swiotlb: store phys address in io_tlb_orig_addr array
authorBecky Bruce <beckyb@kernel.crashing.org>
Mon, 22 Dec 2008 18:26:08 +0000 (10:26 -0800)
committerIngo Molnar <mingo@elte.hu>
Sun, 28 Dec 2008 09:03:12 +0000 (10:03 +0100)
Impact: refactor code, cleanup

When we enable swiotlb for platforms that support HIGHMEM, we
can no longer store the virtual address of the original dma
buffer, because that buffer might not have a permament mapping.

Change the swiotlb code to instead store the physical address of
the original buffer.

Signed-off-by: Becky Bruce <beckyb@kernel.crashing.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
lib/swiotlb.c

index 3657da8..98a7a44 100644 (file)
@@ -89,10 +89,7 @@ static unsigned int io_tlb_index;
  * We need to save away the original address corresponding to a mapped entry
  * for the sync operations.
  */
-static struct swiotlb_phys_addr {
-       struct page *page;
-       unsigned int offset;
-} *io_tlb_orig_addr;
+static phys_addr_t *io_tlb_orig_addr;
 
 /*
  * Protect the above data structures in the map and unmap calls
@@ -204,7 +201,7 @@ swiotlb_init_with_default_size(size_t default_size)
        for (i = 0; i < io_tlb_nslabs; i++)
                io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
        io_tlb_index = 0;
-       io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(struct swiotlb_phys_addr));
+       io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
 
        /*
         * Get the overflow emergency buffer
@@ -278,12 +275,14 @@ swiotlb_late_init_with_default_size(size_t default_size)
                io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
        io_tlb_index = 0;
 
-       io_tlb_orig_addr = (struct swiotlb_phys_addr *)__get_free_pages(GFP_KERNEL,
-                                  get_order(io_tlb_nslabs * sizeof(struct swiotlb_phys_addr)));
+       io_tlb_orig_addr = (phys_addr_t *)
+               __get_free_pages(GFP_KERNEL,
+                                get_order(io_tlb_nslabs *
+                                          sizeof(phys_addr_t)));
        if (!io_tlb_orig_addr)
                goto cleanup3;
 
-       memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(struct swiotlb_phys_addr));
+       memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
 
        /*
         * Get the overflow emergency buffer
@@ -298,8 +297,8 @@ swiotlb_late_init_with_default_size(size_t default_size)
        return 0;
 
 cleanup4:
-       free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs *
-                                                             sizeof(char *)));
+       free_pages((unsigned long)io_tlb_orig_addr,
+                  get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
        io_tlb_orig_addr = NULL;
 cleanup3:
        free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
@@ -330,59 +329,11 @@ static int is_swiotlb_buffer(char *addr)
        return addr >= io_tlb_start && addr < io_tlb_end;
 }
 
-static struct swiotlb_phys_addr swiotlb_bus_to_phys_addr(char *dma_addr)
-{
-       int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
-       struct swiotlb_phys_addr buffer = io_tlb_orig_addr[index];
-       buffer.offset += (long)dma_addr & ((1 << IO_TLB_SHIFT) - 1);
-       buffer.page += buffer.offset >> PAGE_SHIFT;
-       buffer.offset &= PAGE_SIZE - 1;
-       return buffer;
-}
-
-static void
-__sync_single(struct swiotlb_phys_addr buffer, char *dma_addr, size_t size, int dir)
-{
-       if (PageHighMem(buffer.page)) {
-               size_t len, bytes;
-               char *dev, *host, *kmp;
-
-               len = size;
-               while (len != 0) {
-                       unsigned long flags;
-
-                       bytes = len;
-                       if ((bytes + buffer.offset) > PAGE_SIZE)
-                               bytes = PAGE_SIZE - buffer.offset;
-                       local_irq_save(flags); /* protects KM_BOUNCE_READ */
-                       kmp  = kmap_atomic(buffer.page, KM_BOUNCE_READ);
-                       dev  = dma_addr + size - len;
-                       host = kmp + buffer.offset;
-                       if (dir == DMA_FROM_DEVICE)
-                               memcpy(host, dev, bytes);
-                       else
-                               memcpy(dev, host, bytes);
-                       kunmap_atomic(kmp, KM_BOUNCE_READ);
-                       local_irq_restore(flags);
-                       len -= bytes;
-                       buffer.page++;
-                       buffer.offset = 0;
-               }
-       } else {
-               void *v = page_address(buffer.page) + buffer.offset;
-
-               if (dir == DMA_TO_DEVICE)
-                       memcpy(dma_addr, v, size);
-               else
-                       memcpy(v, dma_addr, size);
-       }
-}
-
 /*
  * Allocates bounce buffer and returns its kernel virtual address.
  */
 static void *
-map_single(struct device *hwdev, struct swiotlb_phys_addr buffer, size_t size, int dir)
+map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
 {
        unsigned long flags;
        char *dma_addr;
@@ -392,7 +343,6 @@ map_single(struct device *hwdev, struct swiotlb_phys_addr buffer, size_t size, i
        unsigned long mask;
        unsigned long offset_slots;
        unsigned long max_slots;
-       struct swiotlb_phys_addr slot_buf;
 
        mask = dma_get_seg_boundary(hwdev);
        start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
@@ -477,15 +427,10 @@ found:
         * This is needed when we sync the memory.  Then we sync the buffer if
         * needed.
         */
-       slot_buf = buffer;
-       for (i = 0; i < nslots; i++) {
-               slot_buf.page += slot_buf.offset >> PAGE_SHIFT;
-               slot_buf.offset &= PAGE_SIZE - 1;
-               io_tlb_orig_addr[index+i] = slot_buf;
-               slot_buf.offset += 1 << IO_TLB_SHIFT;
-       }
+       for (i = 0; i < nslots; i++)
+               io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
        if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
-               __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
+               memcpy(dma_addr, phys_to_virt(phys), size);
 
        return dma_addr;
 }
@@ -499,17 +444,17 @@ unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
        unsigned long flags;
        int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
        int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
-       struct swiotlb_phys_addr buffer = swiotlb_bus_to_phys_addr(dma_addr);
+       phys_addr_t phys = io_tlb_orig_addr[index];
 
        /*
         * First, sync the memory before unmapping the entry
         */
-       if ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL))
+       if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
                /*
                 * bounce... copy the data back into the original buffer * and
                 * delete the bounce buffer.
                 */
-               __sync_single(buffer, dma_addr, size, DMA_FROM_DEVICE);
+               memcpy(phys_to_virt(phys), dma_addr, size);
 
        /*
         * Return the buffer to the free list by setting the corresponding
@@ -541,18 +486,21 @@ static void
 sync_single(struct device *hwdev, char *dma_addr, size_t size,
            int dir, int target)
 {
-       struct swiotlb_phys_addr buffer = swiotlb_bus_to_phys_addr(dma_addr);
+       int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
+       phys_addr_t phys = io_tlb_orig_addr[index];
+
+       phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
 
        switch (target) {
        case SYNC_FOR_CPU:
                if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
-                       __sync_single(buffer, dma_addr, size, DMA_FROM_DEVICE);
+                       memcpy(phys_to_virt(phys), dma_addr, size);
                else
                        BUG_ON(dir != DMA_TO_DEVICE);
                break;
        case SYNC_FOR_DEVICE:
                if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
-                       __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
+                       memcpy(dma_addr, phys_to_virt(phys), size);
                else
                        BUG_ON(dir != DMA_FROM_DEVICE);
                break;
@@ -591,10 +539,7 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
                 * swiotlb_map_single(), which will grab memory from
                 * the lowest available address range.
                 */
-               struct swiotlb_phys_addr buffer;
-               buffer.page = virt_to_page(NULL);
-               buffer.offset = 0;
-               ret = map_single(hwdev, buffer, size, DMA_FROM_DEVICE);
+               ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
                if (!ret)
                        return NULL;
        }
@@ -662,7 +607,6 @@ swiotlb_map_single_attrs(struct device *hwdev, void *ptr, size_t size,
 {
        dma_addr_t dev_addr = swiotlb_virt_to_bus(hwdev, ptr);
        void *map;
-       struct swiotlb_phys_addr buffer;
 
        BUG_ON(dir == DMA_NONE);
        /*
@@ -677,9 +621,7 @@ swiotlb_map_single_attrs(struct device *hwdev, void *ptr, size_t size,
        /*
         * Oh well, have to allocate and map a bounce buffer.
         */
-       buffer.page   = virt_to_page(ptr);
-       buffer.offset = (unsigned long)ptr & ~PAGE_MASK;
-       map = map_single(hwdev, buffer, size, dir);
+       map = map_single(hwdev, virt_to_phys(ptr), size, dir);
        if (!map) {
                swiotlb_full(hwdev, size, dir, 1);
                map = io_tlb_overflow_buffer;
@@ -824,20 +766,18 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
                     int dir, struct dma_attrs *attrs)
 {
        struct scatterlist *sg;
-       struct swiotlb_phys_addr buffer;
-       dma_addr_t dev_addr;
        int i;
 
        BUG_ON(dir == DMA_NONE);
 
        for_each_sg(sgl, sg, nelems, i) {
-               dev_addr = swiotlb_sg_to_bus(hwdev, sg);
-               if (range_needs_mapping(sg_virt(sg), sg->length) ||
+               void *addr = sg_virt(sg);
+               dma_addr_t dev_addr = swiotlb_virt_to_bus(hwdev, addr);
+
+               if (range_needs_mapping(addr, sg->length) ||
                    address_needs_mapping(hwdev, dev_addr, sg->length)) {
-                       void *map;
-                       buffer.page   = sg_page(sg);
-                       buffer.offset = sg->offset;
-                       map = map_single(hwdev, buffer, sg->length, dir);
+                       void *map = map_single(hwdev, sg_phys(sg),
+                                              sg->length, dir);
                        if (!map) {
                                /* Don't panic here, we expect map_sg users
                                   to do proper error handling. */