[PATCH v2] mte: Initialize tag storage to KASAN_TAG_INVALID

[Will Deacon]

On Mon, Nov 07, 2022 at 04:47:14PM +0000, Vincenzo Frascino wrote:
> Hi Will,
> 
> On 11/7/22 15:19, Will Deacon wrote:
> > On Wed, Sep 07, 2022 at 12:00:15PM +0100, Vincenzo Frascino wrote:
> >> When the kernel is entered on aarch64, the MTE allocation tags are in an
> >> UNKNOWN state.
> >>
> >> With MTE enabled, the tags are initialized:
> >>  - When a page is allocated and the user maps it with PROT_MTE.
> >>  - On allocation, with in-kernel MTE enabled (HW_TAGS KASAN).
> >>
> >> If the tag pool is zeroed by the hardware at reset, it makes it
> >> difficult to track potential places where the initialization of the
> >> tags was missed.
> >>
> >> This can be observed under QEMU for aarch64, which initializes the MTE
> >> allocation tags to zero.
> >>
> >> Initialize to tag storage to KASAN_TAG_INVALID to catch potential
> >> places where the initialization of the tags was missed.
> >>
> >> This is done introducing a new kernel command line parameter
> >> "mte.tags_init" that enables the debug option.
> >>
> >> Note: The proposed solution should be considered a debug option because
> >> it might have performance impact on large machines at boot.
> >>
> >> Cc: Catalin Marinas <catalin.marinas at arm.com>
> >> Cc: Will Deacon <will at kernel.org>
> >> Signed-off-by: Vincenzo Frascino <vincenzo.frascino at arm.com>
> >> ---
> >>  arch/arm64/kernel/mte.c | 47 +++++++++++++++++++++++++++++++++++++++++
> >>  1 file changed, 47 insertions(+)
> > 
> > I don't really see the point in this change -- who is going to use this
> > option?
> > 
> 
> I think this option can be useful to someone who is trying to debug a problem
> that is related to a missed tag initialization and it is doing it on QEMU.
> 
> QEMU by default would mask this class of problems because it initializes to zero
> the tags at "reset" (which is a valid UNKNOWN STATE according to the architecture).
> 
> I noticed this behavior because I was trying to debug a similar issue which I
> was able to reproduce only on FVP.
> 
> Said that, I originally posted this patch as RFC back in April this year to find
> out if someone else would find it useful, in fact my idea was to keep it locally.
> 
> Please let me know what do you want to do.

I'd prefer to leave the code as-is until we have a concrete ask for this
feature.

Will