Re: [PATCH v1 5/9] sparc64: Use physical address DMA mapping
From: Jason Gunthorpe <jgg@nvidia.com>
Date: 2025-10-03 15:16:16
Also in:
linux-alpha, linux-iommu, linux-mips, lkml, sparclinux, virtualization, xen-devel
On Sun, Sep 28, 2025 at 06:02:25PM +0300, Leon Romanovsky wrote:
quoted hunk ↗ jump to hunk
@@ -273,13 +272,16 @@ static dma_addr_t dma_4u_map_page(struct device *dev, struct page *page, u32 bus_addr, ret; unsigned long iopte_protection; + if (attrs & DMA_ATTR_MMIO) + goto bad_no_ctx; + iommu = dev->archdata.iommu; strbuf = dev->archdata.stc; if (unlikely(direction == DMA_NONE)) goto bad_no_ctx; - oaddr = (unsigned long)(page_address(page) + offset); + oaddr = (unsigned long)(phys_to_virt(phys)); npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK); npages >>= IO_PAGE_SHIFT;
This should be cleaned up some more: oaddr = (unsigned long)(page_address(page) + offset); ret = bus_addr | (oaddr & ~IO_PAGE_MASK); base_paddr = __pa(oaddr & IO_PAGE_MASK); Makes no sense to phys_to_virt() then __pa() on that result. Drop oaddr. Then I would copy and paste the comment from mips about DMA_ATTR_MMIO
quoted hunk ↗ jump to hunk
@@ -367,13 +366,16 @@ static dma_addr_t dma_4v_map_page(struct device *dev, struct page *page, dma_addr_t bus_addr, ret; long entry; + if (attrs & DMA_ATTR_MMIO) + goto bad; + iommu = dev->archdata.iommu; atu = iommu->atu; if (unlikely(direction == DMA_NONE)) goto bad; - oaddr = (unsigned long)(page_address(page) + offset); + oaddr = (unsigned long)(phys_to_virt(phys)); npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK); npages >>= IO_PAGE_SHIFT;
Same remarks here
+static dma_addr_t iounit_map_phys(struct device *dev, phys_addr_t phys,
+ size_t len, enum dma_data_direction dir, unsigned long attrs)
{
- void *vaddr = page_address(page) + offset;
+ void *vaddr = phys_to_virt(phys);
struct iounit_struct *iounit = dev->archdata.iommu;
unsigned long ret, flags;
iounit_get_area() does not seem to need vaddr:
npages = ((vaddr & ~PAGE_MASK) + size + (PAGE_SIZE-1)) >> PAGE_SHIFT;
~PAGE_MASK is page_offset()
iopte = MKIOPTE(__pa(vaddr & PAGE_MASK));
__pa(phys_to_virt(pa)) again:
iopte = MKIOPTE(PAGE_ALIGN(pa));
vaddr = IOUNIT_DMA_BASE + (scan << PAGE_SHIFT) + (vaddr & ~PAGE_MASK);
page_offset, then it replaces vaddr.
So I'd tidy this too.
quoted hunk ↗ jump to hunk
@@ -202,10 +204,10 @@ static dma_addr_t __sbus_iommu_map_page(struct device *dev, struct page *page, * We expect unmapped highmem pages to be not in the cache. * XXX Is this a good assumption? * XXX What if someone else unmaps it here and races us? */
At least ARM32 has code that seems to say these assumptions are not always true.. Oh well. Jason