tile pci: enable IOMMU to support DMA for legacy devices

This change uses the TRIO IOMMU to map the PCI DMA space and physical
memory at different addresses.  We also now use the dma_mapping_ops
to provide support for non-PCI DMA, PCIe DMA (64-bit) and legacy PCI
DMA (32-bit).  We use the kernel's software I/O TLB framework
(i.e. bounce buffers) for the legacy 32-bit PCI device support since
there are a limited number of TLB entries in the IOMMU and it is
non-trivial to handle indexing, searching, matching, etc.  For 32-bit
devices the performance impact of bounce buffers should not be a concern.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

3 files changed, 372 insertions, 145 deletions

diff --git a/arch/tile/kernel/pci-dma.c b/arch/tile/kernel/pci-dma.c
index edd856a000c..b9fe80ec108 100644
--- a/arch/tile/kernel/pci-dma.c
+++ b/arch/tile/kernel/pci-dma.c
@@ -14,6 +14,7 @@
 
 #include <linux/mm.h>
 #include <linux/dma-mapping.h>
+#include <linux/swiotlb.h>
 #include <linux/vmalloc.h>
 #include <linux/export.h>
 #include <asm/tlbflush.h>
@@ -31,10 +32,9 @@
 #define PAGE_HOME_DMA PAGE_HOME_HASH
 #endif
 
-void *dma_alloc_coherent(struct device *dev,
-			 size_t size,
-			 dma_addr_t *dma_handle,
-			 gfp_t gfp)
+static void *tile_dma_alloc_coherent(struct device *dev, size_t size,
+				     dma_addr_t *dma_handle, gfp_t gfp,
+				     struct dma_attrs *attrs)
 {
 	u64 dma_mask = dev->coherent_dma_mask ?: DMA_BIT_MASK(32);
 	int node = dev_to_node(dev);
@@ -68,19 +68,19 @@ void *dma_alloc_coherent(struct device *dev,
 	}
 
 	*dma_handle = addr;
+
 	return page_address(pg);
 }
-EXPORT_SYMBOL(dma_alloc_coherent);
 
 /*
- * Free memory that was allocated with dma_alloc_coherent.
+ * Free memory that was allocated with tile_dma_alloc_coherent.
  */
-void dma_free_coherent(struct device *dev, size_t size,
-		  void *vaddr, dma_addr_t dma_handle)
+static void tile_dma_free_coherent(struct device *dev, size_t size,
+				   void *vaddr, dma_addr_t dma_handle,
+				   struct dma_attrs *attrs)
 {
 	homecache_free_pages((unsigned long)vaddr, get_order(size));
 }
-EXPORT_SYMBOL(dma_free_coherent);
 
 /*
  * The map routines "map" the specified address range for DMA
@@ -199,38 +199,182 @@ static void __dma_complete_pa_range(dma_addr_t dma_addr, size_t size,
 	}
 }
 
+static int tile_dma_map_sg(struct device *dev, struct scatterlist *sglist,
+			   int nents, enum dma_data_direction direction,
+			   struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
 
-/*
- * dma_map_single can be passed any memory address, and there appear
- * to be no alignment constraints.
- *
- * There is a chance that the start of the buffer will share a cache
- * line with some other data that has been touched in the meantime.
- */
-dma_addr_t dma_map_single(struct device *dev, void *ptr, size_t size,
-			  enum dma_data_direction direction)
+	BUG_ON(!valid_dma_direction(direction));
+
+	WARN_ON(nents == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_prep_pa_range(sg->dma_address, sg->length, direction);
+#ifdef CONFIG_NEED_SG_DMA_LENGTH
+		sg->dma_length = sg->length;
+#endif
+	}
+
+	return nents;
+}
+
+static void tile_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
+			      int nents, enum dma_data_direction direction,
+			      struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_complete_pa_range(sg->dma_address, sg->length,
+					direction);
+	}
+}
+
+static dma_addr_t tile_dma_map_page(struct device *dev, struct page *page,
+				    unsigned long offset, size_t size,
+				    enum dma_data_direction direction,
+				    struct dma_attrs *attrs)
 {
-	dma_addr_t dma_addr = __pa(ptr);
+	BUG_ON(!valid_dma_direction(direction));
+
+	BUG_ON(offset + size > PAGE_SIZE);
+	__dma_prep_page(page, offset, size, direction);
+
+	return page_to_pa(page) + offset;
+}
+
+static void tile_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
+				size_t size, enum dma_data_direction direction,
+				struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	__dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
+			    dma_address & PAGE_OFFSET, size, direction);
+}
 
+static void tile_dma_sync_single_for_cpu(struct device *dev,
+					 dma_addr_t dma_handle,
+					 size_t size,
+					 enum dma_data_direction direction)
+{
 	BUG_ON(!valid_dma_direction(direction));
-	WARN_ON(size == 0);
 
-	__dma_prep_pa_range(dma_addr, size, direction);
+	__dma_complete_pa_range(dma_handle, size, direction);
+}
 
-	return dma_addr;
+static void tile_dma_sync_single_for_device(struct device *dev,
+					    dma_addr_t dma_handle, size_t size,
+					    enum dma_data_direction direction)
+{
+	__dma_prep_pa_range(dma_handle, size, direction);
 }
-EXPORT_SYMBOL(dma_map_single);
 
-void dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size,
-		      enum dma_data_direction direction)
+static void tile_dma_sync_sg_for_cpu(struct device *dev,
+				     struct scatterlist *sglist, int nelems,
+				     enum dma_data_direction direction)
 {
+	struct scatterlist *sg;
+	int i;
+
 	BUG_ON(!valid_dma_direction(direction));
-	__dma_complete_pa_range(dma_addr, size, direction);
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_cpu(dev, sg->dma_address,
+					sg_dma_len(sg), direction);
+	}
 }
-EXPORT_SYMBOL(dma_unmap_single);
 
-int dma_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
-	       enum dma_data_direction direction)
+static void tile_dma_sync_sg_for_device(struct device *dev,
+					struct scatterlist *sglist, int nelems,
+					enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_device(dev, sg->dma_address,
+					   sg_dma_len(sg), direction);
+	}
+}
+
+static inline int
+tile_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
+{
+	return 0;
+}
+
+static inline int
+tile_dma_supported(struct device *dev, u64 mask)
+{
+	return 1;
+}
+
+static struct dma_map_ops tile_default_dma_map_ops = {
+	.alloc = tile_dma_alloc_coherent,
+	.free = tile_dma_free_coherent,
+	.map_page = tile_dma_map_page,
+	.unmap_page = tile_dma_unmap_page,
+	.map_sg = tile_dma_map_sg,
+	.unmap_sg = tile_dma_unmap_sg,
+	.sync_single_for_cpu = tile_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_dma_sync_sg_for_device,
+	.mapping_error = tile_dma_mapping_error,
+	.dma_supported = tile_dma_supported
+};
+
+struct dma_map_ops *tile_dma_map_ops = &tile_default_dma_map_ops;
+EXPORT_SYMBOL(tile_dma_map_ops);
+
+/* Generic PCI DMA mapping functions */
+
+static void *tile_pci_dma_alloc_coherent(struct device *dev, size_t size,
+					 dma_addr_t *dma_handle, gfp_t gfp,
+					 struct dma_attrs *attrs)
+{
+	int node = dev_to_node(dev);
+	int order = get_order(size);
+	struct page *pg;
+	dma_addr_t addr;
+
+	gfp |= __GFP_ZERO;
+
+	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
+	if (pg == NULL)
+		return NULL;
+
+	addr = page_to_phys(pg);
+
+	*dma_handle = phys_to_dma(dev, addr);
+
+	return page_address(pg);
+}
+
+/*
+ * Free memory that was allocated with tile_pci_dma_alloc_coherent.
+ */
+static void tile_pci_dma_free_coherent(struct device *dev, size_t size,
+				       void *vaddr, dma_addr_t dma_handle,
+				       struct dma_attrs *attrs)
+{
+	homecache_free_pages((unsigned long)vaddr, get_order(size));
+}
+
+static int tile_pci_dma_map_sg(struct device *dev, struct scatterlist *sglist,
+			       int nents, enum dma_data_direction direction,
+			       struct dma_attrs *attrs)
 {
 	struct scatterlist *sg;
 	int i;
@@ -242,14 +386,20 @@ int dma_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
 	for_each_sg(sglist, sg, nents, i) {
 		sg->dma_address = sg_phys(sg);
 		__dma_prep_pa_range(sg->dma_address, sg->length, direction);
+
+		sg->dma_address = phys_to_dma(dev, sg->dma_address);
+#ifdef CONFIG_NEED_SG_DMA_LENGTH
+		sg->dma_length = sg->length;
+#endif
 	}
 
 	return nents;
 }
-EXPORT_SYMBOL(dma_map_sg);
 
-void dma_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
-		  enum dma_data_direction direction)
+static void tile_pci_dma_unmap_sg(struct device *dev,
+				  struct scatterlist *sglist, int nents,
+				  enum dma_data_direction direction,
+				  struct dma_attrs *attrs)
 {
 	struct scatterlist *sg;
 	int i;
@@ -261,46 +411,60 @@ void dma_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
 					direction);
 	}
 }
-EXPORT_SYMBOL(dma_unmap_sg);
 
-dma_addr_t dma_map_page(struct device *dev, struct page *page,
-			unsigned long offset, size_t size,
-			enum dma_data_direction direction)
+static dma_addr_t tile_pci_dma_map_page(struct device *dev, struct page *page,
+					unsigned long offset, size_t size,
+					enum dma_data_direction direction,
+					struct dma_attrs *attrs)
 {
 	BUG_ON(!valid_dma_direction(direction));
 
 	BUG_ON(offset + size > PAGE_SIZE);
 	__dma_prep_page(page, offset, size, direction);
-	return page_to_pa(page) + offset;
+
+	return phys_to_dma(dev, page_to_pa(page) + offset);
 }
-EXPORT_SYMBOL(dma_map_page);
 
-void dma_unmap_page(struct device *dev, dma_addr_t dma_address, size_t size,
-		    enum dma_data_direction direction)
+static void tile_pci_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
+				    size_t size,
+				    enum dma_data_direction direction,
+				    struct dma_attrs *attrs)
 {
 	BUG_ON(!valid_dma_direction(direction));
+
+	dma_address = dma_to_phys(dev, dma_address);
+
 	__dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
 			    dma_address & PAGE_OFFSET, size, direction);
 }
-EXPORT_SYMBOL(dma_unmap_page);
 
-void dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
-			     size_t size, enum dma_data_direction direction)
+static void tile_pci_dma_sync_single_for_cpu(struct device *dev,
+					     dma_addr_t dma_handle,
+					     size_t size,
+					     enum dma_data_direction direction)
 {
 	BUG_ON(!valid_dma_direction(direction));
+
+	dma_handle = dma_to_phys(dev, dma_handle);
+
 	__dma_complete_pa_range(dma_handle, size, direction);
 }
-EXPORT_SYMBOL(dma_sync_single_for_cpu);
 
-void dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle,
-				size_t size, enum dma_data_direction direction)
+static void tile_pci_dma_sync_single_for_device(struct device *dev,
+						dma_addr_t dma_handle,
+						size_t size,
+						enum dma_data_direction
+						direction)
 {
+	dma_handle = dma_to_phys(dev, dma_handle);
+
 	__dma_prep_pa_range(dma_handle, size, direction);
 }
-EXPORT_SYMBOL(dma_sync_single_for_device);
 
-void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sglist,
-			 int nelems, enum dma_data_direction direction)
+static void tile_pci_dma_sync_sg_for_cpu(struct device *dev,
+					 struct scatterlist *sglist,
+					 int nelems,
+					 enum dma_data_direction direction)
 {
 	struct scatterlist *sg;
 	int i;
@@ -313,10 +477,11 @@ void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sglist,
 					sg_dma_len(sg), direction);
 	}
 }
-EXPORT_SYMBOL(dma_sync_sg_for_cpu);
 
-void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sglist,
-			    int nelems, enum dma_data_direction direction)
+static void tile_pci_dma_sync_sg_for_device(struct device *dev,
+					    struct scatterlist *sglist,
+					    int nelems,
+					    enum dma_data_direction direction)
 {
 	struct scatterlist *sg;
 	int i;
@@ -329,31 +494,93 @@ void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sglist,
 					   sg_dma_len(sg), direction);
 	}
 }
-EXPORT_SYMBOL(dma_sync_sg_for_device);
 
-void dma_sync_single_range_for_cpu(struct device *dev, dma_addr_t dma_handle,
-				   unsigned long offset, size_t size,
-				   enum dma_data_direction direction)
+static inline int
+tile_pci_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
 {
-	dma_sync_single_for_cpu(dev, dma_handle + offset, size, direction);
+	return 0;
 }
-EXPORT_SYMBOL(dma_sync_single_range_for_cpu);
 
-void dma_sync_single_range_for_device(struct device *dev,
-				      dma_addr_t dma_handle,
-				      unsigned long offset, size_t size,
-				      enum dma_data_direction direction)
+static inline int
+tile_pci_dma_supported(struct device *dev, u64 mask)
 {
-	dma_sync_single_for_device(dev, dma_handle + offset, size, direction);
+	return 1;
 }
-EXPORT_SYMBOL(dma_sync_single_range_for_device);
 
-/*
- * dma_alloc_noncoherent() is #defined to return coherent memory,
- * so there's no need to do any flushing here.
- */
-void dma_cache_sync(struct device *dev, void *vaddr, size_t size,
-		    enum dma_data_direction direction)
+static struct dma_map_ops tile_pci_default_dma_map_ops = {
+	.alloc = tile_pci_dma_alloc_coherent,
+	.free = tile_pci_dma_free_coherent,
+	.map_page = tile_pci_dma_map_page,
+	.unmap_page = tile_pci_dma_unmap_page,
+	.map_sg = tile_pci_dma_map_sg,
+	.unmap_sg = tile_pci_dma_unmap_sg,
+	.sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_pci_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
+	.mapping_error = tile_pci_dma_mapping_error,
+	.dma_supported = tile_pci_dma_supported
+};
+
+struct dma_map_ops *gx_pci_dma_map_ops = &tile_pci_default_dma_map_ops;
+EXPORT_SYMBOL(gx_pci_dma_map_ops);
+
+/* PCI DMA mapping functions for legacy PCI devices */
+
+#ifdef CONFIG_SWIOTLB
+static void *tile_swiotlb_alloc_coherent(struct device *dev, size_t size,
+					 dma_addr_t *dma_handle, gfp_t gfp,
+					 struct dma_attrs *attrs)
+{
+	gfp |= GFP_DMA;
+	return swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
+}
+
+static void tile_swiotlb_free_coherent(struct device *dev, size_t size,
+				       void *vaddr, dma_addr_t dma_addr,
+				       struct dma_attrs *attrs)
 {
+	swiotlb_free_coherent(dev, size, vaddr, dma_addr);
 }
-EXPORT_SYMBOL(dma_cache_sync);
+
+static struct dma_map_ops pci_swiotlb_dma_ops = {
+	.alloc = tile_swiotlb_alloc_coherent,
+	.free = tile_swiotlb_free_coherent,
+	.map_page = swiotlb_map_page,
+	.unmap_page = swiotlb_unmap_page,
+	.map_sg = swiotlb_map_sg_attrs,
+	.unmap_sg = swiotlb_unmap_sg_attrs,
+	.sync_single_for_cpu = swiotlb_sync_single_for_cpu,
+	.sync_single_for_device = swiotlb_sync_single_for_device,
+	.sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
+	.sync_sg_for_device = swiotlb_sync_sg_for_device,
+	.dma_supported = swiotlb_dma_supported,
+	.mapping_error = swiotlb_dma_mapping_error,
+};
+
+struct dma_map_ops *gx_legacy_pci_dma_map_ops = &pci_swiotlb_dma_ops;
+#else
+struct dma_map_ops *gx_legacy_pci_dma_map_ops;
+#endif
+EXPORT_SYMBOL(gx_legacy_pci_dma_map_ops);
+
+#ifdef CONFIG_ARCH_HAS_DMA_SET_COHERENT_MASK
+int dma_set_coherent_mask(struct device *dev, u64 mask)
+{
+	struct dma_map_ops *dma_ops = get_dma_ops(dev);
+
+	/* Handle legacy PCI devices with limited memory addressability. */
+	if (((dma_ops == gx_pci_dma_map_ops) ||
+	    (dma_ops == gx_legacy_pci_dma_map_ops)) &&
+	    (mask <= DMA_BIT_MASK(32))) {
+		if (mask > dev->archdata.max_direct_dma_addr)
+			mask = dev->archdata.max_direct_dma_addr;
+	}
+
+	if (!dma_supported(dev, mask))
+		return -EIO;
+	dev->coherent_dma_mask = mask;
+	return 0;
+}
+EXPORT_SYMBOL(dma_set_coherent_mask);
+#endif
diff --git a/arch/tile/kernel/pci_gx.c b/arch/tile/kernel/pci_gx.c
index 1b996bb628f..27f7ab02113 100644
--- a/arch/tile/kernel/pci_gx.c
+++ b/arch/tile/kernel/pci_gx.c
@@ -40,22 +40,8 @@
 #include <arch/sim.h>
 
 /*
- * Initialization flow and process
- * -------------------------------
- *
- * This files containes the routines to search for PCI buses,
+ * This file containes the routines to search for PCI buses,
  * enumerate the buses, and configure any attached devices.
- *
- * There are two entry points here:
- * 1) tile_pci_init
- *    This sets up the pci_controller structs, and opens the
- *    FDs to the hypervisor.  This is called from setup_arch() early
- *    in the boot process.
- * 2) pcibios_init
- *    This probes the PCI bus(es) for any attached hardware.  It's
- *    called by subsys_initcall.  All of the real work is done by the
- *    generic Linux PCI layer.
- *
  */
 
 #define DEBUG_PCI_CFG	0
@@ -110,6 +96,21 @@ static struct pci_ops tile_cfg_ops;
 /* Mask of CPUs that should receive PCIe interrupts. */
 static struct cpumask intr_cpus_map;
 
+/* PCI I/O space support is not implemented. */
+static struct resource pci_ioport_resource = {
+	.name	= "PCI IO",
+	.start	= 0,
+	.end	= 0,
+	.flags	= IORESOURCE_IO,
+};
+
+static struct resource pci_iomem_resource = {
+	.name	= "PCI mem",
+	.start	= TILE_PCI_MEM_START,
+	.end	= TILE_PCI_MEM_END,
+	.flags	= IORESOURCE_MEM,
+};
+
 /*
  * We don't need to worry about the alignment of resources.
  */
@@ -334,8 +335,6 @@ free_irqs:
 }
 
 /*
- * First initialization entry point, called from setup_arch().
- *
  * Find valid controllers and fill in pci_controller structs for each
  * of them.
  *
@@ -583,10 +582,7 @@ static int __devinit setup_pcie_rc_delay(char *str)
 early_param("pcie_rc_delay", setup_pcie_rc_delay);
 
 /*
- * Second PCI initialization entry point, called by subsys_initcall.
- *
- * The controllers have been set up by the time we get here, by a call to
- * tile_pci_init.
+ * PCI initialization entry point, called by subsys_initcall.
  */
 int __init pcibios_init(void)
 {
@@ -594,15 +590,13 @@ int __init pcibios_init(void)
 	LIST_HEAD(resources);
 	int i;
 
+	tile_pci_init();
+
 	if (num_rc_controllers == 0 && num_ep_controllers == 0)
 		return 0;
 
-	pr_info("PCI: Probing PCI hardware\n");
-
 	/*
 	 * We loop over all the TRIO shims and set up the MMIO mappings.
-	 * This step can't be done in tile_pci_init because the MM subsystem
-	 * hasn't been initialized then.
 	 */
 	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
 		gxio_trio_context_t *context = &trio_contexts[i];
@@ -645,9 +639,7 @@ int __init pcibios_init(void)
 		unsigned int class_code_revision;
 		int trio_index;
 		int mac;
-#ifndef USE_SHARED_PCIE_CONFIG_REGION
 		int ret;
-#endif
 
 		if (trio_context->fd < 0)
 			continue;
@@ -802,8 +794,6 @@ int __init pcibios_init(void)
 			pr_err("PCI: PCI CFG PIO alloc failure for mac %d "
 				"on TRIO %d, give up\n", mac, trio_index);
 
-			/* TBD: cleanup ... */
-
 			continue;
 		}
 
@@ -819,8 +809,6 @@ int __init pcibios_init(void)
 			pr_err("PCI: PCI CFG PIO init failure for mac %d "
 				"on TRIO %d, give up\n", mac, trio_index);
 
-			/* TBD: cleanup ... */
-
 			continue;
 		}
 
@@ -837,8 +825,6 @@ int __init pcibios_init(void)
 			pr_err("PCI: PIO map failure for mac %d on TRIO %d\n",
 				mac, trio_index);
 
-			/* TBD: cleanup ... */
-
 			continue;
 		}
 
@@ -852,7 +838,14 @@ int __init pcibios_init(void)
 			continue;
 		}
 
-		pci_add_resource(&resources, &iomem_resource);
+		/*
+		 * The PCI memory resource is located above the PA space.
+		 * The memory range for the PCI root bus should not overlap
+		 * with the physical RAM
+		 */
+		pci_add_resource_offset(&resources, &iomem_resource,
+					1ULL << CHIP_PA_WIDTH());
+
 		bus = pci_scan_root_bus(NULL, 0, controller->ops,
 					controller, &resources);
 		controller->root_bus = bus;
@@ -923,11 +916,6 @@ int __init pcibios_init(void)
 		}
 
 		/*
-		 * We always assign 32-bit PCI bus BAR ranges.
-		 */
-		BUG_ON(bus_address_hi != 0);
-
-		/*
 		 * Alloc a PIO region for PCI memory access for each RC port.
 		 */
 		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
@@ -936,8 +924,6 @@ int __init pcibios_init(void)
 				"give up\n", controller->trio_index,
 				controller->mac);
 
-			/* TBD: cleanup ... */
-
 			continue;
 		}
 
@@ -950,15 +936,13 @@ int __init pcibios_init(void)
 		ret = gxio_trio_init_pio_region_aux(trio_context,
 						    controller->pio_mem_index,
 						    controller->mac,
-						    bus_address_hi,
+						    0,
 						    0);
 		if (ret < 0) {
 			pr_err("PCI: MEM PIO init failure on TRIO %d mac %d, "
 				"give up\n", controller->trio_index,
 				controller->mac);
 
-			/* TBD: cleanup ... */
-
 			continue;
 		}
 
@@ -980,8 +964,6 @@ int __init pcibios_init(void)
 					controller->trio_index,
 					controller->mac, j);
 
-				/* TBD: cleanup ... */
-
 				goto alloc_mem_map_failed;
 			}
 
@@ -991,9 +973,13 @@ int __init pcibios_init(void)
 			 * Initialize the Mem-Map and the I/O MMU so that all
 			 * the physical memory can be accessed by the endpoint
 			 * devices. The base bus address is set to the base CPA
-			 * of this memory controller, so is the base VA. The
+			 * of this memory controller plus an offset (see pci.h).
+			 * The region's base VA is set to the base CPA. The
 			 * I/O MMU table essentially translates the CPA to
-			 * the real PA.
+			 * the real PA. Implicitly, for node 0, we create
+			 * a separate Mem-Map region that serves as the inbound
+			 * window for legacy 32-bit devices. This is a direct
+			 * map of the low 4GB CPA space.
 			 */
 			ret = gxio_trio_init_memory_map_mmu_aux(trio_context,
 				controller->mem_maps[j],
@@ -1001,7 +987,8 @@ int __init pcibios_init(void)
 				nr_pages << PAGE_SHIFT,
 				trio_context->asid,
 				controller->mac,
-				start_pfn << PAGE_SHIFT,
+				(start_pfn << PAGE_SHIFT) +
+				TILE_PCI_MEM_MAP_BASE_OFFSET,
 				j,
 				GXIO_TRIO_ORDER_MODE_UNORDERED);
 			if (ret < 0) {
@@ -1010,11 +997,8 @@ int __init pcibios_init(void)
 					controller->trio_index,
 					controller->mac, j);
 
-				/* TBD: cleanup ... */
-
 				goto alloc_mem_map_failed;
 			}
-
 			continue;
 
 alloc_mem_map_failed:
@@ -1028,11 +1012,19 @@ alloc_mem_map_failed:
 subsys_initcall(pcibios_init);
 
 /*
- * No bus fixups needed.
+ * PCI scan code calls the arch specific pcibios_fixup_bus() each time it scans
+ * a new bridge. Called after each bus is probed, but before its children are
+ * examined.
  */
 void __devinit pcibios_fixup_bus(struct pci_bus *bus)
 {
-	/* Nothing needs to be done. */
+	struct pci_dev *dev = bus->self;
+
+	if (!dev) {
+		/* This is the root bus. */
+		bus->resource[0] = &pci_ioport_resource;
+		bus->resource[1] = &pci_iomem_resource;
+	}
 }
 
 /*
@@ -1069,6 +1061,17 @@ int pcibios_enable_device(struct pci_dev *dev, int mask)
 	return pci_enable_resources(dev, mask);
 }
 
+/* Called for each device after PCI setup is done. */
+static void __init
+pcibios_fixup_final(struct pci_dev *pdev)
+{
+	set_dma_ops(&pdev->dev, gx_pci_dma_map_ops);
+	set_dma_offset(&pdev->dev, TILE_PCI_MEM_MAP_BASE_OFFSET);
+	pdev->dev.archdata.max_direct_dma_addr =
+		TILE_PCI_MAX_DIRECT_DMA_ADDRESS;
+}
+DECLARE_PCI_FIXUP_FINAL(PCI_ANY_ID, PCI_ANY_ID, pcibios_fixup_final);
+
 /* Map a PCI MMIO bus address into VA space. */
 void __iomem *ioremap(resource_size_t phys_addr, unsigned long size)
 {
@@ -1127,7 +1130,7 @@ got_it:
 	 * We need to keep the PCI bus address's in-page offset in the VA.
 	 */
 	return iorpc_ioremap(trio_fd, offset, size) +
-					(phys_addr & (PAGE_SIZE - 1));
+		(phys_addr & (PAGE_SIZE - 1));
 }
 EXPORT_SYMBOL(ioremap);
 
diff --git a/arch/tile/kernel/setup.c b/arch/tile/kernel/setup.c
index fdde3b6986e..2b8b689e596 100644
--- a/arch/tile/kernel/setup.c
+++ b/arch/tile/kernel/setup.c
@@ -23,6 +23,7 @@
 #include <linux/irq.h>
 #include <linux/kexec.h>
 #include <linux/pci.h>
+#include <linux/swiotlb.h>
 #include <linux/initrd.h>
 #include <linux/io.h>
 #include <linux/highmem.h>
@@ -109,7 +110,7 @@ static unsigned int __initdata maxnodemem_pfn[MAX_NUMNODES] = {
 };
 static nodemask_t __initdata isolnodes;
 
-#ifdef CONFIG_PCI
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 enum { DEFAULT_PCI_RESERVE_MB = 64 };
 static unsigned int __initdata pci_reserve_mb = DEFAULT_PCI_RESERVE_MB;
 unsigned long __initdata pci_reserve_start_pfn = -1U;
@@ -160,7 +161,7 @@ static int __init setup_isolnodes(char *str)
 }
 early_param("isolnodes", setup_isolnodes);
 
-#ifdef CONFIG_PCI
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 static int __init setup_pci_reserve(char* str)
 {
 	unsigned long mb;
@@ -171,7 +172,7 @@ static int __init setup_pci_reserve(char* str)
 
 	pci_reserve_mb = mb;
 	pr_info("Reserving %dMB for PCIE root complex mappings\n",
-	       pci_reserve_mb);
+		pci_reserve_mb);
 	return 0;
 }
 early_param("pci_reserve", setup_pci_reserve);
@@ -411,7 +412,7 @@ static void __init setup_memory(void)
 			continue;
 		}
 #endif
-#ifdef CONFIG_PCI
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 		/*
 		 * Blocks that overlap the pci reserved region must
 		 * have enough space to hold the maximum percpu data
@@ -604,11 +605,9 @@ static void __init setup_bootmem_allocator_node(int i)
 	/* Free all the space back into the allocator. */
 	free_bootmem(PFN_PHYS(start), PFN_PHYS(end - start));
 
-#if defined(CONFIG_PCI)
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 	/*
-	 * Throw away any memory aliased by the PCI region.  FIXME: this
-	 * is a temporary hack to work around bug 10502, and needs to be
-	 * fixed properly.
+	 * Throw away any memory aliased by the PCI region.
 	 */
 	if (pci_reserve_start_pfn < end && pci_reserve_end_pfn > start)
 		reserve_bootmem(PFN_PHYS(pci_reserve_start_pfn),
@@ -1353,8 +1352,7 @@ void __init setup_arch(char **cmdline_p)
 	setup_cpu_maps();
 
 
-#ifdef CONFIG_PCI
-#if !defined (__tilegx__)
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 	/*
 	 * Initialize the PCI structures.  This is done before memory
 	 * setup so that we know whether or not a pci_reserve region
@@ -1362,7 +1360,6 @@ void __init setup_arch(char **cmdline_p)
 	 */
 	if (tile_pci_init() == 0)
 		pci_reserve_mb = 0;
-#endif
 
 	/* PCI systems reserve a region just below 4GB for mapping iomem. */
 	pci_reserve_end_pfn  = (1 << (32 - PAGE_SHIFT));
@@ -1384,6 +1381,10 @@ void __init setup_arch(char **cmdline_p)
 	 * any memory using the bootmem allocator.
 	 */
 
+#ifdef CONFIG_SWIOTLB
+	swiotlb_init(0);
+#endif
+
 	paging_init();
 	setup_numa_mapping();
 	zone_sizes_init();
@@ -1391,10 +1392,6 @@ void __init setup_arch(char **cmdline_p)
 	setup_cpu(1);
 	setup_clock();
 	load_hv_initrd();
-
-#if defined(CONFIG_PCI) && defined (__tilegx__)
-	tile_pci_init();
-#endif
 }
 
 
@@ -1538,11 +1535,11 @@ static struct resource code_resource = {
 };
 
 /*
- * We reserve all resources above 4GB so that PCI won't try to put
+ * On Pro, we reserve all resources above 4GB so that PCI won't try to put
  * mappings above 4GB; the standard allows that for some devices but
  * the probing code trunates values to 32 bits.
  */
-#ifdef CONFIG_PCI
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 static struct resource* __init
 insert_non_bus_resource(void)
 {
@@ -1588,7 +1585,7 @@ static int __init request_standard_resources(void)
 	enum { CODE_DELTA = MEM_SV_INTRPT - PAGE_OFFSET };
 
 	iomem_resource.end = -1LL;
-#ifdef CONFIG_PCI
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 	insert_non_bus_resource();
 #endif
 
@@ -1596,7 +1593,7 @@ static int __init request_standard_resources(void)
 		u64 start_pfn = node_start_pfn[i];
 		u64 end_pfn = node_end_pfn[i];
 
-#ifdef CONFIG_PCI
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
 		if (start_pfn <= pci_reserve_start_pfn &&
 		    end_pfn > pci_reserve_start_pfn) {
 			if (end_pfn > pci_reserve_end_pfn)