MIPS: Octeon: Rewrite DMA mapping functions.

All Octeon chips can support more than 4GB of RAM.  Also due to how Octeon
PCI is setup, even some configurations with less than 4GB of RAM will have
portions that are not accessible from 32-bit devices.

Enable the swiotlb code to handle the cases where a device cannot directly
do DMA.  This is a complete rewrite of the Octeon DMA mapping code.
Signed-off-by: David Daney <ddaney@caviumnetworks.com>
Patchwork: http://patchwork.linux-mips.org/patch/1639/Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

arch/mips/cavium-octeon/Kconfig

@@ -87,3 +87,15 @@ config ARCH_SPARSEMEM_ENABLE
 config CAVIUM_OCTEON_HELPER
 	def_bool y
 	depends on OCTEON_ETHERNET || PCI
+
+config IOMMU_HELPER
+	bool
+
+config NEED_SG_DMA_LENGTH
+	bool
+
+config SWIOTLB
+	def_bool y
+	depends on CPU_CAVIUM_OCTEON
+	select IOMMU_HELPER
+	select NEED_SG_DMA_LENGTH

arch/mips/include/asm/mach-cavium-octeon/dma-coherence.h

@@ -15,13 +15,12 @@
 
 struct device;
 
-dma_addr_t octeon_map_dma_mem(struct device *, void *, size_t);
-void octeon_unmap_dma_mem(struct device *, dma_addr_t);
+extern void octeon_pci_dma_init(void);
 
 static inline dma_addr_t plat_map_dma_mem(struct device *dev, void *addr,
 	size_t size)
 {
-	return octeon_map_dma_mem(dev, addr, size);
+	BUG();
 }
 
 static inline dma_addr_t plat_map_dma_mem_page(struct device *dev,
@@ -33,23 +32,23 @@ static inline dma_addr_t plat_map_dma_mem_page(struct device *dev,
 static inline unsigned long plat_dma_addr_to_phys(struct device *dev,
 	dma_addr_t dma_addr)
 {
-	return dma_addr;
+	BUG();
 }
 
 static inline void plat_unmap_dma_mem(struct device *dev, dma_addr_t dma_addr,
 	size_t size, enum dma_data_direction direction)
 {
-	octeon_unmap_dma_mem(dev, dma_addr);
+	BUG();
 }
 
 static inline int plat_dma_supported(struct device *dev, u64 mask)
 {
-	return 1;
+	BUG();
 }
 
 static inline void plat_extra_sync_for_device(struct device *dev)
 {
-	mb();
+	BUG();
 }
 
 static inline int plat_device_is_coherent(struct device *dev)
@@ -60,7 +59,14 @@ static inline int plat_device_is_coherent(struct device *dev)
 static inline int plat_dma_mapping_error(struct device *dev,
 					 dma_addr_t dma_addr)
 {
-	return dma_addr == -1;
+	BUG();
 }
 
+dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr);
+phys_addr_t dma_to_phys(struct device *dev, dma_addr_t daddr);
+
+struct dma_map_ops;
+extern struct dma_map_ops *octeon_pci_dma_map_ops;
+extern char *octeon_swiotlb;
+
 #endif /* __ASM_MACH_CAVIUM_OCTEON_DMA_COHERENCE_H */

arch/mips/include/asm/octeon/pci-octeon.h

@@ -35,6 +35,16 @@
 extern int (*octeon_pcibios_map_irq)(const struct pci_dev *dev,
 				     u8 slot, u8 pin);
 
+/*
+ * For PCI (not PCIe) the BAR2 base address.
+ */
+#define OCTEON_BAR2_PCI_ADDRESS 0x8000000000ull
+
+/*
+ * For PCI (not PCIe) the base of the memory mapped by BAR1
+ */
+extern u64 octeon_bar1_pci_phys;
+
 /*
  * The following defines are used when octeon_dma_bar_type =
  * OCTEON_DMA_BAR_TYPE_BIG

arch/mips/pci/pci-octeon.c

@@ -11,6 +11,7 @@
 #include <linux/interrupt.h>
 #include <linux/time.h>
 #include <linux/delay.h>
+#include <linux/swiotlb.h>
 
 #include <asm/time.h>
 
@@ -19,6 +20,8 @@
 #include <asm/octeon/cvmx-pci-defs.h>
 #include <asm/octeon/pci-octeon.h>
 
+#include <dma-coherence.h>
+
 #define USE_OCTEON_INTERNAL_ARBITER
 
 /*
@@ -32,6 +35,8 @@
 /* Octeon't PCI controller uses did=3, subdid=3 for PCI memory. */
 #define OCTEON_PCI_MEMSPACE_OFFSET  (0x00011b0000000000ull)
 
+u64 octeon_bar1_pci_phys;
+
 /**
  * This is the bit decoding used for the Octeon PCI controller addresses
  */
@@ -170,6 +175,8 @@ int pcibios_plat_dev_init(struct pci_dev *dev)
 		pci_write_config_dword(dev, pos + PCI_ERR_ROOT_STATUS, dconfig);
 	}
 
+	dev->dev.archdata.dma_ops = octeon_pci_dma_map_ops;
+
 	return 0;
 }
 
@@ -618,12 +625,10 @@ static int __init octeon_pci_setup(void)
 	 * before the readl()'s below. We don't want BAR2 overlapping
 	 * with BAR0/BAR1 during these reads.
 	 */
-	octeon_npi_write32(CVMX_NPI_PCI_CFG08, 0);
-	octeon_npi_write32(CVMX_NPI_PCI_CFG09, 0x80);
-
-	/* Disable the BAR1 movable mappings */
-	for (index = 0; index < 32; index++)
-		octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index), 0);
+	octeon_npi_write32(CVMX_NPI_PCI_CFG08,
+			   (u32)(OCTEON_BAR2_PCI_ADDRESS & 0xffffffffull));
+	octeon_npi_write32(CVMX_NPI_PCI_CFG09,
+			   (u32)(OCTEON_BAR2_PCI_ADDRESS >> 32));
 
 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
 		/* Remap the Octeon BAR 0 to 0-2GB */
@@ -637,6 +642,25 @@ static int __init octeon_pci_setup(void)
 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 2ul << 30);
 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
 
+		/* BAR1 movable mappings set for identity mapping */
+		octeon_bar1_pci_phys = 0x80000000ull;
+		for (index = 0; index < 32; index++) {
+			union cvmx_pci_bar1_indexx bar1_index;
+
+			bar1_index.u32 = 0;
+			/* Address bits[35:22] sent to L2C */
+			bar1_index.s.addr_idx =
+				(octeon_bar1_pci_phys >> 22) + index;
+			/* Don't put PCI accesses in L2. */
+			bar1_index.s.ca = 1;
+			/* Endian Swap Mode */
+			bar1_index.s.end_swp = 1;
+			/* Set '1' when the selected address range is valid. */
+			bar1_index.s.addr_v = 1;
+			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
+					   bar1_index.u32);
+		}
+
 		/* Devices go after BAR1 */
 		octeon_pci_mem_resource.start =
 			OCTEON_PCI_MEMSPACE_OFFSET + (4ul << 30) -
@@ -652,6 +676,27 @@ static int __init octeon_pci_setup(void)
 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 0);
 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
 
+		/* BAR1 movable regions contiguous to cover the swiotlb */
+		octeon_bar1_pci_phys =
+			virt_to_phys(octeon_swiotlb) & ~((1ull << 22) - 1);
+
+		for (index = 0; index < 32; index++) {
+			union cvmx_pci_bar1_indexx bar1_index;
+
+			bar1_index.u32 = 0;
+			/* Address bits[35:22] sent to L2C */
+			bar1_index.s.addr_idx =
+				(octeon_bar1_pci_phys >> 22) + index;
+			/* Don't put PCI accesses in L2. */
+			bar1_index.s.ca = 1;
+			/* Endian Swap Mode */
+			bar1_index.s.end_swp = 1;
+			/* Set '1' when the selected address range is valid. */
+			bar1_index.s.addr_v = 1;
+			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
+					   bar1_index.u32);
+		}
+
 		/* Devices go after BAR0 */
 		octeon_pci_mem_resource.start =
 			OCTEON_PCI_MEMSPACE_OFFSET + (128ul << 20) +
@@ -667,6 +712,9 @@ static int __init octeon_pci_setup(void)
 	 * was setup properly.
 	 */
 	cvmx_write_csr(CVMX_NPI_PCI_INT_SUM2, -1);
+
+	octeon_pci_dma_init();
+
 	return 0;
 }
 

arch/mips/pci/pcie-octeon.c

@@ -75,6 +75,8 @@ union cvmx_pcie_address {
 	} mem;
 };
 
+#include <dma-coherence.h>
+
 /**
  * Return the Core virtual base address for PCIe IO access. IOs are
  * read/written as an offset from this address.
@@ -1391,6 +1393,9 @@ static int __init octeon_pcie_setup(void)
 			cvmx_pcie_get_io_size(1) - 1;
 		register_pci_controller(&octeon_pcie1_controller);
 	}
+
+	octeon_pci_dma_init();
+
 	return 0;
 }