dma/ia64: update ia64 machvecs, swiotlb.c

Change all ia64 machvecs to use the new dma_*map*_attrs() interfaces.
Implement the old dma_*map_*() interfaces in terms of the corresponding new
interfaces.  For ia64/sn, make use of one dma attribute,
DMA_ATTR_WRITE_BARRIER.  Introduce swiotlb_*map*_attrs() functions.

Signed-off-by: Arthur Kepner <akepner@sgi.com>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Cc: Jes Sorensen <jes@sgi.com>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: David Miller <davem@davemloft.net>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Grant Grundler <grundler@parisc-linux.org>
Cc: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

3 files changed, 129 insertions, 77 deletions

diff --git a/arch/ia64/hp/common/hwsw_iommu.c b/arch/ia64/hp/common/hwsw_iommu.c
index 8f6bcfe1dada..1c44ec2a1d58 100644
--- a/arch/ia64/hp/common/hwsw_iommu.c
+++ b/arch/ia64/hp/common/hwsw_iommu.c
@@ -20,10 +20,10 @@
 extern int swiotlb_late_init_with_default_size (size_t size);
 extern ia64_mv_dma_alloc_coherent	swiotlb_alloc_coherent;
 extern ia64_mv_dma_free_coherent	swiotlb_free_coherent;
-extern ia64_mv_dma_map_single		swiotlb_map_single;
-extern ia64_mv_dma_unmap_single		swiotlb_unmap_single;
-extern ia64_mv_dma_map_sg		swiotlb_map_sg;
-extern ia64_mv_dma_unmap_sg		swiotlb_unmap_sg;
+extern ia64_mv_dma_map_single_attrs	swiotlb_map_single_attrs;
+extern ia64_mv_dma_unmap_single_attrs	swiotlb_unmap_single_attrs;
+extern ia64_mv_dma_map_sg_attrs		swiotlb_map_sg_attrs;
+extern ia64_mv_dma_unmap_sg_attrs	swiotlb_unmap_sg_attrs;
 extern ia64_mv_dma_supported		swiotlb_dma_supported;
 extern ia64_mv_dma_mapping_error	swiotlb_dma_mapping_error;
 
@@ -31,19 +31,19 @@ extern ia64_mv_dma_mapping_error	swiotlb_dma_mapping_error;
 
 extern ia64_mv_dma_alloc_coherent	sba_alloc_coherent;
 extern ia64_mv_dma_free_coherent	sba_free_coherent;
-extern ia64_mv_dma_map_single		sba_map_single;
-extern ia64_mv_dma_unmap_single		sba_unmap_single;
-extern ia64_mv_dma_map_sg		sba_map_sg;
-extern ia64_mv_dma_unmap_sg		sba_unmap_sg;
+extern ia64_mv_dma_map_single_attrs	sba_map_single_attrs;
+extern ia64_mv_dma_unmap_single_attrs	sba_unmap_single_attrs;
+extern ia64_mv_dma_map_sg_attrs		sba_map_sg_attrs;
+extern ia64_mv_dma_unmap_sg_attrs	sba_unmap_sg_attrs;
 extern ia64_mv_dma_supported		sba_dma_supported;
 extern ia64_mv_dma_mapping_error	sba_dma_mapping_error;
 
 #define hwiommu_alloc_coherent		sba_alloc_coherent
 #define hwiommu_free_coherent		sba_free_coherent
-#define hwiommu_map_single		sba_map_single
-#define hwiommu_unmap_single		sba_unmap_single
-#define hwiommu_map_sg			sba_map_sg
-#define hwiommu_unmap_sg		sba_unmap_sg
+#define hwiommu_map_single_attrs	sba_map_single_attrs
+#define hwiommu_unmap_single_attrs	sba_unmap_single_attrs
+#define hwiommu_map_sg_attrs		sba_map_sg_attrs
+#define hwiommu_unmap_sg_attrs		sba_unmap_sg_attrs
 #define hwiommu_dma_supported		sba_dma_supported
 #define hwiommu_dma_mapping_error	sba_dma_mapping_error
 #define hwiommu_sync_single_for_cpu	machvec_dma_sync_single
@@ -98,41 +98,48 @@ hwsw_free_coherent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma
 }
 
 dma_addr_t
-hwsw_map_single (struct device *dev, void *addr, size_t size, int dir)
+hwsw_map_single_attrs(struct device *dev, void *addr, size_t size, int dir,
+		       struct dma_attrs *attrs)
 {
 	if (use_swiotlb(dev))
-		return swiotlb_map_single(dev, addr, size, dir);
+		return swiotlb_map_single_attrs(dev, addr, size, dir, attrs);
 	else
-		return hwiommu_map_single(dev, addr, size, dir);
+		return hwiommu_map_single_attrs(dev, addr, size, dir, attrs);
 }
+EXPORT_SYMBOL(hwsw_map_single_attrs);
 
 void
-hwsw_unmap_single (struct device *dev, dma_addr_t iova, size_t size, int dir)
+hwsw_unmap_single_attrs(struct device *dev, dma_addr_t iova, size_t size,
+			 int dir, struct dma_attrs *attrs)
 {
 	if (use_swiotlb(dev))
-		return swiotlb_unmap_single(dev, iova, size, dir);
+		return swiotlb_unmap_single_attrs(dev, iova, size, dir, attrs);
 	else
-		return hwiommu_unmap_single(dev, iova, size, dir);
+		return hwiommu_unmap_single_attrs(dev, iova, size, dir, attrs);
 }
-
+EXPORT_SYMBOL(hwsw_unmap_single_attrs);
 
 int
-hwsw_map_sg (struct device *dev, struct scatterlist *sglist, int nents, int dir)
+hwsw_map_sg_attrs(struct device *dev, struct scatterlist *sglist, int nents,
+		   int dir, struct dma_attrs *attrs)
 {
 	if (use_swiotlb(dev))
-		return swiotlb_map_sg(dev, sglist, nents, dir);
+		return swiotlb_map_sg_attrs(dev, sglist, nents, dir, attrs);
 	else
-		return hwiommu_map_sg(dev, sglist, nents, dir);
+		return hwiommu_map_sg_attrs(dev, sglist, nents, dir, attrs);
 }
+EXPORT_SYMBOL(hwsw_map_sg_attrs);
 
 void
-hwsw_unmap_sg (struct device *dev, struct scatterlist *sglist, int nents, int dir)
+hwsw_unmap_sg_attrs(struct device *dev, struct scatterlist *sglist, int nents,
+		     int dir, struct dma_attrs *attrs)
 {
 	if (use_swiotlb(dev))
-		return swiotlb_unmap_sg(dev, sglist, nents, dir);
+		return swiotlb_unmap_sg_attrs(dev, sglist, nents, dir, attrs);
 	else
-		return hwiommu_unmap_sg(dev, sglist, nents, dir);
+		return hwiommu_unmap_sg_attrs(dev, sglist, nents, dir, attrs);
 }
+EXPORT_SYMBOL(hwsw_unmap_sg_attrs);
 
 void
 hwsw_sync_single_for_cpu (struct device *dev, dma_addr_t addr, size_t size, int dir)
@@ -185,10 +192,6 @@ hwsw_dma_mapping_error (dma_addr_t dma_addr)
 }
 
 EXPORT_SYMBOL(hwsw_dma_mapping_error);
-EXPORT_SYMBOL(hwsw_map_single);
-EXPORT_SYMBOL(hwsw_unmap_single);
-EXPORT_SYMBOL(hwsw_map_sg);
-EXPORT_SYMBOL(hwsw_unmap_sg);
 EXPORT_SYMBOL(hwsw_dma_supported);
 EXPORT_SYMBOL(hwsw_alloc_coherent);
 EXPORT_SYMBOL(hwsw_free_coherent);
diff --git a/arch/ia64/hp/common/sba_iommu.c b/arch/ia64/hp/common/sba_iommu.c
index 9409de5c9441..6ce729f46de8 100644
--- a/arch/ia64/hp/common/sba_iommu.c
+++ b/arch/ia64/hp/common/sba_iommu.c
@@ -899,16 +899,18 @@ sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
 }
 
 /**
- * sba_map_single - map one buffer and return IOVA for DMA
+ * sba_map_single_attrs - map one buffer and return IOVA for DMA
  * @dev: instance of PCI owned by the driver that's asking.
  * @addr:  driver buffer to map.
  * @size:  number of bytes to map in driver buffer.
  * @dir:  R/W or both.
+ * @attrs: optional dma attributes
  *
  * See Documentation/DMA-mapping.txt
  */
 dma_addr_t
-sba_map_single(struct device *dev, void *addr, size_t size, int dir)
+sba_map_single_attrs(struct device *dev, void *addr, size_t size, int dir,
+		     struct dma_attrs *attrs)
 {
 	struct ioc *ioc;
 	dma_addr_t iovp;
@@ -932,7 +934,8 @@ sba_map_single(struct device *dev, void *addr, size_t size, int dir)
  		** Device is bit capable of DMA'ing to the buffer...
 		** just return the PCI address of ptr
  		*/
-		DBG_BYPASS("sba_map_single() bypass mask/addr: 0x%lx/0x%lx\n",
+		DBG_BYPASS("sba_map_single_attrs() bypass mask/addr: "
+			   "0x%lx/0x%lx\n",
 		           to_pci_dev(dev)->dma_mask, pci_addr);
 		return pci_addr;
 	}
@@ -953,7 +956,7 @@ sba_map_single(struct device *dev, void *addr, size_t size, int dir)
 
 #ifdef ASSERT_PDIR_SANITY
 	spin_lock_irqsave(&ioc->res_lock, flags);
-	if (sba_check_pdir(ioc,"Check before sba_map_single()"))
+	if (sba_check_pdir(ioc,"Check before sba_map_single_attrs()"))
 		panic("Sanity check failed");
 	spin_unlock_irqrestore(&ioc->res_lock, flags);
 #endif
@@ -982,11 +985,12 @@ sba_map_single(struct device *dev, void *addr, size_t size, int dir)
 	/* form complete address */
 #ifdef ASSERT_PDIR_SANITY
 	spin_lock_irqsave(&ioc->res_lock, flags);
-	sba_check_pdir(ioc,"Check after sba_map_single()");
+	sba_check_pdir(ioc,"Check after sba_map_single_attrs()");
 	spin_unlock_irqrestore(&ioc->res_lock, flags);
 #endif
 	return SBA_IOVA(ioc, iovp, offset);
 }
+EXPORT_SYMBOL(sba_map_single_attrs);
 
 #ifdef ENABLE_MARK_CLEAN
 static SBA_INLINE void
@@ -1013,15 +1017,17 @@ sba_mark_clean(struct ioc *ioc, dma_addr_t iova, size_t size)
 #endif
 
 /**
- * sba_unmap_single - unmap one IOVA and free resources
+ * sba_unmap_single_attrs - unmap one IOVA and free resources
  * @dev: instance of PCI owned by the driver that's asking.
  * @iova:  IOVA of driver buffer previously mapped.
  * @size:  number of bytes mapped in driver buffer.
  * @dir:  R/W or both.
+ * @attrs: optional dma attributes
  *
  * See Documentation/DMA-mapping.txt
  */
-void sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, int dir)
+void sba_unmap_single_attrs(struct device *dev, dma_addr_t iova, size_t size,
+			    int dir, struct dma_attrs *attrs)
 {
 	struct ioc *ioc;
 #if DELAYED_RESOURCE_CNT > 0
@@ -1038,7 +1044,8 @@ void sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, int dir)
 		/*
 		** Address does not fall w/in IOVA, must be bypassing
 		*/
-		DBG_BYPASS("sba_unmap_single() bypass addr: 0x%lx\n", iova);
+		DBG_BYPASS("sba_unmap_single_atttrs() bypass addr: 0x%lx\n",
+			   iova);
 
 #ifdef ENABLE_MARK_CLEAN
 		if (dir == DMA_FROM_DEVICE) {
@@ -1087,7 +1094,7 @@ void sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, int dir)
 	spin_unlock_irqrestore(&ioc->res_lock, flags);
 #endif /* DELAYED_RESOURCE_CNT == 0 */
 }
-
+EXPORT_SYMBOL(sba_unmap_single_attrs);
 
 /**
  * sba_alloc_coherent - allocate/map shared mem for DMA
@@ -1144,7 +1151,8 @@ sba_alloc_coherent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp
 	 * If device can't bypass or bypass is disabled, pass the 32bit fake
 	 * device to map single to get an iova mapping.
 	 */
-	*dma_handle = sba_map_single(&ioc->sac_only_dev->dev, addr, size, 0);
+	*dma_handle = sba_map_single_attrs(&ioc->sac_only_dev->dev, addr,
+					   size, 0, NULL);
 
 	return addr;
 }
@@ -1161,7 +1169,7 @@ sba_alloc_coherent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp
  */
 void sba_free_coherent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle)
 {
-	sba_unmap_single(dev, dma_handle, size, 0);
+	sba_unmap_single_attrs(dev, dma_handle, size, 0, NULL);
 	free_pages((unsigned long) vaddr, get_order(size));
 }
 
@@ -1410,10 +1418,12 @@ sba_coalesce_chunks(struct ioc *ioc, struct device *dev,
  * @sglist:  array of buffer/length pairs
  * @nents:  number of entries in list
  * @dir:  R/W or both.
+ * @attrs: optional dma attributes
  *
  * See Documentation/DMA-mapping.txt
  */
-int sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, int dir)
+int sba_map_sg_attrs(struct device *dev, struct scatterlist *sglist, int nents,
+		     int dir, struct dma_attrs *attrs)
 {
 	struct ioc *ioc;
 	int coalesced, filled = 0;
@@ -1441,16 +1451,16 @@ int sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, int di
 	/* Fast path single entry scatterlists. */
 	if (nents == 1) {
 		sglist->dma_length = sglist->length;
-		sglist->dma_address = sba_map_single(dev, sba_sg_address(sglist), sglist->length, dir);
+		sglist->dma_address = sba_map_single_attrs(dev, sba_sg_address(sglist), sglist->length, dir, attrs);
 		return 1;
 	}
 
 #ifdef ASSERT_PDIR_SANITY
 	spin_lock_irqsave(&ioc->res_lock, flags);
-	if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
+	if (sba_check_pdir(ioc,"Check before sba_map_sg_attrs()"))
 	{
 		sba_dump_sg(ioc, sglist, nents);
-		panic("Check before sba_map_sg()");
+		panic("Check before sba_map_sg_attrs()");
 	}
 	spin_unlock_irqrestore(&ioc->res_lock, flags);
 #endif
@@ -1479,10 +1489,10 @@ int sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, int di
 
 #ifdef ASSERT_PDIR_SANITY
 	spin_lock_irqsave(&ioc->res_lock, flags);
-	if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
+	if (sba_check_pdir(ioc,"Check after sba_map_sg_attrs()"))
 	{
 		sba_dump_sg(ioc, sglist, nents);
-		panic("Check after sba_map_sg()\n");
+		panic("Check after sba_map_sg_attrs()\n");
 	}
 	spin_unlock_irqrestore(&ioc->res_lock, flags);
 #endif
@@ -1492,18 +1502,20 @@ int sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, int di
 
 	return filled;
 }
-
+EXPORT_SYMBOL(sba_map_sg_attrs);
 
 /**
- * sba_unmap_sg - unmap Scatter/Gather list
+ * sba_unmap_sg_attrs - unmap Scatter/Gather list
  * @dev: instance of PCI owned by the driver that's asking.
  * @sglist:  array of buffer/length pairs
  * @nents:  number of entries in list
  * @dir:  R/W or both.
+ * @attrs: optional dma attributes
  *
  * See Documentation/DMA-mapping.txt
  */
-void sba_unmap_sg (struct device *dev, struct scatterlist *sglist, int nents, int dir)
+void sba_unmap_sg_attrs(struct device *dev, struct scatterlist *sglist,
+			int nents, int dir, struct dma_attrs *attrs)
 {
 #ifdef ASSERT_PDIR_SANITY
 	struct ioc *ioc;
@@ -1518,13 +1530,14 @@ void sba_unmap_sg (struct device *dev, struct scatterlist *sglist, int nents, in
 	ASSERT(ioc);
 
 	spin_lock_irqsave(&ioc->res_lock, flags);
-	sba_check_pdir(ioc,"Check before sba_unmap_sg()");
+	sba_check_pdir(ioc,"Check before sba_unmap_sg_attrs()");
 	spin_unlock_irqrestore(&ioc->res_lock, flags);
 #endif
 
 	while (nents && sglist->dma_length) {
 
-		sba_unmap_single(dev, sglist->dma_address, sglist->dma_length, dir);
+		sba_unmap_single_attrs(dev, sglist->dma_address,
+				       sglist->dma_length, dir, attrs);
 		sglist = sg_next(sglist);
 		nents--;
 	}
@@ -1533,11 +1546,12 @@ void sba_unmap_sg (struct device *dev, struct scatterlist *sglist, int nents, in
 
 #ifdef ASSERT_PDIR_SANITY
 	spin_lock_irqsave(&ioc->res_lock, flags);
-	sba_check_pdir(ioc,"Check after sba_unmap_sg()");
+	sba_check_pdir(ioc,"Check after sba_unmap_sg_attrs()");
 	spin_unlock_irqrestore(&ioc->res_lock, flags);
 #endif
 
 }
+EXPORT_SYMBOL(sba_unmap_sg_attrs);
 
 /**************************************************************
 *
@@ -2166,10 +2180,6 @@ sba_page_override(char *str)
 __setup("sbapagesize=",sba_page_override);
 
 EXPORT_SYMBOL(sba_dma_mapping_error);
-EXPORT_SYMBOL(sba_map_single);
-EXPORT_SYMBOL(sba_unmap_single);
-EXPORT_SYMBOL(sba_map_sg);
-EXPORT_SYMBOL(sba_unmap_sg);
 EXPORT_SYMBOL(sba_dma_supported);
 EXPORT_SYMBOL(sba_alloc_coherent);
 EXPORT_SYMBOL(sba_free_coherent);
diff --git a/arch/ia64/sn/pci/pci_dma.c b/arch/ia64/sn/pci/pci_dma.c
index 18b94b792d54..52175af299a0 100644
--- a/arch/ia64/sn/pci/pci_dma.c
+++ b/arch/ia64/sn/pci/pci_dma.c
@@ -10,6 +10,7 @@
  */
 
 #include <linux/module.h>
+#include <linux/dma-attrs.h>
 #include <asm/dma.h>
 #include <asm/sn/intr.h>
 #include <asm/sn/pcibus_provider_defs.h>
@@ -149,11 +150,12 @@ void sn_dma_free_coherent(struct device *dev, size_t size, void *cpu_addr,
 EXPORT_SYMBOL(sn_dma_free_coherent);
 
 /**
- * sn_dma_map_single - map a single page for DMA
+ * sn_dma_map_single_attrs - map a single page for DMA
  * @dev: device to map for
  * @cpu_addr: kernel virtual address of the region to map
  * @size: size of the region
  * @direction: DMA direction
+ * @attrs: optional dma attributes
  *
  * Map the region pointed to by @cpu_addr for DMA and return the
  * DMA address.
@@ -163,42 +165,59 @@ EXPORT_SYMBOL(sn_dma_free_coherent);
  * no way of saving the dmamap handle from the alloc to later free
  * (which is pretty much unacceptable).
  *
+ * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
+ * dma_map_consistent() so that writes force a flush of pending DMA.
+ * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
+ * Document Number: 007-4763-001)
+ *
  * TODO: simplify our interface;
  *       figure out how to save dmamap handle so can use two step.
  */
-dma_addr_t sn_dma_map_single(struct device *dev, void *cpu_addr, size_t size,
-			     int direction)
+dma_addr_t sn_dma_map_single_attrs(struct device *dev, void *cpu_addr,
+				   size_t size, int direction,
+				   struct dma_attrs *attrs)
 {
 	dma_addr_t dma_addr;
 	unsigned long phys_addr;
 	struct pci_dev *pdev = to_pci_dev(dev);
 	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+	int dmabarr;
+
+	dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs);
 
 	BUG_ON(dev->bus != &pci_bus_type);
 
 	phys_addr = __pa(cpu_addr);
-	dma_addr = provider->dma_map(pdev, phys_addr, size, SN_DMA_ADDR_PHYS);
+	if (dmabarr)
+		dma_addr = provider->dma_map_consistent(pdev, phys_addr,
+							size, SN_DMA_ADDR_PHYS);
+	else
+		dma_addr = provider->dma_map(pdev, phys_addr, size,
+					     SN_DMA_ADDR_PHYS);
+
 	if (!dma_addr) {
 		printk(KERN_ERR "%s: out of ATEs\n", __func__);
 		return 0;
 	}
 	return dma_addr;
 }
-EXPORT_SYMBOL(sn_dma_map_single);
+EXPORT_SYMBOL(sn_dma_map_single_attrs);
 
 /**
- * sn_dma_unmap_single - unamp a DMA mapped page
+ * sn_dma_unmap_single_attrs - unamp a DMA mapped page
  * @dev: device to sync
  * @dma_addr: DMA address to sync
  * @size: size of region
  * @direction: DMA direction
+ * @attrs: optional dma attributes
  *
  * This routine is supposed to sync the DMA region specified
  * by @dma_handle into the coherence domain.  On SN, we're always cache
  * coherent, so we just need to free any ATEs associated with this mapping.
  */
-void sn_dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size,
-			 int direction)
+void sn_dma_unmap_single_attrs(struct device *dev, dma_addr_t dma_addr,
+			       size_t size, int direction,
+			       struct dma_attrs *attrs)
 {
 	struct pci_dev *pdev = to_pci_dev(dev);
 	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
@@ -207,19 +226,21 @@ void sn_dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size,
 
 	provider->dma_unmap(pdev, dma_addr, direction);
 }
-EXPORT_SYMBOL(sn_dma_unmap_single);
+EXPORT_SYMBOL(sn_dma_unmap_single_attrs);
 
 /**
- * sn_dma_unmap_sg - unmap a DMA scatterlist
+ * sn_dma_unmap_sg_attrs - unmap a DMA scatterlist
  * @dev: device to unmap
  * @sg: scatterlist to unmap
  * @nhwentries: number of scatterlist entries
  * @direction: DMA direction
+ * @attrs: optional dma attributes
  *
  * Unmap a set of streaming mode DMA translations.
  */
-void sn_dma_unmap_sg(struct device *dev, struct scatterlist *sgl,
-		     int nhwentries, int direction)
+void sn_dma_unmap_sg_attrs(struct device *dev, struct scatterlist *sgl,
+			   int nhwentries, int direction,
+			   struct dma_attrs *attrs)
 {
 	int i;
 	struct pci_dev *pdev = to_pci_dev(dev);
@@ -234,25 +255,34 @@ void sn_dma_unmap_sg(struct device *dev, struct scatterlist *sgl,
 		sg->dma_length = 0;
 	}
 }
-EXPORT_SYMBOL(sn_dma_unmap_sg);
+EXPORT_SYMBOL(sn_dma_unmap_sg_attrs);
 
 /**
- * sn_dma_map_sg - map a scatterlist for DMA
+ * sn_dma_map_sg_attrs - map a scatterlist for DMA
  * @dev: device to map for
  * @sg: scatterlist to map
  * @nhwentries: number of entries
  * @direction: direction of the DMA transaction
+ * @attrs: optional dma attributes
+ *
+ * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
+ * dma_map_consistent() so that writes force a flush of pending DMA.
+ * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
+ * Document Number: 007-4763-001)
  *
  * Maps each entry of @sg for DMA.
  */
-int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl, int nhwentries,
-		  int direction)
+int sn_dma_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
+			int nhwentries, int direction, struct dma_attrs *attrs)
 {
 	unsigned long phys_addr;
 	struct scatterlist *saved_sg = sgl, *sg;
 	struct pci_dev *pdev = to_pci_dev(dev);
 	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
 	int i;
+	int dmabarr;
+
+	dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs);
 
 	BUG_ON(dev->bus != &pci_bus_type);
 
@@ -260,11 +290,19 @@ int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl, int nhwentries,
 	 * Setup a DMA address for each entry in the scatterlist.
 	 */
 	for_each_sg(sgl, sg, nhwentries, i) {
+		dma_addr_t dma_addr;
 		phys_addr = SG_ENT_PHYS_ADDRESS(sg);
-		sg->dma_address = provider->dma_map(pdev,
-						    phys_addr, sg->length,
-						    SN_DMA_ADDR_PHYS);
+		if (dmabarr)
+			dma_addr = provider->dma_map_consistent(pdev,
+								phys_addr,
+								sg->length,
+								SN_DMA_ADDR_PHYS);
+		else
+			dma_addr = provider->dma_map(pdev, phys_addr,
+						     sg->length,
+						     SN_DMA_ADDR_PHYS);
 
+		sg->dma_address = dma_addr;
 		if (!sg->dma_address) {
 			printk(KERN_ERR "%s: out of ATEs\n", __func__);
 
@@ -272,7 +310,8 @@ int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl, int nhwentries,
 			 * Free any successfully allocated entries.
 			 */
 			if (i > 0)
-				sn_dma_unmap_sg(dev, saved_sg, i, direction);
+				sn_dma_unmap_sg_attrs(dev, saved_sg, i,
+						      direction, attrs);
 			return 0;
 		}
 
@@ -281,7 +320,7 @@ int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl, int nhwentries,
 
 	return nhwentries;
 }
-EXPORT_SYMBOL(sn_dma_map_sg);
+EXPORT_SYMBOL(sn_dma_map_sg_attrs);
 
 void sn_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
 				size_t size, int direction)