swiotlb: Make io_tlb_overflow_buffer a physical address
authorAlexander Duyck <alexander.h.duyck@intel.com>
Mon, 15 Oct 2012 17:19:34 +0000 (10:19 -0700)
committerKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Tue, 30 Oct 2012 13:32:06 +0000 (09:32 -0400)
This change makes it so that we can avoid virt_to_phys overhead when using the
io_tlb_overflow_buffer.  My original plan was to completely remove the value
and replace it with a constant but I had seen that there were recent patches
that stated this couldn't be done until all device drivers that depended on
that functionality be updated.

Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
lib/swiotlb.c

index 8c4791f..f8c0d4e 100644 (file)
@@ -70,7 +70,7 @@ static unsigned long io_tlb_nslabs;
  */
 static unsigned long io_tlb_overflow = 32*1024;
 
-static void *io_tlb_overflow_buffer;
+static phys_addr_t io_tlb_overflow_buffer;
 
 /*
  * This is a free list describing the number of free entries available from
@@ -138,6 +138,7 @@ void swiotlb_print_info(void)
 
 void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
 {
+       void *v_overflow_buffer;
        unsigned long i, bytes;
 
        bytes = nslabs << IO_TLB_SHIFT;
@@ -147,6 +148,15 @@ void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
        io_tlb_end = io_tlb_start + bytes;
 
        /*
+        * Get the overflow emergency buffer
+        */
+       v_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
+       if (!v_overflow_buffer)
+               panic("Cannot allocate SWIOTLB overflow buffer!\n");
+
+       io_tlb_overflow_buffer = __pa(v_overflow_buffer);
+
+       /*
         * Allocate and initialize the free list array.  This array is used
         * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
         * between io_tlb_start and io_tlb_end.
@@ -157,12 +167,6 @@ void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
        io_tlb_index = 0;
        io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
 
-       /*
-        * Get the overflow emergency buffer
-        */
-       io_tlb_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
-       if (!io_tlb_overflow_buffer)
-               panic("Cannot allocate SWIOTLB overflow buffer!\n");
        if (verbose)
                swiotlb_print_info();
 }
@@ -252,6 +256,7 @@ int
 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
 {
        unsigned long i, bytes;
+       unsigned char *v_overflow_buffer;
 
        bytes = nslabs << IO_TLB_SHIFT;
 
@@ -262,6 +267,16 @@ swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
        memset(tlb, 0, bytes);
 
        /*
+        * Get the overflow emergency buffer
+        */
+       v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
+                                                    get_order(io_tlb_overflow));
+       if (!v_overflow_buffer)
+               goto cleanup2;
+
+       io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
+
+       /*
         * Allocate and initialize the free list array.  This array is used
         * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
         * between io_tlb_start and io_tlb_end.
@@ -269,7 +284,7 @@ swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
        io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
                                      get_order(io_tlb_nslabs * sizeof(int)));
        if (!io_tlb_list)
-               goto cleanup2;
+               goto cleanup3;
 
        for (i = 0; i < io_tlb_nslabs; i++)
                io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
@@ -280,18 +295,10 @@ swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
                                 get_order(io_tlb_nslabs *
                                           sizeof(phys_addr_t)));
        if (!io_tlb_orig_addr)
-               goto cleanup3;
+               goto cleanup4;
 
        memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
 
-       /*
-        * Get the overflow emergency buffer
-        */
-       io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
-                                                 get_order(io_tlb_overflow));
-       if (!io_tlb_overflow_buffer)
-               goto cleanup4;
-
        swiotlb_print_info();
 
        late_alloc = 1;
@@ -299,13 +306,13 @@ swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
        return 0;
 
 cleanup4:
-       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 *
                                                         sizeof(int)));
        io_tlb_list = NULL;
+cleanup3:
+       free_pages((unsigned long)v_overflow_buffer,
+                  get_order(io_tlb_overflow));
+       io_tlb_overflow_buffer = 0;
 cleanup2:
        io_tlb_end = 0;
        io_tlb_start = 0;
@@ -315,11 +322,11 @@ cleanup2:
 
 void __init swiotlb_free(void)
 {
-       if (!io_tlb_overflow_buffer)
+       if (!io_tlb_orig_addr)
                return;
 
        if (late_alloc) {
-               free_pages((unsigned long)io_tlb_overflow_buffer,
+               free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
                           get_order(io_tlb_overflow));
                free_pages((unsigned long)io_tlb_orig_addr,
                           get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
@@ -328,7 +335,7 @@ void __init swiotlb_free(void)
                free_pages((unsigned long)phys_to_virt(io_tlb_start),
                           get_order(io_tlb_nslabs << IO_TLB_SHIFT));
        } else {
-               free_bootmem_late(__pa(io_tlb_overflow_buffer),
+               free_bootmem_late(io_tlb_overflow_buffer,
                                  PAGE_ALIGN(io_tlb_overflow));
                free_bootmem_late(__pa(io_tlb_orig_addr),
                                  PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
@@ -698,7 +705,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
        map = map_single(dev, phys, size, dir);
        if (!map) {
                swiotlb_full(dev, size, dir, 1);
-               map = io_tlb_overflow_buffer;
+               return phys_to_dma(dev, io_tlb_overflow_buffer);
        }
 
        dev_addr = swiotlb_virt_to_bus(dev, map);
@@ -708,7 +715,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
         */
        if (!dma_capable(dev, dev_addr, size)) {
                swiotlb_tbl_unmap_single(dev, map, size, dir);
-               dev_addr = swiotlb_virt_to_bus(dev, io_tlb_overflow_buffer);
+               return phys_to_dma(dev, io_tlb_overflow_buffer);
        }
 
        return dev_addr;
@@ -927,7 +934,7 @@ EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
 int
 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
 {
-       return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
+       return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
 }
 EXPORT_SYMBOL(swiotlb_dma_mapping_error);