swiotlb: refactor coherent buffer allocation
authorChristoph Hellwig <hch@lst.de>
Tue, 9 Jan 2018 21:15:30 +0000 (22:15 +0100)
committerChristoph Hellwig <hch@lst.de>
Mon, 15 Jan 2018 08:35:49 +0000 (09:35 +0100)
Factor out a new swiotlb_alloc_buffer helper that allocates DMA coherent
memory from the swiotlb bounce buffer.

This allows to simplify the swiotlb_alloc implemenation that uses
dma_direct_alloc to try to allocate a reachable buffer first.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Christian König <christian.koenig@amd.com>
lib/swiotlb.c

index 1a147f1354a18c68bbc2673d1ef87c7209a95980..0039b7c5e690255c9e534b89008475a4b0108110 100644 (file)
@@ -709,75 +709,79 @@ void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
 }
 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
 
-void *
-swiotlb_alloc_coherent(struct device *hwdev, size_t size,
-                      dma_addr_t *dma_handle, gfp_t flags)
+static inline bool dma_coherent_ok(struct device *dev, dma_addr_t addr,
+               size_t size)
 {
-       bool warn = !(flags & __GFP_NOWARN);
-       dma_addr_t dev_addr;
-       void *ret;
-       int order = get_order(size);
-       u64 dma_mask = DMA_BIT_MASK(32);
+       u64 mask = DMA_BIT_MASK(32);
 
-       if (hwdev && hwdev->coherent_dma_mask)
-               dma_mask = hwdev->coherent_dma_mask;
+       if (dev && dev->coherent_dma_mask)
+               mask = dev->coherent_dma_mask;
+       return addr + size - 1 <= mask;
+}
 
-       ret = (void *)__get_free_pages(flags, order);
-       if (ret) {
-               dev_addr = swiotlb_virt_to_bus(hwdev, ret);
-               if (dev_addr + size - 1 > dma_mask) {
-                       /*
-                        * The allocated memory isn't reachable by the device.
-                        */
-                       free_pages((unsigned long) ret, order);
-                       ret = NULL;
-               }
-       }
-       if (!ret) {
-               /*
-                * We are either out of memory or the device can't DMA to
-                * GFP_DMA memory; fall back on map_single(), which
-                * will grab memory from the lowest available address range.
-                */
-               phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE,
-                                              warn ? 0 : DMA_ATTR_NO_WARN);
-               if (paddr == SWIOTLB_MAP_ERROR)
-                       goto err_warn;
+static void *
+swiotlb_alloc_buffer(struct device *dev, size_t size, dma_addr_t *dma_handle,
+               unsigned long attrs)
+{
+       phys_addr_t phys_addr;
 
-               ret = phys_to_virt(paddr);
-               dev_addr = swiotlb_phys_to_dma(hwdev, paddr);
+       if (swiotlb_force == SWIOTLB_NO_FORCE)
+               goto out_warn;
 
-               /* Confirm address can be DMA'd by device */
-               if (dev_addr + size - 1 > dma_mask) {
-                       printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
-                              (unsigned long long)dma_mask,
-                              (unsigned long long)dev_addr);
+       phys_addr = swiotlb_tbl_map_single(dev,
+                       swiotlb_phys_to_dma(dev, io_tlb_start),
+                       0, size, DMA_FROM_DEVICE, 0);
+       if (phys_addr == SWIOTLB_MAP_ERROR)
+               goto out_warn;
 
-                       /*
-                        * DMA_TO_DEVICE to avoid memcpy in unmap_single.
-                        * The DMA_ATTR_SKIP_CPU_SYNC is optional.
-                        */
-                       swiotlb_tbl_unmap_single(hwdev, paddr,
-                                                size, DMA_TO_DEVICE,
-                                                DMA_ATTR_SKIP_CPU_SYNC);
-                       goto err_warn;
-               }
-       }
+       *dma_handle = swiotlb_phys_to_dma(dev, phys_addr);
+       if (dma_coherent_ok(dev, *dma_handle, size))
+               goto out_unmap;
 
-       *dma_handle = dev_addr;
-       memset(ret, 0, size);
+       memset(phys_to_virt(phys_addr), 0, size);
+       return phys_to_virt(phys_addr);
 
-       return ret;
+out_unmap:
+       dev_warn(dev, "hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
+               (unsigned long long)(dev ? dev->coherent_dma_mask : 0),
+               (unsigned long long)*dma_handle);
 
-err_warn:
-       if (warn && printk_ratelimit()) {
-               pr_warn("swiotlb: coherent allocation failed for device %s size=%zu\n",
-                       dev_name(hwdev), size);
+       /*
+        * DMA_TO_DEVICE to avoid memcpy in unmap_single.
+        * DMA_ATTR_SKIP_CPU_SYNC is optional.
+        */
+       swiotlb_tbl_unmap_single(dev, phys_addr, size, DMA_TO_DEVICE,
+                       DMA_ATTR_SKIP_CPU_SYNC);
+out_warn:
+       if ((attrs & DMA_ATTR_NO_WARN) && printk_ratelimit()) {
+               dev_warn(dev,
+                       "swiotlb: coherent allocation failed, size=%zu\n",
+                       size);
                dump_stack();
        }
-
        return NULL;
 }
+
+void *
+swiotlb_alloc_coherent(struct device *hwdev, size_t size,
+                      dma_addr_t *dma_handle, gfp_t flags)
+{
+       int order = get_order(size);
+       unsigned long attrs = (flags & __GFP_NOWARN) ? DMA_ATTR_NO_WARN : 0;
+       void *ret;
+
+       ret = (void *)__get_free_pages(flags, order);
+       if (ret) {
+               *dma_handle = swiotlb_virt_to_bus(hwdev, ret);
+               if (dma_coherent_ok(hwdev, *dma_handle, size)) {
+                       memset(ret, 0, size);
+                       return ret;
+               }
+               free_pages((unsigned long)ret, order);
+       }
+
+       return swiotlb_alloc_buffer(hwdev, size, dma_handle, attrs);
+}
 EXPORT_SYMBOL(swiotlb_alloc_coherent);
 
 static bool swiotlb_free_buffer(struct device *dev, size_t size,
@@ -1103,6 +1107,10 @@ void *swiotlb_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 {
        void *vaddr;
 
+       /* temporary workaround: */
+       if (gfp & __GFP_NOWARN)
+               attrs |= DMA_ATTR_NO_WARN;
+
        /*
         * Don't print a warning when the first allocation attempt fails.
         * swiotlb_alloc_coherent() will print a warning when the DMA memory
@@ -1112,7 +1120,7 @@ void *swiotlb_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 
        vaddr = dma_direct_alloc(dev, size, dma_handle, gfp, attrs);
        if (!vaddr)
-               vaddr = swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
+               vaddr = swiotlb_alloc_buffer(dev, size, dma_handle, attrs);
        return vaddr;
 }