metag: Memory handling

Meta has instructions for accessing:
 - bytes        - GETB (1 byte)
 - words        - GETW (2 bytes)
 - doublewords  - GETD (4 bytes)
 - longwords    - GETL (8 bytes)

All accesses must be aligned. Unaligned accesses can be detected and
made to fault on Meta2, however it isn't possible to fix up unaligned
writes so we don't bother fixing up reads either.

This patch adds metag memory handling code including:
 - I/O memory (io.h, ioremap.c): Actually any virtual memory can be
   accessed with these helpers. A part of the non-MMUable address space
   is used for memory mapped I/O. The ioremap() function is implemented
   one to one for non-MMUable addresses.
 - User memory (uaccess.h, usercopy.c): User memory is directly
   accessible from privileged code.
 - Kernel memory (maccess.c): probe_kernel_write() needs to be
   overwridden to use the I/O functions when doing a simple aligned
   write to non-writecombined memory, otherwise the write may be split
   by the generic version.

Note that due to the fact that a portion of the virtual address space is
non-MMUable, and therefore always maps directly to the physical address
space, metag specific I/O functions are made available (metag_in32,
metag_out32 etc). These cast the address argument to a pointer so that
they can be used with raw physical addresses. These accessors are only
to be used for accessing fixed core Meta architecture registers in the
non-MMU region, and not for any SoC/peripheral registers.

Signed-off-by: James Hogan <james.hogan@imgtec.com>

2 files changed, 157 insertions, 0 deletions

diff --git a/arch/metag/mm/ioremap.c b/arch/metag/mm/ioremap.c
new file mode 100644
index 000000000000..a136a435fdaa
--- /dev/null
+++ b/arch/metag/mm/ioremap.c
@@ -0,0 +1,89 @@
+/*
+ * Re-map IO memory to kernel address space so that we can access it.
+ * Needed for memory-mapped I/O devices mapped outside our normal DRAM
+ * window (that is, all memory-mapped I/O devices).
+ *
+ * Copyright (C) 1995,1996 Linus Torvalds
+ *
+ * Meta port based on CRIS-port by Axis Communications AB
+ */
+
+#include <linux/vmalloc.h>
+#include <linux/io.h>
+#include <linux/export.h>
+#include <linux/slab.h>
+#include <linux/mm.h>
+
+#include <asm/pgtable.h>
+
+/*
+ * Remap an arbitrary physical address space into the kernel virtual
+ * address space. Needed when the kernel wants to access high addresses
+ * directly.
+ *
+ * NOTE! We need to allow non-page-aligned mappings too: we will obviously
+ * have to convert them into an offset in a page-aligned mapping, but the
+ * caller shouldn't need to know that small detail.
+ */
+void __iomem *__ioremap(unsigned long phys_addr, size_t size,
+			unsigned long flags)
+{
+	unsigned long addr;
+	struct vm_struct *area;
+	unsigned long offset, last_addr;
+	pgprot_t prot;
+
+	/* Don't allow wraparound or zero size */
+	last_addr = phys_addr + size - 1;
+	if (!size || last_addr < phys_addr)
+		return NULL;
+
+	/* Custom region addresses are accessible and uncached by default. */
+	if (phys_addr >= LINSYSCUSTOM_BASE &&
+	    phys_addr < (LINSYSCUSTOM_BASE + LINSYSCUSTOM_LIMIT))
+		return (__force void __iomem *) phys_addr;
+
+	/*
+	 * Mappings have to be page-aligned
+	 */
+	offset = phys_addr & ~PAGE_MASK;
+	phys_addr &= PAGE_MASK;
+	size = PAGE_ALIGN(last_addr+1) - phys_addr;
+	prot = __pgprot(_PAGE_PRESENT | _PAGE_WRITE | _PAGE_DIRTY |
+			_PAGE_ACCESSED | _PAGE_KERNEL | _PAGE_CACHE_WIN0 |
+			flags);
+
+	/*
+	 * Ok, go for it..
+	 */
+	area = get_vm_area(size, VM_IOREMAP);
+	if (!area)
+		return NULL;
+	area->phys_addr = phys_addr;
+	addr = (unsigned long) area->addr;
+	if (ioremap_page_range(addr, addr + size, phys_addr, prot)) {
+		vunmap((void *) addr);
+		return NULL;
+	}
+	return (__force void __iomem *) (offset + (char *)addr);
+}
+EXPORT_SYMBOL(__ioremap);
+
+void __iounmap(void __iomem *addr)
+{
+	struct vm_struct *p;
+
+	if ((__force unsigned long)addr >= LINSYSCUSTOM_BASE &&
+	    (__force unsigned long)addr < (LINSYSCUSTOM_BASE +
+					   LINSYSCUSTOM_LIMIT))
+		return;
+
+	p = remove_vm_area((void *)(PAGE_MASK & (unsigned long __force)addr));
+	if (unlikely(!p)) {
+		pr_err("iounmap: bad address %p\n", addr);
+		return;
+	}
+
+	kfree(p);
+}
+EXPORT_SYMBOL(__iounmap);
diff --git a/arch/metag/mm/maccess.c b/arch/metag/mm/maccess.c
new file mode 100644
index 000000000000..eba2cfc935b1
--- /dev/null
+++ b/arch/metag/mm/maccess.c
@@ -0,0 +1,68 @@
+/*
+ * safe read and write memory routines callable while atomic
+ *
+ * Copyright 2012 Imagination Technologies
+ */
+
+#include <linux/uaccess.h>
+#include <asm/io.h>
+
+/*
+ * The generic probe_kernel_write() uses the user copy code which can split the
+ * writes if the source is unaligned, and repeats writes to make exceptions
+ * precise. We override it here to avoid these things happening to memory mapped
+ * IO memory where they could have undesired effects.
+ * Due to the use of CACHERD instruction this only works on Meta2 onwards.
+ */
+#ifdef CONFIG_METAG_META21
+long probe_kernel_write(void *dst, const void *src, size_t size)
+{
+	unsigned long ldst = (unsigned long)dst;
+	void __iomem *iodst = (void __iomem *)dst;
+	unsigned long lsrc = (unsigned long)src;
+	const u8 *psrc = (u8 *)src;
+	unsigned int pte, i;
+	u8 bounce[8] __aligned(8);
+
+	if (!size)
+		return 0;
+
+	/* Use the write combine bit to decide is the destination is MMIO. */
+	pte = __builtin_meta2_cacherd(dst);
+
+	/* Check the mapping is valid and writeable. */
+	if ((pte & (MMCU_ENTRY_WR_BIT | MMCU_ENTRY_VAL_BIT))
+	    != (MMCU_ENTRY_WR_BIT | MMCU_ENTRY_VAL_BIT))
+		return -EFAULT;
+
+	/* Fall back to generic version for cases we're not interested in. */
+	if (pte & MMCU_ENTRY_WRC_BIT	|| /* write combined memory */
+	    (ldst & (size - 1))		|| /* destination unaligned */
+	    size > 8			|| /* more than max write size */
+	    (size & (size - 1)))	   /* non power of 2 size */
+		return __probe_kernel_write(dst, src, size);
+
+	/* If src is unaligned, copy to the aligned bounce buffer first. */
+	if (lsrc & (size - 1)) {
+		for (i = 0; i < size; ++i)
+			bounce[i] = psrc[i];
+		psrc = bounce;
+	}
+
+	switch (size) {
+	case 1:
+		writeb(*psrc, iodst);
+		break;
+	case 2:
+		writew(*(const u16 *)psrc, iodst);
+		break;
+	case 4:
+		writel(*(const u32 *)psrc, iodst);
+		break;
+	case 8:
+		writeq(*(const u64 *)psrc, iodst);
+		break;
+	}
+	return 0;
+}
+#endif