On Tue, Apr 21, 2020 at 03:26:01PM +0100, Catalin Marinas wrote:
> Even if the ID_AA64PFR1_EL1 register advertises the presence of MTE, it
> is not guaranteed that the memory system on the SoC supports the
> feature. In the absence of system-wide MTE support, the behaviour is
> undefined and the kernel should not enable the MTE memory type in
> MAIR_EL1.
>
> For FDT, add an 'arm,armv8.5-memtag' property to the /memory nodes and
> check for its presence during MTE probing. For example:
>
> memory@80000000 {
> device_type = "memory";
> arm,armv8.5-memtag;
> reg = <0x00000000 0x80000000 0 0x80000000>,
> <0x00000008 0x80000000 0 0x80000000>;
> };
>
> If the /memory nodes are not present in DT or if at least one node does
> not support MTE, the feature will be disabled. On EFI systems, it is
> assumed that the memory description matches the EFI memory map (if not,
> it is considered a firmware bug).
>
> MTE is not currently supported on ACPI systems.
>
> Signed-off-by: Catalin Marinas <catalin.marinas@xxxxxxx>
> Cc: Rob Herring <Rob.Herring@xxxxxxx>
> Cc: Mark Rutland <mark.rutland@xxxxxxx>
> Cc: Will Deacon <will@xxxxxxxxxx>
> Cc: Suzuki K Poulose <Suzuki.Poulose@xxxxxxx>
This patch turns out to be incomplete. While it does not expose the
HWCAP2_MTE to user when the above DT property is not present, it still
allows user access to the ID_AA64PFR1_EL1.MTE field (via MRS emulations)
since it is marked as visible.
--
Catalin
