]> www.infradead.org Git - users/hch/misc.git/commitdiff
dma-mapping: Implement link/unlink ranges API
authorLeon Romanovsky <leonro@nvidia.com>
Mon, 5 May 2025 07:01:44 +0000 (10:01 +0300)
committerMarek Szyprowski <m.szyprowski@samsung.com>
Tue, 6 May 2025 06:36:53 +0000 (08:36 +0200)
Introduce new DMA APIs to perform DMA linkage of buffers
in layers higher than DMA.

In proposed API, the callers will perform the following steps.
In map path:
if (dma_can_use_iova(...))
    dma_iova_alloc()
    for (page in range)
       dma_iova_link_next(...)
    dma_iova_sync(...)
else
     /* Fallback to legacy map pages */
             for (all pages)
       dma_map_page(...)

In unmap path:
if (dma_can_use_iova(...))
     dma_iova_destroy()
else
     for (all pages)
dma_unmap_page(...)

Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Marek Szyprowski <m.szyprowski@samsung.com>
drivers/iommu/dma-iommu.c
include/linux/dma-mapping.h

index d2c298083e0aab2cb8f53c42107639f1fb4e0dbf..1569efe711189b91ed9154f281823b52f80fd935 100644 (file)
@@ -1175,6 +1175,17 @@ static phys_addr_t iommu_dma_map_swiotlb(struct device *dev, phys_addr_t phys,
        return phys;
 }
 
+/*
+ * Checks if a physical buffer has unaligned boundaries with respect to
+ * the IOMMU granule. Returns non-zero if either the start or end
+ * address is not aligned to the granule boundary.
+ */
+static inline size_t iova_unaligned(struct iova_domain *iovad, phys_addr_t phys,
+                                   size_t size)
+{
+       return iova_offset(iovad, phys | size);
+}
+
 dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
              unsigned long offset, size_t size, enum dma_data_direction dir,
              unsigned long attrs)
@@ -1192,7 +1203,7 @@ dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
         * we don't need to use a bounce page.
         */
        if (dev_use_swiotlb(dev, size, dir) &&
-           iova_offset(iovad, phys | size)) {
+           iova_unaligned(iovad, phys, size)) {
                phys = iommu_dma_map_swiotlb(dev, phys, size, dir, attrs);
                if (phys == (phys_addr_t)DMA_MAPPING_ERROR)
                        return DMA_MAPPING_ERROR;
@@ -1818,6 +1829,268 @@ void dma_iova_free(struct device *dev, struct dma_iova_state *state)
 }
 EXPORT_SYMBOL_GPL(dma_iova_free);
 
+static int __dma_iova_link(struct device *dev, dma_addr_t addr,
+               phys_addr_t phys, size_t size, enum dma_data_direction dir,
+               unsigned long attrs)
+{
+       bool coherent = dev_is_dma_coherent(dev);
+
+       if (!coherent && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
+               arch_sync_dma_for_device(phys, size, dir);
+
+       return iommu_map_nosync(iommu_get_dma_domain(dev), addr, phys, size,
+                       dma_info_to_prot(dir, coherent, attrs), GFP_ATOMIC);
+}
+
+static int iommu_dma_iova_bounce_and_link(struct device *dev, dma_addr_t addr,
+               phys_addr_t phys, size_t bounce_len,
+               enum dma_data_direction dir, unsigned long attrs,
+               size_t iova_start_pad)
+{
+       struct iommu_domain *domain = iommu_get_dma_domain(dev);
+       struct iova_domain *iovad = &domain->iova_cookie->iovad;
+       phys_addr_t bounce_phys;
+       int error;
+
+       bounce_phys = iommu_dma_map_swiotlb(dev, phys, bounce_len, dir, attrs);
+       if (bounce_phys == DMA_MAPPING_ERROR)
+               return -ENOMEM;
+
+       error = __dma_iova_link(dev, addr - iova_start_pad,
+                       bounce_phys - iova_start_pad,
+                       iova_align(iovad, bounce_len), dir, attrs);
+       if (error)
+               swiotlb_tbl_unmap_single(dev, bounce_phys, bounce_len, dir,
+                               attrs);
+       return error;
+}
+
+static int iommu_dma_iova_link_swiotlb(struct device *dev,
+               struct dma_iova_state *state, phys_addr_t phys, size_t offset,
+               size_t size, enum dma_data_direction dir, unsigned long attrs)
+{
+       struct iommu_domain *domain = iommu_get_dma_domain(dev);
+       struct iommu_dma_cookie *cookie = domain->iova_cookie;
+       struct iova_domain *iovad = &cookie->iovad;
+       size_t iova_start_pad = iova_offset(iovad, phys);
+       size_t iova_end_pad = iova_offset(iovad, phys + size);
+       dma_addr_t addr = state->addr + offset;
+       size_t mapped = 0;
+       int error;
+
+       if (iova_start_pad) {
+               size_t bounce_len = min(size, iovad->granule - iova_start_pad);
+
+               error = iommu_dma_iova_bounce_and_link(dev, addr, phys,
+                               bounce_len, dir, attrs, iova_start_pad);
+               if (error)
+                       return error;
+               state->__size |= DMA_IOVA_USE_SWIOTLB;
+
+               mapped += bounce_len;
+               size -= bounce_len;
+               if (!size)
+                       return 0;
+       }
+
+       size -= iova_end_pad;
+       error = __dma_iova_link(dev, addr + mapped, phys + mapped, size, dir,
+                       attrs);
+       if (error)
+               goto out_unmap;
+       mapped += size;
+
+       if (iova_end_pad) {
+               error = iommu_dma_iova_bounce_and_link(dev, addr + mapped,
+                               phys + mapped, iova_end_pad, dir, attrs, 0);
+               if (error)
+                       goto out_unmap;
+               state->__size |= DMA_IOVA_USE_SWIOTLB;
+       }
+
+       return 0;
+
+out_unmap:
+       dma_iova_unlink(dev, state, 0, mapped, dir, attrs);
+       return error;
+}
+
+/**
+ * dma_iova_link - Link a range of IOVA space
+ * @dev: DMA device
+ * @state: IOVA state
+ * @phys: physical address to link
+ * @offset: offset into the IOVA state to map into
+ * @size: size of the buffer
+ * @dir: DMA direction
+ * @attrs: attributes of mapping properties
+ *
+ * Link a range of IOVA space for the given IOVA state without IOTLB sync.
+ * This function is used to link multiple physical addresses in contiguous
+ * IOVA space without performing costly IOTLB sync.
+ *
+ * The caller is responsible to call to dma_iova_sync() to sync IOTLB at
+ * the end of linkage.
+ */
+int dma_iova_link(struct device *dev, struct dma_iova_state *state,
+               phys_addr_t phys, size_t offset, size_t size,
+               enum dma_data_direction dir, unsigned long attrs)
+{
+       struct iommu_domain *domain = iommu_get_dma_domain(dev);
+       struct iommu_dma_cookie *cookie = domain->iova_cookie;
+       struct iova_domain *iovad = &cookie->iovad;
+       size_t iova_start_pad = iova_offset(iovad, phys);
+
+       if (WARN_ON_ONCE(iova_start_pad && offset > 0))
+               return -EIO;
+
+       if (dev_use_swiotlb(dev, size, dir) &&
+           iova_unaligned(iovad, phys, size))
+               return iommu_dma_iova_link_swiotlb(dev, state, phys, offset,
+                               size, dir, attrs);
+
+       return __dma_iova_link(dev, state->addr + offset - iova_start_pad,
+                       phys - iova_start_pad,
+                       iova_align(iovad, size + iova_start_pad), dir, attrs);
+}
+EXPORT_SYMBOL_GPL(dma_iova_link);
+
+/**
+ * dma_iova_sync - Sync IOTLB
+ * @dev: DMA device
+ * @state: IOVA state
+ * @offset: offset into the IOVA state to sync
+ * @size: size of the buffer
+ *
+ * Sync IOTLB for the given IOVA state. This function should be called on
+ * the IOVA-contiguous range created by one ore more dma_iova_link() calls
+ * to sync the IOTLB.
+ */
+int dma_iova_sync(struct device *dev, struct dma_iova_state *state,
+               size_t offset, size_t size)
+{
+       struct iommu_domain *domain = iommu_get_dma_domain(dev);
+       struct iommu_dma_cookie *cookie = domain->iova_cookie;
+       struct iova_domain *iovad = &cookie->iovad;
+       dma_addr_t addr = state->addr + offset;
+       size_t iova_start_pad = iova_offset(iovad, addr);
+
+       return iommu_sync_map(domain, addr - iova_start_pad,
+                     iova_align(iovad, size + iova_start_pad));
+}
+EXPORT_SYMBOL_GPL(dma_iova_sync);
+
+static void iommu_dma_iova_unlink_range_slow(struct device *dev,
+               dma_addr_t addr, size_t size, enum dma_data_direction dir,
+               unsigned long attrs)
+{
+       struct iommu_domain *domain = iommu_get_dma_domain(dev);
+       struct iommu_dma_cookie *cookie = domain->iova_cookie;
+       struct iova_domain *iovad = &cookie->iovad;
+       size_t iova_start_pad = iova_offset(iovad, addr);
+       dma_addr_t end = addr + size;
+
+       do {
+               phys_addr_t phys;
+               size_t len;
+
+               phys = iommu_iova_to_phys(domain, addr);
+               if (WARN_ON(!phys))
+                       /* Something very horrible happen here */
+                       return;
+
+               len = min_t(size_t,
+                       end - addr, iovad->granule - iova_start_pad);
+
+               if (!dev_is_dma_coherent(dev) &&
+                   !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
+                       arch_sync_dma_for_cpu(phys, len, dir);
+
+               swiotlb_tbl_unmap_single(dev, phys, len, dir, attrs);
+
+               addr += len;
+               iova_start_pad = 0;
+       } while (addr < end);
+}
+
+static void __iommu_dma_iova_unlink(struct device *dev,
+               struct dma_iova_state *state, size_t offset, size_t size,
+               enum dma_data_direction dir, unsigned long attrs,
+               bool free_iova)
+{
+       struct iommu_domain *domain = iommu_get_dma_domain(dev);
+       struct iommu_dma_cookie *cookie = domain->iova_cookie;
+       struct iova_domain *iovad = &cookie->iovad;
+       dma_addr_t addr = state->addr + offset;
+       size_t iova_start_pad = iova_offset(iovad, addr);
+       struct iommu_iotlb_gather iotlb_gather;
+       size_t unmapped;
+
+       if ((state->__size & DMA_IOVA_USE_SWIOTLB) ||
+           (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)))
+               iommu_dma_iova_unlink_range_slow(dev, addr, size, dir, attrs);
+
+       iommu_iotlb_gather_init(&iotlb_gather);
+       iotlb_gather.queued = free_iova && READ_ONCE(cookie->fq_domain);
+
+       size = iova_align(iovad, size + iova_start_pad);
+       addr -= iova_start_pad;
+       unmapped = iommu_unmap_fast(domain, addr, size, &iotlb_gather);
+       WARN_ON(unmapped != size);
+
+       if (!iotlb_gather.queued)
+               iommu_iotlb_sync(domain, &iotlb_gather);
+       if (free_iova)
+               iommu_dma_free_iova(domain, addr, size, &iotlb_gather);
+}
+
+/**
+ * dma_iova_unlink - Unlink a range of IOVA space
+ * @dev: DMA device
+ * @state: IOVA state
+ * @offset: offset into the IOVA state to unlink
+ * @size: size of the buffer
+ * @dir: DMA direction
+ * @attrs: attributes of mapping properties
+ *
+ * Unlink a range of IOVA space for the given IOVA state.
+ */
+void dma_iova_unlink(struct device *dev, struct dma_iova_state *state,
+               size_t offset, size_t size, enum dma_data_direction dir,
+               unsigned long attrs)
+{
+        __iommu_dma_iova_unlink(dev, state, offset, size, dir, attrs, false);
+}
+EXPORT_SYMBOL_GPL(dma_iova_unlink);
+
+/**
+ * dma_iova_destroy - Finish a DMA mapping transaction
+ * @dev: DMA device
+ * @state: IOVA state
+ * @mapped_len: number of bytes to unmap
+ * @dir: DMA direction
+ * @attrs: attributes of mapping properties
+ *
+ * Unlink the IOVA range up to @mapped_len and free the entire IOVA space. The
+ * range of IOVA from dma_addr to @mapped_len must all be linked, and be the
+ * only linked IOVA in state.
+ */
+void dma_iova_destroy(struct device *dev, struct dma_iova_state *state,
+               size_t mapped_len, enum dma_data_direction dir,
+               unsigned long attrs)
+{
+       if (mapped_len)
+               __iommu_dma_iova_unlink(dev, state, 0, mapped_len, dir, attrs,
+                               true);
+       else
+               /*
+                * We can be here if first call to dma_iova_link() failed and
+                * there is nothing to unlink, so let's be more clear.
+                */
+               dma_iova_free(dev, state);
+}
+EXPORT_SYMBOL_GPL(dma_iova_destroy);
+
 void iommu_setup_dma_ops(struct device *dev)
 {
        struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
index de7f73810d5481d3fa082a708626935734d12f06..a71e110f1e9dad5d2c9c4cd1e0ccf12b6f1e6223 100644 (file)
@@ -309,6 +309,17 @@ static inline bool dma_use_iova(struct dma_iova_state *state)
 bool dma_iova_try_alloc(struct device *dev, struct dma_iova_state *state,
                phys_addr_t phys, size_t size);
 void dma_iova_free(struct device *dev, struct dma_iova_state *state);
+void dma_iova_destroy(struct device *dev, struct dma_iova_state *state,
+               size_t mapped_len, enum dma_data_direction dir,
+               unsigned long attrs);
+int dma_iova_sync(struct device *dev, struct dma_iova_state *state,
+               size_t offset, size_t size);
+int dma_iova_link(struct device *dev, struct dma_iova_state *state,
+               phys_addr_t phys, size_t offset, size_t size,
+               enum dma_data_direction dir, unsigned long attrs);
+void dma_iova_unlink(struct device *dev, struct dma_iova_state *state,
+               size_t offset, size_t size, enum dma_data_direction dir,
+               unsigned long attrs);
 #else /* CONFIG_IOMMU_DMA */
 static inline bool dma_use_iova(struct dma_iova_state *state)
 {
@@ -323,6 +334,27 @@ static inline void dma_iova_free(struct device *dev,
                struct dma_iova_state *state)
 {
 }
+static inline void dma_iova_destroy(struct device *dev,
+               struct dma_iova_state *state, size_t mapped_len,
+               enum dma_data_direction dir, unsigned long attrs)
+{
+}
+static inline int dma_iova_sync(struct device *dev,
+               struct dma_iova_state *state, size_t offset, size_t size)
+{
+       return -EOPNOTSUPP;
+}
+static inline int dma_iova_link(struct device *dev,
+               struct dma_iova_state *state, phys_addr_t phys, size_t offset,
+               size_t size, enum dma_data_direction dir, unsigned long attrs)
+{
+       return -EOPNOTSUPP;
+}
+static inline void dma_iova_unlink(struct device *dev,
+               struct dma_iova_state *state, size_t offset, size_t size,
+               enum dma_data_direction dir, unsigned long attrs)
+{
+}
 #endif /* CONFIG_IOMMU_DMA */
 
 #if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC)