diff --git a/arch/powerpc/sysdev/dart_iommu.c b/arch/powerpc/sysdev/dart_iommu.c
index 559db2b..17cf15e 100644
--- a/arch/powerpc/sysdev/dart_iommu.c
+++ b/arch/powerpc/sysdev/dart_iommu.c
@@ -70,6 +70,8 @@
 static int dart_dirty;
 static int dart_is_u4;
 
+#define DART_U4_BYPASS_BASE	0x8000000000ull
+
 #define DBG(...)
 
 static inline void dart_tlb_invalidate_all(void)
@@ -292,12 +294,20 @@
 	set_bit(iommu_table_dart.it_size - 1, iommu_table_dart.it_map);
 }
 
-static void pci_dma_dev_setup_dart(struct pci_dev *dev)
+static void dma_dev_setup_dart(struct device *dev)
 {
 	/* We only have one iommu table on the mac for now, which makes
 	 * things simple. Setup all PCI devices to point to this table
 	 */
-	set_iommu_table_base(&dev->dev, &iommu_table_dart);
+	if (get_dma_ops(dev) == &dma_direct_ops)
+		set_dma_offset(dev, DART_U4_BYPASS_BASE);
+	else
+		set_iommu_table_base(dev, &iommu_table_dart);
+}
+
+static void pci_dma_dev_setup_dart(struct pci_dev *dev)
+{
+	dma_dev_setup_dart(&dev->dev);
 }
 
 static void pci_dma_bus_setup_dart(struct pci_bus *bus)
@@ -315,6 +325,45 @@
 		PCI_DN(dn)->iommu_table = &iommu_table_dart;
 }
 
+static bool dart_device_on_pcie(struct device *dev)
+{
+	struct device_node *np = of_node_get(dev->of_node);
+
+	while(np) {
+		if (of_device_is_compatible(np, "U4-pcie") ||
+		    of_device_is_compatible(np, "u4-pcie")) {
+			of_node_put(np);
+			return true;
+		}
+		np = of_get_next_parent(np);
+	}
+	return false;
+}
+
+static int dart_dma_set_mask(struct device *dev, u64 dma_mask)
+{
+	if (!dev->dma_mask || !dma_supported(dev, dma_mask))
+		return -EIO;
+
+	/* U4 supports a DART bypass, we use it for 64-bit capable
+	 * devices to improve performances. However, that only works
+	 * for devices connected to U4 own PCIe interface, not bridged
+	 * through hypertransport. We need the device to support at
+	 * least 40 bits of addresses.
+	 */
+	if (dart_device_on_pcie(dev) && dma_mask >= DMA_BIT_MASK(40)) {
+		dev_info(dev, "Using 64-bit DMA iommu bypass\n");
+		set_dma_ops(dev, &dma_direct_ops);
+	} else {
+		dev_info(dev, "Using 32-bit DMA via iommu\n");
+		set_dma_ops(dev, &dma_iommu_ops);
+	}
+	dma_dev_setup_dart(dev);
+
+	*dev->dma_mask = dma_mask;
+	return 0;
+}
+
 void __init iommu_init_early_dart(void)
 {
 	struct device_node *dn;
@@ -328,20 +377,25 @@
 		dart_is_u4 = 1;
 	}
 
+	/* Initialize the DART HW */
+	if (dart_init(dn) != 0)
+		goto bail;
+
 	/* Setup low level TCE operations for the core IOMMU code */
 	ppc_md.tce_build = dart_build;
 	ppc_md.tce_free  = dart_free;
 	ppc_md.tce_flush = dart_flush;
 
-	/* Initialize the DART HW */
-	if (dart_init(dn) == 0) {
-		ppc_md.pci_dma_dev_setup = pci_dma_dev_setup_dart;
-		ppc_md.pci_dma_bus_setup = pci_dma_bus_setup_dart;
+	/* Setup bypass if supported */
+	if (dart_is_u4)
+		ppc_md.dma_set_mask = dart_dma_set_mask;
 
-		/* Setup pci_dma ops */
-		set_pci_dma_ops(&dma_iommu_ops);
-		return;
-	}
+	ppc_md.pci_dma_dev_setup = pci_dma_dev_setup_dart;
+	ppc_md.pci_dma_bus_setup = pci_dma_bus_setup_dart;
+
+	/* Setup pci_dma ops */
+	set_pci_dma_ops(&dma_iommu_ops);
+	return;
 
  bail:
 	/* If init failed, use direct iommu and null setup functions */