alloc_pages will use the right gfp mask, as well as keeping the code for that common.

---
 kernel/dma/direct.c | 41 ++++++++++++++++++++---------------------
 1 file changed, 20 insertions(+), 21 deletions(-)
kernel/dma/direct.c +20 -21
index b5d56810130b22..ace9159c992f65 100644
@@ -147,6 +147,22 @@ static struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
 	return page;
 }
 
+static void *dma_direct_alloc_from_pool(struct device *dev, size_t size,
+		dma_addr_t *dma_handle, gfp_t gfp)
+{
+	struct page *page;
+	u64 phys_mask;
+	void *ret;
+
+	gfp |= dma_direct_optimal_gfp_mask(dev, dev->coherent_dma_mask,
+					   &phys_mask);
+	page = dma_alloc_from_pool(dev, size, &ret, gfp, dma_coherent_ok);
+	if (!page)
+		return NULL;
+	*dma_handle = phys_to_dma_direct(dev, page_to_phys(page));
+	return ret;
+}
+
 void *dma_direct_alloc(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
 {
@@ -163,17 +179,8 @@ void *dma_direct_alloc(struct device *dev, size_t size,
 	if (attrs & DMA_ATTR_NO_WARN)
 		gfp |= __GFP_NOWARN;
 
-	if (dma_should_alloc_from_pool(dev, gfp, attrs)) {
-		u64 phys_mask;
-
-		gfp |= dma_direct_optimal_gfp_mask(dev, dev->coherent_dma_mask,
-				&phys_mask);
-		page = dma_alloc_from_pool(dev, size, &ret, gfp,
-				dma_coherent_ok);
-		if (!page)
-			return NULL;
-		goto done;
-	}
+	if (dma_should_alloc_from_pool(dev, gfp, attrs))
+		return dma_direct_alloc_from_pool(dev, size, dma_handle, gfp);
 
 	/* we always manually zero the memory once we are done */
 	page = __dma_direct_alloc_pages(dev, size, gfp & ~__GFP_ZERO);
@@ -297,15 +304,8 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 {
 	struct page *page;
 
-	if (dma_should_alloc_from_pool(dev, gfp, 0)) {
-		void *ret;
-
-		page = dma_alloc_from_pool(dev, size, &ret, gfp,
-				dma_coherent_ok);
-		if (!page)
-			return NULL;
-		goto done;
-	}
+	if (dma_should_alloc_from_pool(dev, gfp, 0))
+		return dma_direct_alloc_from_pool(dev, size, dma_handle, gfp);
 
 	page = __dma_direct_alloc_pages(dev, size, gfp | __GFP_ZERO);
 	if (!page)
@@ -326,7 +326,6 @@ struct page *dma_direct_alloc_pages(struct device *dev, size_t size,
 				1 << get_order(size)))
 			goto out_free_pages;
 	}
-done:
 	*dma_handle = phys_to_dma_direct(dev, page_to_phys(page));
 	return page;
 out_free_pages:
Signature · 6 lines
2.28.0

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#7 lore raw

PATCH 6/8 dma-direct: simplify the DMA_ATTR_NO_KERNEL_MAPPING handling

In reply to Christoph Hellwig

Use and entirely separate code path for the DMA_ATTR_NO_KERNEL_MAPPING path. This avoids any confusion about the ret type, and avoids lots of attr checks and helpers that can be significantly simplified now.

It also ensures that common handling is applied to architetures still using the arch alloc/free hooks.

---
 include/linux/dma-noncoherent.h |  13 -----
 kernel/dma/direct.c             | 100 +++++++++++++-------------------
 2 files changed, 39 insertions(+), 74 deletions(-)
include/linux/dma-noncoherent.h +0 -13
index e61283e06576a8..73ac149fa181b4 100644
@@ -21,19 +21,6 @@ static inline bool dev_is_dma_coherent(struct device *dev)
 }
 #endif /* CONFIG_ARCH_HAS_DMA_COHERENCE_H */
 
-/*
- * Check if an allocation needs to be marked uncached to be coherent.
- */
-static __always_inline bool dma_alloc_need_uncached(struct device *dev,
-		unsigned long attrs)
-{
-	if (dev_is_dma_coherent(dev))
-		return false;
-	if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
-		return false;
-	return true;
-}
-
 void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 		gfp_t gfp, unsigned long attrs);
 void arch_dma_free(struct device *dev, size_t size, void *cpu_addr,
kernel/dma/direct.c +39 -61

159 lines not on this page · raw message

Signature · 1 line
2.28.0
#8 lore raw

PATCH 7/8 dma-iommu: remove __iommu_dma_mmap

In reply to Christoph Hellwig

The function has a single caller, so open code it there and take advantage of the precalculated page count variable.

---
 drivers/iommu/dma-iommu.c | 17 +----------------
 1 file changed, 1 insertion(+), 16 deletions(-)
drivers/iommu/dma-iommu.c +1 -16
index b363b20a9f41ce..7922f545cd5eef 100644
@@ -656,21 +656,6 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
 	return NULL;
 }
 
-/**
- * __iommu_dma_mmap - Map a buffer into provided user VMA
- * @pages: Array representing buffer from __iommu_dma_alloc()
- * @size: Size of buffer in bytes
- * @vma: VMA describing requested userspace mapping
- *
- * Maps the pages of the buffer in @pages into @vma. The caller is responsible
- * for verifying the correct size and protection of @vma beforehand.
- */
-static int __iommu_dma_mmap(struct page **pages, size_t size,
-		struct vm_area_struct *vma)
-{
-	return vm_map_pages(vma, pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
-}
-
 static void iommu_dma_sync_single_for_cpu(struct device *dev,
 		dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
 {
@@ -1075,7 +1060,7 @@ static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,
 		struct page **pages = dma_common_find_pages(cpu_addr);
 
 		if (pages)
-			return __iommu_dma_mmap(pages, size, vma);
+			return vm_map_pages(vma, pages, nr_pages);
 		pfn = vmalloc_to_pfn(cpu_addr);
 	} else {
 		pfn = page_to_pfn(virt_to_page(cpu_addr));
Signature · 6 lines
2.28.0

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
#9 lore raw

PATCH 8/8 WIP: add a dma_alloc_contiguous API

In reply to Christoph Hellwig

Add a new API that returns a virtually non-contigous array of pages and dma address. This API is only implemented for dma-iommu and will not be implemented for non-iommu DMA API instances that have to allocate contiguous memory. It is up to the caller to check if the API is available.

The intent is that media drivers can use this API if either:

- no kernel mapping or only temporary kernel mappings are required. That is as a better replacement for DMA_ATTR_NO_KERNEL_MAPPING
- a kernel mapping is required for cached and DMA mapped pages, but the driver also needs the pages to e.g. map them to userspace. In that sense it is a replacement for some aspects of the recently removed and never fully implemented DMA_ATTR_NON_CONSISTENT

---
 drivers/iommu/dma-iommu.c   | 73 +++++++++++++++++++++++++------------
 include/linux/dma-mapping.h |  9 +++++
 kernel/dma/mapping.c        | 35 ++++++++++++++++++
 3 files changed, 93 insertions(+), 24 deletions(-)
drivers/iommu/dma-iommu.c +49 -24
index 7922f545cd5eef..158026a856622c 100644
@@ -565,23 +565,12 @@ static struct page **__iommu_dma_alloc_pages(struct device *dev,
 	return pages;
 }
 
-/**
- * iommu_dma_alloc_remap - Allocate and map a buffer contiguous in IOVA space
- * @dev: Device to allocate memory for. Must be a real device
- *	 attached to an iommu_dma_domain
- * @size: Size of buffer in bytes
- * @dma_handle: Out argument for allocated DMA handle
- * @gfp: Allocation flags
- * @prot: pgprot_t to use for the remapped mapping
- * @attrs: DMA attributes for this allocation
- *
- * If @size is less than PAGE_SIZE, then a full CPU page will be allocated,
+/*
+ * If size is less than PAGE_SIZE, then a full CPU page will be allocated,
  * but an IOMMU which supports smaller pages might not map the whole thing.
- *
- * Return: Mapped virtual address, or NULL on failure.
  */
-static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
-		dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
+static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
+		size_t size, dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
 		unsigned long attrs)
 {
 	struct iommu_domain *domain = iommu_get_dma_domain(dev);
@@ -593,7 +582,6 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
 	struct page **pages;
 	struct sg_table sgt;
 	dma_addr_t iova;
-	void *vaddr;
 
 	*dma_handle = DMA_MAPPING_ERROR;
 
@@ -636,17 +624,10 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
 			< size)
 		goto out_free_sg;
 
-	vaddr = dma_common_pages_remap(pages, size, prot,
-			__builtin_return_address(0));
-	if (!vaddr)
-		goto out_unmap;
-
 	*dma_handle = iova;
 	sg_free_table(&sgt);
-	return vaddr;
+	return pages;
 
-out_unmap:
-	__iommu_dma_unmap(dev, iova, size);
 out_free_sg:
 	sg_free_table(&sgt);
 out_free_iova:
@@ -656,6 +637,46 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
 	return NULL;
 }
 
+static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
+		dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
+		unsigned long attrs)
+{
+	struct page **pages;
+	void *vaddr;
+
+	pages = __iommu_dma_alloc_noncontiguous(dev, size, dma_handle, gfp,
+						prot, attrs);
+	if (!pages)
+		return NULL;
+	vaddr = dma_common_pages_remap(pages, size, prot,
+			__builtin_return_address(0));
+	if (!vaddr)
+		goto out_unmap;
+	return vaddr;
+
+out_unmap:
+	__iommu_dma_unmap(dev, *dma_handle, size);
+	__iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
+	return NULL;
+}
+
+#ifdef CONFIG_DMA_REMAP
+static struct page **iommu_dma_alloc_noncontiguous(struct device *dev,
+		size_t size, dma_addr_t *dma_handle, gfp_t gfp,
+		unsigned long attrs)
+{
+	return __iommu_dma_alloc_noncontiguous(dev, size, dma_handle, gfp,
+					       PAGE_KERNEL, attrs);
+}
+
+static void iommu_dma_free_noncontiguous(struct device *dev, size_t size,
+		struct page **pages, dma_addr_t dma_handle)
+{
+	__iommu_dma_unmap(dev, dma_handle, size);
+	__iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
+}
+#endif
+
 static void iommu_dma_sync_single_for_cpu(struct device *dev,
 		dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
 {
@@ -1110,6 +1131,10 @@ static const struct dma_map_ops iommu_dma_ops = {
 	.free			= iommu_dma_free,
 	.alloc_pages		= dma_common_alloc_pages,
 	.free_pages		= dma_common_free_pages,
+#ifdef CONFIG_DMA_REMAP
+	.alloc_noncontiguous	= iommu_dma_alloc_noncontiguous,
+	.free_noncontiguous	= iommu_dma_free_noncontiguous,
+#endif
 	.mmap			= iommu_dma_mmap,
 	.get_sgtable		= iommu_dma_get_sgtable,
 	.map_page		= iommu_dma_map_page,
include/linux/dma-mapping.h +9 -0
index 4b9b1d64f5ec9e..51bbc32365bb8d 100644
@@ -74,6 +74,10 @@ struct dma_map_ops {
 			gfp_t gfp);
 	void (*free_pages)(struct device *dev, size_t size, struct page *vaddr,
 			dma_addr_t dma_handle, enum dma_data_direction dir);
+	struct page **(*alloc_noncontiguous)(struct device *dev, size_t size,
+			dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs);
+	void (*free_noncontiguous)(struct device *dev, size_t size,
+			struct page **pages, dma_addr_t dma_handle);
 	int (*mmap)(struct device *, struct vm_area_struct *,
 			  void *, dma_addr_t, size_t,
 			  unsigned long attrs);
@@ -384,6 +388,11 @@ void *dma_alloc_noncoherent(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, enum dma_data_direction dir, gfp_t gfp);
 void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
 		dma_addr_t dma_handle, enum dma_data_direction dir);
+bool dma_can_alloc_noncontiguous(struct device *dev);
+struct page **dma_alloc_noncontiguous(struct device *dev, size_t size,
+		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs);
+void dma_free_noncontiguous(struct device *dev, size_t size,
+		struct page **pages, dma_addr_t dma_handle);
 
 static inline dma_addr_t dma_map_single_attrs(struct device *dev, void *ptr,
 		size_t size, enum dma_data_direction dir, unsigned long attrs)
kernel/dma/mapping.c +35 -0
index 06115f59f4ffbf..6d975d1a20dd72 100644
@@ -529,6 +529,41 @@ void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
 }
 EXPORT_SYMBOL_GPL(dma_free_noncoherent);
 
+bool dma_can_alloc_noncontiguous(struct device *dev)
+{
+	const struct dma_map_ops *ops = get_dma_ops(dev);
+
+	return ops && ops->free_noncontiguous;
+}
+EXPORT_SYMBOL_GPL(dma_can_alloc_noncontiguous);
+
+struct page **dma_alloc_noncontiguous(struct device *dev, size_t size,
+		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
+{
+	const struct dma_map_ops *ops = get_dma_ops(dev);
+
+	if (WARN_ON_ONCE(!dma_can_alloc_noncontiguous(dev)))
+		return NULL;
+	if (attrs & ~DMA_ATTR_ALLOC_SINGLE_PAGES) {
+		dev_warn(dev, "invalid flags (0x%lx) for %s\n",
+			 attrs, __func__);
+		return NULL;
+	}
+	return ops->alloc_noncontiguous(dev, size, dma_handle, gfp, attrs);
+}
+EXPORT_SYMBOL_GPL(dma_alloc_noncontiguous);
+
+void dma_free_noncontiguous(struct device *dev, size_t size,
+		struct page **pages, dma_addr_t dma_handle)
+{
+	const struct dma_map_ops *ops = get_dma_ops(dev);
+
+	if (WARN_ON_ONCE(!dma_can_alloc_noncontiguous(dev)))
+		return;
+	ops->free_noncontiguous(dev, size, pages, dma_handle);
+}
+EXPORT_SYMBOL_GPL(dma_free_noncontiguous);
+
 int dma_supported(struct device *dev, u64 mask)
 {
 	const struct dma_map_ops *ops = get_dma_ops(dev);
Signature · 1 line
2.28.0
#10 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Hi Christoph,

Quoted 5 lines · Christoph Hellwig

Add a new API that returns a virtually non-contigous array of pages and dma address. This API is only implemented for dma-iommu and will not be implemented for non-iommu DMA API instances that have to allocate contiguous memory. It is up to the caller to check if the API is available.

Would you mind scheding some more light on what made the previous attempt not work well? I liked the previous API because it was more consistent with the regular dma_alloc_coherent().

Quoted 2 lines

The intent is that media drivers can use this API if either:

FWIW, the USB subsystem also has similar needs, and so do some DRM drivers using DMA API rather than IOMMU API directly. Basically I believe that all the users removed in your previous series relied on custom downstream patches to make DMA_ATTR_NON_CONSISTENT work and could be finally made work in upstream using this API.

Quoted 7 lines

- no kernel mapping or only temporary kernel mappings are required. That is as a better replacement for DMA_ATTR_NO_KERNEL_MAPPING
- a kernel mapping is required for cached and DMA mapped pages, but the driver also needs the pages to e.g. map them to userspace. In that sense it is a replacement for some aspects of the recently removed and never fully implemented DMA_ATTR_NON_CONSISTENT

What's the expected allocation and mapping flow with the latter? Would that be

pages = dma_alloc_noncoherent(...)
vaddr = vmap(pages, ...);

?

Would one just use the usual dma_sync_for_{cpu,device}() for cache invallidate/clean, while keeping the mapping in place?

Best regards,
Tomasz

#11 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 9 lines · Tomasz Figa

Hi Christoph,

Quoted 5 lines · Christoph Hellwig

Add a new API that returns a virtually non-contigous array of pages and dma address. This API is only implemented for dma-iommu and will not be implemented for non-iommu DMA API instances that have to allocate contiguous memory. It is up to the caller to check if the API is available.

Quoted 3 lines

Would you mind scheding some more light on what made the previous attempt not work well? I liked the previous API because it was more consistent with the regular dma_alloc_coherent().

The problem is that with a dma_alloc_noncoherent that can return pages not in the kernel mapping we can't just use virt_to_page to fill in scatterlists or mmap the buffer to userspace, but would need new helpers and another two methods.

Quoted 7 lines
Quoted 6 lines

- no kernel mapping or only temporary kernel mappings are required. That is as a better replacement for DMA_ATTR_NO_KERNEL_MAPPING
- a kernel mapping is required for cached and DMA mapped pages, but the driver also needs the pages to e.g. map them to userspace. In that sense it is a replacement for some aspects of the recently removed and never fully implemented DMA_ATTR_NON_CONSISTENT

Quoted 6 lines

What's the expected allocation and mapping flow with the latter? Would that be

pages = dma_alloc_noncoherent(...)
vaddr = vmap(pages, ...);

?

Yes. Witht the vmap step optional for replacements of DMA_ATTR_NO_KERNEL_MAPPING, which is another nightmare to deal with.

Quoted 2 lines

Would one just use the usual dma_sync_for_{cpu,device}() for cache invallidate/clean, while keeping the mapping in place?

Yes. And make sure the API isn't implemented when VIVT caches are used, but that isn't really different from the current interface.

#12 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 34 lines · Christoph Hellwig
Quoted 9 lines · Tomasz Figa

Hi Christoph,

Quoted 5 lines · Christoph Hellwig

Add a new API that returns a virtually non-contigous array of pages and dma address. This API is only implemented for dma-iommu and will not be implemented for non-iommu DMA API instances that have to allocate contiguous memory. It is up to the caller to check if the API is available.

Quoted 3 lines

Would you mind scheding some more light on what made the previous attempt not work well? I liked the previous API because it was more consistent with the regular dma_alloc_coherent().

The problem is that with a dma_alloc_noncoherent that can return pages not in the kernel mapping we can't just use virt_to_page to fill in scatterlists or mmap the buffer to userspace, but would need new helpers and another two methods.

Quoted 7 lines
Quoted 6 lines

- no kernel mapping or only temporary kernel mappings are required. That is as a better replacement for DMA_ATTR_NO_KERNEL_MAPPING
- a kernel mapping is required for cached and DMA mapped pages, but the driver also needs the pages to e.g. map them to userspace. In that sense it is a replacement for some aspects of the recently removed and never fully implemented DMA_ATTR_NON_CONSISTENT

Quoted 6 lines

What's the expected allocation and mapping flow with the latter? Would that be

pages = dma_alloc_noncoherent(...)
vaddr = vmap(pages, ...);

?

Quoted 8 lines

Yes. Witht the vmap step optional for replacements of DMA_ATTR_NO_KERNEL_MAPPING, which is another nightmare to deal with.

Quoted 2 lines

Would one just use the usual dma_sync_for_{cpu,device}() for cache invallidate/clean, while keeping the mapping in place?

Yes. And make sure the API isn't implemented when VIVT caches are used, but that isn't really different from the current interface.

Okay, thanks. Let's see if we can make necessary changes to the videobuf2.

+Sergey Senozhatsky for awareness too.

Best regrards,
Tomasz

#13 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 6 lines · Tomasz Figa
Quoted 2 lines

Yes. And make sure the API isn't implemented when VIVT caches are used, but that isn't really different from the current interface.

Okay, thanks. Let's see if we can make necessary changes to the videobuf2.

+Sergey Senozhatsky for awareness too.

I can defer the changes a bit to see if you'd really much prefer the former interface. I think for now the most important thing is that it works properly for the potential users, and the prime one is videobuf2 for now. drm also seems like a big potential users, but I had a really hard time getting the developers to engage in API development.

#14 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 7 lines · Christoph Hellwig
Quoted 5 lines · Tomasz Figa
Quoted 2 lines

Yes. And make sure the API isn't implemented when VIVT caches are used, but that isn't really different from the current interface.

Okay, thanks. Let's see if we can make necessary changes to the videobuf2.

Quoted 8 lines
Quoted 1 line

+Sergey Senozhatsky for awareness too.

I can defer the changes a bit to see if you'd really much prefer the former interface. I think for now the most important thing is that it works properly for the potential users, and the prime one is videobuf2 for now. drm also seems like a big potential users, but I had a really hard time getting the developers to engage in API development.

My initial feeling is that it should work, but we'll give you a definitive answer once we prototype it. :)

We might actually give it a try in the USB HCD subsystem as well, to implement usb_alloc_noncoherent(), as an optimization for drivers which have to perform multiple random accesses to the URB buffers. I think you might recall discussing this by the way of the pwc and uvcvideo camera drivers.

Best regards,
Tomasz
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#15 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 8 lines · Tomasz Figa

My initial feeling is that it should work, but we'll give you a definitive answer once we prototype it. :)

We might actually give it a try in the USB HCD subsystem as well, to implement usb_alloc_noncoherent(), as an optimization for drivers which have to perform multiple random accesses to the URB buffers. I think you might recall discussing this by the way of the pwc and uvcvideo camera drivers.

Yes. I guess for usb the dma_alloc_noncoherent as-is in linux-next
might be better. So if you have the cycles please prototype it
either way, although for that we'd also need at least a mmap_noncoherent method, and probaby a get_sgtable_noncoherent one.

#16 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

In reply to Christoph Hellwig

+CC Ricardo who will be looking into using this in the USB stack (UVC camera driver).

Quoted 216 lines · Christoph Hellwig

Add a new API that returns a virtually non-contigous array of pages and dma address. This API is only implemented for dma-iommu and will not be implemented for non-iommu DMA API instances that have to allocate contiguous memory. It is up to the caller to check if the API is available.

The intent is that media drivers can use this API if either:

- no kernel mapping or only temporary kernel mappings are required. That is as a better replacement for DMA_ATTR_NO_KERNEL_MAPPING
- a kernel mapping is required for cached and DMA mapped pages, but the driver also needs the pages to e.g. map them to userspace. In that sense it is a replacement for some aspects of the recently removed and never fully implemented DMA_ATTR_NON_CONSISTENT

---
 drivers/iommu/dma-iommu.c   | 73 +++++++++++++++++++++++++------------
 include/linux/dma-mapping.h |  9 +++++
 kernel/dma/mapping.c        | 35 ++++++++++++++++++
 3 files changed, 93 insertions(+), 24 deletions(-)
drivers/iommu/dma-iommu.c +49 -24
index 7922f545cd5eef..158026a856622c 100644
@@ -565,23 +565,12 @@ static struct page **__iommu_dma_alloc_pages(struct device *dev,
        return pages;
 }
 
-/**
- * iommu_dma_alloc_remap - Allocate and map a buffer contiguous in IOVA space
- * @dev: Device to allocate memory for. Must be a real device
- *      attached to an iommu_dma_domain
- * @size: Size of buffer in bytes
- * @dma_handle: Out argument for allocated DMA handle
- * @gfp: Allocation flags
- * @prot: pgprot_t to use for the remapped mapping
- * @attrs: DMA attributes for this allocation
- *
- * If @size is less than PAGE_SIZE, then a full CPU page will be allocated,
+/*
+ * If size is less than PAGE_SIZE, then a full CPU page will be allocated,
  * but an IOMMU which supports smaller pages might not map the whole thing.
- *
- * Return: Mapped virtual address, or NULL on failure.
  */
-static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
-               dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
+static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
+               size_t size, dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
                unsigned long attrs)
 {
        struct iommu_domain *domain = iommu_get_dma_domain(dev);
@@ -593,7 +582,6 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
        struct page **pages;
        struct sg_table sgt;
        dma_addr_t iova;
-       void *vaddr;
 
        *dma_handle = DMA_MAPPING_ERROR;
 
@@ -636,17 +624,10 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
                        < size)
                goto out_free_sg;
 
-       vaddr = dma_common_pages_remap(pages, size, prot,
-                       __builtin_return_address(0));
-       if (!vaddr)
-               goto out_unmap;
-
        *dma_handle = iova;
        sg_free_table(&sgt);
-       return vaddr;
+       return pages;
 
-out_unmap:
-       __iommu_dma_unmap(dev, iova, size);
 out_free_sg:
        sg_free_table(&sgt);
 out_free_iova:
@@ -656,6 +637,46 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
        return NULL;
 }
 
+static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
+               dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
+               unsigned long attrs)
+{
+       struct page **pages;
+       void *vaddr;
+
+       pages = __iommu_dma_alloc_noncontiguous(dev, size, dma_handle, gfp,
+                                               prot, attrs);
+       if (!pages)
+               return NULL;
+       vaddr = dma_common_pages_remap(pages, size, prot,
+                       __builtin_return_address(0));
+       if (!vaddr)
+               goto out_unmap;
+       return vaddr;
+
+out_unmap:
+       __iommu_dma_unmap(dev, *dma_handle, size);
+       __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
+       return NULL;
+}
+
+#ifdef CONFIG_DMA_REMAP
+static struct page **iommu_dma_alloc_noncontiguous(struct device *dev,
+               size_t size, dma_addr_t *dma_handle, gfp_t gfp,
+               unsigned long attrs)
+{
+       return __iommu_dma_alloc_noncontiguous(dev, size, dma_handle, gfp,
+                                              PAGE_KERNEL, attrs);
+}
+
+static void iommu_dma_free_noncontiguous(struct device *dev, size_t size,
+               struct page **pages, dma_addr_t dma_handle)
+{
+       __iommu_dma_unmap(dev, dma_handle, size);
+       __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
+}
+#endif
+
 static void iommu_dma_sync_single_for_cpu(struct device *dev,
                dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
 {
@@ -1110,6 +1131,10 @@ static const struct dma_map_ops iommu_dma_ops = {
        .free                   = iommu_dma_free,
        .alloc_pages            = dma_common_alloc_pages,
        .free_pages             = dma_common_free_pages,
+#ifdef CONFIG_DMA_REMAP
+       .alloc_noncontiguous    = iommu_dma_alloc_noncontiguous,
+       .free_noncontiguous     = iommu_dma_free_noncontiguous,
+#endif
        .mmap                   = iommu_dma_mmap,
        .get_sgtable            = iommu_dma_get_sgtable,
        .map_page               = iommu_dma_map_page,
include/linux/dma-mapping.h +9 -0
index 4b9b1d64f5ec9e..51bbc32365bb8d 100644
@@ -74,6 +74,10 @@ struct dma_map_ops {
                        gfp_t gfp);
        void (*free_pages)(struct device *dev, size_t size, struct page *vaddr,
                        dma_addr_t dma_handle, enum dma_data_direction dir);
+       struct page **(*alloc_noncontiguous)(struct device *dev, size_t size,
+                       dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs);
+       void (*free_noncontiguous)(struct device *dev, size_t size,
+                       struct page **pages, dma_addr_t dma_handle);
        int (*mmap)(struct device *, struct vm_area_struct *,
                          void *, dma_addr_t, size_t,
                          unsigned long attrs);
@@ -384,6 +388,11 @@ void *dma_alloc_noncoherent(struct device *dev, size_t size,
                dma_addr_t *dma_handle, enum dma_data_direction dir, gfp_t gfp);
 void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
                dma_addr_t dma_handle, enum dma_data_direction dir);
+bool dma_can_alloc_noncontiguous(struct device *dev);
+struct page **dma_alloc_noncontiguous(struct device *dev, size_t size,
+               dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs);
+void dma_free_noncontiguous(struct device *dev, size_t size,
+               struct page **pages, dma_addr_t dma_handle);
 
 static inline dma_addr_t dma_map_single_attrs(struct device *dev, void *ptr,
                size_t size, enum dma_data_direction dir, unsigned long attrs)
kernel/dma/mapping.c +35 -1
index 06115f59f4ffbf..6d975d1a20dd72 100644
@@ -529,6 +529,41 @@ void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
 }
 EXPORT_SYMBOL_GPL(dma_free_noncoherent);
 
+bool dma_can_alloc_noncontiguous(struct device *dev)
+{
+       const struct dma_map_ops *ops = get_dma_ops(dev);
+
+       return ops && ops->free_noncontiguous;
+}
+EXPORT_SYMBOL_GPL(dma_can_alloc_noncontiguous);
+
+struct page **dma_alloc_noncontiguous(struct device *dev, size_t size,
+               dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
+{
+       const struct dma_map_ops *ops = get_dma_ops(dev);
+
+       if (WARN_ON_ONCE(!dma_can_alloc_noncontiguous(dev)))
+               return NULL;
+       if (attrs & ~DMA_ATTR_ALLOC_SINGLE_PAGES) {
+               dev_warn(dev, "invalid flags (0x%lx) for %s\n",
+                        attrs, __func__);
+               return NULL;
+       }
+       return ops->alloc_noncontiguous(dev, size, dma_handle, gfp, attrs);
+}
+EXPORT_SYMBOL_GPL(dma_alloc_noncontiguous);
+
+void dma_free_noncontiguous(struct device *dev, size_t size,
+               struct page **pages, dma_addr_t dma_handle)
+{
+       const struct dma_map_ops *ops = get_dma_ops(dev);
+
+       if (WARN_ON_ONCE(!dma_can_alloc_noncontiguous(dev)))
+               return;
+       ops->free_noncontiguous(dev, size, pages, dma_handle);
+}
+EXPORT_SYMBOL_GPL(dma_free_noncontiguous);
+
 int dma_supported(struct device *dev, u64 mask)
 {
        const struct dma_map_ops *ops = get_dma_ops(dev);
--

2.28.0

#17 lore raw

RE: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 7 lines · Christoph Hellwig
Quoted 6 lines

Add a new API that returns a virtually non-contigous array of pages and dma address. This API is only implemented for dma-iommu and will not be implemented for non-iommu DMA API instances that have to allocate contiguous memory. It is up to the caller to check if the API is available.

Isn't there already a flag that is only implemented for ARM systems with iommu that forces pages to actually be physically contiguous?

So isn't this some kind of strange opposite?

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#18 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

In reply to Tomasz Figa

Hi Christoph

I have started now to give a try to your patchset. Sorry for the delay.

For uvc I have prepared this patch:
https://github.com/ribalda/linux/commit/9094fe223fe38f8c8ff21366d893b43cbbdf0113

I have tested successfully in a x86_64 noteboot..., yes I know there is no change for that platform :).
I am trying to get hold of an arm device that can run the latest kernel from upstream.

On the meanwhile if you could take a look to the patch to verify that this the way that you expect the drivers to use your api I would appreciate it

Thanks

Quoted 221 lines · Tomasz Figa

+CC Ricardo who will be looking into using this in the USB stack (UVC camera driver).

Quoted 216 lines · Christoph Hellwig

Add a new API that returns a virtually non-contigous array of pages and dma address. This API is only implemented for dma-iommu and will not be implemented for non-iommu DMA API instances that have to allocate contiguous memory. It is up to the caller to check if the API is available.

The intent is that media drivers can use this API if either:

- no kernel mapping or only temporary kernel mappings are required. That is as a better replacement for DMA_ATTR_NO_KERNEL_MAPPING
- a kernel mapping is required for cached and DMA mapped pages, but the driver also needs the pages to e.g. map them to userspace. In that sense it is a replacement for some aspects of the recently removed and never fully implemented DMA_ATTR_NON_CONSISTENT

---
 drivers/iommu/dma-iommu.c   | 73 +++++++++++++++++++++++++------------
 include/linux/dma-mapping.h |  9 +++++
 kernel/dma/mapping.c        | 35 ++++++++++++++++++
 3 files changed, 93 insertions(+), 24 deletions(-)
drivers/iommu/dma-iommu.c +49 -24
index 7922f545cd5eef..158026a856622c 100644
@@ -565,23 +565,12 @@ static struct page **__iommu_dma_alloc_pages(struct device *dev,
        return pages;
 }
 
-/**
- * iommu_dma_alloc_remap - Allocate and map a buffer contiguous in IOVA space
- * @dev: Device to allocate memory for. Must be a real device
- *      attached to an iommu_dma_domain
- * @size: Size of buffer in bytes
- * @dma_handle: Out argument for allocated DMA handle
- * @gfp: Allocation flags
- * @prot: pgprot_t to use for the remapped mapping
- * @attrs: DMA attributes for this allocation
- *
- * If @size is less than PAGE_SIZE, then a full CPU page will be allocated,
+/*
+ * If size is less than PAGE_SIZE, then a full CPU page will be allocated,
  * but an IOMMU which supports smaller pages might not map the whole thing.
- *
- * Return: Mapped virtual address, or NULL on failure.
  */
-static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
-               dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
+static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
+               size_t size, dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
                unsigned long attrs)
 {
        struct iommu_domain *domain = iommu_get_dma_domain(dev);
@@ -593,7 +582,6 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
        struct page **pages;
        struct sg_table sgt;
        dma_addr_t iova;
-       void *vaddr;
 
        *dma_handle = DMA_MAPPING_ERROR;
 
@@ -636,17 +624,10 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
                        < size)
                goto out_free_sg;
 
-       vaddr = dma_common_pages_remap(pages, size, prot,
-                       __builtin_return_address(0));
-       if (!vaddr)
-               goto out_unmap;
-
        *dma_handle = iova;
        sg_free_table(&sgt);
-       return vaddr;
+       return pages;
 
-out_unmap:
-       __iommu_dma_unmap(dev, iova, size);
 out_free_sg:
        sg_free_table(&sgt);
 out_free_iova:
@@ -656,6 +637,46 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
        return NULL;
 }
 
+static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
+               dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
+               unsigned long attrs)
+{
+       struct page **pages;
+       void *vaddr;
+
+       pages = __iommu_dma_alloc_noncontiguous(dev, size, dma_handle, gfp,
+                                               prot, attrs);
+       if (!pages)
+               return NULL;
+       vaddr = dma_common_pages_remap(pages, size, prot,
+                       __builtin_return_address(0));
+       if (!vaddr)
+               goto out_unmap;
+       return vaddr;
+
+out_unmap:
+       __iommu_dma_unmap(dev, *dma_handle, size);
+       __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
+       return NULL;
+}
+
+#ifdef CONFIG_DMA_REMAP
+static struct page **iommu_dma_alloc_noncontiguous(struct device *dev,
+               size_t size, dma_addr_t *dma_handle, gfp_t gfp,
+               unsigned long attrs)
+{
+       return __iommu_dma_alloc_noncontiguous(dev, size, dma_handle, gfp,
+                                              PAGE_KERNEL, attrs);
+}
+
+static void iommu_dma_free_noncontiguous(struct device *dev, size_t size,
+               struct page **pages, dma_addr_t dma_handle)
+{
+       __iommu_dma_unmap(dev, dma_handle, size);
+       __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
+}
+#endif
+
 static void iommu_dma_sync_single_for_cpu(struct device *dev,
                dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
 {
@@ -1110,6 +1131,10 @@ static const struct dma_map_ops iommu_dma_ops = {
        .free                   = iommu_dma_free,
        .alloc_pages            = dma_common_alloc_pages,
        .free_pages             = dma_common_free_pages,
+#ifdef CONFIG_DMA_REMAP
+       .alloc_noncontiguous    = iommu_dma_alloc_noncontiguous,
+       .free_noncontiguous     = iommu_dma_free_noncontiguous,
+#endif
        .mmap                   = iommu_dma_mmap,
        .get_sgtable            = iommu_dma_get_sgtable,
        .map_page               = iommu_dma_map_page,
include/linux/dma-mapping.h +9 -0
index 4b9b1d64f5ec9e..51bbc32365bb8d 100644
@@ -74,6 +74,10 @@ struct dma_map_ops {
                        gfp_t gfp);
        void (*free_pages)(struct device *dev, size_t size, struct page *vaddr,
                        dma_addr_t dma_handle, enum dma_data_direction dir);
+       struct page **(*alloc_noncontiguous)(struct device *dev, size_t size,
+                       dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs);
+       void (*free_noncontiguous)(struct device *dev, size_t size,
+                       struct page **pages, dma_addr_t dma_handle);
        int (*mmap)(struct device *, struct vm_area_struct *,
                          void *, dma_addr_t, size_t,
                          unsigned long attrs);
@@ -384,6 +388,11 @@ void *dma_alloc_noncoherent(struct device *dev, size_t size,
                dma_addr_t *dma_handle, enum dma_data_direction dir, gfp_t gfp);
 void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
                dma_addr_t dma_handle, enum dma_data_direction dir);
+bool dma_can_alloc_noncontiguous(struct device *dev);
+struct page **dma_alloc_noncontiguous(struct device *dev, size_t size,
+               dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs);
+void dma_free_noncontiguous(struct device *dev, size_t size,
+               struct page **pages, dma_addr_t dma_handle);
 
 static inline dma_addr_t dma_map_single_attrs(struct device *dev, void *ptr,
                size_t size, enum dma_data_direction dir, unsigned long attrs)
kernel/dma/mapping.c +35 -1
index 06115f59f4ffbf..6d975d1a20dd72 100644
@@ -529,6 +529,41 @@ void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
 }
 EXPORT_SYMBOL_GPL(dma_free_noncoherent);
 
+bool dma_can_alloc_noncontiguous(struct device *dev)
+{
+       const struct dma_map_ops *ops = get_dma_ops(dev);
+
+       return ops && ops->free_noncontiguous;
+}
+EXPORT_SYMBOL_GPL(dma_can_alloc_noncontiguous);
+
+struct page **dma_alloc_noncontiguous(struct device *dev, size_t size,
+               dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
+{
+       const struct dma_map_ops *ops = get_dma_ops(dev);
+
+       if (WARN_ON_ONCE(!dma_can_alloc_noncontiguous(dev)))
+               return NULL;
+       if (attrs & ~DMA_ATTR_ALLOC_SINGLE_PAGES) {
+               dev_warn(dev, "invalid flags (0x%lx) for %s\n",
+                        attrs, __func__);
+               return NULL;
+       }
+       return ops->alloc_noncontiguous(dev, size, dma_handle, gfp, attrs);
+}
+EXPORT_SYMBOL_GPL(dma_alloc_noncontiguous);
+
+void dma_free_noncontiguous(struct device *dev, size_t size,
+               struct page **pages, dma_addr_t dma_handle)
+{
+       const struct dma_map_ops *ops = get_dma_ops(dev);
+
+       if (WARN_ON_ONCE(!dma_can_alloc_noncontiguous(dev)))
+               return;
+       ops->free_noncontiguous(dev, size, pages, dma_handle);
+}
+EXPORT_SYMBOL_GPL(dma_free_noncontiguous);
+
 int dma_supported(struct device *dev, u64 mask)
 {
        const struct dma_map_ops *ops = get_dma_ops(dev);
--

2.28.0

Signature · 1 line
Ricardo Ribalda
#19 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 7 lines · Ricardo Ribalda

Hi Christoph

I have started now to give a try to your patchset. Sorry for the delay.

For uvc I have prepared this patch:
https://github.com/ribalda/linux/commit/9094fe223fe38f8c8ff21366d893b43cbbdf0113

Quoted 8 lines

I have tested successfully in a x86_64 noteboot..., yes I know there is no change for that platform :).
I am trying to get hold of an arm device that can run the latest kernel from upstream.

On the meanwhile if you could take a look to the patch to verify that this the way that you expect the drivers to use your api I would appreciate it

This looks pretty reaosnable.

Note that ifdef CONFIG_DMA_NONCOHERENT in the old code doesn't actually work, as that option is an internal thing just for mips and sh..

#20 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Hi Christoph

Quoted 14 lines · Christoph Hellwig
Quoted 12 lines · Ricardo Ribalda

Hi Christoph

I have started now to give a try to your patchset. Sorry for the delay.

For uvc I have prepared this patch:
https://github.com/ribalda/linux/commit/9094fe223fe38f8c8ff21366d893b43cbbdf0113

I have tested successfully in a x86_64 noteboot..., yes I know there is no change for that platform :).
I am trying to get hold of an arm device that can run the latest kernel from upstream.

Quoted 6 lines
Quoted 3 lines

On the meanwhile if you could take a look to the patch to verify that this the way that you expect the drivers to use your api I would appreciate it

This looks pretty reaosnable.

Great

Also FYI, I managed to boot an ARM device with that tree. But I could not test the uvc driver (it was a remote device with no usb device attached)

Hopefully I will be able to test it for real this week.

Any suggestions for how to measure performance difference?

Thanks!

Quoted 2 lines

Note that ifdef CONFIG_DMA_NONCOHERENT in the old code doesn't actually work, as that option is an internal thing just for mips and sh..

Signature · 1 line
Ricardo Ribalda
#21 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 7 lines · Ricardo Ribalda

Also FYI, I managed to boot an ARM device with that tree. But I could not test the uvc driver (it was a remote device with no usb device attached)

Hopefully I will be able to test it for real this week.

Any suggestions for how to measure performance difference?

I have to admit I don't know at all how uvc works. But the main problem with dma_alloc_coherent is that all access is uncached. So anything that does larger and/or many data transfers to and from it will be glacially slow. With the dma streaming API we still have to pay for cache flushes, but only before and after the transfers, and in many cases in a somewhat optimized fashion.

#22 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

In reply to Ricardo Ribalda
Quoted 18 lines · Ricardo Ribalda

Hi Christoph

Quoted 14 lines · Christoph Hellwig
Quoted 12 lines · Ricardo Ribalda

Hi Christoph

I have started now to give a try to your patchset. Sorry for the delay.

For uvc I have prepared this patch:
https://github.com/ribalda/linux/commit/9094fe223fe38f8c8ff21366d893b43cbbdf0113

I have tested successfully in a x86_64 noteboot..., yes I know there is no change for that platform :).
I am trying to get hold of an arm device that can run the latest kernel from upstream.

Quoted 9 lines
Quoted 6 lines
Quoted 3 lines

On the meanwhile if you could take a look to the patch to verify that this the way that you expect the drivers to use your api I would appreciate it

This looks pretty reaosnable.

Great

Thanks Christoph for taking a look quickly.

Quoted 7 lines

Also FYI, I managed to boot an ARM device with that tree. But I could not test the uvc driver (it was a remote device with no usb device attached)

Hopefully I will be able to test it for real this week.

Any suggestions for how to measure performance difference?

Back in time Kieran (+CC) shared a patch to add extra statistics for packet processing and payload assembly, with results of various approaches summarized in a spreadsheet:
https://docs.google.com/spreadsheets/d/1uPdbdVcebO9OQ0LQ8hR2LGIEySWgSnGwwhzv7LPXAlU/edit#gid=0

That and just simple CPU usage comparison would be enough.

Quoted 5 lines

Thanks!

Quoted 2 lines

Note that ifdef CONFIG_DMA_NONCOHERENT in the old code doesn't actually work, as that option is an internal thing just for mips and sh..

In what terms it doesn't actually work? Last time I checked some platforms actually defined CONFIG_DMA_NONCOHERENT, so those would instead use the kmalloc() + dma_map() path. I don't have any background on why that was added and whether it needs to be preserved, though. Kieran, Laurent, do you have any insight?

Best regards,
Tomasz
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#23 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Quoted 5 lines · Tomasz Figa

In what terms it doesn't actually work? Last time I checked some platforms actually defined CONFIG_DMA_NONCOHERENT, so those would instead use the kmalloc() + dma_map() path. I don't have any background on why that was added and whether it needs to be preserved, though. Kieran, Laurent, do you have any insight?

CONFIG_DMA_NONCOHERENT is set on sh and mips for platforms that may support non-coherent DMA at compile time (but at least for mips that doesn't actually means this gets used). Using that ifdef to decide on using usb_alloc_coherent vs letting the usb layer map the data seems at best odd, and if we are unlucky papering over a bug somewhere.

#24 lore raw

Re: [PATCH 8/8] WIP: add a dma_alloc_contiguous API

Hi Christoph

I have been testing with real hardware on arm64 your patchset. And uvc performs 20 times better using Kieran's test

https://github.com/ribalda/linux/tree/uvc-noncontiguous

These are the result of running yavta --capture=1000

dma_alloc_noncontiguous

frames:  999
packets: 999
empty:   0 (0 %)
errors:  0
invalid: 0
pts: 0 early, 0 initial, 999 ok
scr: 0 count ok, 0 diff ok
sof: 2048 <= sof <= 0, freq 0.000 kHz
bytes 78466000 : duration 33303
FPS: 29.99
URB: 418105/5000 uS/qty: 83.621 avg 98.783 std 17.396 min 1264.688 max (uS)
header: 100040/5000 uS/qty: 20.008 avg 19.458 std 2.969 min 454.167 max (uS)
latency: 347653/5000 uS/qty: 69.530 avg 98.937 std 9.114 min 1256.875 max (uS)
decode: 70452/5000 uS/qty: 14.090 avg 11.547 std 6.146 min 271.510 max (uS)
raw decode speed: 8.967 Gbits/s
raw URB handling speed: 1.501 Gbits/s
throughput: 18.848 Mbits/s

URB decode CPU usage 0.211500 %

usb_alloc_coherent

frames:  999
packets: 999
empty:   0 (0 %)
errors:  0
invalid: 0
pts: 0 early, 0 initial, 999 ok
scr: 0 count ok, 0 diff ok
sof: 2048 <= sof <= 0, freq 0.000 kHz
bytes 70501712 : duration 33319
FPS: 29.98
URB: 1854128/5000 uS/qty: 370.825 avg 417.133 std 14.539 min 2875.760 max (uS)
header: 98765/5000 uS/qty: 19.753 avg 30.714 std 1.042 min 573.463 max (uS)
latency: 453316/5000 uS/qty: 90.663 avg 114.987 std 4.065 min 860.795 max (uS)
decode: 1400811/5000 uS/qty: 280.162 avg 330.786 std 6.305 min 2758.202 max (uS)
raw decode speed: 402.866 Mbits/s
raw URB handling speed: 304.214 Mbits/s
throughput: 16.927 Mbits/s

URB decode CPU usage 4.204200 %

Best regards

Quoted 13 lines · Christoph Hellwig
Quoted 5 lines · Tomasz Figa

In what terms it doesn't actually work? Last time I checked some platforms actually defined CONFIG_DMA_NONCOHERENT, so those would instead use the kmalloc() + dma_map() path. I don't have any background on why that was added and whether it needs to be preserved, though. Kieran, Laurent, do you have any insight?

CONFIG_DMA_NONCOHERENT is set on sh and mips for platforms that may support non-coherent DMA at compile time (but at least for mips that doesn't actually means this gets used). Using that ifdef to decide on using usb_alloc_coherent vs letting the usb layer map the data seems at best odd, and if we are unlucky papering over a bug somewhere.

Signature · 1 line
Ricardo Ribalda
#25 lore raw

PATCH 1/1 WIP! media: uvcvideo: Use dma_alloc_noncontiguos API

In reply to Christoph Hellwig

On architectures where the is no coherent caching such as ARM use the dma_alloc_noncontiguos API and handle manually the cache flushing using dma_sync_single().

With this patch on the affected architectures we can measure up to 20x performance improvement in uvc_video_copy_data_work().

---

This patch depends on dma_alloc_contiguous API1315351diffmboxseries

https://lore.kernel.org/patchwork/patch/1315351/#1535182

 drivers/media/usb/uvc/uvc_video.c | 69 +++++++++++++++++++++++++------
 drivers/media/usb/uvc/uvcvideo.h  |  1 +
 2 files changed, 58 insertions(+), 12 deletions(-)
drivers/media/usb/uvc/uvc_video.c +57 -12
index ff624bb857d3..ef1b029b8576 100644
@@ -1641,6 +1641,11 @@ static void uvc_video_encode_bulk(struct uvc_urb *uvc_urb,
 	urb->transfer_buffer_length = stream->urb_size - len;
 }
 
+static inline struct device *stream_to_dmadev(struct uvc_streaming *stream)
+{
+	return stream->dev->udev->bus->controller->parent;
+}
+
 static void uvc_video_complete(struct urb *urb)
 {
 	struct uvc_urb *uvc_urb = urb->context;
@@ -1693,6 +1698,11 @@ static void uvc_video_complete(struct urb *urb)
 	 * Process the URB headers, and optionally queue expensive memcpy tasks
 	 * to be deferred to a work queue.
 	 */
+	if (uvc_urb->pages)
+		dma_sync_single_for_cpu(stream_to_dmadev(stream),
+					urb->transfer_dma,
+					urb->transfer_buffer_length,
+					DMA_FROM_DEVICE);
 	stream->decode(uvc_urb, buf, buf_meta);
 
 	/* If no async work is needed, resubmit the URB immediately. */
@@ -1723,8 +1733,15 @@ static void uvc_free_urb_buffers(struct uvc_streaming *stream)
 			continue;
 
 #ifndef CONFIG_DMA_NONCOHERENT
-		usb_free_coherent(stream->dev->udev, stream->urb_size,
-				  uvc_urb->buffer, uvc_urb->dma);
+		if (uvc_urb->pages) {
+			vunmap(uvc_urb->buffer);
+			dma_free_noncontiguous(stream_to_dmadev(stream),
+					       stream->urb_size,
+					       uvc_urb->pages, uvc_urb->dma);
+		} else {
+			usb_free_coherent(stream->dev->udev, stream->urb_size,
+					  uvc_urb->buffer, uvc_urb->dma);
+		}
 #else
 		kfree(uvc_urb->buffer);
 #endif
@@ -1734,6 +1751,42 @@ static void uvc_free_urb_buffers(struct uvc_streaming *stream)
 	stream->urb_size = 0;
 }
 
+#ifndef CONFIG_DMA_NONCOHERENT
+static bool uvc_alloc_urb_buffer(struct uvc_streaming *stream, struct uvc_urb *uvc_urb,
+				 gfp_t gfp_flags)
+{
+	struct device *dma_dev = dma_dev = stream_to_dmadev(stream);
+
+	if (!dma_can_alloc_noncontiguous(dma_dev)) {
+		uvc_urb->buffer = usb_alloc_coherent(stream->dev->udev, stream->urb_size,
+						     gfp_flags | __GFP_NOWARN, &uvc_urb->dma);
+		return uvc_urb->buffer != NULL;
+	}
+
+	uvc_urb->pages = dma_alloc_noncontiguous(dma_dev, stream->urb_size,
+						 &uvc_urb->dma, gfp_flags | __GFP_NOWARN, 0);
+	if (!uvc_urb->pages)
+		return false;
+
+	uvc_urb->buffer = vmap(uvc_urb->pages, PAGE_ALIGN(stream->urb_size) >> PAGE_SHIFT,
+			       VM_DMA_COHERENT, PAGE_KERNEL);
+	if (!uvc_urb->buffer) {
+		dma_free_noncontiguous(dma_dev, stream->urb_size, uvc_urb->pages, uvc_urb->dma);
+		return false;
+	}
+
+	return true;
+}
+#else
+static bool uvc_alloc_urb_buffer(struct uvc_streaming *stream, struct uvc_urb *uvc_urb,
+				 gfp_t gfp_flags)
+{
+	uvc_urb->buffer = kmalloc(stream->urb_size, gfp_flags | __GFP_NOWARN);
+
+	return uvc_urb->buffer != NULL;
+}
+#endif
+
 /*
  * Allocate transfer buffers. This function can be called with buffers
  * already allocated when resuming from suspend, in which case it will
@@ -1764,19 +1817,11 @@ static int uvc_alloc_urb_buffers(struct uvc_streaming *stream,
 
 	/* Retry allocations until one succeed. */
 	for (; npackets > 1; npackets /= 2) {
+		stream->urb_size = psize * npackets;
 		for (i = 0; i < UVC_URBS; ++i) {
 			struct uvc_urb *uvc_urb = &stream->uvc_urb[i];
 
-			stream->urb_size = psize * npackets;
-#ifndef CONFIG_DMA_NONCOHERENT
-			uvc_urb->buffer = usb_alloc_coherent(
-				stream->dev->udev, stream->urb_size,
-				gfp_flags | __GFP_NOWARN, &uvc_urb->dma);
-#else
-			uvc_urb->buffer =
-			    kmalloc(stream->urb_size, gfp_flags | __GFP_NOWARN);
-#endif
-			if (!uvc_urb->buffer) {
+			if (!uvc_alloc_urb_buffer(stream, uvc_urb, gfp_flags)) {
 				uvc_free_urb_buffers(stream);
 				break;
 			}
drivers/media/usb/uvc/uvcvideo.h +1 -0
index 60d830d74ac1..80eeeaf3cd06 100644
@@ -544,6 +544,7 @@ struct uvc_urb {
 
 	char *buffer;
 	dma_addr_t dma;
+	struct page **pages;
 
 	unsigned int async_operations;
 	struct uvc_copy_op copy_operations[UVC_MAX_PACKETS];
Signature · 1 line
2.29.2.299.gdc1121823c-goog
#26 lore raw

Re: [PATCH] WIP! media: uvcvideo: Use dma_alloc_noncontiguos API

Quoted 6 lines · Ricardo Ribalda

On architectures where the is no coherent caching such as ARM use the dma_alloc_noncontiguos API and handle manually the cache flushing using dma_sync_single().

With this patch on the affected architectures we can measure up to 20x performance improvement in uvc_video_copy_data_work().

This has a bunch of crazy long lines, but otherwise looks fine to me.

Quoted 5 lines

---

This patch depends on dma_alloc_contiguous API1315351diffmboxseries

How do we want to proceed? Do the media maintainers want to pick up that patch? Should I pick up the media patch in the dma-mapping tree?

Can you respost a combined series to get started?

#27 lore raw

Re: [PATCH] WIP! media: uvcvideo: Use dma_alloc_noncontiguos API

HI Christoph

Quoted 10 lines · Christoph Hellwig
Quoted 6 lines · Ricardo Ribalda

On architectures where the is no coherent caching such as ARM use the dma_alloc_noncontiguos API and handle manually the cache flushing using dma_sync_single().

With this patch on the affected architectures we can measure up to 20x performance improvement in uvc_video_copy_data_work().

This has a bunch of crazy long lines, but otherwise looks fine to me.

That is easy to solve :)

https://github.com/ribalda/linux/commit/17ab65a08302e845ad7ae7775ce54b387a58a887

Quoted 9 lines
Quoted 6 lines

---

This patch depends on dma_alloc_contiguous API1315351diffmboxseries

How do we want to proceed? Do the media maintainers want to pick up that patch? Should I pick up the media patch in the dma-mapping tree?

I was hoping that you could answer that question :).

Do you have other use-cases than linux-media in mind?

I think Sergey wants to experiment also with vb2, to figure out how much it affects it.
His change will be much more complicated than mine thought, there are more cornercases there.

Quoted 2 lines

Can you respost a combined series to get started?

Sure. Shall I also include the profiling patch?

Best regards

Signature · 1 line
Ricardo Ribalda
#28 lore raw

Re: [PATCH] WIP! media: uvcvideo: Use dma_alloc_noncontiguos API

Quoted 8 lines · Ricardo Ribalda

I was hoping that you could answer that question :).

Do you have other use-cases than linux-media in mind?

I think Sergey wants to experiment also with vb2, to figure out how much it affects it.
His change will be much more complicated than mine thought, there are more cornercases there.

I don't have anything urgend lined up, although I think there are plenty other potential use cases.

Quoted 3 lines
Quoted 1 line

Can you respost a combined series to get started?

Sure. Shall I also include the profiling patch?

That is in the media code, right? I don't really care too much.