	         MEMORY ATTRIBUTE ALIASING ON IA-64

			   Bjorn Helgaas
		       <bjorn.helgaas@hp.com>
			    May 4, 2006


MEMORY ATTRIBUTES

    Itanium supports several attributes for virtual memory references.
    The attribute is part of the virtual translation, i.e., it is
    contained in the TLB entry.  The ones of most interest to the Linux
    kernel are:

	WB		Write-back (cacheable)
	UC		Uncacheable
	WC		Write-coalescing

    System memory typically uses the WB attribute.  The UC attribute is
    used for memory-mapped I/O devices.  The WC attribute is uncacheable
    like UC is, but writes may be delayed and combined to increase
    performance for things like frame buffers.

    The Itanium architecture requires that we avoid accessing the same
    page with both a cacheable mapping and an uncacheable mapping[1].

    The design of the chipset determines which attributes are supported
    on which regions of the address space.  For example, some chipsets
    support either WB or UC access to main memory, while others support
    only WB access.

MEMORY MAP

    Platform firmware describes the physical memory map and the
    supported attributes for each region.  At boot-time, the kernel uses
    the EFI GetMemoryMap() interface.  ACPI can also describe memory
    devices and the attributes they support, but Linux/ia64 currently
    doesn't use this information.

    The kernel uses the efi_memmap table returned from GetMemoryMap() to
    learn the attributes supported by each region of physical address
    space.  Unfortunately, this table does not completely describe the
    address space because some machines omit some or all of the MMIO
    regions from the map.

    The kernel maintains another table, kern_memmap, which describes the
    memory Linux is actually using and the attribute for each region.
    This contains only system memory; it does not contain MMIO space.

    The kern_memmap table typically contains only a subset of the system
    memory described by the efi_memmap.  Linux/ia64 can't use all memory
    in the system because of constraints imposed by the identity mapping
    scheme.

    The efi_memmap table is preserved unmodified because the original
    boot-time information is required for kexec.

KERNEL IDENTITY MAPPINGS

    Linux/ia64 identity mappings are done with large pages, currently
    either 16MB or 64MB, referred to as "granules."  Cacheable mappings
    are speculative[2], so the processor can read any location in the
    page at any time, independent of the programmer's intentions.  This
    means that to avoid attribute aliasing, Linux can create a cacheable
    identity mapping only when the entire granule supports cacheable
    access.

    Therefore, kern_memmap contains only full granule-sized regions that
    can referenced safely by an identity mapping.

    Uncacheable mappings are not speculative, so the processor will
    generate UC accesses only to locations explicitly referenced by
    software.  This allows UC identity mappings to cover granules that
    are only partially populated, or populated with a combination of UC
    and WB regions.

USER MAPPINGS

    User mappings are typically done with 16K or 64K pages.  The smaller
    page size allows more flexibility because only 16K or 64K has to be
    homogeneous with respect to memory attributes.

POTENTIAL ATTRIBUTE ALIASING CASES

    There are several ways the kernel creates new mappings:

    mmap of /dev/mem

	This uses remap_pfn_range(), which creates user mappings.  These
	mappings may be either WB or UC.  If the region being mapped
	happens to be in kern_memmap, meaning that it may also be mapped
	by a kernel identity mapping, the user mapping must use the same
	attribute as the kernel mapping.

	If the region is not in kern_memmap, the user mapping should use
	an attribute reported as being supported in the EFI memory map.

	Since the EFI memory map does not describe MMIO on some
	machines, this should use an uncacheable mapping as a fallback.

    mmap of /sys/class/pci_bus/.../legacy_mem

	This is very similar to mmap of /dev/mem, except that legacy_mem
	only allows mmap of the one megabyte "legacy MMIO" area for a
	specific PCI bus.  Typically this is the first megabyte of
	physical address space, but it may be different on machines with
	several VGA devices.

	"X" uses this to access VGA frame buffers.  Using legacy_mem
	rather than /dev/mem allows multiple instances of X to talk to
	different VGA cards.

	The /dev/mem mmap constraints apply.

	However, since this is for mapping legacy MMIO space, WB access
	does not make sense.  This matters on machines without legacy
	VGA support: these machines may have WB memory for the entire
	first megabyte (or even the entire first granule).

	On these machines, we could mmap legacy_mem as WB, which would
	be safe in terms of attribute aliasing, but X has no way of
	knowing that it is accessing regular memory, not a frame buffer,
	so the kernel should fail the mmap rather than doing it with WB.

    read/write of /dev/mem

	This uses copy_from_user(), which implicitly uses a kernel
	identity mapping.  This is obviously safe for things in
	kern_memmap.

	There may be corner cases of things that are not in kern_memmap,
	but could be accessed this way.  For example, registers in MMIO
	space are not in kern_memmap, but could be accessed with a UC
	mapping.  This would not cause attribute aliasing.  But
	registers typically can be accessed only with four-byte or
	eight-byte accesses, and the copy_from_user() path doesn't allow
	any control over the access size, so this would be dangerous.

    ioremap()

	This returns a kernel identity mapping for use inside the
	kernel.

	If the region is in kern_memmap, we should use the attribute
	specified there.  Otherwise, if the EFI memory map reports that
	the entire granule supports WB, we should use that (granules
	that are partially reserved or occupied by firmware do not appear
	in kern_memmap).  Otherwise, we should use a UC mapping.

PAST PROBLEM CASES

    mmap of various MMIO regions from /dev/mem by "X" on Intel platforms

      The EFI memory map may not report these MMIO regions.

      These must be allowed so that X will work.  This means that
      when the EFI memory map is incomplete, every /dev/mem mmap must
      succeed.  It may create either WB or UC user mappings, depending
      on whether the region is in kern_memmap or the EFI memory map.

    mmap of 0x0-0xA0000 /dev/mem by "hwinfo" on HP sx1000 with VGA enabled

      See https://bugzilla.novell.com/show_bug.cgi?id=140858.

      The EFI memory map reports the following attributes:
        0x00000-0x9FFFF WB only
        0xA0000-0xBFFFF UC only (VGA frame buffer)
        0xC0000-0xFFFFF WB only

      This mmap is done with user pages, not kernel identity mappings,
      so it is safe to use WB mappings.

      The kernel VGA driver may ioremap the VGA frame buffer at 0xA0000,
      which will use a granule-sized UC mapping covering 0-0xFFFFF.  This
      granule covers some WB-only memory, but since UC is non-speculative,
      the processor will never generate an uncacheable reference to the
      WB-only areas unless the driver explicitly touches them.

    mmap of 0x0-0xFFFFF legacy_mem by "X"

      If the EFI memory map reports this entire range as WB, there
      is no VGA MMIO hole, and the mmap should fail or be done with
      a WB mapping.

      There's no easy way for X to determine whether the 0xA0000-0xBFFFF
      region is a frame buffer or just memory, so I think it's best to
      just fail this mmap request rather than using a WB mapping.  As
      far as I know, there's no need to map legacy_mem with WB
      mappings.

      Otherwise, a UC mapping of the entire region is probably safe.
      The VGA hole means the region will not be in kern_memmap.  The
      HP sx1000 chipset doesn't support UC access to the memory surrounding
      the VGA hole, but X doesn't need that area anyway and should not
      reference it.

    mmap of 0xA0000-0xBFFFF legacy_mem by "X" on HP sx1000 with VGA disabled

      The EFI memory map reports the following attributes:
        0x00000-0xFFFFF WB only (no VGA MMIO hole)

      This is a special case of the previous case, and the mmap should
      fail for the same reason as above.

NOTES

    [1] SDM rev 2.2, vol 2, sec 4.4.1.
    [2] SDM rev 2.2, vol 2, sec 4.4.6.