[PATCH] x86_64: Use function pointers to call DMA mapping functions

AK: I hacked Muli's original patch a lot and there were a lot
of changes - all bugs are probably to blame on me now.
There were also some changes in the fall back behaviour
for swiotlb - in particular it doesn't try to use GFP_DMA
now anymore. Also all DMA mapping operations use the
same core dma_alloc_coherent code with proper fallbacks now.
And various other changes and cleanups.

Known problems: iommu=force swiotlb=force together breaks
                needs more testing.

This patch cleans up x86_64's DMA mapping dispatching code. Right now
we have three possible IOMMU types: AGP GART, swiotlb and nommu, and
in the future we will also have Xen's x86_64 swiotlb and other HW
IOMMUs for x86_64. In order to support all of them cleanly, this
patch:

- introduces a struct dma_mapping_ops with function pointers for each
  of the DMA mapping operations of gart (AMD HW IOMMU), swiotlb
  (software IOMMU) and nommu (no IOMMU).

- gets rid of:

  if (swiotlb)
      return swiotlb_xxx();

- PCI_DMA_BUS_IS_PHYS is now checked against the dma_ops being set
This makes swiotlb faster by avoiding double copying in some cases.

Signed-Off-By: Muli Ben-Yehuda <mulix@mulix.org>
Signed-Off-By: Jon D. Mason <jdmason@us.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

6 files changed, 464 insertions, 433 deletions

diff --git a/arch/x86_64/kernel/Makefile b/arch/x86_64/kernel/Makefile
index 4601fdafb7cb..7df09edb36a1 100644
--- a/arch/x86_64/kernel/Makefile
+++ b/arch/x86_64/kernel/Makefile
@@ -8,7 +8,7 @@ obj-y	:= process.o signal.o entry.o traps.o irq.o \
                 ptrace.o time.o ioport.o ldt.o setup.o i8259.o sys_x86_64.o \
                 x8664_ksyms.o i387.o syscall.o vsyscall.o \
                 setup64.o bootflag.o e820.o reboot.o quirks.o i8237.o \
-                dmi_scan.o
+                dmi_scan.o pci-dma.o pci-nommu.o
 
 obj-$(CONFIG_X86_MCE)         += mce.o
 obj-$(CONFIG_X86_MCE_INTEL)     += mce_intel.o
@@ -29,7 +29,7 @@ obj-$(CONFIG_SOFTWARE_SUSPEND)	+= suspend_asm.o
 obj-$(CONFIG_CPU_FREQ)          += cpufreq/
 obj-$(CONFIG_EARLY_PRINTK)      += early_printk.o
 obj-$(CONFIG_GART_IOMMU)        += pci-gart.o aperture.o
-obj-$(CONFIG_DUMMY_IOMMU)       += pci-nommu.o pci-dma.o
+obj-$(CONFIG_SWIOTLB)           += pci-swiotlb.o
 obj-$(CONFIG_KPROBES)           += kprobes.o
 obj-$(CONFIG_X86_PM_TIMER)      += pmtimer.o
 
diff --git a/arch/x86_64/kernel/pci-dma.c b/arch/x86_64/kernel/pci-dma.c
index cab471cf3edb..2f5d8328e2b9 100644
--- a/arch/x86_64/kernel/pci-dma.c
+++ b/arch/x86_64/kernel/pci-dma.c
@@ -8,53 +8,259 @@
 #include <linux/pci.h>
 #include <linux/module.h>
 #include <asm/io.h>
+#include <asm/proto.h>
 
-/* Map a set of buffers described by scatterlist in streaming
- * mode for DMA.  This is the scatter-gather version of the
- * above pci_map_single interface.  Here the scatter gather list
- * elements are each tagged with the appropriate dma address
- * and length.  They are obtained via sg_dma_{address,length}(SG).
- *
- * NOTE: An implementation may be able to use a smaller number of
- *       DMA address/length pairs than there are SG table elements.
- *       (for example via virtual mapping capabilities)
- *       The routine returns the number of addr/length pairs actually
- *       used, at most nents.
- *
- * Device ownership issues as mentioned above for pci_map_single are
- * the same here.
+int iommu_merge __read_mostly = 0;
+EXPORT_SYMBOL(iommu_merge);
+
+dma_addr_t bad_dma_address __read_mostly;
+EXPORT_SYMBOL(bad_dma_address);
+
+/* This tells the BIO block layer to assume merging. Default to off
+   because we cannot guarantee merging later. */
+int iommu_bio_merge __read_mostly = 0;
+EXPORT_SYMBOL(iommu_bio_merge);
+
+int iommu_sac_force __read_mostly = 0;
+EXPORT_SYMBOL(iommu_sac_force);
+
+int no_iommu __read_mostly;
+#ifdef CONFIG_IOMMU_DEBUG
+int panic_on_overflow __read_mostly = 1;
+int force_iommu __read_mostly = 1;
+#else
+int panic_on_overflow __read_mostly = 0;
+int force_iommu __read_mostly= 0;
+#endif
+
+/* Dummy device used for NULL arguments (normally ISA). Better would
+   be probably a smaller DMA mask, but this is bug-to-bug compatible
+   to i386. */
+struct device fallback_dev = {
+        .bus_id = "fallback device",
+        .coherent_dma_mask = 0xffffffff,
+        .dma_mask = &fallback_dev.coherent_dma_mask,
+};
+
+/* Allocate DMA memory on node near device */
+noinline static void *
+dma_alloc_pages(struct device *dev, gfp_t gfp, unsigned order)
+{
+        struct page *page;
+        int node;
+        if (dev->bus == &pci_bus_type)
+                node = pcibus_to_node(to_pci_dev(dev)->bus);
+        else
+                node = numa_node_id();
+        page = alloc_pages_node(node, gfp, order);
+        return page ? page_address(page) : NULL;
+}
+
+/*
+ * Allocate memory for a coherent mapping.
  */
-int dma_map_sg(struct device *hwdev, struct scatterlist *sg,
-               int nents, int direction)
+void *
+dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
+                   gfp_t gfp)
 {
-        int i;
-
-        BUG_ON(direction == DMA_NONE);
-        for (i = 0; i < nents; i++ ) {
-                struct scatterlist *s = &sg[i];
-                BUG_ON(!s->page); 
-                s->dma_address = virt_to_bus(page_address(s->page) +s->offset);
-                s->dma_length = s->length;
+        void *memory;
+        unsigned long dma_mask = 0;
+        u64 bus;
+
+        if (!dev)
+                dev = &fallback_dev;
+        dma_mask = dev->coherent_dma_mask;
+        if (dma_mask == 0)
+                dma_mask = 0xffffffff;
+
+        /* Kludge to make it bug-to-bug compatible with i386. i386
+           uses the normal dma_mask for alloc_coherent. */
+        dma_mask &= *dev->dma_mask;
+
+        /* Why <=? Even when the mask is smaller than 4GB it is often
+           larger than 16MB and in this case we have a chance of
+           finding fitting memory in the next higher zone first. If
+           not retry with true GFP_DMA. -AK */
+        if (dma_mask <= 0xffffffff)
+                gfp |= GFP_DMA32;
+
+ again:
+        memory = dma_alloc_pages(dev, gfp, get_order(size));
+        if (memory == NULL)
+                return NULL;
+
+        {
+                int high, mmu;
+                bus = virt_to_bus(memory);
+                high = (bus + size) >= dma_mask;
+                mmu = high;
+                if (force_iommu && !(gfp & GFP_DMA))
+                        mmu = 1;
+                else if (high) {
+                        free_pages((unsigned long)memory,
+                                   get_order(size));
+
+                        /* Don't use the 16MB ZONE_DMA unless absolutely
+                           needed. It's better to use remapping first. */
+                        if (dma_mask < 0xffffffff && !(gfp & GFP_DMA)) {
+                                gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
+                                goto again;
+                        }
+
+                        if (dma_ops->alloc_coherent)
+                                return dma_ops->alloc_coherent(dev, size,
+                                                           dma_handle, gfp);
+                        return NULL;
+                }
+
+                memset(memory, 0, size);
+                if (!mmu) {
+                        *dma_handle = virt_to_bus(memory);
+                        return memory;
+                }
+        }
+
+        if (dma_ops->alloc_coherent) {
+                free_pages((unsigned long)memory, get_order(size));
+                gfp &= ~(GFP_DMA|GFP_DMA32);
+                return dma_ops->alloc_coherent(dev, size, dma_handle, gfp);
+        }
+
+        if (dma_ops->map_simple) {
+                *dma_handle = dma_ops->map_simple(dev, memory,
+                                              size,
+                                              PCI_DMA_BIDIRECTIONAL);
+                if (*dma_handle != bad_dma_address)
+                        return memory;
         }
-        return nents;
-}
 
-EXPORT_SYMBOL(dma_map_sg);
+        if (panic_on_overflow)
+                panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n",size);
+        free_pages((unsigned long)memory, get_order(size));
+        return NULL;
+}
+EXPORT_SYMBOL(dma_alloc_coherent);
 
-/* Unmap a set of streaming mode DMA translations.
- * Again, cpu read rules concerning calls here are the same as for
- * pci_unmap_single() above.
+/*
+ * Unmap coherent memory.
+ * The caller must ensure that the device has finished accessing the mapping.
  */
-void dma_unmap_sg(struct device *dev, struct scatterlist *sg,
-                  int nents, int dir)
+void dma_free_coherent(struct device *dev, size_t size,
+                         void *vaddr, dma_addr_t bus)
+{
+        if (dma_ops->unmap_single)
+                dma_ops->unmap_single(dev, bus, size, 0);
+        free_pages((unsigned long)vaddr, get_order(size));
+}
+EXPORT_SYMBOL(dma_free_coherent);
+
+int dma_supported(struct device *dev, u64 mask)
+{
+        if (dma_ops->dma_supported)
+                return dma_ops->dma_supported(dev, mask);
+
+        /* Copied from i386. Doesn't make much sense, because it will
+           only work for pci_alloc_coherent.
+           The caller just has to use GFP_DMA in this case. */
+        if (mask < 0x00ffffff)
+                return 0;
+
+        /* Tell the device to use SAC when IOMMU force is on.  This
+           allows the driver to use cheaper accesses in some cases.
+
+           Problem with this is that if we overflow the IOMMU area and
+           return DAC as fallback address the device may not handle it
+           correctly.
+
+           As a special case some controllers have a 39bit address
+           mode that is as efficient as 32bit (aic79xx). Don't force
+           SAC for these.  Assume all masks <= 40 bits are of this
+           type. Normally this doesn't make any difference, but gives
+           more gentle handling of IOMMU overflow. */
+        if (iommu_sac_force && (mask >= 0xffffffffffULL)) {
+                printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
+                return 0;
+        }
+
+        return 1;
+}
+EXPORT_SYMBOL(dma_supported);
+
+int dma_set_mask(struct device *dev, u64 mask)
 {
-        int i;
-        for (i = 0; i < nents; i++) { 
-                struct scatterlist *s = &sg[i];
-                BUG_ON(s->page == NULL); 
-                BUG_ON(s->dma_address == 0); 
-                dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
-        } 
+        if (!dev->dma_mask || !dma_supported(dev, mask))
+                return -EIO;
+        *dev->dma_mask = mask;
+        return 0;
 }
+EXPORT_SYMBOL(dma_set_mask);
+
+/* iommu=[size][,noagp][,off][,force][,noforce][,leak][,memaper[=order]][,merge]
+         [,forcesac][,fullflush][,nomerge][,biomerge]
+   size  set size of iommu (in bytes)
+   noagp don't initialize the AGP driver and use full aperture.
+   off   don't use the IOMMU
+   leak  turn on simple iommu leak tracing (only when CONFIG_IOMMU_LEAK is on)
+   memaper[=order] allocate an own aperture over RAM with size 32MB^order.
+   noforce don't force IOMMU usage. Default.
+   force  Force IOMMU.
+   merge  Do lazy merging. This may improve performance on some block devices.
+          Implies force (experimental)
+   biomerge Do merging at the BIO layer. This is more efficient than merge,
+            but should be only done with very big IOMMUs. Implies merge,force.
+   nomerge Don't do SG merging.
+   forcesac For SAC mode for masks <40bits  (experimental)
+   fullflush Flush IOMMU on each allocation (default)
+   nofullflush Don't use IOMMU fullflush
+   allowed  overwrite iommu off workarounds for specific chipsets.
+   soft  Use software bounce buffering (default for Intel machines)
+   noaperture Don't touch the aperture for AGP.
+*/
+__init int iommu_setup(char *p)
+{
+    iommu_merge = 1;
 
-EXPORT_SYMBOL(dma_unmap_sg);
+    while (*p) {
+            if (!strncmp(p,"off",3))
+                    no_iommu = 1;
+            /* gart_parse_options has more force support */
+            if (!strncmp(p,"force",5))
+                    force_iommu = 1;
+            if (!strncmp(p,"noforce",7)) {
+                    iommu_merge = 0;
+                    force_iommu = 0;
+            }
+
+            if (!strncmp(p, "biomerge",8)) {
+                    iommu_bio_merge = 4096;
+                    iommu_merge = 1;
+                    force_iommu = 1;
+            }
+            if (!strncmp(p, "panic",5))
+                    panic_on_overflow = 1;
+            if (!strncmp(p, "nopanic",7))
+                    panic_on_overflow = 0;
+            if (!strncmp(p, "merge",5)) {
+                    iommu_merge = 1;
+                    force_iommu = 1;
+            }
+            if (!strncmp(p, "nomerge",7))
+                    iommu_merge = 0;
+            if (!strncmp(p, "forcesac",8))
+                    iommu_sac_force = 1;
+
+#ifdef CONFIG_SWIOTLB
+            if (!strncmp(p, "soft",4))
+                    swiotlb = 1;
+#endif
+
+#ifdef CONFIG_GART_IOMMU
+            gart_parse_options(p);
+#endif
+
+            p += strcspn(p, ",");
+            if (*p == ',')
+                    ++p;
+    }
+    return 1;
+}
diff --git a/arch/x86_64/kernel/pci-gart.c b/arch/x86_64/kernel/pci-gart.c
index 116e00f1bb7b..c37fc7726ba6 100644
--- a/arch/x86_64/kernel/pci-gart.c
+++ b/arch/x86_64/kernel/pci-gart.c
@@ -30,8 +30,8 @@
 #include <asm/proto.h>
 #include <asm/cacheflush.h>
 #include <asm/kdebug.h>
-
-dma_addr_t bad_dma_address;
+#include <asm/swiotlb.h>
+#include <asm/dma.h>
 
 unsigned long iommu_bus_base;   /* GART remapping area (physical) */
 static unsigned long iommu_size;        /* size of remapping area bytes */
@@ -39,18 +39,6 @@ static unsigned long iommu_pages;	/* .. and in pages */
 
 u32 *iommu_gatt_base;           /* Remapping table */
 
-int no_iommu; 
-static int no_agp; 
-#ifdef CONFIG_IOMMU_DEBUG
-int panic_on_overflow = 1; 
-int force_iommu = 1;
-#else
-int panic_on_overflow = 0;
-int force_iommu = 0;
-#endif
-int iommu_merge = 1;
-int iommu_sac_force = 0; 
-
 /* If this is disabled the IOMMU will use an optimized flushing strategy
    of only flushing when an mapping is reused. With it true the GART is flushed 
    for every mapping. Problem is that doing the lazy flush seems to trigger
@@ -58,10 +46,6 @@ int iommu_sac_force = 0;
    also seen with Qlogic at least). */
 int iommu_fullflush = 1;
 
-/* This tells the BIO block layer to assume merging. Default to off
-   because we cannot guarantee merging later. */
-int iommu_bio_merge = 0;
-
 #define MAX_NB 8
 
 /* Allocation bitmap for the remapping area */ 
@@ -102,16 +86,6 @@ AGPEXTERN __u32 *agp_gatt_table;
 
 static unsigned long next_bit;  /* protected by iommu_bitmap_lock */
 static int need_flush;          /* global flush state. set for each gart wrap */
-static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
-                               size_t size, int dir, int do_panic);
-
-/* Dummy device used for NULL arguments (normally ISA). Better would
-   be probably a smaller DMA mask, but this is bug-to-bug compatible to i386. */
-static struct device fallback_dev = {
-        .bus_id = "fallback device",
-        .coherent_dma_mask = 0xffffffff,
-        .dma_mask = &fallback_dev.coherent_dma_mask,
-};
 
 static unsigned long alloc_iommu(int size) 
 {       
@@ -185,114 +159,7 @@ static void flush_gart(struct device *dev)
         spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
 } 
 
-/* Allocate DMA memory on node near device */
-noinline
-static void *dma_alloc_pages(struct device *dev, gfp_t gfp, unsigned order)
-{
-        struct page *page;
-        int node;
-        if (dev->bus == &pci_bus_type)
-                node = pcibus_to_node(to_pci_dev(dev)->bus);
-        else
-                node = numa_node_id();
-        page = alloc_pages_node(node, gfp, order);
-        return page ? page_address(page) : NULL;
-}
-
-/* 
- * Allocate memory for a coherent mapping.
- */
-void *
-dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
-                   gfp_t gfp)
-{
-        void *memory;
-        unsigned long dma_mask = 0;
-        u64 bus;
 
-        if (!dev)
-                dev = &fallback_dev;
-        dma_mask = dev->coherent_dma_mask;
-        if (dma_mask == 0) 
-                dma_mask = 0xffffffff; 
-
-        /* Kludge to make it bug-to-bug compatible with i386. i386
-           uses the normal dma_mask for alloc_coherent. */
-        dma_mask &= *dev->dma_mask;
-
-        /* Why <=? Even when the mask is smaller than 4GB it is often larger 
-           than 16MB and in this case we have a chance of finding fitting memory 
-           in the next higher zone first. If not retry with true GFP_DMA. -AK */
-        if (dma_mask <= 0xffffffff)
-                gfp |= GFP_DMA32;
-
- again:
-        memory = dma_alloc_pages(dev, gfp, get_order(size));
-        if (memory == NULL)
-                return NULL;
-
-        {
-                int high, mmu;
-                bus = virt_to_bus(memory);
-                high = (bus + size) >= dma_mask;
-                mmu = high;
-                if (force_iommu && !(gfp & GFP_DMA)) 
-                        mmu = 1;
-                if (no_iommu || dma_mask < 0xffffffffUL) { 
-                        if (high) {
-                                free_pages((unsigned long)memory,
-                                           get_order(size));
-
-                                if (swiotlb) {
-                                        return
-                                        swiotlb_alloc_coherent(dev, size,
-                                                               dma_handle,
-                                                               gfp);
-                                }
-
-                                if (!(gfp & GFP_DMA)) { 
-                                        gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
-                                        goto again;
-                                }
-                                return NULL;
-                        }
-                        mmu = 0; 
-                }       
-                memset(memory, 0, size); 
-                if (!mmu) { 
-                        *dma_handle = virt_to_bus(memory);
-                        return memory;
-                }
-        } 
-
-        *dma_handle = dma_map_area(dev, bus, size, PCI_DMA_BIDIRECTIONAL, 0);
-        if (*dma_handle == bad_dma_address)
-                goto error; 
-        flush_gart(dev);
-        return memory; 
-        
-error:
-        if (panic_on_overflow)
-                panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n", size);
-        free_pages((unsigned long)memory, get_order(size)); 
-        return NULL; 
-}
-
-/* 
- * Unmap coherent memory.
- * The caller must ensure that the device has finished accessing the mapping.
- */
-void dma_free_coherent(struct device *dev, size_t size,
-                         void *vaddr, dma_addr_t bus)
-{
-        if (swiotlb) {
-                swiotlb_free_coherent(dev, size, vaddr, bus);
-                return;
-        }
-
-        dma_unmap_single(dev, bus, size, 0);
-        free_pages((unsigned long)vaddr, get_order(size));              
-}
 
 #ifdef CONFIG_IOMMU_LEAK
 
@@ -326,7 +193,7 @@ void dump_leak(void)
 #define CLEAR_LEAK(x)
 #endif
 
-static void iommu_full(struct device *dev, size_t size, int dir, int do_panic)
+static void iommu_full(struct device *dev, size_t size, int dir)
 {
         /* 
          * Ran out of IOMMU space for this operation. This is very bad.
@@ -342,11 +209,11 @@ static void iommu_full(struct device *dev, size_t size, int dir, int do_panic)
   "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
                size, dev->bus_id);
 
-        if (size > PAGE_SIZE*EMERGENCY_PAGES && do_panic) {
+        if (size > PAGE_SIZE*EMERGENCY_PAGES) {
                 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
                         panic("PCI-DMA: Memory would be corrupted\n");
                 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL) 
-                        panic("PCI-DMA: Random memory would be DMAed\n");
+                        panic(KERN_ERR "PCI-DMA: Random memory would be DMAed\n");
         } 
 
 #ifdef CONFIG_IOMMU_LEAK
@@ -385,8 +252,8 @@ static inline int nonforced_iommu(struct device *dev, unsigned long addr, size_t
 /* Map a single continuous physical area into the IOMMU.
  * Caller needs to check if the iommu is needed and flush.
  */
-static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
-                                size_t size, int dir, int do_panic)
+static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
+                                size_t size, int dir)
 { 
         unsigned long npages = to_pages(phys_mem, size);
         unsigned long iommu_page = alloc_iommu(npages);
@@ -396,7 +263,7 @@ static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
                         return phys_mem; 
                 if (panic_on_overflow)
                         panic("dma_map_area overflow %lu bytes\n", size);
-                iommu_full(dev, size, dir, do_panic);
+                iommu_full(dev, size, dir);
                 return bad_dma_address;
         }
 
@@ -408,15 +275,21 @@ static dma_addr_t dma_map_area(struct device *dev, unsigned long phys_mem,
         return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
 }
 
+static dma_addr_t gart_map_simple(struct device *dev, char *buf,
+                                 size_t size, int dir)
+{
+        dma_addr_t map = dma_map_area(dev, virt_to_bus(buf), size, dir);
+        flush_gart(dev);
+        return map;
+}
+
 /* Map a single area into the IOMMU */
-dma_addr_t dma_map_single(struct device *dev, void *addr, size_t size, int dir)
+dma_addr_t gart_map_single(struct device *dev, void *addr, size_t size, int dir)
 {
         unsigned long phys_mem, bus;
 
         BUG_ON(dir == DMA_NONE);
 
-        if (swiotlb)
-                return swiotlb_map_single(dev,addr,size,dir);
         if (!dev)
                 dev = &fallback_dev;
 
@@ -424,10 +297,24 @@ dma_addr_t dma_map_single(struct device *dev, void *addr, size_t size, int dir)
         if (!need_iommu(dev, phys_mem, size))
                 return phys_mem; 
 
-        bus = dma_map_area(dev, phys_mem, size, dir, 1);
-        flush_gart(dev); 
+        bus = gart_map_simple(dev, addr, size, dir);
         return bus; 
-} 
+}
+
+/*
+ * Wrapper for pci_unmap_single working with scatterlists.
+ */
+void gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
+{
+        int i;
+
+        for (i = 0; i < nents; i++) {
+                struct scatterlist *s = &sg[i];
+                if (!s->dma_length || !s->length)
+                        break;
+                dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
+        }
+}
 
 /* Fallback for dma_map_sg in case of overflow */
 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
@@ -443,10 +330,10 @@ static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
                 struct scatterlist *s = &sg[i];
                 unsigned long addr = page_to_phys(s->page) + s->offset; 
                 if (nonforced_iommu(dev, addr, s->length)) { 
-                        addr = dma_map_area(dev, addr, s->length, dir, 0);
+                        addr = dma_map_area(dev, addr, s->length, dir);
                         if (addr == bad_dma_address) { 
                                 if (i > 0) 
-                                        dma_unmap_sg(dev, sg, i, dir);
+                                        gart_unmap_sg(dev, sg, i, dir);
                                 nents = 0; 
                                 sg[0].dma_length = 0;
                                 break;
@@ -515,7 +402,7 @@ static inline int dma_map_cont(struct scatterlist *sg, int start, int stopat,
  * DMA map all entries in a scatterlist.
  * Merge chunks that have page aligned sizes into a continuous mapping. 
  */
-int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
+int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
 {
         int i;
         int out;
@@ -527,8 +414,6 @@ int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
         if (nents == 0) 
                 return 0;
 
-        if (swiotlb)
-                return swiotlb_map_sg(dev,sg,nents,dir);
         if (!dev)
                 dev = &fallback_dev;
 
@@ -571,13 +456,13 @@ int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
 
 error:
         flush_gart(NULL);
-        dma_unmap_sg(dev, sg, nents, dir);
+        gart_unmap_sg(dev, sg, nents, dir);
         /* When it was forced try again unforced */
         if (force_iommu) 
                 return dma_map_sg_nonforce(dev, sg, nents, dir);
         if (panic_on_overflow)
                 panic("dma_map_sg: overflow on %lu pages\n", pages);
-        iommu_full(dev, pages << PAGE_SHIFT, dir, 0);
+        iommu_full(dev, pages << PAGE_SHIFT, dir);
         for (i = 0; i < nents; i++)
                 sg[i].dma_address = bad_dma_address;
         return 0;
@@ -586,18 +471,13 @@ error:
 /*
  * Free a DMA mapping.
  */ 
-void dma_unmap_single(struct device *dev, dma_addr_t dma_addr,
+void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
                       size_t size, int direction)
 {
         unsigned long iommu_page; 
         int npages;
         int i;
 
-        if (swiotlb) {
-                swiotlb_unmap_single(dev,dma_addr,size,direction);
-                return;
-        }
-
         if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE || 
             dma_addr >= iommu_bus_base + iommu_size)
                 return;
@@ -610,68 +490,7 @@ void dma_unmap_single(struct device *dev, dma_addr_t dma_addr,
         free_iommu(iommu_page, npages);
 }
 
-/* 
- * Wrapper for pci_unmap_single working with scatterlists.
- */ 
-void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
-{
-        int i;
-        if (swiotlb) {
-                swiotlb_unmap_sg(dev,sg,nents,dir);
-                return;
-        }
-        for (i = 0; i < nents; i++) { 
-                struct scatterlist *s = &sg[i];
-                if (!s->dma_length || !s->length) 
-                        break;
-                dma_unmap_single(dev, s->dma_address, s->dma_length, dir);
-        }
-}
-
-int dma_supported(struct device *dev, u64 mask)
-{
-        /* Copied from i386. Doesn't make much sense, because it will 
-           only work for pci_alloc_coherent.
-           The caller just has to use GFP_DMA in this case. */
-        if (mask < 0x00ffffff)
-                return 0;
-
-        /* Tell the device to use SAC when IOMMU force is on. 
-           This allows the driver to use cheaper accesses in some cases.
-
-           Problem with this is that if we overflow the IOMMU area
-           and return DAC as fallback address the device may not handle it correctly.
-           
-           As a special case some controllers have a 39bit address mode 
-           that is as efficient as 32bit (aic79xx). Don't force SAC for these.
-           Assume all masks <= 40 bits are of this type. Normally this doesn't
-           make any difference, but gives more gentle handling of IOMMU overflow. */
-        if (iommu_sac_force && (mask >= 0xffffffffffULL)) { 
-                printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
-                return 0; 
-        }
-
-        return 1;
-} 
-
-int dma_get_cache_alignment(void)
-{
-        return boot_cpu_data.x86_clflush_size;
-}
-
-EXPORT_SYMBOL(dma_unmap_sg);
-EXPORT_SYMBOL(dma_map_sg);
-EXPORT_SYMBOL(dma_map_single);
-EXPORT_SYMBOL(dma_unmap_single);
-EXPORT_SYMBOL(dma_supported);
-EXPORT_SYMBOL(no_iommu);
-EXPORT_SYMBOL(force_iommu); 
-EXPORT_SYMBOL(bad_dma_address);
-EXPORT_SYMBOL(iommu_bio_merge);
-EXPORT_SYMBOL(iommu_sac_force);
-EXPORT_SYMBOL(dma_get_cache_alignment);
-EXPORT_SYMBOL(dma_alloc_coherent);
-EXPORT_SYMBOL(dma_free_coherent);
+static int no_agp;
 
 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
 { 
@@ -778,6 +597,21 @@ static __init int init_k8_gatt(struct agp_kern_info *info)
 
 extern int agp_amd64_init(void);
 
+static struct dma_mapping_ops gart_dma_ops = {
+        .mapping_error = NULL,
+        .map_single = gart_map_single,
+        .map_simple = gart_map_simple,
+        .unmap_single = gart_unmap_single,
+        .sync_single_for_cpu = NULL,
+        .sync_single_for_device = NULL,
+        .sync_single_range_for_cpu = NULL,
+        .sync_single_range_for_device = NULL,
+        .sync_sg_for_cpu = NULL,
+        .sync_sg_for_device = NULL,
+        .map_sg = gart_map_sg,
+        .unmap_sg = gart_unmap_sg,
+};
+
 static int __init pci_iommu_init(void)
 { 
         struct agp_kern_info info;
@@ -799,16 +633,15 @@ static int __init pci_iommu_init(void)
 
         if (swiotlb) { 
                 no_iommu = 1;
-                printk(KERN_INFO "PCI-DMA: Using software bounce buffering for IO (SWIOTLB)\n");
                 return -1; 
         } 
         
         if (no_iommu ||
-            (!force_iommu && (end_pfn-1) < 0xffffffff>>PAGE_SHIFT) ||
+            (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
             !iommu_aperture ||
             (no_agp && init_k8_gatt(&info) < 0)) {
-                printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n"); 
                 no_iommu = 1;
+                no_iommu_init();
                 return -1;
         }
 
@@ -885,100 +718,50 @@ static int __init pci_iommu_init(void)
                      
         flush_gart(NULL);
 
+        printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
+        dma_ops = &gart_dma_ops;
+
         return 0;
 } 
 
 /* Must execute after PCI subsystem */
 fs_initcall(pci_iommu_init);
 
-/* iommu=[size][,noagp][,off][,force][,noforce][,leak][,memaper[=order]][,merge]
-         [,forcesac][,fullflush][,nomerge][,biomerge]
-   size  set size of iommu (in bytes) 
-   noagp don't initialize the AGP driver and use full aperture.
-   off   don't use the IOMMU
-   leak  turn on simple iommu leak tracing (only when CONFIG_IOMMU_LEAK is on)
-   memaper[=order] allocate an own aperture over RAM with size 32MB^order.  
-   noforce don't force IOMMU usage. Default.
-   force  Force IOMMU.
-   merge  Do lazy merging. This may improve performance on some block devices.
-          Implies force (experimental)
-   biomerge Do merging at the BIO layer. This is more efficient than merge,
-            but should be only done with very big IOMMUs. Implies merge,force.
-   nomerge Don't do SG merging.
-   forcesac For SAC mode for masks <40bits  (experimental)
-   fullflush Flush IOMMU on each allocation (default) 
-   nofullflush Don't use IOMMU fullflush
-   allowed  overwrite iommu off workarounds for specific chipsets.
-   soft  Use software bounce buffering (default for Intel machines)
-   noaperture Don't touch the aperture for AGP.
-*/
-__init int iommu_setup(char *p)
-{ 
-    int arg;
-
-    while (*p) {
-            if (!strncmp(p,"noagp",5))
-                    no_agp = 1;
-            if (!strncmp(p,"off",3))
-                    no_iommu = 1;
-            if (!strncmp(p,"force",5)) {
-                    force_iommu = 1;
-                    iommu_aperture_allowed = 1;
-            }
-            if (!strncmp(p,"allowed",7))
-                    iommu_aperture_allowed = 1;
-            if (!strncmp(p,"noforce",7)) {
-                    iommu_merge = 0;
-                    force_iommu = 0;
-            }
-            if (!strncmp(p, "memaper", 7)) {
-                    fallback_aper_force = 1; 
-                    p += 7; 
-                    if (*p == '=') {
-                            ++p;
-                            if (get_option(&p, &arg))
-                                    fallback_aper_order = arg;
-                    }
-            } 
-            if (!strncmp(p, "biomerge",8)) {
-                    iommu_bio_merge = 4096;
-                    iommu_merge = 1;
-                    force_iommu = 1;
-            }
-            if (!strncmp(p, "panic",5))
-                    panic_on_overflow = 1;
-            if (!strncmp(p, "nopanic",7))
-                    panic_on_overflow = 0;          
-            if (!strncmp(p, "merge",5)) {
-                    iommu_merge = 1;
-                    force_iommu = 1; 
-            }
-            if (!strncmp(p, "nomerge",7))
-                    iommu_merge = 0;
-            if (!strncmp(p, "forcesac",8))
-                    iommu_sac_force = 1;
-            if (!strncmp(p, "fullflush",8))
-                    iommu_fullflush = 1;
-            if (!strncmp(p, "nofullflush",11))
-                    iommu_fullflush = 0;
-            if (!strncmp(p, "soft",4))
-                    swiotlb = 1;
-            if (!strncmp(p, "noaperture",10))
-                    fix_aperture = 0;
+void gart_parse_options(char *p)
+{
+        int arg;
+
 #ifdef CONFIG_IOMMU_LEAK
-            if (!strncmp(p,"leak",4)) {
-                    leak_trace = 1;
-                    p += 4; 
-                    if (*p == '=') ++p;
-                    if (isdigit(*p) && get_option(&p, &arg))
-                            iommu_leak_pages = arg;
-            } else
+        if (!strncmp(p,"leak",4)) {
+                leak_trace = 1;
+                p += 4;
+                if (*p == '=') ++p;
+                if (isdigit(*p) && get_option(&p, &arg))
+                        iommu_leak_pages = arg;
+        }
 #endif
-            if (isdigit(*p) && get_option(&p, &arg)) 
-                    iommu_size = arg;
-            p += strcspn(p, ",");
-            if (*p == ',')
-                    ++p;
-    }
-    return 1;
-} 
+        if (isdigit(*p) && get_option(&p, &arg))
+                iommu_size = arg;
+        if (!strncmp(p, "fullflush",8))
+                iommu_fullflush = 1;
+        if (!strncmp(p, "nofullflush",11))
+                iommu_fullflush = 0;
+        if (!strncmp(p,"noagp",5))
+                no_agp = 1;
+        if (!strncmp(p, "noaperture",10))
+                fix_aperture = 0;
+        /* duplicated from pci-dma.c */
+        if (!strncmp(p,"force",5))
+                iommu_aperture_allowed = 1;
+        if (!strncmp(p,"allowed",7))
+                iommu_aperture_allowed = 1;
+        if (!strncmp(p, "memaper", 7)) {
+                fallback_aper_force = 1;
+                p += 7;
+                if (*p == '=') {
+                        ++p;
+                        if (get_option(&p, &arg))
+                                fallback_aper_order = arg;
+                }
+        }
+}
diff --git a/arch/x86_64/kernel/pci-nommu.c b/arch/x86_64/kernel/pci-nommu.c
index 5a981dca87ff..3209adac3272 100644
--- a/arch/x86_64/kernel/pci-nommu.c
+++ b/arch/x86_64/kernel/pci-nommu.c
@@ -6,89 +6,93 @@
 #include <linux/string.h>
 #include <asm/proto.h>
 #include <asm/processor.h>
+#include <asm/dma.h>
 
-int iommu_merge = 0;
-EXPORT_SYMBOL(iommu_merge);
-
-dma_addr_t bad_dma_address;
-EXPORT_SYMBOL(bad_dma_address);
-
-int iommu_bio_merge = 0;
-EXPORT_SYMBOL(iommu_bio_merge);
-
-int iommu_sac_force = 0;
-EXPORT_SYMBOL(iommu_sac_force);
-
-/* 
- * Dummy IO MMU functions
- */
-
-void *dma_alloc_coherent(struct device *hwdev, size_t size,
-                         dma_addr_t *dma_handle, gfp_t gfp)
+static int
+check_addr(char *name, struct device *hwdev, dma_addr_t bus, size_t size)
 {
-        void *ret;
-        u64 mask;
-        int order = get_order(size);
-
-        if (hwdev)
-                mask = hwdev->coherent_dma_mask & *hwdev->dma_mask;
-        else
-                mask = 0xffffffff;
-        for (;;) {
-                ret = (void *)__get_free_pages(gfp, order);
-                if (ret == NULL)
-                        return NULL;
-                *dma_handle = virt_to_bus(ret);
-                if ((*dma_handle & ~mask) == 0)
-                        break;
-                free_pages((unsigned long)ret, order);
-                if (gfp & GFP_DMA)
-                        return NULL;
-                gfp |= GFP_DMA;
+        if (hwdev && bus + size > *hwdev->dma_mask) {
+                printk(KERN_ERR
+                    "nommu_%s: overflow %Lx+%lu of device mask %Lx\n",
+               name, (long long)bus, size, (long long)*hwdev->dma_mask);
+                return 0;
         }
+        return 1;
+}
 
-        memset(ret, 0, size);
-        return ret;
+static dma_addr_t
+nommu_map_single(struct device *hwdev, void *ptr, size_t size,
+               int direction)
+{
+        dma_addr_t bus = virt_to_bus(ptr);
+        if (!check_addr("map_single", hwdev, bus, size))
+                                return bad_dma_address;
+        return bus;
 }
-EXPORT_SYMBOL(dma_alloc_coherent);
 
-void dma_free_coherent(struct device *hwdev, size_t size,
-                         void *vaddr, dma_addr_t dma_handle)
+void nommu_unmap_single(struct device *dev, dma_addr_t addr,size_t size,
+                        int direction)
 {
-        free_pages((unsigned long)vaddr, get_order(size));
 }
-EXPORT_SYMBOL(dma_free_coherent);
 
-int dma_supported(struct device *hwdev, u64 mask)
+/* Map a set of buffers described by scatterlist in streaming
+ * mode for DMA.  This is the scatter-gather version of the
+ * above pci_map_single interface.  Here the scatter gather list
+ * elements are each tagged with the appropriate dma address
+ * and length.  They are obtained via sg_dma_{address,length}(SG).
+ *
+ * NOTE: An implementation may be able to use a smaller number of
+ *       DMA address/length pairs than there are SG table elements.
+ *       (for example via virtual mapping capabilities)
+ *       The routine returns the number of addr/length pairs actually
+ *       used, at most nents.
+ *
+ * Device ownership issues as mentioned above for pci_map_single are
+ * the same here.
+ */
+int nommu_map_sg(struct device *hwdev, struct scatterlist *sg,
+               int nents, int direction)
 {
-        /*
-         * we fall back to GFP_DMA when the mask isn't all 1s,
-         * so we can't guarantee allocations that must be
-         * within a tighter range than GFP_DMA..
-         * RED-PEN this won't work for pci_map_single. Caller has to
-         * use GFP_DMA in the first place.
-         */
-        if (mask < 0x00ffffff)
-                return 0;
+        int i;
 
-        return 1;
-} 
-EXPORT_SYMBOL(dma_supported);
+        BUG_ON(direction == DMA_NONE);
+        for (i = 0; i < nents; i++ ) {
+                struct scatterlist *s = &sg[i];
+                BUG_ON(!s->page);
+                s->dma_address = virt_to_bus(page_address(s->page) +s->offset);
+                if (!check_addr("map_sg", hwdev, s->dma_address, s->length))
+                        return 0;
+                s->dma_length = s->length;
+        }
+        return nents;
+}
 
-int dma_get_cache_alignment(void)
+/* Unmap a set of streaming mode DMA translations.
+ * Again, cpu read rules concerning calls here are the same as for
+ * pci_unmap_single() above.
+ */
+void nommu_unmap_sg(struct device *dev, struct scatterlist *sg,
+                  int nents, int dir)
 {
-        return boot_cpu_data.x86_clflush_size;
 }
-EXPORT_SYMBOL(dma_get_cache_alignment);
 
-static int __init check_ram(void) 
-{ 
-        if (end_pfn >= 0xffffffff>>PAGE_SHIFT) { 
-                printk(
-                KERN_ERR "WARNING more than 4GB of memory but IOMMU not compiled in.\n"
-                KERN_ERR "WARNING 32bit PCI may malfunction.\n");
-        } 
-        return 0;
-} 
-__initcall(check_ram);
+struct dma_mapping_ops nommu_dma_ops = {
+        .map_single = nommu_map_single,
+        .unmap_single = nommu_unmap_single,
+        .map_sg = nommu_map_sg,
+        .unmap_sg = nommu_unmap_sg,
+        .is_phys = 1,
+};
 
+void __init no_iommu_init(void)
+{
+        if (dma_ops)
+                return;
+        printk(KERN_INFO "PCI-DMA: Disabling IOMMU.\n");
+        dma_ops = &nommu_dma_ops;
+        if (end_pfn > MAX_DMA32_PFN) {
+                printk(KERN_ERR
+                       "WARNING more than 4GB of memory but IOMMU disabled.\n"
+                       KERN_ERR "WARNING 32bit PCI may malfunction.\n");
+        }
+}
diff --git a/arch/x86_64/kernel/pci-swiotlb.c b/arch/x86_64/kernel/pci-swiotlb.c
new file mode 100644
index 000000000000..3569a25ad7fb
--- /dev/null
+++ b/arch/x86_64/kernel/pci-swiotlb.c
@@ -0,0 +1,42 @@
+/* Glue code to lib/swiotlb.c */
+
+#include <linux/pci.h>
+#include <linux/cache.h>
+#include <linux/module.h>
+#include <asm/dma-mapping.h>
+#include <asm/proto.h>
+#include <asm/swiotlb.h>
+#include <asm/dma.h>
+
+int swiotlb __read_mostly;
+EXPORT_SYMBOL(swiotlb);
+
+struct dma_mapping_ops swiotlb_dma_ops = {
+        .mapping_error = swiotlb_dma_mapping_error,
+        .alloc_coherent = swiotlb_alloc_coherent,
+        .free_coherent = swiotlb_free_coherent,
+        .map_single = swiotlb_map_single,
+        .unmap_single = swiotlb_unmap_single,
+        .sync_single_for_cpu = swiotlb_sync_single_for_cpu,
+        .sync_single_for_device = swiotlb_sync_single_for_device,
+        .sync_single_range_for_cpu = swiotlb_sync_single_range_for_cpu,
+        .sync_single_range_for_device = swiotlb_sync_single_range_for_device,
+        .sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
+        .sync_sg_for_device = swiotlb_sync_sg_for_device,
+        .map_sg = swiotlb_map_sg,
+        .unmap_sg = swiotlb_unmap_sg,
+        .dma_supported = NULL,
+};
+
+void pci_swiotlb_init(void)
+{
+        /* don't initialize swiotlb if iommu=off (no_iommu=1) */
+        if (!iommu_aperture && !no_iommu &&
+            (end_pfn > MAX_DMA32_PFN || force_iommu))
+               swiotlb = 1;
+        if (swiotlb) {
+                swiotlb_init();
+                printk(KERN_INFO "PCI-DMA: Using software bounce buffering for IO (SWIOTLB)\n");
+                dma_ops = &swiotlb_dma_ops;
+        }
+}
diff --git a/arch/x86_64/kernel/setup.c b/arch/x86_64/kernel/setup.c
index 6810690d6045..e7a4d2cd3968 100644
--- a/arch/x86_64/kernel/setup.c
+++ b/arch/x86_64/kernel/setup.c
@@ -45,6 +45,7 @@
 #include <linux/kexec.h>
 #include <linux/cpufreq.h>
 #include <linux/dmi.h>
+#include <linux/dma-mapping.h>
 
 #include <asm/mtrr.h>
 #include <asm/uaccess.h>
@@ -63,7 +64,9 @@
 #include <asm/setup.h>
 #include <asm/mach_apic.h>
 #include <asm/numa.h>
+#include <asm/swiotlb.h>
 #include <asm/sections.h>
+#include <asm/gart-mapping.h>
 
 /*
  * Machine setup..
@@ -88,11 +91,6 @@ int bootloader_type;
 
 unsigned long saved_video_mode;
 
-#ifdef CONFIG_SWIOTLB
-int swiotlb;
-EXPORT_SYMBOL(swiotlb);
-#endif
-
 /*
  * Setup options
  */
@@ -389,11 +387,9 @@ static __init void parse_cmdline_early (char ** cmdline_p)
                         numa_setup(from+5); 
 #endif
 
-#ifdef CONFIG_GART_IOMMU 
                 if (!memcmp(from,"iommu=",6)) { 
                         iommu_setup(from+6); 
                 }
-#endif
 
                 if (!memcmp(from,"oops=panic", 10))
                         panic_on_oops = 1;