MIPS: PowerTV: Use O(1) algorthm for phys_to_dma/dma_to_phys

Replace phys_to_dma()/dma_to_phys() looping algorithm with an O(1) algorithm
The approach taken is inspired by the sparse memory implementation: take a
certain number of high-order bits off the address them, use this as an
index into a table containing an offset to the desired address and add
it to the original value. There is a table for mapping physical addresses
to DMA addresses and another one for the reverse mapping. The table sizes
depend on how fine-grained the mappings need to be; Coarser granularity
less to smaller tables.  On a processor with 32-bit physical and DMA
addresses, with 4 MIB granularity, memory usage is two 2048-byte arrays.
Each 32-byte cache line thus covers 64 MiB of address space.

Also, renames phys_to_bus() to phys_to_dma() and bus_to_phys() to
dma_to_phys() to align with kernel usage.

[Ralf: Fixed silly build breakage due to stackoverflow warning caused by
huge array on stack.]

Signed-off-by: David VomLehn <dvomlehn@cisco.com>
To: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/1257/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

2 files changed, 125 insertions, 48 deletions

diff --git a/arch/mips/include/asm/mach-powertv/dma-coherence.h b/arch/mips/include/asm/mach-powertv/dma-coherence.h
index 5b8d5ebeb83..f76029c2406 100644
--- a/arch/mips/include/asm/mach-powertv/dma-coherence.h
+++ b/arch/mips/include/asm/mach-powertv/dma-coherence.h
@@ -65,21 +65,21 @@ static inline dma_addr_t plat_map_dma_mem(struct device *dev, void *addr,
 	size_t size)
 {
 	if (is_kseg2(addr))
-		return phys_to_bus(virt_to_phys_from_pte(addr));
+		return phys_to_dma(virt_to_phys_from_pte(addr));
 	else
-		return phys_to_bus(virt_to_phys(addr));
+		return phys_to_dma(virt_to_phys(addr));
 }
 
 static inline dma_addr_t plat_map_dma_mem_page(struct device *dev,
 	struct page *page)
 {
-	return phys_to_bus(page_to_phys(page));
+	return phys_to_dma(page_to_phys(page));
 }
 
 static inline unsigned long plat_dma_addr_to_phys(struct device *dev,
 	dma_addr_t dma_addr)
 {
-	return bus_to_phys(dma_addr);
+	return dma_to_phys(dma_addr);
 }
 
 static inline void plat_unmap_dma_mem(struct device *dev, dma_addr_t dma_addr,
diff --git a/arch/mips/include/asm/mach-powertv/ioremap.h b/arch/mips/include/asm/mach-powertv/ioremap.h
index e6276d5146e..076f2eeaa57 100644
--- a/arch/mips/include/asm/mach-powertv/ioremap.h
+++ b/arch/mips/include/asm/mach-powertv/ioremap.h
@@ -10,64 +10,101 @@
 #define __ASM_MACH_POWERTV_IOREMAP_H
 
 #include <linux/types.h>
+#include <linux/log2.h>
+#include <linux/compiler.h>
 
-#define LOW_MEM_BOUNDARY_PHYS	0x20000000
-#define LOW_MEM_BOUNDARY_MASK	(~(LOW_MEM_BOUNDARY_PHYS - 1))
+#include <asm/pgtable-bits.h>
+#include <asm/addrspace.h>
+
+/* We're going to mess with bits, so get sizes */
+#define IOR_BPC			8			/* Bits per char */
+#define IOR_PHYS_BITS		(IOR_BPC * sizeof(phys_addr_t))
+#define IOR_DMA_BITS		(IOR_BPC * sizeof(dma_addr_t))
 
 /*
- * The bus addresses are different than the physical addresses that
- * the processor sees by an offset. This offset varies by ASIC
- * version. Define a variable to hold the offset and some macros to
- * make the conversion simpler. */
-extern unsigned long phys_to_bus_offset;
-
-#ifdef CONFIG_HIGHMEM
-#define MEM_GAP_PHYS		0x60000000
+ * Define the granularity of physical/DMA mapping in terms of the number
+ * of bits that defines the offset within a grain. These will be the
+ * least significant bits of the address. The rest of a physical or DMA
+ * address will be used to index into an appropriate table to find the
+ * offset to add to the address to yield the corresponding DMA or physical
+ * address, respectively.
+ */
+#define IOR_LSBITS		22			/* Bits in a grain */
+
 /*
- * TODO: We will use the hard code for conversion between physical and
- * bus until the bootloader releases their device tree to us.
+ * Compute the number of most significant address bits after removing those
+ * used for the offset within a grain and then compute the number of table
+ * entries for the conversion.
  */
-#define phys_to_bus(x) (((x) < LOW_MEM_BOUNDARY_PHYS) ? \
-	((x) + phys_to_bus_offset) : (x))
-#define bus_to_phys(x) (((x) < MEM_GAP_PHYS_ADDR) ? \
-	((x) - phys_to_bus_offset) : (x))
-#else
-#define phys_to_bus(x) ((x) + phys_to_bus_offset)
-#define bus_to_phys(x) ((x) - phys_to_bus_offset)
-#endif
+#define IOR_PHYS_MSBITS		(IOR_PHYS_BITS - IOR_LSBITS)
+#define IOR_NUM_PHYS_TO_DMA	((phys_addr_t) 1 << IOR_PHYS_MSBITS)
+
+#define IOR_DMA_MSBITS		(IOR_DMA_BITS - IOR_LSBITS)
+#define IOR_NUM_DMA_TO_PHYS	((dma_addr_t) 1 << IOR_DMA_MSBITS)
 
 /*
- * Determine whether the address we are given is for an ASIC device
- * Params:  addr    Address to check
- * Returns: Zero if the address is not for ASIC devices, non-zero
- *      if it is.
+ * Define data structures used as elements in the arrays for the conversion
+ * between physical and DMA addresses. We do some slightly fancy math to
+ * compute the width of the offset element of the conversion tables so
+ * that we can have the smallest conversion tables. Next, round up the
+ * sizes to the next higher power of two, i.e. the offset element will have
+ * 8, 16, 32, 64, etc. bits. This eliminates the need to mask off any
+ * bits.  Finally, we compute a shift value that puts the most significant
+ * bits of the offset into the most significant bits of the offset element.
+ * This makes it more efficient on processors without barrel shifters and
+ * easier to see the values if the conversion table is dumped in binary.
  */
-static inline int asic_is_device_addr(phys_t addr)
+#define _IOR_OFFSET_WIDTH(n)	(1 << order_base_2(n))
+#define IOR_OFFSET_WIDTH(n) \
+	(_IOR_OFFSET_WIDTH(n) < 8 ? 8 : _IOR_OFFSET_WIDTH(n))
+
+#define IOR_PHYS_OFFSET_BITS	IOR_OFFSET_WIDTH(IOR_PHYS_MSBITS)
+#define IOR_PHYS_SHIFT		(IOR_PHYS_BITS - IOR_PHYS_OFFSET_BITS)
+
+#define IOR_DMA_OFFSET_BITS	IOR_OFFSET_WIDTH(IOR_DMA_MSBITS)
+#define IOR_DMA_SHIFT		(IOR_DMA_BITS - IOR_DMA_OFFSET_BITS)
+
+struct ior_phys_to_dma {
+	dma_addr_t offset:IOR_DMA_OFFSET_BITS __packed
+		__aligned((IOR_DMA_OFFSET_BITS / IOR_BPC));
+};
+
+struct ior_dma_to_phys {
+	dma_addr_t offset:IOR_PHYS_OFFSET_BITS __packed
+		__aligned((IOR_PHYS_OFFSET_BITS / IOR_BPC));
+};
+
+extern struct ior_phys_to_dma _ior_phys_to_dma[IOR_NUM_PHYS_TO_DMA];
+extern struct ior_dma_to_phys _ior_dma_to_phys[IOR_NUM_DMA_TO_PHYS];
+
+static inline dma_addr_t _phys_to_dma_offset_raw(phys_addr_t phys)
 {
-	return !((phys_t)addr & (phys_t) LOW_MEM_BOUNDARY_MASK);
+	return (dma_addr_t)_ior_phys_to_dma[phys >> IOR_LSBITS].offset;
 }
 
-/*
- * Determine whether the address we are given is external RAM mappable
- * into KSEG1.
- * Params:  addr    Address to check
- * Returns: Zero if the address is not for external RAM and
- */
-static inline int asic_is_lowmem_ram_addr(phys_t addr)
+static inline dma_addr_t _dma_to_phys_offset_raw(dma_addr_t dma)
 {
-	/*
-	 * The RAM always starts at the following address in the processor's
-	 * physical address space
-	 */
-	static const phys_t phys_ram_base = 0x10000000;
-	phys_t bus_ram_base;
+	return (dma_addr_t)_ior_dma_to_phys[dma >> IOR_LSBITS].offset;
+}
 
-	bus_ram_base = phys_to_bus_offset + phys_ram_base;
+/* These are not portable and should not be used in drivers. Drivers should
+ * be using ioremap() and friends to map physical addreses to virtual
+ * addresses and dma_map*() and friends to map virtual addresses into DMA
+ * addresses and back.
+ */
+static inline dma_addr_t phys_to_dma(phys_addr_t phys)
+{
+	return phys + (_phys_to_dma_offset_raw(phys) << IOR_PHYS_SHIFT);
+}
 
-	return addr >= bus_ram_base &&
-		addr < (bus_ram_base + (LOW_MEM_BOUNDARY_PHYS - phys_ram_base));
+static inline phys_addr_t dma_to_phys(dma_addr_t dma)
+{
+	return dma + (_dma_to_phys_offset_raw(dma) << IOR_DMA_SHIFT);
 }
 
+extern void ioremap_add_map(dma_addr_t phys, phys_addr_t alias,
+	dma_addr_t size);
+
 /*
  * Allow physical addresses to be fixed up to help peripherals located
  * outside the low 32-bit range -- generic pass-through version.
@@ -77,10 +114,50 @@ static inline phys_t fixup_bigphys_addr(phys_t phys_addr, phys_t size)
 	return phys_addr;
 }
 
-static inline void __iomem *plat_ioremap(phys_t offset, unsigned long size,
+/*
+ * Handle the special case of addresses the area aliased into the first
+ * 512 MiB of the processor's physical address space. These turn into either
+ * kseg0 or kseg1 addresses, depending on flags.
+ */
+static inline void __iomem *plat_ioremap(phys_t start, unsigned long size,
 	unsigned long flags)
 {
-	return NULL;
+	phys_addr_t start_offset;
+	void __iomem *result = NULL;
+
+	/* Start by checking to see whether this is an aliased address */
+	start_offset = _dma_to_phys_offset_raw(start);
+
+	/*
+	 * If:
+	 * o	the memory is aliased into the first 512 MiB, and
+	 * o	the start and end are in the same RAM bank, and
+	 * o	we don't have a zero size or wrap around, and
+	 * o	we are supposed to create an uncached mapping,
+	 *	handle this is a kseg0 or kseg1 address
+	 */
+	if (start_offset != 0) {
+		phys_addr_t last;
+		dma_addr_t dma_to_phys_offset;
+
+		last = start + size - 1;
+		dma_to_phys_offset =
+			_dma_to_phys_offset_raw(last) << IOR_DMA_SHIFT;
+
+		if (dma_to_phys_offset == start_offset &&
+			size != 0 && start <= last) {
+			phys_t adjusted_start;
+			adjusted_start = start + start_offset;
+			if (flags == _CACHE_UNCACHED)
+				result = (void __iomem *) (unsigned long)
+					CKSEG1ADDR(adjusted_start);
+			else
+				result = (void __iomem *) (unsigned long)
+					CKSEG0ADDR(adjusted_start);
+		}
+	}
+
+	return result;
 }
 
 static inline int plat_iounmap(const volatile void __iomem *addr)