ARM: tegra: dvfs: Add dvfs thermal dependency
[linux-3.10.git] / lib / swiotlb.c
index b237e64..d23762e 100644 (file)
@@ -20,7 +20,7 @@
 #include <linux/cache.h>
 #include <linux/dma-mapping.h>
 #include <linux/mm.h>
-#include <linux/module.h>
+#include <linux/export.h>
 #include <linux/spinlock.h>
 #include <linux/string.h>
 #include <linux/swiotlb.h>
@@ -57,10 +57,10 @@ int swiotlb_force;
  * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
  * API.
  */
-static char *io_tlb_start, *io_tlb_end;
+static phys_addr_t io_tlb_start, io_tlb_end;
 
 /*
- * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
+ * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
  * io_tlb_end.  This is command line adjustable via setup_io_tlb_npages.
  */
 static unsigned long io_tlb_nslabs;
@@ -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
@@ -105,11 +105,28 @@ setup_io_tlb_npages(char *str)
        if (!strcmp(str, "force"))
                swiotlb_force = 1;
 
-       return 1;
+       return 0;
 }
-__setup("swiotlb=", setup_io_tlb_npages);
+early_param("swiotlb", setup_io_tlb_npages);
 /* make io_tlb_overflow tunable too? */
 
+unsigned long swiotlb_nr_tbl(void)
+{
+       return io_tlb_nslabs;
+}
+EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
+
+/* default to 64MB */
+#define IO_TLB_DEFAULT_SIZE (64UL<<20)
+unsigned long swiotlb_size_or_default(void)
+{
+       unsigned long size;
+
+       size = io_tlb_nslabs << IO_TLB_SHIFT;
+
+       return size ? size : (IO_TLB_DEFAULT_SIZE);
+}
+
 /* Note that this doesn't work with highmem page */
 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
                                      volatile void *address)
@@ -117,59 +134,74 @@ static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
        return phys_to_dma(hwdev, virt_to_phys(address));
 }
 
+static bool no_iotlb_memory;
+
 void swiotlb_print_info(void)
 {
        unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
-       phys_addr_t pstart, pend;
+       unsigned char *vstart, *vend;
 
-       pstart = virt_to_phys(io_tlb_start);
-       pend = virt_to_phys(io_tlb_end);
+       if (no_iotlb_memory) {
+               pr_warn("software IO TLB: No low mem\n");
+               return;
+       }
+
+       vstart = phys_to_virt(io_tlb_start);
+       vend = phys_to_virt(io_tlb_end);
 
-       printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
-              bytes >> 20, io_tlb_start, io_tlb_end);
-       printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
-              (unsigned long long)pstart,
-              (unsigned long long)pend);
+       printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
+              (unsigned long long)io_tlb_start,
+              (unsigned long long)io_tlb_end,
+              bytes >> 20, vstart, vend - 1);
 }
 
-void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
+int __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;
 
        io_tlb_nslabs = nslabs;
-       io_tlb_start = tlb;
+       io_tlb_start = __pa(tlb);
        io_tlb_end = io_tlb_start + bytes;
 
        /*
+        * Get the overflow emergency buffer
+        */
+       v_overflow_buffer = alloc_bootmem_low_pages_nopanic(
+                                               PAGE_ALIGN(io_tlb_overflow));
+       if (!v_overflow_buffer)
+               return -ENOMEM;
+
+       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.
         */
-       io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
+       io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
        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(phys_addr_t));
+       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(io_tlb_overflow);
-       if (!io_tlb_overflow_buffer)
-               panic("Cannot allocate SWIOTLB overflow buffer!\n");
        if (verbose)
                swiotlb_print_info();
+
+       return 0;
 }
 
 /*
  * Statically reserve bounce buffer space and initialize bounce buffer data
  * structures for the software IO TLB used to implement the DMA API.
  */
-void __init
-swiotlb_init_with_default_size(size_t default_size, int verbose)
+void  __init
+swiotlb_init(int verbose)
 {
+       size_t default_size = IO_TLB_DEFAULT_SIZE;
+       unsigned char *vstart;
        unsigned long bytes;
 
        if (!io_tlb_nslabs) {
@@ -179,20 +211,16 @@ swiotlb_init_with_default_size(size_t default_size, int verbose)
 
        bytes = io_tlb_nslabs << IO_TLB_SHIFT;
 
-       /*
-        * Get IO TLB memory from the low pages
-        */
-       io_tlb_start = alloc_bootmem_low_pages(bytes);
-       if (!io_tlb_start)
-               panic("Cannot allocate SWIOTLB buffer");
-
-       swiotlb_init_with_tbl(io_tlb_start, io_tlb_nslabs, verbose);
-}
+       /* Get IO TLB memory from the low pages */
+       vstart = alloc_bootmem_low_pages_nopanic(PAGE_ALIGN(bytes));
+       if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
+               return;
 
-void __init
-swiotlb_init(int verbose)
-{
-       swiotlb_init_with_default_size(64 * (1<<20), verbose);  /* default to 64MB */
+       if (io_tlb_start)
+               free_bootmem(io_tlb_start,
+                                PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
+       pr_warn("Cannot allocate SWIOTLB buffer");
+       no_iotlb_memory = true;
 }
 
 /*
@@ -203,8 +231,10 @@ swiotlb_init(int verbose)
 int
 swiotlb_late_init_with_default_size(size_t default_size)
 {
-       unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
+       unsigned long bytes, req_nslabs = io_tlb_nslabs;
+       unsigned char *vstart = NULL;
        unsigned int order;
+       int rc = 0;
 
        if (!io_tlb_nslabs) {
                io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
@@ -219,24 +249,51 @@ swiotlb_late_init_with_default_size(size_t default_size)
        bytes = io_tlb_nslabs << IO_TLB_SHIFT;
 
        while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
-               io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
-                                                       order);
-               if (io_tlb_start)
+               vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
+                                                 order);
+               if (vstart)
                        break;
                order--;
        }
 
-       if (!io_tlb_start)
-               goto cleanup1;
-
+       if (!vstart) {
+               io_tlb_nslabs = req_nslabs;
+               return -ENOMEM;
+       }
        if (order != get_order(bytes)) {
                printk(KERN_WARNING "Warning: only able to allocate %ld MB "
                       "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
                io_tlb_nslabs = SLABS_PER_PAGE << order;
-               bytes = io_tlb_nslabs << IO_TLB_SHIFT;
        }
+       rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
+       if (rc)
+               free_pages((unsigned long)vstart, order);
+       return rc;
+}
+
+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;
+
+       io_tlb_nslabs = nslabs;
+       io_tlb_start = virt_to_phys(tlb);
        io_tlb_end = io_tlb_start + bytes;
-       memset(io_tlb_start, 0, bytes);
+
+       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
@@ -246,7 +303,7 @@ swiotlb_late_init_with_default_size(size_t default_size)
        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);
@@ -257,18 +314,10 @@ swiotlb_late_init_with_default_size(size_t default_size)
                                 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;
@@ -276,65 +325,64 @@ 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(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 = NULL;
-       free_pages((unsigned long)io_tlb_start, order);
-       io_tlb_start = NULL;
-cleanup1:
-       io_tlb_nslabs = req_nslabs;
+       io_tlb_end = 0;
+       io_tlb_start = 0;
+       io_tlb_nslabs = 0;
        return -ENOMEM;
 }
 
 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)));
                free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
                                                                 sizeof(int)));
-               free_pages((unsigned long)io_tlb_start,
+               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),
-                                 io_tlb_overflow);
+               free_bootmem_late(io_tlb_overflow_buffer,
+                                 PAGE_ALIGN(io_tlb_overflow));
                free_bootmem_late(__pa(io_tlb_orig_addr),
-                                 io_tlb_nslabs * sizeof(phys_addr_t));
+                                 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
                free_bootmem_late(__pa(io_tlb_list),
-                                 io_tlb_nslabs * sizeof(int));
-               free_bootmem_late(__pa(io_tlb_start),
-                                 io_tlb_nslabs << IO_TLB_SHIFT);
+                                 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
+               free_bootmem_late(io_tlb_start,
+                                 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
        }
+       io_tlb_nslabs = 0;
 }
 
 static int is_swiotlb_buffer(phys_addr_t paddr)
 {
-       return paddr >= virt_to_phys(io_tlb_start) &&
-               paddr < virt_to_phys(io_tlb_end);
+       return paddr >= io_tlb_start && paddr < io_tlb_end;
 }
 
 /*
  * Bounce: copy the swiotlb buffer back to the original dma location
  */
-void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
-                   enum dma_data_direction dir)
+static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
+                          size_t size, enum dma_data_direction dir)
 {
-       unsigned long pfn = PFN_DOWN(phys);
+       unsigned long pfn = PFN_DOWN(orig_addr);
+       unsigned char *vaddr = phys_to_virt(tlb_addr);
 
        if (PageHighMem(pfn_to_page(pfn))) {
                /* The buffer does not have a mapping.  Map it in and copy */
-               unsigned int offset = phys & ~PAGE_MASK;
+               unsigned int offset = orig_addr & ~PAGE_MASK;
                char *buffer;
                unsigned int sz = 0;
                unsigned long flags;
@@ -343,41 +391,42 @@ void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
                        sz = min_t(size_t, PAGE_SIZE - offset, size);
 
                        local_irq_save(flags);
-                       buffer = kmap_atomic(pfn_to_page(pfn),
-                                            KM_BOUNCE_READ);
+                       buffer = kmap_atomic(pfn_to_page(pfn));
                        if (dir == DMA_TO_DEVICE)
-                               memcpy(dma_addr, buffer + offset, sz);
+                               memcpy(vaddr, buffer + offset, sz);
                        else
-                               memcpy(buffer + offset, dma_addr, sz);
-                       kunmap_atomic(buffer, KM_BOUNCE_READ);
+                               memcpy(buffer + offset, vaddr, sz);
+                       kunmap_atomic(buffer);
                        local_irq_restore(flags);
 
                        size -= sz;
                        pfn++;
-                       dma_addr += sz;
+                       vaddr += sz;
                        offset = 0;
                }
+       } else if (dir == DMA_TO_DEVICE) {
+               memcpy(vaddr, phys_to_virt(orig_addr), size);
        } else {
-               if (dir == DMA_TO_DEVICE)
-                       memcpy(dma_addr, phys_to_virt(phys), size);
-               else
-                       memcpy(phys_to_virt(phys), dma_addr, size);
+               memcpy(phys_to_virt(orig_addr), vaddr, size);
        }
 }
-EXPORT_SYMBOL_GPL(swiotlb_bounce);
 
-void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
-                            phys_addr_t phys, size_t size,
-                            enum dma_data_direction dir)
+phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
+                                  dma_addr_t tbl_dma_addr,
+                                  phys_addr_t orig_addr, size_t size,
+                                  enum dma_data_direction dir)
 {
        unsigned long flags;
-       char *dma_addr;
+       phys_addr_t tlb_addr;
        unsigned int nslots, stride, index, wrap;
        int i;
        unsigned long mask;
        unsigned long offset_slots;
        unsigned long max_slots;
 
+       if (no_iotlb_memory)
+               panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
+
        mask = dma_get_seg_boundary(hwdev);
 
        tbl_dma_addr &= mask;
@@ -435,7 +484,7 @@ void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
                                io_tlb_list[i] = 0;
                        for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
                                io_tlb_list[i] = ++count;
-                       dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
+                       tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
 
                        /*
                         * Update the indices to avoid searching in the next
@@ -453,7 +502,7 @@ void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
 
 not_found:
        spin_unlock_irqrestore(&io_tlb_lock, flags);
-       return NULL;
+       return SWIOTLB_MAP_ERROR;
 found:
        spin_unlock_irqrestore(&io_tlb_lock, flags);
 
@@ -463,11 +512,11 @@ found:
         * needed.
         */
        for (i = 0; i < nslots; i++)
-               io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
+               io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
        if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
-               swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
+               swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
 
-       return dma_addr;
+       return tlb_addr;
 }
 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
 
@@ -475,11 +524,10 @@ EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
  * Allocates bounce buffer and returns its kernel virtual address.
  */
 
-static void *
-map_single(struct device *hwdev, phys_addr_t phys, size_t size,
-          enum dma_data_direction dir)
+phys_addr_t map_single(struct device *hwdev, phys_addr_t phys, size_t size,
+                      enum dma_data_direction dir)
 {
-       dma_addr_t start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start);
+       dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
 
        return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
 }
@@ -487,20 +535,19 @@ map_single(struct device *hwdev, phys_addr_t phys, size_t size,
 /*
  * dma_addr is the kernel virtual address of the bounce buffer to unmap.
  */
-void
-swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
-                       enum dma_data_direction dir)
+void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
+                             size_t size, enum dma_data_direction 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;
-       phys_addr_t phys = io_tlb_orig_addr[index];
+       int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
+       phys_addr_t orig_addr = io_tlb_orig_addr[index];
 
        /*
         * First, sync the memory before unmapping the entry
         */
-       if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
-               swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
+       if (orig_addr && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
+               swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
 
        /*
         * Return the buffer to the free list by setting the corresponding
@@ -529,26 +576,27 @@ swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
 }
 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
 
-void
-swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
-                       enum dma_data_direction dir,
-                       enum dma_sync_target target)
+void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
+                            size_t size, enum dma_data_direction dir,
+                            enum dma_sync_target target)
 {
-       int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
-       phys_addr_t phys = io_tlb_orig_addr[index];
+       int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
+       phys_addr_t orig_addr = io_tlb_orig_addr[index];
 
-       phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
+       orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
 
        switch (target) {
        case SYNC_FOR_CPU:
                if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
-                       swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
+                       swiotlb_bounce(orig_addr, tlb_addr,
+                                      size, DMA_FROM_DEVICE);
                else
                        BUG_ON(dir != DMA_TO_DEVICE);
                break;
        case SYNC_FOR_DEVICE:
                if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
-                       swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
+                       swiotlb_bounce(orig_addr, tlb_addr,
+                                      size, DMA_TO_DEVICE);
                else
                        BUG_ON(dir != DMA_FROM_DEVICE);
                break;
@@ -571,12 +619,15 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
                dma_mask = hwdev->coherent_dma_mask;
 
        ret = (void *)__get_free_pages(flags, order);
-       if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
-               /*
-                * The allocated memory isn't reachable by the device.
-                */
-               free_pages((unsigned long) ret, order);
-               ret = NULL;
+       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) {
                /*
@@ -584,25 +635,29 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
                 * GFP_DMA memory; fall back on map_single(), which
                 * will grab memory from the lowest available address range.
                 */
-               ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
-               if (!ret)
+               phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
+               if (paddr == SWIOTLB_MAP_ERROR)
                        return NULL;
-       }
 
-       memset(ret, 0, size);
-       dev_addr = swiotlb_virt_to_bus(hwdev, ret);
+               ret = phys_to_virt(paddr);
+               dev_addr = phys_to_dma(hwdev, paddr);
 
-       /* 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);
+               /* 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);
 
-               /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
-               swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
-               return NULL;
+                       /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
+                       swiotlb_tbl_unmap_single(hwdev, paddr,
+                                                size, DMA_TO_DEVICE);
+                       return NULL;
+               }
        }
+
        *dma_handle = dev_addr;
+       memset(ret, 0, size);
+
        return ret;
 }
 EXPORT_SYMBOL(swiotlb_alloc_coherent);
@@ -618,7 +673,7 @@ swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
                free_pages((unsigned long)vaddr, get_order(size));
        else
                /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
-               swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
+               swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
 }
 EXPORT_SYMBOL(swiotlb_free_coherent);
 
@@ -659,9 +714,8 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
                            enum dma_data_direction dir,
                            struct dma_attrs *attrs)
 {
-       phys_addr_t phys = page_to_phys(page) + offset;
+       phys_addr_t map, phys = page_to_phys(page) + offset;
        dma_addr_t dev_addr = phys_to_dma(dev, phys);
-       void *map;
 
        BUG_ON(dir == DMA_NONE);
        /*
@@ -672,22 +726,20 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
        if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
                return dev_addr;
 
-       /*
-        * Oh well, have to allocate and map a bounce buffer.
-        */
+       /* Oh well, have to allocate and map a bounce buffer. */
        map = map_single(dev, phys, size, dir);
-       if (!map) {
+       if (map == SWIOTLB_MAP_ERROR) {
                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);
+       dev_addr = phys_to_dma(dev, map);
 
-       /*
-        * Ensure that the address returned is DMA'ble
-        */
-       if (!dma_capable(dev, dev_addr, size))
-               panic("map_single: bounce buffer is not DMA'ble");
+       /* Ensure that the address returned is DMA'ble */
+       if (!dma_capable(dev, dev_addr, size)) {
+               swiotlb_tbl_unmap_single(dev, map, size, dir);
+               return phys_to_dma(dev, io_tlb_overflow_buffer);
+       }
 
        return dev_addr;
 }
@@ -709,7 +761,7 @@ static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
        BUG_ON(dir == DMA_NONE);
 
        if (is_swiotlb_buffer(paddr)) {
-               swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
+               swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
                return;
        }
 
@@ -753,8 +805,7 @@ swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
        BUG_ON(dir == DMA_NONE);
 
        if (is_swiotlb_buffer(paddr)) {
-               swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
-                                      target);
+               swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
                return;
        }
 
@@ -811,9 +862,9 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
 
                if (swiotlb_force ||
                    !dma_capable(hwdev, dev_addr, sg->length)) {
-                       void *map = map_single(hwdev, sg_phys(sg),
-                                              sg->length, dir);
-                       if (!map) {
+                       phys_addr_t map = map_single(hwdev, sg_phys(sg),
+                                                    sg->length, dir);
+                       if (map == SWIOTLB_MAP_ERROR) {
                                /* Don't panic here, we expect map_sg users
                                   to do proper error handling. */
                                swiotlb_full(hwdev, sg->length, dir, 0);
@@ -822,7 +873,7 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
                                sgl[0].dma_length = 0;
                                return 0;
                        }
-                       sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
+                       sg->dma_address = phys_to_dma(hwdev, map);
                } else
                        sg->dma_address = dev_addr;
                sg->dma_length = sg->length;
@@ -905,7 +956,7 @@ 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);
 
@@ -918,6 +969,6 @@ EXPORT_SYMBOL(swiotlb_dma_mapping_error);
 int
 swiotlb_dma_supported(struct device *hwdev, u64 mask)
 {
-       return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
+       return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
 }
 EXPORT_SYMBOL(swiotlb_dma_supported);