[PATCH v5 19/25] arm64: mte: Add PTRACE_{PEEK,POKE}MTETAGS support

Catalin Marinas catalin.marinas at arm.com
Fri Jul 3 09:18:37 EDT 2020


On Wed, Jul 01, 2020 at 02:32:43PM -0300, Luis Machado wrote:
> On 7/1/20 2:16 PM, Catalin Marinas wrote:
> > On Thu, Jun 25, 2020 at 02:10:10PM -0300, Luis Machado wrote:
> > > I have one point below I wanted to clarify regarding
> > > PEEKMTETAGS/POKEMTETAGS.
> > > 
> > > But before that, I've pushed v2 of the MTE series for GDB here:
> > > 
> > > https://sourceware.org/git/?p=binutils-gdb.git;a=shortlog;h=refs/heads/users/luisgpm/aarch64-mte-v2
> > > 
> > > That series adds sctlr and gcr registers to the NT_ARM_MTE (still using a
> > > dummy value of 0x407) register set. It would be nice if the Linux Kernel and
> > > the debuggers were in sync in terms of supporting this new register set. GDB
> > > assumes the register set exists if HWCAP2_MTE is there.
> > > 
> > > So, if we want to adjust the register set, we should probably consider doing
> > > that now. That prevents the situation where debuggers would need to do
> > > another check to confirm NT_ARM_MTE is exported. I'd rather avoid that.
> > 
> > I'm happy to do this before merging, though we need to agree on the
> > semantics.
> > 
> > Do you need both read and write access? Also wondering whether the
> 
> If I recall the previous discussion correctly, Kevin thought access to both
> of these would be interesting to the user. It sounded like having read-only
> access was enough. If so,...
> 
> > prctl() value would be a better option than the raw register bits (well,
> > not entirely raw, masking out the irrelevant part).
> 
> ... then exposing the most useful bits to the user would be better, and up
> to you to define.
> 
> I can tweak the GDB patches to turn the sctlr and gcr values into flag
> fields. Then GDB can just show those in a more meaningful way. I just need
> to know what the bits would look like.

We may have some software only behaviour added to these bits at some
point (e.g. deliver signal on return from syscall for faults on the
uaccess routines). They would not be represented in the SCTLR/GCR
registers.

> I'd rather not make these values writable if we don't think there is a good
> use case for it. Better avoid giving developers more knobs than they need?

There's the CRIU use-case for restoring this but I don't think we do it
for other prctl() controls.

-- 
Catalin



More information about the linux-arm-kernel mailing list