Merge git://git.kernel.org/pub/scm/linux/kernel/git/cmetcalf/linux-tile

Pull arch/tile updates from Chris Metcalf:
 "These changes provide support for PCIe root complex and USB host mode
  for tilegx's on-chip I/Os.

  In addition, this pull provides the required underpinning for the
  on-chip networking support that was pulled into 3.5.  The changes have
  all been through LKML (with several rounds for PCIe RC) and on
  linux-next."

* git://git.kernel.org/pub/scm/linux/kernel/git/cmetcalf/linux-tile:
  tile: updates to pci root complex from community feedback
  bounce: allow use of bounce pool via config option
  usb: add host support for the tilegx architecture
  arch/tile: provide kernel support for the tilegx USB shim
  tile pci: enable IOMMU to support DMA for legacy devices
  arch/tile: enable ZONE_DMA for tilegx
  tilegx pci: support I/O to arbitrarily-cached pages
  tile: remove unused header
  arch/tile: tilegx PCI root complex support
  arch/tile: provide kernel support for the tilegx TRIO shim
  arch/tile: break out the "csum a long" function to <asm/checksum.h>
  arch/tile: provide kernel support for the tilegx mPIPE shim
  arch/tile: common DMA code for the GXIO IORPC subsystem
  arch/tile: support MMIO-based readb/writeb etc.
  arch/tile: introduce GXIO IORPC framework for tilegx

5 files changed, 2080 insertions, 118 deletions

diff --git a/arch/tile/kernel/Makefile b/arch/tile/kernel/Makefile
index 5de99248d8df..5334be8e2538 100644
--- a/arch/tile/kernel/Makefile
+++ b/arch/tile/kernel/Makefile
@@ -14,4 +14,9 @@ obj-$(CONFIG_SMP)		+= smpboot.o smp.o tlb.o
 obj-$(CONFIG_MODULES)		+= module.o
 obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
 obj-$(CONFIG_KEXEC)		+= machine_kexec.o relocate_kernel_$(BITS).o
+ifdef CONFIG_TILEGX
+obj-$(CONFIG_PCI)		+= pci_gx.o
+else
 obj-$(CONFIG_PCI)		+= pci.o
+endif
+obj-$(CONFIG_TILE_USB)		+= usb.o
diff --git a/arch/tile/kernel/pci-dma.c b/arch/tile/kernel/pci-dma.c
index b3ed19f8779c..b9fe80ec1089 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>
@@ -22,13 +23,18 @@
 /* Generic DMA mapping functions: */
 
 /*
- * Allocate what Linux calls "coherent" memory, which for us just
- * means uncached.
+ * Allocate what Linux calls "coherent" memory.  On TILEPro this is
+ * uncached memory; on TILE-Gx it is hash-for-home memory.
  */
-void *dma_alloc_coherent(struct device *dev,
-			 size_t size,
-			 dma_addr_t *dma_handle,
-			 gfp_t gfp)
+#ifdef __tilepro__
+#define PAGE_HOME_DMA PAGE_HOME_UNCACHED
+#else
+#define PAGE_HOME_DMA PAGE_HOME_HASH
+#endif
+
+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);
@@ -39,39 +45,42 @@ void *dma_alloc_coherent(struct device *dev,
 	gfp |= __GFP_ZERO;
 
 	/*
-	 * By forcing NUMA node 0 for 32-bit masks we ensure that the
-	 * high 32 bits of the resulting PA will be zero.  If the mask
-	 * size is, e.g., 24, we may still not be able to guarantee a
-	 * suitable memory address, in which case we will return NULL.
-	 * But such devices are uncommon.
+	 * If the mask specifies that the memory be in the first 4 GB, then
+	 * we force the allocation to come from the DMA zone.  We also
+	 * force the node to 0 since that's the only node where the DMA
+	 * zone isn't empty.  If the mask size is smaller than 32 bits, we
+	 * may still not be able to guarantee a suitable memory address, in
+	 * which case we will return NULL.  But such devices are uncommon.
 	 */
-	if (dma_mask <= DMA_BIT_MASK(32))
+	if (dma_mask <= DMA_BIT_MASK(32)) {
+		gfp |= GFP_DMA;
 		node = 0;
+	}
 
-	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_UNCACHED);
+	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
 	if (pg == NULL)
 		return NULL;
 
 	addr = page_to_phys(pg);
 	if (addr + size > dma_mask) {
-		homecache_free_pages(addr, order);
+		__homecache_free_pages(pg, order);
 		return NULL;
 	}
 
 	*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
@@ -87,52 +96,285 @@ EXPORT_SYMBOL(dma_free_coherent);
  * can count on nothing having been touched.
  */
 
-/* Flush a PA range from cache page by page. */
-static void __dma_map_pa_range(dma_addr_t dma_addr, size_t size)
+/* Set up a single page for DMA access. */
+static void __dma_prep_page(struct page *page, unsigned long offset,
+			    size_t size, enum dma_data_direction direction)
+{
+	/*
+	 * Flush the page from cache if necessary.
+	 * On tilegx, data is delivered to hash-for-home L3; on tilepro,
+	 * data is delivered direct to memory.
+	 *
+	 * NOTE: If we were just doing DMA_TO_DEVICE we could optimize
+	 * this to be a "flush" not a "finv" and keep some of the
+	 * state in cache across the DMA operation, but it doesn't seem
+	 * worth creating the necessary flush_buffer_xxx() infrastructure.
+	 */
+	int home = page_home(page);
+	switch (home) {
+	case PAGE_HOME_HASH:
+#ifdef __tilegx__
+		return;
+#endif
+		break;
+	case PAGE_HOME_UNCACHED:
+#ifdef __tilepro__
+		return;
+#endif
+		break;
+	case PAGE_HOME_IMMUTABLE:
+		/* Should be going to the device only. */
+		BUG_ON(direction == DMA_FROM_DEVICE ||
+		       direction == DMA_BIDIRECTIONAL);
+		return;
+	case PAGE_HOME_INCOHERENT:
+		/* Incoherent anyway, so no need to work hard here. */
+		return;
+	default:
+		BUG_ON(home < 0 || home >= NR_CPUS);
+		break;
+	}
+	homecache_finv_page(page);
+
+#ifdef DEBUG_ALIGNMENT
+	/* Warn if the region isn't cacheline aligned. */
+	if (offset & (L2_CACHE_BYTES - 1) || (size & (L2_CACHE_BYTES - 1)))
+		pr_warn("Unaligned DMA to non-hfh memory: PA %#llx/%#lx\n",
+			PFN_PHYS(page_to_pfn(page)) + offset, size);
+#endif
+}
+
+/* Make the page ready to be read by the core. */
+static void __dma_complete_page(struct page *page, unsigned long offset,
+				size_t size, enum dma_data_direction direction)
+{
+#ifdef __tilegx__
+	switch (page_home(page)) {
+	case PAGE_HOME_HASH:
+		/* I/O device delivered data the way the cpu wanted it. */
+		break;
+	case PAGE_HOME_INCOHERENT:
+		/* Incoherent anyway, so no need to work hard here. */
+		break;
+	case PAGE_HOME_IMMUTABLE:
+		/* Extra read-only copies are not a problem. */
+		break;
+	default:
+		/* Flush the bogus hash-for-home I/O entries to memory. */
+		homecache_finv_map_page(page, PAGE_HOME_HASH);
+		break;
+	}
+#endif
+}
+
+static void __dma_prep_pa_range(dma_addr_t dma_addr, size_t size,
+				enum dma_data_direction direction)
 {
 	struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
-	size_t bytesleft = PAGE_SIZE - (dma_addr & (PAGE_SIZE - 1));
+	unsigned long offset = dma_addr & (PAGE_SIZE - 1);
+	size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
+
+	while (size != 0) {
+		__dma_prep_page(page, offset, bytes, direction);
+		size -= bytes;
+		++page;
+		offset = 0;
+		bytes = min((size_t)PAGE_SIZE, size);
+	}
+}
 
-	while ((ssize_t)size > 0) {
-		/* Flush the page. */
-		homecache_flush_cache(page++, 0);
+static void __dma_complete_pa_range(dma_addr_t dma_addr, size_t size,
+				    enum dma_data_direction direction)
+{
+	struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
+	unsigned long offset = dma_addr & (PAGE_SIZE - 1);
+	size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
+
+	while (size != 0) {
+		__dma_complete_page(page, offset, bytes, direction);
+		size -= bytes;
+		++page;
+		offset = 0;
+		bytes = min((size_t)PAGE_SIZE, 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;
+
+	BUG_ON(!valid_dma_direction(direction));
+
+	WARN_ON(nents == 0 || sglist->length == 0);
 
-		/* Figure out if we need to continue on the next page. */
-		size -= bytesleft;
-		bytesleft = PAGE_SIZE;
+	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;
 }
 
-/*
- * 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)
+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));
-	WARN_ON(size == 0);
 
-	__dma_map_pa_range(dma_addr, size);
+	__dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
+			    dma_address & PAGE_OFFSET, size, direction);
+}
 
-	return dma_addr;
+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));
+
+	__dma_complete_pa_range(dma_handle, size, direction);
+}
+
+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));
+	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);
+	}
+}
+
+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));
 }
-EXPORT_SYMBOL(dma_unmap_single);
 
-int dma_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
-	   enum dma_data_direction direction)
+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;
@@ -143,73 +385,103 @@ 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_map_pa_range(sg->dma_address, sg->length);
+		__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 *sg, int nhwentries,
-	     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;
+
 	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);
+	}
 }
-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);
-	homecache_flush_cache(page, 0);
+	__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)
 {
-	unsigned long start = PFN_DOWN(dma_handle);
-	unsigned long end = PFN_DOWN(dma_handle + size - 1);
-	unsigned long i;
+	dma_handle = dma_to_phys(dev, dma_handle);
 
-	BUG_ON(!valid_dma_direction(direction));
-	for (i = start; i <= end; ++i)
-		homecache_flush_cache(pfn_to_page(i), 0);
+	__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 *sg, 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;
+
 	BUG_ON(!valid_dma_direction(direction));
-	WARN_ON(nelems == 0 || sg[0].length == 0);
+	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_sync_sg_for_cpu);
 
-/*
- * Flush and invalidate cache for scatterlist.
- */
-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;
@@ -222,31 +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() returns non-cacheable 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);
+}
+
+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_cache_sync);
+EXPORT_SYMBOL(dma_set_coherent_mask);
+#endif
diff --git a/arch/tile/kernel/pci_gx.c b/arch/tile/kernel/pci_gx.c
new file mode 100644
index 000000000000..fa75264a82ae
--- /dev/null
+++ b/arch/tile/kernel/pci_gx.c
@@ -0,0 +1,1543 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/mmzone.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/string.h>
+#include <linux/init.h>
+#include <linux/capability.h>
+#include <linux/sched.h>
+#include <linux/errno.h>
+#include <linux/irq.h>
+#include <linux/msi.h>
+#include <linux/io.h>
+#include <linux/uaccess.h>
+#include <linux/ctype.h>
+
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/byteorder.h>
+
+#include <gxio/iorpc_globals.h>
+#include <gxio/kiorpc.h>
+#include <gxio/trio.h>
+#include <gxio/iorpc_trio.h>
+#include <hv/drv_trio_intf.h>
+
+#include <arch/sim.h>
+
+/*
+ * This file containes the routines to search for PCI buses,
+ * enumerate the buses, and configure any attached devices.
+ */
+
+#define DEBUG_PCI_CFG	0
+
+#if DEBUG_PCI_CFG
+#define TRACE_CFG_WR(size, val, bus, dev, func, offset) \
+	pr_info("CFG WR %d-byte VAL %#x to bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#define TRACE_CFG_RD(size, val, bus, dev, func, offset) \
+	pr_info("CFG RD %d-byte VAL %#x from bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#else
+#define TRACE_CFG_WR(...)
+#define TRACE_CFG_RD(...)
+#endif
+
+static int __devinitdata pci_probe = 1;
+
+/* Information on the PCIe RC ports configuration. */
+static int __devinitdata pcie_rc[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/*
+ * On some platforms with one or more Gx endpoint ports, we need to
+ * delay the PCIe RC port probe for a few seconds to work around
+ * a HW PCIe link-training bug. The exact delay is specified with
+ * a kernel boot argument in the form of "pcie_rc_delay=T,P,S",
+ * where T is the TRIO instance number, P is the port number and S is
+ * the delay in seconds. If the delay is not provided, the value
+ * will be DEFAULT_RC_DELAY.
+ */
+static int __devinitdata rc_delay[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/* Default number of seconds that the PCIe RC port probe can be delayed. */
+#define DEFAULT_RC_DELAY	10
+
+/* Max number of seconds that the PCIe RC port probe can be delayed. */
+#define MAX_RC_DELAY		20
+
+/* Array of the PCIe ports configuration info obtained from the BIB. */
+struct pcie_port_property pcie_ports[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/* All drivers share the TRIO contexts defined here. */
+gxio_trio_context_t trio_contexts[TILEGX_NUM_TRIO];
+
+/* Pointer to an array of PCIe RC controllers. */
+struct pci_controller pci_controllers[TILEGX_NUM_TRIO * TILEGX_TRIO_PCIES];
+int num_rc_controllers;
+static int num_ep_controllers;
+
+static struct pci_ops tile_cfg_ops;
+
+/* Mask of CPUs that should receive PCIe interrupts. */
+static struct cpumask intr_cpus_map;
+
+/*
+ * We don't need to worry about the alignment of resources.
+ */
+resource_size_t pcibios_align_resource(void *data, const struct resource *res,
+				resource_size_t size, resource_size_t align)
+{
+	return res->start;
+}
+EXPORT_SYMBOL(pcibios_align_resource);
+
+
+/*
+ * Pick a CPU to receive and handle the PCIe interrupts, based on the IRQ #.
+ * For now, we simply send interrupts to non-dataplane CPUs.
+ * We may implement methods to allow user to specify the target CPUs,
+ * e.g. via boot arguments.
+ */
+static int tile_irq_cpu(int irq)
+{
+	unsigned int count;
+	int i = 0;
+	int cpu;
+
+	count = cpumask_weight(&intr_cpus_map);
+	if (unlikely(count == 0)) {
+		pr_warning("intr_cpus_map empty, interrupts will be"
+			   " delievered to dataplane tiles\n");
+		return irq % (smp_height * smp_width);
+	}
+
+	count = irq % count;
+	for_each_cpu(cpu, &intr_cpus_map) {
+		if (i++ == count)
+			break;
+	}
+	return cpu;
+}
+
+/*
+ * Open a file descriptor to the TRIO shim.
+ */
+static int __devinit tile_pcie_open(int trio_index)
+{
+	gxio_trio_context_t *context = &trio_contexts[trio_index];
+	int ret;
+
+	/*
+	 * This opens a file descriptor to the TRIO shim.
+	 */
+	ret = gxio_trio_init(context, trio_index);
+	if (ret < 0)
+		return ret;
+
+	/*
+	 * Allocate an ASID for the kernel.
+	 */
+	ret = gxio_trio_alloc_asids(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: ASID alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto asid_alloc_failure;
+	}
+
+	context->asid = ret;
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+	/*
+	 * Alloc a PIO region for config access, shared by all MACs per TRIO.
+	 * This shouldn't fail since the kernel is supposed to the first
+	 * client of the TRIO's PIO regions.
+	 */
+	ret = gxio_trio_alloc_pio_regions(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+
+	context->pio_cfg_index = ret;
+
+	/*
+	 * For PIO CFG, the bus_address_hi parameter is 0. The mac parameter
+	 * is also 0 because it is specified in PIO_REGION_SETUP_CFG_ADDR.
+	 */
+	ret = gxio_trio_init_pio_region_aux(context, context->pio_cfg_index,
+		0, 0, HV_TRIO_PIO_FLAG_CONFIG_SPACE);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO init failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+#endif
+
+	return ret;
+
+asid_alloc_failure:
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+pio_alloc_failure:
+#endif
+	hv_dev_close(context->fd);
+
+	return ret;
+}
+
+static void
+tilegx_legacy_irq_ack(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_EVENT_RESET_K, 1UL << d->irq);
+}
+
+static void
+tilegx_legacy_irq_mask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_SET_K, 1UL << d->irq);
+}
+
+static void
+tilegx_legacy_irq_unmask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_RESET_K, 1UL << d->irq);
+}
+
+static struct irq_chip tilegx_legacy_irq_chip = {
+	.name			= "tilegx_legacy_irq",
+	.irq_ack		= tilegx_legacy_irq_ack,
+	.irq_mask		= tilegx_legacy_irq_mask,
+	.irq_unmask		= tilegx_legacy_irq_unmask,
+
+	/* TBD: support set_affinity. */
+};
+
+/*
+ * This is a wrapper function of the kernel level-trigger interrupt
+ * handler handle_level_irq() for PCI legacy interrupts. The TRIO
+ * is configured such that only INTx Assert interrupts are proxied
+ * to Linux which just calls handle_level_irq() after clearing the
+ * MAC INTx Assert status bit associated with this interrupt.
+ */
+static void
+trio_handle_level_irq(unsigned int irq, struct irq_desc *desc)
+{
+	struct pci_controller *controller = irq_desc_get_handler_data(desc);
+	gxio_trio_context_t *trio_context = controller->trio;
+	uint64_t intx = (uint64_t)irq_desc_get_chip_data(desc);
+	int mac = controller->mac;
+	unsigned int reg_offset;
+	uint64_t level_mask;
+
+	handle_level_irq(irq, desc);
+
+	/*
+	 * Clear the INTx Level status, otherwise future interrupts are
+	 * not sent.
+	 */
+	reg_offset = (TRIO_PCIE_INTFC_MAC_INT_STS <<
+		TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+		TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	level_mask = TRIO_PCIE_INTFC_MAC_INT_STS__INT_LEVEL_MASK << intx;
+
+	__gxio_mmio_write(trio_context->mmio_base_mac + reg_offset, level_mask);
+}
+
+/*
+ * Create kernel irqs and set up the handlers for the legacy interrupts.
+ * Also some minimum initialization for the MSI support.
+ */
+static int __devinit tile_init_irqs(struct pci_controller *controller)
+{
+	int i;
+	int j;
+	int irq;
+	int result;
+
+	cpumask_copy(&intr_cpus_map, cpu_online_mask);
+
+
+	for (i = 0; i < 4; i++) {
+		gxio_trio_context_t *context = controller->trio;
+		int cpu;
+
+		/* Ask the kernel to allocate an IRQ. */
+		irq = create_irq();
+		if (irq < 0) {
+			pr_err("PCI: no free irq vectors, failed for %d\n", i);
+
+			goto free_irqs;
+		}
+		controller->irq_intx_table[i] = irq;
+
+		/* Distribute the 4 IRQs to different tiles. */
+		cpu = tile_irq_cpu(irq);
+
+		/* Configure the TRIO intr binding for this IRQ. */
+		result = gxio_trio_config_legacy_intr(context, cpu_x(cpu),
+						      cpu_y(cpu), KERNEL_PL,
+						      irq, controller->mac, i);
+		if (result < 0) {
+			pr_err("PCI: MAC intx config failed for %d\n", i);
+
+			goto free_irqs;
+		}
+
+		/*
+		 * Register the IRQ handler with the kernel.
+		 */
+		irq_set_chip_and_handler(irq, &tilegx_legacy_irq_chip,
+					trio_handle_level_irq);
+		irq_set_chip_data(irq, (void *)(uint64_t)i);
+		irq_set_handler_data(irq, controller);
+	}
+
+	return 0;
+
+free_irqs:
+	for (j = 0; j < i; j++)
+		destroy_irq(controller->irq_intx_table[j]);
+
+	return -1;
+}
+
+/*
+ * Find valid controllers and fill in pci_controller structs for each
+ * of them.
+ *
+ * Returns the number of controllers discovered.
+ */
+int __init tile_pci_init(void)
+{
+	int num_trio_shims = 0;
+	int ctl_index = 0;
+	int i, j;
+
+	if (!pci_probe) {
+		pr_info("PCI: disabled by boot argument\n");
+		return 0;
+	}
+
+	pr_info("PCI: Searching for controllers...\n");
+
+	/*
+	 * We loop over all the TRIO shims.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		int ret;
+
+		ret = tile_pcie_open(i);
+		if (ret < 0)
+			continue;
+
+		num_trio_shims++;
+	}
+
+	if (num_trio_shims == 0 || sim_is_simulator())
+		return 0;
+
+	/*
+	 * Now determine which PCIe ports are configured to operate in RC mode.
+	 * We look at the Board Information Block first and then see if there
+	 * are any overriding configuration by the HW strapping pin.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		gxio_trio_context_t *context = &trio_contexts[i];
+		int ret;
+
+		if (context->fd < 0)
+			continue;
+
+		ret = hv_dev_pread(context->fd, 0,
+			(HV_VirtAddr)&pcie_ports[i][0],
+			sizeof(struct pcie_port_property) * TILEGX_TRIO_PCIES,
+			GXIO_TRIO_OP_GET_PORT_PROPERTY);
+		if (ret < 0) {
+			pr_err("PCI: PCIE_GET_PORT_PROPERTY failure, error %d,"
+				" on TRIO %d\n", ret, i);
+			continue;
+		}
+
+		for (j = 0; j < TILEGX_TRIO_PCIES; j++) {
+			if (pcie_ports[i][j].allow_rc) {
+				pcie_rc[i][j] = 1;
+				num_rc_controllers++;
+			}
+			else if (pcie_ports[i][j].allow_ep) {
+				num_ep_controllers++;
+			}