[IA64] swiotlb abstraction (e.g. for Xen)
authorJan Beulich <jbeulich@novell.com>
Tue, 6 Feb 2007 02:53:04 +0000 (18:53 -0800)
committerTony Luck <tony.luck@intel.com>
Tue, 6 Feb 2007 02:53:04 +0000 (18:53 -0800)
Add abstraction so that the file can be used by environments other than IA64
and EM64T, namely for Xen.

Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Tony Luck <tony.luck@intel.com>
include/asm-ia64/swiotlb.h [new file with mode: 0644]
include/asm-x86_64/swiotlb.h
lib/swiotlb.c

diff --git a/include/asm-ia64/swiotlb.h b/include/asm-ia64/swiotlb.h
new file mode 100644 (file)
index 0000000..452c162
--- /dev/null
@@ -0,0 +1,9 @@
+#ifndef _ASM_SWIOTLB_H
+#define _ASM_SWIOTLB_H 1
+
+#include <asm/machvec.h>
+
+#define SWIOTLB_ARCH_NEED_LATE_INIT
+#define SWIOTLB_ARCH_NEED_ALLOC
+
+#endif /* _ASM_SWIOTLB_H */
index f9c589539a82b7c11fae1425c5782b5189db9cc1..ab913ffcad569201bb921bc6e7c4bfc1db5f450b 100644 (file)
@@ -44,6 +44,7 @@ extern void swiotlb_init(void);
 extern int swiotlb_force;
 
 #ifdef CONFIG_SWIOTLB
+#define SWIOTLB_ARCH_NEED_ALLOC
 extern int swiotlb;
 #else
 #define swiotlb 0
index 067eed5b275864f3ece19ad6c31b96e6c6a3ed76..50a43801018259422345f815b39876b6a1a6a78f 100644 (file)
@@ -28,6 +28,7 @@
 #include <asm/io.h>
 #include <asm/dma.h>
 #include <asm/scatterlist.h>
+#include <asm/swiotlb.h>
 
 #include <linux/init.h>
 #include <linux/bootmem.h>
 #define OFFSET(val,align) ((unsigned long)     \
                           ( (val) & ( (align) - 1)))
 
+#ifndef SG_ENT_VIRT_ADDRESS
 #define SG_ENT_VIRT_ADDRESS(sg)        (page_address((sg)->page) + (sg)->offset)
 #define SG_ENT_PHYS_ADDRESS(sg)        virt_to_bus(SG_ENT_VIRT_ADDRESS(sg))
+#endif
 
 /*
  * Maximum allowable number of contiguous slabs to map,
@@ -101,13 +104,25 @@ static unsigned int io_tlb_index;
  * We need to save away the original address corresponding to a mapped entry
  * for the sync operations.
  */
-static unsigned char **io_tlb_orig_addr;
+#ifndef SWIOTLB_ARCH_HAS_IO_TLB_ADDR_T
+typedef char *io_tlb_addr_t;
+#define swiotlb_orig_addr_null(buffer) (!(buffer))
+#define ptr_to_io_tlb_addr(ptr) (ptr)
+#define page_to_io_tlb_addr(pg, off) (page_address(pg) + (off))
+#define sg_to_io_tlb_addr(sg) SG_ENT_VIRT_ADDRESS(sg)
+#endif
+static io_tlb_addr_t *io_tlb_orig_addr;
 
 /*
  * Protect the above data structures in the map and unmap calls
  */
 static DEFINE_SPINLOCK(io_tlb_lock);
 
+#ifdef SWIOTLB_EXTRA_VARIABLES
+SWIOTLB_EXTRA_VARIABLES;
+#endif
+
+#ifndef SWIOTLB_ARCH_HAS_SETUP_IO_TLB_NPAGES
 static int __init
 setup_io_tlb_npages(char *str)
 {
@@ -122,9 +137,25 @@ setup_io_tlb_npages(char *str)
                swiotlb_force = 1;
        return 1;
 }
+#endif
 __setup("swiotlb=", setup_io_tlb_npages);
 /* make io_tlb_overflow tunable too? */
 
+#ifndef swiotlb_adjust_size
+#define swiotlb_adjust_size(size) ((void)0)
+#endif
+
+#ifndef swiotlb_adjust_seg
+#define swiotlb_adjust_seg(start, size) ((void)0)
+#endif
+
+#ifndef swiotlb_print_info
+#define swiotlb_print_info(bytes) \
+       printk(KERN_INFO "Placing %luMB software IO TLB between 0x%lx - " \
+              "0x%lx\n", bytes >> 20, \
+              virt_to_bus(io_tlb_start), virt_to_bus(io_tlb_end))
+#endif
+
 /*
  * Statically reserve bounce buffer space and initialize bounce buffer data
  * structures for the software IO TLB used to implement the DMA API.
@@ -138,6 +169,8 @@ swiotlb_init_with_default_size(size_t default_size)
                io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
                io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
        }
+       swiotlb_adjust_size(io_tlb_nslabs);
+       swiotlb_adjust_size(io_tlb_overflow);
 
        bytes = io_tlb_nslabs << IO_TLB_SHIFT;
 
@@ -155,10 +188,14 @@ swiotlb_init_with_default_size(size_t default_size)
         * between io_tlb_start and io_tlb_end.
         */
        io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
-       for (i = 0; i < io_tlb_nslabs; i++)
+       for (i = 0; i < io_tlb_nslabs; i++) {
+               if ( !(i % IO_TLB_SEGSIZE) )
+                       swiotlb_adjust_seg(io_tlb_start + (i << IO_TLB_SHIFT),
+                               IO_TLB_SEGSIZE << IO_TLB_SHIFT);
                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(char *));
+       io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(io_tlb_addr_t));
 
        /*
         * Get the overflow emergency buffer
@@ -166,17 +203,21 @@ swiotlb_init_with_default_size(size_t default_size)
        io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
        if (!io_tlb_overflow_buffer)
                panic("Cannot allocate SWIOTLB overflow buffer!\n");
+       swiotlb_adjust_seg(io_tlb_overflow_buffer, io_tlb_overflow);
 
-       printk(KERN_INFO "Placing software IO TLB between 0x%lx - 0x%lx\n",
-              virt_to_bus(io_tlb_start), virt_to_bus(io_tlb_end));
+       swiotlb_print_info(bytes);
 }
+#ifndef __swiotlb_init_with_default_size
+#define __swiotlb_init_with_default_size swiotlb_init_with_default_size
+#endif
 
 void __init
 swiotlb_init(void)
 {
-       swiotlb_init_with_default_size(64 * (1<<20));   /* default to 64MB */
+       __swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
 }
 
+#ifdef SWIOTLB_ARCH_NEED_LATE_INIT
 /*
  * Systems with larger DMA zones (those that don't support ISA) can
  * initialize the swiotlb later using the slab allocator if needed.
@@ -234,12 +275,12 @@ 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 = (unsigned char **)__get_free_pages(GFP_KERNEL,
-                                  get_order(io_tlb_nslabs * sizeof(char *)));
+       io_tlb_orig_addr = (io_tlb_addr_t *)__get_free_pages(GFP_KERNEL,
+                                  get_order(io_tlb_nslabs * sizeof(io_tlb_addr_t)));
        if (!io_tlb_orig_addr)
                goto cleanup3;
 
-       memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(char *));
+       memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(io_tlb_addr_t));
 
        /*
         * Get the overflow emergency buffer
@@ -249,19 +290,17 @@ swiotlb_late_init_with_default_size(size_t default_size)
        if (!io_tlb_overflow_buffer)
                goto cleanup4;
 
-       printk(KERN_INFO "Placing %luMB software IO TLB between 0x%lx - "
-              "0x%lx\n", bytes >> 20,
-              virt_to_bus(io_tlb_start), virt_to_bus(io_tlb_end));
+       swiotlb_print_info(bytes);
 
        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(io_tlb_addr_t)));
        io_tlb_orig_addr = NULL;
 cleanup3:
-       free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
-                                                        sizeof(int)));
+       free_pages((unsigned long)io_tlb_list,
+                  get_order(io_tlb_nslabs * sizeof(int)));
        io_tlb_list = NULL;
 cleanup2:
        io_tlb_end = NULL;
@@ -271,7 +310,9 @@ cleanup1:
        io_tlb_nslabs = req_nslabs;
        return -ENOMEM;
 }
+#endif
 
+#ifndef SWIOTLB_ARCH_HAS_NEEDS_MAPPING
 static inline int
 address_needs_mapping(struct device *hwdev, dma_addr_t addr)
 {
@@ -282,11 +323,35 @@ address_needs_mapping(struct device *hwdev, dma_addr_t addr)
        return (addr & ~mask) != 0;
 }
 
+static inline int range_needs_mapping(const void *ptr, size_t size)
+{
+       return swiotlb_force;
+}
+
+static inline int order_needs_mapping(unsigned int order)
+{
+       return 0;
+}
+#endif
+
+static void
+__sync_single(io_tlb_addr_t buffer, char *dma_addr, size_t size, int dir)
+{
+#ifndef SWIOTLB_ARCH_HAS_SYNC_SINGLE
+       if (dir == DMA_TO_DEVICE)
+               memcpy(dma_addr, buffer, size);
+       else
+               memcpy(buffer, dma_addr, size);
+#else
+       __swiotlb_arch_sync_single(buffer, dma_addr, size, dir);
+#endif
+}
+
 /*
  * Allocates bounce buffer and returns its kernel virtual address.
  */
 static void *
-map_single(struct device *hwdev, char *buffer, size_t size, int dir)
+map_single(struct device *hwdev, io_tlb_addr_t buffer, size_t size, int dir)
 {
        unsigned long flags;
        char *dma_addr;
@@ -359,7 +424,7 @@ map_single(struct device *hwdev, char *buffer, size_t size, int dir)
         */
        io_tlb_orig_addr[index] = buffer;
        if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
-               memcpy(dma_addr, buffer, size);
+               __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
 
        return dma_addr;
 }
@@ -373,17 +438,18 @@ 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;
-       char *buffer = io_tlb_orig_addr[index];
+       io_tlb_addr_t buffer = io_tlb_orig_addr[index];
 
        /*
         * First, sync the memory before unmapping the entry
         */
-       if (buffer && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
+       if (!swiotlb_orig_addr_null(buffer)
+           && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
                /*
                 * bounce... copy the data back into the original buffer * and
                 * delete the bounce buffer.
                 */
-               memcpy(buffer, dma_addr, size);
+               __sync_single(buffer, dma_addr, size, DMA_FROM_DEVICE);
 
        /*
         * Return the buffer to the free list by setting the corresponding
@@ -416,18 +482,18 @@ sync_single(struct device *hwdev, char *dma_addr, size_t size,
            int dir, int target)
 {
        int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
-       char *buffer = io_tlb_orig_addr[index];
+       io_tlb_addr_t buffer = io_tlb_orig_addr[index];
 
        switch (target) {
        case SYNC_FOR_CPU:
                if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
-                       memcpy(buffer, dma_addr, size);
+                       __sync_single(buffer, dma_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))
-                       memcpy(dma_addr, buffer, size);
+                       __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
                else
                        BUG_ON(dir != DMA_FROM_DEVICE);
                break;
@@ -436,6 +502,8 @@ sync_single(struct device *hwdev, char *dma_addr, size_t size,
        }
 }
 
+#ifdef SWIOTLB_ARCH_NEED_ALLOC
+
 void *
 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
                       dma_addr_t *dma_handle, gfp_t flags)
@@ -451,7 +519,10 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
         */
        flags |= GFP_DMA;
 
-       ret = (void *)__get_free_pages(flags, order);
+       if (!order_needs_mapping(order))
+               ret = (void *)__get_free_pages(flags, order);
+       else
+               ret = NULL;
        if (ret && address_needs_mapping(hwdev, virt_to_bus(ret))) {
                /*
                 * The allocated memory isn't reachable by the device.
@@ -489,6 +560,7 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
        *dma_handle = dev_addr;
        return ret;
 }
+EXPORT_SYMBOL(swiotlb_alloc_coherent);
 
 void
 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
@@ -501,6 +573,9 @@ swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
                /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
                swiotlb_unmap_single (hwdev, dma_handle, size, DMA_TO_DEVICE);
 }
+EXPORT_SYMBOL(swiotlb_free_coherent);
+
+#endif
 
 static void
 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
@@ -542,13 +617,14 @@ swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir)
         * we can safely return the device addr and not worry about bounce
         * buffering it.
         */
-       if (!address_needs_mapping(hwdev, dev_addr) && !swiotlb_force)
+       if (!range_needs_mapping(ptr, size)
+           && !address_needs_mapping(hwdev, dev_addr))
                return dev_addr;
 
        /*
         * Oh well, have to allocate and map a bounce buffer.
         */
-       map = map_single(hwdev, ptr, size, dir);
+       map = map_single(hwdev, ptr_to_io_tlb_addr(ptr), size, dir);
        if (!map) {
                swiotlb_full(hwdev, size, dir, 1);
                map = io_tlb_overflow_buffer;
@@ -676,17 +752,16 @@ int
 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sg, int nelems,
               int dir)
 {
-       void *addr;
        dma_addr_t dev_addr;
        int i;
 
        BUG_ON(dir == DMA_NONE);
 
        for (i = 0; i < nelems; i++, sg++) {
-               addr = SG_ENT_VIRT_ADDRESS(sg);
-               dev_addr = virt_to_bus(addr);
-               if (swiotlb_force || address_needs_mapping(hwdev, dev_addr)) {
-                       void *map = map_single(hwdev, addr, sg->length, dir);
+               dev_addr = SG_ENT_PHYS_ADDRESS(sg);
+               if (range_needs_mapping(SG_ENT_VIRT_ADDRESS(sg), sg->length)
+                   || address_needs_mapping(hwdev, dev_addr)) {
+                       void *map = map_single(hwdev, sg_to_io_tlb_addr(sg), sg->length, dir);
                        if (!map) {
                                /* Don't panic here, we expect map_sg users
                                   to do proper error handling. */
@@ -760,6 +835,44 @@ swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
        swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
 }
 
+#ifdef SWIOTLB_ARCH_NEED_MAP_PAGE
+
+dma_addr_t
+swiotlb_map_page(struct device *hwdev, struct page *page,
+                unsigned long offset, size_t size,
+                enum dma_data_direction direction)
+{
+       dma_addr_t dev_addr;
+       char *map;
+
+       dev_addr = page_to_bus(page) + offset;
+       if (address_needs_mapping(hwdev, dev_addr)) {
+               map = map_single(hwdev, page_to_io_tlb_addr(page, offset), size, direction);
+               if (!map) {
+                       swiotlb_full(hwdev, size, direction, 1);
+                       map = io_tlb_overflow_buffer;
+               }
+               dev_addr = virt_to_bus(map);
+       }
+
+       return dev_addr;
+}
+
+void
+swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
+                  size_t size, enum dma_data_direction direction)
+{
+       char *dma_addr = bus_to_virt(dev_addr);
+
+       BUG_ON(direction == DMA_NONE);
+       if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
+               unmap_single(hwdev, dma_addr, size, direction);
+       else if (direction == DMA_FROM_DEVICE)
+               dma_mark_clean(dma_addr, size);
+}
+
+#endif
+
 int
 swiotlb_dma_mapping_error(dma_addr_t dma_addr)
 {
@@ -772,10 +885,13 @@ swiotlb_dma_mapping_error(dma_addr_t dma_addr)
  * during bus mastering, then you would pass 0x00ffffff as the mask to
  * this function.
  */
+#ifndef __swiotlb_dma_supported
+#define __swiotlb_dma_supported(hwdev, mask) (virt_to_bus(io_tlb_end - 1) <= (mask))
+#endif
 int
 swiotlb_dma_supported(struct device *hwdev, u64 mask)
 {
-       return virt_to_bus(io_tlb_end - 1) <= mask;
+       return __swiotlb_dma_supported(hwdev, mask);
 }
 
 EXPORT_SYMBOL(swiotlb_init);
@@ -790,6 +906,4 @@ EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
-EXPORT_SYMBOL(swiotlb_alloc_coherent);
-EXPORT_SYMBOL(swiotlb_free_coherent);
 EXPORT_SYMBOL(swiotlb_dma_supported);