search for: out_free_pages

On 2020-06-22 16:38, Ralph Campbell wrote: > The patch to add zero page migration to GPU memory inadvertantly included inadvertently > part of a future change which broke normal page migration to GPU memory > by copying too much data and corrupting GPU memory. > Fix this by only copying one page instead of a byte count. > > Fixes: 9d4296a7d4b3 ("drm/nouveau/nouveau/hmm:

When calling OpenCL clEnqueueSVMMigrateMem() on a region of memory that is backed by pte_none() or zero pages, migrate_vma_setup() will fill the source PFN array with an entry indicating the source page is zero. Use this to optimize migration to device private memory by allocating GPU memory and zero filling it instead of failing to migrate the page. Signed-off-by: Ralph Campbell <rcampbell at

...R) { - dma_direct_free_pages(dev, size, ret, *dma_handle, attrs); + size = PAGE_ALIGN(size); + page = alloc_pages(gfp, get_order(size)); + if (!page) return NULL; - } + ret = page_address(page); + *dma_handle = vdma_alloc(virt_to_phys(ret), size); + if (*dma_handle == DMA_MAPPING_ERROR) + goto out_free_pages; + + if (attrs & DMA_ATTR_NON_CONSISTENT) + return ret; + arch_dma_prep_coherent(page, size); + return (void *)(UNCAC_BASE + __pa(ret)); - return ret; +out_free_pages: + __free_pages(page, get_order(size)); + return NULL; } static void jazz_dma_free(struct device *dev, size_t size, void...

These are based on 5.8.0-rc2 and intended for Ben Skeggs' nouveau tree. I believe the changes can be queued for 5.8-rcX after being reviewed. These were part of a larger series but I'm resending them separately as suggested by Jason Gunthorpe. https://lore.kernel.org/linux-mm/20200619215649.32297-1-rcampbell at nvidia.com/ Note that in order to exercise/test patch 2 here, you will need a

On 2020-06-22 16:38, Ralph Campbell wrote: > The OpenCL function clEnqueueSVMMigrateMem(), without any flags, will > migrate memory in the given address range to device private memory. The > source pages might already have been migrated to device private memory. > In that case, the source struct page is not checked to see if it is > a device private page and incorrectly computes the

Factor the main copy page to vram routine out into a helper that acts on a single page and which doesn't require the nouveau_dmem_migrate structure for argument passing. As an added benefit the new version only allocates the dma address array once and reuses it for each subsequent chunk of work. Signed-off-by: Christoph Hellwig <hch at lst.de> ---

When memory is migrated to the GPU it is likely to be accessed by GPU code soon afterwards. Instead of waiting for a GPU fault, map the migrated memory into the GPU page tables with the same access permissions as the source CPU page table entries. This preserves copy on write semantics. Signed-off-by: Ralph Campbell <rcampbell at nvidia.com> Cc: Christoph Hellwig <hch at lst.de> Cc:

...letion(-) diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c index 8c2788633c1766..e4bf1bb159f7c7 100644 --- a/drivers/iommu/dma-iommu.c +++ b/drivers/iommu/dma-iommu.c @@ -822,7 +822,7 @@ static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev, if (!iova) goto out_free_pages; - if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, GFP_KERNEL)) + if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, gfp)) goto out_free_iova; if (!(ioprot & IOMMU_CACHE)) { -- 2.39.0

...letion(-) diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c index 8c2788633c1766..e4bf1bb159f7c7 100644 --- a/drivers/iommu/dma-iommu.c +++ b/drivers/iommu/dma-iommu.c @@ -822,7 +822,7 @@ static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev, if (!iova) goto out_free_pages; - if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, GFP_KERNEL)) + if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, gfp)) goto out_free_iova; if (!(ioprot & IOMMU_CACHE)) { -- 2.39.0

On Wed, Aug 07, 2019 at 08:02:14AM -0700, Ralph Campbell wrote: > When memory is migrated to the GPU it is likely to be accessed by GPU > code soon afterwards. Instead of waiting for a GPU fault, map the > migrated memory into the GPU page tables with the same access permissions > as the source CPU page table entries. This preserves copy on write > semantics. > >

...z/jazzdma.c index 2bf849caf507b1..f53bc043334c01 100644 --- a/arch/mips/jazz/jazzdma.c +++ b/arch/mips/jazz/jazzdma.c @@ -506,9 +506,6 @@ static void *jazz_dma_alloc(struct device *dev, size_t size, *dma_handle = vdma_alloc(virt_to_phys(ret), size); if (*dma_handle == DMA_MAPPING_ERROR) goto out_free_pages; - - if (attrs & DMA_ATTR_NON_CONSISTENT) - return ret; arch_dma_prep_coherent(page, size); return (void *)(UNCAC_BASE + __pa(ret)); @@ -521,8 +518,6 @@ static void jazz_dma_free(struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle, unsigned long attrs) { vdma_free(d...

The patch to add zero page migration to GPU memory inadvertantly included part of a future change which broke normal page migration to GPU memory by copying too much data and corrupting GPU memory. Fix this by only copying one page instead of a byte count. Fixes: 9d4296a7d4b3 ("drm/nouveau/nouveau/hmm: fix migrate zero page to GPU") Signed-off-by: Ralph Campbell <rcampbell at

The patch to add zero page migration to GPU memory inadvertantly included part of a future change which broke normal page migration to GPU memory by copying too much data and corrupting GPU memory. Fix this by only copying one page instead of a byte count. Fixes: 9d4296a7d4b3 ("drm/nouveau/nouveau/hmm: fix migrate zero page to GPU") Signed-off-by: Ralph Campbell <rcampbell at

The OpenCL function clEnqueueSVMMigrateMem(), without any flags, will migrate memory in the given address range to device private memory. The source pages might already have been migrated to device private memory. In that case, the source struct page is not checked to see if it is a device private page and incorrectly computes the GPU's physical address of local memory leading to data

On Tue, 23 Jun 2020 at 10:51, John Hubbard <jhubbard at nvidia.com> wrote: > > On 2020-06-22 16:38, Ralph Campbell wrote: > > The patch to add zero page migration to GPU memory inadvertantly included > > inadvertently > > > part of a future change which broke normal page migration to GPU memory > > by copying too much data and corrupting GPU memory. > >

...iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c > index 8c2788633c1766..e4bf1bb159f7c7 100644 > --- a/drivers/iommu/dma-iommu.c > +++ b/drivers/iommu/dma-iommu.c > @@ -822,7 +822,7 @@ static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev, > if (!iova) > goto out_free_pages; > > - if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, GFP_KERNEL)) > + if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, gfp)) > goto out_free_iova; > > if (!(ioprot & IOMMU_CACHE)) {

When memory is migrated to the GPU, it is likely to be accessed by GPU code soon afterwards. Instead of waiting for a GPU fault, map the migrated memory into the GPU page tables with the same access permissions as the source CPU page table entries. This preserves copy on write semantics. Signed-off-by: Ralph Campbell <rcampbell at nvidia.com> Cc: Christoph Hellwig <hch at lst.de> Cc:

Originally patch 4 was targeted for Jason's rdma tree since other HMM related changes were queued there. Now that those have been merged, these patches just contain changes to nouveau so they could go through any tree. I guess Ben Skeggs' tree would be appropriate. Changes since v2: Added patches 1-3 to fix some minor issues. Eliminated nouveau_find_svmm() since it is easily found.

Use the new MMU_NOTIFY_MIGRATE event to skip GPU MMU invalidations of device private memory and handle the invalidation in the driver as part of migrating device private memory. Signed-off-by: Ralph Campbell <rcampbell at nvidia.com> --- drivers/gpu/drm/nouveau/nouveau_dmem.c | 15 ++++++++++++--- drivers/gpu/drm/nouveau/nouveau_svm.c | 21 +++++++++------------

...z/jazzdma.c index d0b5a2ba2b1a8a..0f9a9cb7fe7a95 100644 --- a/arch/mips/jazz/jazzdma.c +++ b/arch/mips/jazz/jazzdma.c @@ -505,9 +505,6 @@ static void *jazz_dma_alloc(struct device *dev, size_t size, *dma_handle = vdma_alloc(virt_to_phys(ret), size); if (*dma_handle == DMA_MAPPING_ERROR) goto out_free_pages; - - if (attrs & DMA_ATTR_NON_CONSISTENT) - return ret; arch_dma_prep_coherent(page, size); return (void *)(UNCAC_BASE + __pa(ret)); @@ -520,8 +517,6 @@ static void jazz_dma_free(struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle, unsigned long attrs) { vdma_free(d...