Skip to contentSkip to conversation

dma: fix DMA_ATTR_NO_KERNEL_MAPPING for no-IOMMU platforms

PATCH Superseded Carlo Caione 3 participants 6 messages Feb 3, 2015 – Feb 4, 2015

Subsystems (4)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover dma: fix DMA_ATTR_NO_KERNEL_MAPPING for no-IOMMU platforms

The DMA_ATTR_NO_KERNEL_MAPPING attribute is used to notify dma-mapping core that the driver will not use kernel mapping for the allocated buffer at all, so the core can skip creating it.

Unfortunately at the moment this attribute is only valid for IOMMU setups. In non-IOMMU setups the codepath doesn't obey DMA_ATTR_NO_KERNEL_MAPPING so when the CMA region is in high-memory all the buffers created with this attribute that do not require a kernel virtual address space still put pressure on the vmalloc area (for reference see
http://lists.infradead.org/pipermail/linux-arm-kernel/2014-August/279325.html)

This patchset is composed by two patches.

The first patch fixes the Exynos DRM driver so that it keeps working when the non-IOMMU DMA layer is fixed. The Exynos DRM driver doesn't follow the recommendations in DMA-attributes.txt so in the non-IOMMU case when DMA_ATTR_NO_KERNEL_MAPPING is used the driver directly uses the returned kernel virtual address which it explicitly requested not to be assigned. That must be fixed before the underlying DMA subsystem is improved to obey DMA_ATTR_NO_KERNEL_MAPPING.

The second patch implements DMA_ATTR_NO_KERNEL_MAPPING for non-IOMMU capable platforms.

Please note that:

* The first patch must be applied before the second one to avoid breaking the Exynos DRM driver. This patch without the DMA layer fix works fine (we save git-bisect) but it's not fully correct (we have double mapping in kernel address space).

* The second patch breaks the Exynos DRM driver unless the first patch is
* applied first

Carlo Caione (1):
drm/exynos: fix DMA_ATTR_NO_KERNEL_MAPPING usage

Jasper St. Pierre (1):
arm/dma-mapping: Respect NO_KERNEL_MAPPING when we don't have an IOMMU

 arch/arm/mm/dma-mapping.c                 | 67 +++++++++++++++++++------------
 drivers/gpu/drm/exynos/exynos_drm_buf.c   |  6 +--
 drivers/gpu/drm/exynos/exynos_drm_fbdev.c | 29 +++++--------
 drivers/gpu/drm/exynos/exynos_drm_gem.h   |  2 +
 4 files changed, 55 insertions(+), 49 deletions(-)
Signature · 1 line
2.2.2
#1 lore raw

PATCH 1/2 drm/exynos: fix DMA_ATTR_NO_KERNEL_MAPPING usage

The Exynos DRM driver doesn't follow the correct API when dealing with dma_{alloc, mmap, free}_attrs functions and the DMA_ATTR_NO_KERNEL_MAPPING attribute.

When a IOMMU is not available and the DMA_ATTR_NO_KERNEL_MAPPING is used, the driver should use the pointer returned by dma_alloc_attr() as a cookie.

The Exynos DRM driver directly uses the non-requested virtual kernel address returned by the DMA mapping subsystem. This just works now because the non-IOMMU codepath doesn't obey DMA_ATTR_NO_KERNEL_MAPPING but we need to fix it before fixing the DMA layer.

---
 drivers/gpu/drm/exynos/exynos_drm_buf.c   |  6 +++---
 drivers/gpu/drm/exynos/exynos_drm_fbdev.c | 29 +++++++++--------------------
 drivers/gpu/drm/exynos/exynos_drm_gem.h   |  2 ++
 3 files changed, 14 insertions(+), 23 deletions(-)
drivers/gpu/drm/exynos/exynos_drm_buf.c +3 -3
index 9c80884..24994ba 100644
@@ -63,11 +63,11 @@ static int lowlevel_buffer_allocate(struct drm_device *dev,
 			return -ENOMEM;
 		}
 
-		buf->kvaddr = (void __iomem *)dma_alloc_attrs(dev->dev,
+		buf->cookie = dma_alloc_attrs(dev->dev,
 					buf->size,
 					&buf->dma_addr, GFP_KERNEL,
 					&buf->dma_attrs);
-		if (!buf->kvaddr) {
+		if (!buf->cookie) {
 			DRM_ERROR("failed to allocate buffer.\n");
 			ret = -ENOMEM;
 			goto err_free;
@@ -132,7 +132,7 @@ static void lowlevel_buffer_deallocate(struct drm_device *dev,
 	buf->sgt = NULL;
 
 	if (!is_drm_iommu_supported(dev)) {
-		dma_free_attrs(dev->dev, buf->size, buf->kvaddr,
+		dma_free_attrs(dev->dev, buf->size, buf->cookie,
 				(dma_addr_t)buf->dma_addr, &buf->dma_attrs);
 		drm_free_large(buf->pages);
 	} else
drivers/gpu/drm/exynos/exynos_drm_fbdev.c +9 -20
index e12ea90..84f8dfe 100644
@@ -79,9 +79,9 @@ static int exynos_drm_fbdev_update(struct drm_fb_helper *helper,
 				     struct drm_framebuffer *fb)
 {
 	struct fb_info *fbi = helper->fbdev;
-	struct drm_device *dev = helper->dev;
 	struct exynos_drm_gem_buf *buffer;
 	unsigned int size = fb->width * fb->height * (fb->bits_per_pixel >> 3);
+	unsigned int nr_pages;
 	unsigned long offset;
 
 	drm_fb_helper_fill_fix(fbi, fb->pitches[0], fb->depth);
@@ -94,25 +94,14 @@ static int exynos_drm_fbdev_update(struct drm_fb_helper *helper,
 		return -EFAULT;
 	}
 
-	/* map pages with kernel virtual space. */
+	nr_pages = buffer->size >> PAGE_SHIFT;
+
+	buffer->kvaddr = (void __iomem *) vmap(buffer->pages,
+			nr_pages, VM_MAP,
+			pgprot_writecombine(PAGE_KERNEL));
 	if (!buffer->kvaddr) {
-		if (is_drm_iommu_supported(dev)) {
-			unsigned int nr_pages = buffer->size >> PAGE_SHIFT;
-
-			buffer->kvaddr = (void __iomem *) vmap(buffer->pages,
-					nr_pages, VM_MAP,
-					pgprot_writecombine(PAGE_KERNEL));
-		} else {
-			phys_addr_t dma_addr = buffer->dma_addr;
-			if (dma_addr)
-				buffer->kvaddr = (void __iomem *)phys_to_virt(dma_addr);
-			else
-				buffer->kvaddr = (void __iomem *)NULL;
-		}
-		if (!buffer->kvaddr) {
-			DRM_ERROR("failed to map pages to kernel space.\n");
-			return -EIO;
-		}
+		DRM_ERROR("failed to map pages to kernel space.\n");
+		return -EIO;
 	}
 
 	/* buffer count to framebuffer always is 1 at booting time. */
@@ -313,7 +302,7 @@ static void exynos_drm_fbdev_destroy(struct drm_device *dev,
 	struct exynos_drm_gem_obj *exynos_gem_obj = exynos_fbd->exynos_gem_obj;
 	struct drm_framebuffer *fb;
 
-	if (is_drm_iommu_supported(dev) && exynos_gem_obj->buffer->kvaddr)
+	if (exynos_gem_obj->buffer->kvaddr)
 		vunmap(exynos_gem_obj->buffer->kvaddr);
 
 	/* release drm framebuffer and real buffer */
drivers/gpu/drm/exynos/exynos_drm_gem.h +2 -0
index ec58fe9..308173c 100644
@@ -22,6 +22,7 @@
 /*
  * exynos drm gem buffer structure.
  *
+ * @cookie: cookie returned by dma_alloc_attrs
  * @kvaddr: kernel virtual address to allocated memory region.
  * *userptr: user space address.
  * @dma_addr: bus address(accessed by dma) to allocated memory region.
@@ -35,6 +36,7 @@
  *	VM_PFNMAP or not.
  */
 struct exynos_drm_gem_buf {
+	void 			*cookie;
 	void __iomem		*kvaddr;
 	unsigned long		userptr;
 	dma_addr_t		dma_addr;
Signature · 1 line
2.2.2
#2 lore raw

PATCH 1/2 drm/exynos: fix DMA_ATTR_NO_KERNEL_MAPPING usage

Hi,

Quoted 14 lines · Carlo Caione

The Exynos DRM driver doesn't follow the correct API when dealing with dma_{alloc, mmap, free}_attrs functions and the DMA_ATTR_NO_KERNEL_MAPPING attribute.

When a IOMMU is not available and the DMA_ATTR_NO_KERNEL_MAPPING is used, the driver should use the pointer returned by dma_alloc_attr() as a cookie.

The Exynos DRM driver directly uses the non-requested virtual kernel address returned by the DMA mapping subsystem. This just works now because the non-IOMMU codepath doesn't obey DMA_ATTR_NO_KERNEL_MAPPING but we need to fix it before fixing the DMA layer.

Quoted 7 lines
---
 drivers/gpu/drm/exynos/exynos_drm_buf.c   |  6 +++---
 drivers/gpu/drm/exynos/exynos_drm_fbdev.c | 29 +++++++++--------------------
 drivers/gpu/drm/exynos/exynos_drm_gem.h   |  2 ++
 3 files changed, 14 insertions(+), 23 deletions(-)

Thanks.

#3 lore raw

PATCH 2/2 arm/dma-mapping: Respect NO_KERNEL_MAPPING when we don't have an IOMMU

In reply to Carlo Caione

From: "Jasper St. Pierre" <[email protected]>

Even without an iommu, NO_KERNEL_MAPPING is still convenient to save on kernel address space in places where we don't need a kernel mapping. Implement support for it in the two places where we're creating an expensive mapping.

__alloc_from_pool uses an internal pool from which we already have virtual addresses, so it's not relevant, and __alloc_simple_buffer uses alloc_pages, which will always return a lowmem page, which is already mapped into kernel space, so we can't prevent a mapping for it in that case.

---
 arch/arm/mm/dma-mapping.c | 67 +++++++++++++++++++++++++++++------------------
 1 file changed, 41 insertions(+), 26 deletions(-)
arch/arm/mm/dma-mapping.c +41 -26
index a673c7f..6843293 100644
@@ -289,11 +289,11 @@ static void __dma_free_buffer(struct page *page, size_t size)
 
 static void *__alloc_from_contiguous(struct device *dev, size_t size,
 				     pgprot_t prot, struct page **ret_page,
-				     const void *caller);
+				     const void *caller, bool want_vaddr);
 
 static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
 				 pgprot_t prot, struct page **ret_page,
-				 const void *caller);
+				 const void *caller, bool want_vaddr);
 
 static void *
 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot,
@@ -357,10 +357,10 @@ static int __init atomic_pool_init(void)
 
 	if (dev_get_cma_area(NULL))
 		ptr = __alloc_from_contiguous(NULL, atomic_pool_size, prot,
-					      &page, atomic_pool_init);
+					      &page, atomic_pool_init, true);
 	else
 		ptr = __alloc_remap_buffer(NULL, atomic_pool_size, gfp, prot,
-					   &page, atomic_pool_init);
+					   &page, atomic_pool_init, true);
 	if (ptr) {
 		int ret;
 
@@ -467,13 +467,15 @@ static void __dma_remap(struct page *page, size_t size, pgprot_t prot)
 
 static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
 				 pgprot_t prot, struct page **ret_page,
-				 const void *caller)
+				 const void *caller, bool want_vaddr)
 {
 	struct page *page;
-	void *ptr;
+	void *ptr = NULL;
 	page = __dma_alloc_buffer(dev, size, gfp);
 	if (!page)
 		return NULL;
+	if (!want_vaddr)
+		goto out;
 
 	ptr = __dma_alloc_remap(page, size, gfp, prot, caller);
 	if (!ptr) {
@@ -481,6 +483,7 @@ static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
 		return NULL;
 	}
 
+ out:
 	*ret_page = page;
 	return ptr;
 }
@@ -523,12 +526,12 @@ static int __free_from_pool(void *start, size_t size)
 
 static void *__alloc_from_contiguous(struct device *dev, size_t size,
 				     pgprot_t prot, struct page **ret_page,
-				     const void *caller)
+				     const void *caller, bool want_vaddr)
 {
 	unsigned long order = get_order(size);
 	size_t count = size >> PAGE_SHIFT;
 	struct page *page;
-	void *ptr;
+	void *ptr = NULL;
 
 	page = dma_alloc_from_contiguous(dev, count, order);
 	if (!page)
@@ -536,6 +539,9 @@ static void *__alloc_from_contiguous(struct device *dev, size_t size,
 
 	__dma_clear_buffer(page, size);
 
+	if (!want_vaddr)
+		goto out;
+
 	if (PageHighMem(page)) {
 		ptr = __dma_alloc_remap(page, size, GFP_KERNEL, prot, caller);
 		if (!ptr) {
@@ -546,17 +552,21 @@ static void *__alloc_from_contiguous(struct device *dev, size_t size,
 		__dma_remap(page, size, prot);
 		ptr = page_address(page);
 	}
+
+ out:
 	*ret_page = page;
 	return ptr;
 }
 
 static void __free_from_contiguous(struct device *dev, struct page *page,
-				   void *cpu_addr, size_t size)
+				   void *cpu_addr, size_t size, bool want_vaddr)
 {
-	if (PageHighMem(page))
-		__dma_free_remap(cpu_addr, size);
-	else
-		__dma_remap(page, size, PAGE_KERNEL);
+	if (want_vaddr) {
+		if (PageHighMem(page))
+			__dma_free_remap(cpu_addr, size);
+		else
+			__dma_remap(page, size, PAGE_KERNEL);
+	}
 	dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
 }
 
@@ -574,12 +584,12 @@ static inline pgprot_t __get_dma_pgprot(struct dma_attrs *attrs, pgprot_t prot)
 
 #define nommu() 1
 
-#define __get_dma_pgprot(attrs, prot)	__pgprot(0)
-#define __alloc_remap_buffer(dev, size, gfp, prot, ret, c)	NULL
+#define __get_dma_pgprot(attrs, prot)				__pgprot(0)
+#define __alloc_remap_buffer(dev, size, gfp, prot, ret, c, wv)	NULL
 #define __alloc_from_pool(size, ret_page)			NULL
-#define __alloc_from_contiguous(dev, size, prot, ret, c)	NULL
+#define __alloc_from_contiguous(dev, size, prot, ret, c, wv)	NULL
 #define __free_from_pool(cpu_addr, size)			0
-#define __free_from_contiguous(dev, page, cpu_addr, size)	do { } while (0)
+#define __free_from_contiguous(dev, page, cpu_addr, size, wv)	do { } while (0)
 #define __dma_free_remap(cpu_addr, size)			do { } while (0)
 
 #endif	/* CONFIG_MMU */
@@ -599,11 +609,13 @@ static void *__alloc_simple_buffer(struct device *dev, size_t size, gfp_t gfp,
 
 
 static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
-			 gfp_t gfp, pgprot_t prot, bool is_coherent, const void *caller)
+			 gfp_t gfp, pgprot_t prot, bool is_coherent,
+			 struct dma_attrs *attrs, const void *caller)
 {
 	u64 mask = get_coherent_dma_mask(dev);
 	struct page *page = NULL;
 	void *addr;
+	bool want_vaddr;
 
 #ifdef CONFIG_DMA_API_DEBUG
 	u64 limit = (mask + 1) & ~mask;
@@ -631,20 +643,21 @@ static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
 
 	*handle = DMA_ERROR_CODE;
 	size = PAGE_ALIGN(size);
+	want_vaddr = !dma_get_attr(DMA_ATTR_NO_KERNEL_MAPPING, attrs);
 
 	if (is_coherent || nommu())
 		addr = __alloc_simple_buffer(dev, size, gfp, &page);
 	else if (!(gfp & __GFP_WAIT))
 		addr = __alloc_from_pool(size, &page);
 	else if (!dev_get_cma_area(dev))
-		addr = __alloc_remap_buffer(dev, size, gfp, prot, &page, caller);
+		addr = __alloc_remap_buffer(dev, size, gfp, prot, &page, caller, want_vaddr);
 	else
-		addr = __alloc_from_contiguous(dev, size, prot, &page, caller);
+		addr = __alloc_from_contiguous(dev, size, prot, &page, caller, want_vaddr);
 
-	if (addr)
+	if (page)
 		*handle = pfn_to_dma(dev, page_to_pfn(page));
 
-	return addr;
+	return want_vaddr ? addr : &page;
 }
 
 /*
@@ -661,7 +674,7 @@ void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
 		return memory;
 
 	return __dma_alloc(dev, size, handle, gfp, prot, false,
-			   __builtin_return_address(0));
+			   attrs, __builtin_return_address(0));
 }
 
 static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
@@ -674,7 +687,7 @@ static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
 		return memory;
 
 	return __dma_alloc(dev, size, handle, gfp, prot, true,
-			   __builtin_return_address(0));
+			   attrs, __builtin_return_address(0));
 }
 
 /*
@@ -715,6 +728,7 @@ static void __arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
 			   bool is_coherent)
 {
 	struct page *page = pfn_to_page(dma_to_pfn(dev, handle));
+	bool want_vaddr = !dma_get_attr(DMA_ATTR_NO_KERNEL_MAPPING, attrs);
 
 	if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
 		return;
@@ -726,14 +740,15 @@ static void __arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
 	} else if (__free_from_pool(cpu_addr, size)) {
 		return;
 	} else if (!dev_get_cma_area(dev)) {
-		__dma_free_remap(cpu_addr, size);
+		if (want_vaddr)
+			__dma_free_remap(cpu_addr, size);
 		__dma_free_buffer(page, size);
 	} else {
 		/*
 		 * Non-atomic allocations cannot be freed with IRQs disabled
 		 */
 		WARN_ON(irqs_disabled());
-		__free_from_contiguous(dev, page, cpu_addr, size);
+		__free_from_contiguous(dev, page, cpu_addr, size, want_vaddr);
 	}
 }
 
Signature · 1 line
2.2.2
#4 lore raw

PATCH 2/2 arm/dma-mapping: Respect NO_KERNEL_MAPPING when we don't have an IOMMU

Quoted 177 lines · Carlo Caione

From: "Jasper St. Pierre" <[email protected]>

Even without an iommu, NO_KERNEL_MAPPING is still convenient to save on kernel address space in places where we don't need a kernel mapping. Implement support for it in the two places where we're creating an expensive mapping.

__alloc_from_pool uses an internal pool from which we already have virtual addresses, so it's not relevant, and __alloc_simple_buffer uses alloc_pages, which will always return a lowmem page, which is already mapped into kernel space, so we can't prevent a mapping for it in that case.

---
  arch/arm/mm/dma-mapping.c | 67 +++++++++++++++++++++++++++++------------------
  1 file changed, 41 insertions(+), 26 deletions(-)
arch/arm/mm/dma-mapping.c +32 -20
index a673c7f..6843293 100644
@@ -289,11 +289,11 @@ static void __dma_free_buffer(struct page *page, size_t size)
 
  static void *__alloc_from_contiguous(struct device *dev, size_t size,
  				     pgprot_t prot, struct page **ret_page,
-				     const void *caller);
+				     const void *caller, bool want_vaddr);
 
  static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
  				 pgprot_t prot, struct page **ret_page,
-				 const void *caller);
+				 const void *caller, bool want_vaddr);
 
  static void *
  __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot,
@@ -357,10 +357,10 @@ static int __init atomic_pool_init(void)
 
  	if (dev_get_cma_area(NULL))
  		ptr = __alloc_from_contiguous(NULL, atomic_pool_size, prot,
-					      &page, atomic_pool_init);
+					      &page, atomic_pool_init, true);
  	else
  		ptr = __alloc_remap_buffer(NULL, atomic_pool_size, gfp, prot,
-					   &page, atomic_pool_init);
+					   &page, atomic_pool_init, true);
  	if (ptr) {
  		int ret;
 
@@ -467,13 +467,15 @@ static void __dma_remap(struct page *page, size_t size, pgprot_t prot)
 
  static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
  				 pgprot_t prot, struct page **ret_page,
-				 const void *caller)
+				 const void *caller, bool want_vaddr)
  {
  	struct page *page;
-	void *ptr;
+	void *ptr = NULL;
  	page = __dma_alloc_buffer(dev, size, gfp);
  	if (!page)
  		return NULL;
+	if (!want_vaddr)
+		goto out;
 
  	ptr = __dma_alloc_remap(page, size, gfp, prot, caller);
  	if (!ptr) {
@@ -481,6 +483,7 @@ static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
  		return NULL;
  	}
 
+ out:
  	*ret_page = page;
  	return ptr;
  }
@@ -523,12 +526,12 @@ static int __free_from_pool(void *start, size_t size)
 
  static void *__alloc_from_contiguous(struct device *dev, size_t size,
  				     pgprot_t prot, struct page **ret_page,
-				     const void *caller)
+				     const void *caller, bool want_vaddr)
  {
  	unsigned long order = get_order(size);
  	size_t count = size >> PAGE_SHIFT;
  	struct page *page;
-	void *ptr;
+	void *ptr = NULL;
 
  	page = dma_alloc_from_contiguous(dev, count, order);
  	if (!page)
@@ -536,6 +539,9 @@ static void *__alloc_from_contiguous(struct device *dev, size_t size,
 
  	__dma_clear_buffer(page, size);
 
+	if (!want_vaddr)
+		goto out;
+
  	if (PageHighMem(page)) {
  		ptr = __dma_alloc_remap(page, size, GFP_KERNEL, prot, caller);
  		if (!ptr) {
@@ -546,17 +552,21 @@ static void *__alloc_from_contiguous(struct device *dev, size_t size,
  		__dma_remap(page, size, prot);
  		ptr = page_address(page);
  	}
+
+ out:
  	*ret_page = page;
  	return ptr;
  }
 
  static void __free_from_contiguous(struct device *dev, struct page *page,
-				   void *cpu_addr, size_t size)
+				   void *cpu_addr, size_t size, bool want_vaddr)
  {
-	if (PageHighMem(page))
-		__dma_free_remap(cpu_addr, size);
-	else
-		__dma_remap(page, size, PAGE_KERNEL);
+	if (want_vaddr) {
+		if (PageHighMem(page))
+			__dma_free_remap(cpu_addr, size);
+		else
+			__dma_remap(page, size, PAGE_KERNEL);
+	}
  	dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
  }
 
@@ -574,12 +584,12 @@ static inline pgprot_t __get_dma_pgprot(struct dma_attrs *attrs, pgprot_t prot)
 
  #define nommu() 1
 
-#define __get_dma_pgprot(attrs, prot)	__pgprot(0)
-#define __alloc_remap_buffer(dev, size, gfp, prot, ret, c)	NULL
+#define __get_dma_pgprot(attrs, prot)				__pgprot(0)
+#define __alloc_remap_buffer(dev, size, gfp, prot, ret, c, wv)	NULL
  #define __alloc_from_pool(size, ret_page)			NULL
-#define __alloc_from_contiguous(dev, size, prot, ret, c)	NULL
+#define __alloc_from_contiguous(dev, size, prot, ret, c, wv)	NULL
  #define __free_from_pool(cpu_addr, size)			0
-#define __free_from_contiguous(dev, page, cpu_addr, size)	do { } while (0)
+#define __free_from_contiguous(dev, page, cpu_addr, size, wv)	do { } while (0)
  #define __dma_free_remap(cpu_addr, size)			do { } while (0)
 
  #endif	/* CONFIG_MMU */
@@ -599,11 +609,13 @@ static void *__alloc_simple_buffer(struct device *dev, size_t size, gfp_t gfp,
 
 
  static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
-			 gfp_t gfp, pgprot_t prot, bool is_coherent, const void *caller)
+			 gfp_t gfp, pgprot_t prot, bool is_coherent,
+			 struct dma_attrs *attrs, const void *caller)
  {
  	u64 mask = get_coherent_dma_mask(dev);
  	struct page *page = NULL;
  	void *addr;
+	bool want_vaddr;
 
  #ifdef CONFIG_DMA_API_DEBUG
  	u64 limit = (mask + 1) & ~mask;
@@ -631,20 +643,21 @@ static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
 
  	*handle = DMA_ERROR_CODE;
  	size = PAGE_ALIGN(size);
+	want_vaddr = !dma_get_attr(DMA_ATTR_NO_KERNEL_MAPPING, attrs);
 
  	if (is_coherent || nommu())
  		addr = __alloc_simple_buffer(dev, size, gfp, &page);
  	else if (!(gfp & __GFP_WAIT))
  		addr = __alloc_from_pool(size, &page);
  	else if (!dev_get_cma_area(dev))
-		addr = __alloc_remap_buffer(dev, size, gfp, prot, &page, caller);
+		addr = __alloc_remap_buffer(dev, size, gfp, prot, &page, caller, want_vaddr);
  	else
-		addr = __alloc_from_contiguous(dev, size, prot, &page, caller);

arch/arm/mm/dma-mapping.c static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, · line 662

Quoted 10 lines
arch/arm/mm/dma-mapping.c +3 -2
+		addr = __alloc_from_contiguous(dev, size, prot, &page, caller, want_vaddr);
 
-	if (addr)
+	if (page)
  		*handle = pfn_to_dma(dev, page_to_pfn(page));
 
-	return addr;
+	return want_vaddr ? addr : &page;
  }
 

What happens if __alloc_remap_buffer or __alloc_from_contiguous fails to allocate?
From this, its seems like we will return &page which will always be non-null
so it will look like a success.

Thanks,
Laura

Signature · 3 lines
Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
#5 lore raw

PATCH 2/2 arm/dma-mapping: Respect NO_KERNEL_MAPPING when we don't have an IOMMU

Quoted 216 lines · Laura Abbott
Quoted 215 lines · Carlo Caione

From: "Jasper St. Pierre" <[email protected]>

Even without an iommu, NO_KERNEL_MAPPING is still convenient to save on kernel address space in places where we don't need a kernel mapping. Implement support for it in the two places where we're creating an expensive mapping.

__alloc_from_pool uses an internal pool from which we already have virtual addresses, so it's not relevant, and __alloc_simple_buffer uses alloc_pages, which will always return a lowmem page, which is already mapped into kernel space, so we can't prevent a mapping for it in that case.

---
  arch/arm/mm/dma-mapping.c | 67
+++++++++++++++++++++++++++++------------------
  1 file changed, 41 insertions(+), 26 deletions(-)
arch/arm/mm/dma-mapping.c +2 -2
index a673c7f..6843293 100644
@@ -289,11 +289,11 @@ static void __dma_free_buffer(struct page *page,
 size_t size)
 
  static void *__alloc_from_contiguous(struct device *dev, size_t size,
                                     pgprot_t prot, struct page
 **ret_page,
-                                    const void *caller);
+                                    const void *caller, bool want_vaddr);
 
  static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t
 gfp,
                                 pgprot_t prot, struct page **ret_page,
-                                const void *caller);
+                                const void *caller, bool want_vaddr);

static void *
__dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t
prot,

arch/arm/mm/dma-mapping.c +2 -2
@@ -357,10 +357,10 @@ static int __init atomic_pool_init(void)
 
        if (dev_get_cma_area(NULL))
                ptr = __alloc_from_contiguous(NULL, atomic_pool_size,
 prot,
-                                             &page, atomic_pool_init);
+                                             &page, atomic_pool_init,
 true);
        else
                ptr = __alloc_remap_buffer(NULL, atomic_pool_size, gfp,
 prot,
-                                          &page, atomic_pool_init);
+                                          &page, atomic_pool_init, true);
        if (ptr) {
                int ret;
arch/arm/mm/dma-mapping.c +4 -2
@@ -467,13 +467,15 @@ static void __dma_remap(struct page *page, size_t
 size, pgprot_t prot)
 
  static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t
 gfp,
                                 pgprot_t prot, struct page **ret_page,
-                                const void *caller)
+                                const void *caller, bool want_vaddr)
  {
        struct page *page;
-       void *ptr;
+       void *ptr = NULL;
        page = __dma_alloc_buffer(dev, size, gfp);
        if (!page)
                return NULL;
+       if (!want_vaddr)
+               goto out;
 
        ptr = __dma_alloc_remap(page, size, gfp, prot, caller);
        if (!ptr) {
arch/arm/mm/dma-mapping.c +1 -0
@@ -481,6 +483,7 @@ static void *__alloc_remap_buffer(struct device *dev,
 size_t size, gfp_t gfp,
                return NULL;
        }
 
+ out:
        *ret_page = page;
        return ptr;
  }
arch/arm/mm/dma-mapping.c +2 -2
@@ -523,12 +526,12 @@ static int __free_from_pool(void *start, size_t
 size)
 
  static void *__alloc_from_contiguous(struct device *dev, size_t size,
                                     pgprot_t prot, struct page
 **ret_page,
-                                    const void *caller)
+                                    const void *caller, bool want_vaddr)
  {
        unsigned long order = get_order(size);
        size_t count = size >> PAGE_SHIFT;
        struct page *page;
-       void *ptr;
+       void *ptr = NULL;
 
        page = dma_alloc_from_contiguous(dev, count, order);
        if (!page)
arch/arm/mm/dma-mapping.c +3 -0
@@ -536,6 +539,9 @@ static void *__alloc_from_contiguous(struct device
 *dev, size_t size,
 
        __dma_clear_buffer(page, size);
 
+       if (!want_vaddr)
+               goto out;
+
        if (PageHighMem(page)) {
                ptr = __dma_alloc_remap(page, size, GFP_KERNEL, prot,
caller);
                if (!ptr) {
arch/arm/mm/dma-mapping.c +9 -5
@@ -546,17 +552,21 @@ static void *__alloc_from_contiguous(struct device
 *dev, size_t size,
                __dma_remap(page, size, prot);
                ptr = page_address(page);
        }
+
+ out:
        *ret_page = page;
        return ptr;
  }
 
  static void __free_from_contiguous(struct device *dev, struct page
 *page,
-                                  void *cpu_addr, size_t size)
+                                  void *cpu_addr, size_t size, bool
 want_vaddr)
  {
-       if (PageHighMem(page))
-               __dma_free_remap(cpu_addr, size);
-       else
-               __dma_remap(page, size, PAGE_KERNEL);
+       if (want_vaddr) {
+               if (PageHighMem(page))
+                       __dma_free_remap(cpu_addr, size);
+               else
+                       __dma_remap(page, size, PAGE_KERNEL);
+       }
        dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);

}

arch/arm/mm/dma-mapping.c +4 -4
@@ -574,12 +584,12 @@ static inline pgprot_t __get_dma_pgprot(struct
 dma_attrs *attrs, pgprot_t prot)
 
  #define nommu() 1
 
-#define __get_dma_pgprot(attrs, prot)  __pgprot(0)
-#define __alloc_remap_buffer(dev, size, gfp, prot, ret, c)     NULL
+#define __get_dma_pgprot(attrs, prot)
 __pgprot(0)
+#define __alloc_remap_buffer(dev, size, gfp, prot, ret, c, wv) NULL
  #define __alloc_from_pool(size, ret_page)                     NULL
-#define __alloc_from_contiguous(dev, size, prot, ret, c)       NULL
+#define __alloc_from_contiguous(dev, size, prot, ret, c, wv)   NULL
  #define __free_from_pool(cpu_addr, size)                      0
-#define __free_from_contiguous(dev, page, cpu_addr, size)      do { }
 while (0)
+#define __free_from_contiguous(dev, page, cpu_addr, size, wv)  do { }

while (0)
#define __dma_free_remap(cpu_addr, size) do { }
while (0)

  #endif        /* CONFIG_MMU */
arch/arm/mm/dma-mapping.c +3 -1
@@ -599,11 +609,13 @@ static void *__alloc_simple_buffer(struct device
 *dev, size_t size, gfp_t gfp,
 
 
  static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t
 *handle,
-                        gfp_t gfp, pgprot_t prot, bool is_coherent, const
 void *caller)
+                        gfp_t gfp, pgprot_t prot, bool is_coherent,
+                        struct dma_attrs *attrs, const void *caller)
  {
        u64 mask = get_coherent_dma_mask(dev);
        struct page *page = NULL;
        void *addr;
+       bool want_vaddr;
  #ifdef CONFIG_DMA_API_DEBUG
        u64 limit = (mask + 1) & ~mask;
arch/arm/mm/dma-mapping.c +4 -3
@@ -631,20 +643,21 @@ static void *__dma_alloc(struct device *dev, size_t
 size, dma_addr_t *handle,
 
        *handle = DMA_ERROR_CODE;
        size = PAGE_ALIGN(size);
+       want_vaddr = !dma_get_attr(DMA_ATTR_NO_KERNEL_MAPPING, attrs);
 
        if (is_coherent || nommu())
                addr = __alloc_simple_buffer(dev, size, gfp, &page);
        else if (!(gfp & __GFP_WAIT))
                addr = __alloc_from_pool(size, &page);
        else if (!dev_get_cma_area(dev))
-               addr = __alloc_remap_buffer(dev, size, gfp, prot, &page,
 caller);
+               addr = __alloc_remap_buffer(dev, size, gfp, prot, &page,
 caller, want_vaddr);
        else
-               addr = __alloc_from_contiguous(dev, size, prot, &page,
 caller);
+               addr = __alloc_from_contiguous(dev, size, prot, &page,
 caller, want_vaddr);
 
-       if (addr)
+       if (page)
                *handle = pfn_to_dma(dev, page_to_pfn(page));
Quoted 8 lines

arch/arm/mm/dma-mapping.c static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, · line 661

Quoted 4 lines
arch/arm/mm/dma-mapping.c +1 -1
-       return addr;
+       return want_vaddr ? addr : &page;
  }

What happens if __alloc_remap_buffer or __alloc_from_contiguous fails to
allocate?
From this, its seems like we will return &page which will always be non-null so it will look like a success.

Good catch. I'll fix it in v2.

Thank you,

Signature · 1 line
Carlo Caione