On Mon, Jun 22, 2020 at 12:00:48PM -0700, Peter Collingbourne wrote:
> On Mon, Jun 22, 2020 at 10:17 AM Catalin Marinas
> <catalin.marinas@xxxxxxx> wrote:
> > On Mon, Dec 16, 2019 at 09:30:36AM -0800, Peter Collingbourne wrote:
> > > On Mon, Dec 16, 2019 at 6:20 AM Kevin Brodsky <kevin.brodsky@xxxxxxx> wrote:
> > > > In this patch, the default exclusion mask remains 0 (i.e. all tags can be generated).
> > > > After some more discussions, Branislav and I think that it would be better to start
> > > > with the reverse, i.e. all tags but 0 excluded (mask = 0xfe or 0xff).
> > > >
> > > > This should simplify the MTE setup in the early C runtime quite a bit. Indeed, if all
> > > > tags can be generated, doing any heap or stack tagging before the
> > > > PR_SET_TAGGED_ADDR_CTRL prctl() is issued can cause problems, notably because tagged
> > > > addresses could end up being passed to syscalls. Conversely, if IRG and ADDG never
> > > > set the top byte by default, then tagging operations should be no-ops until the
> > > > prctl() is issued. This would be particularly useful given that it may not be
> > > > straightforward for the C runtime to issue the prctl() before doing anything else.
> > > >
> > > > Additionally, since the default tag checking mode is PR_MTE_TCF_NONE, it would make
> > > > perfect sense not to generate tags by default.
> > >
> > > This would indeed allow the early C runtime startup code to pass
> > > tagged addresses to syscalls,
[...]
> > > but I don't think it would entirely free
> > > the code from the burden of worrying about stack tagging. Either way,
> > > any stack frames that are active at the point when the prctl() is
> > > issued would need to be compiled without stack tagging, because
> > > otherwise those stack frames may use ADDG to rematerialize a stack
> > > object address, which may produce a different address post-prctl.
[...]
> > > Setting the exclude mask to 0xffff would at least make it more likely
> > > for this problem to be detected, though.
> >
> > I thought it would be detected if we didn't have a 0xffff default
> > exclude mask. With only tag 0 generated, any such problem could be
> > hidden.
>
> I don't think that's the case, as long as you aren't using 0 as a
> catch-all tag. Imagine that you have some hypothetical startup code
> that looks like this:
>
> void init() {
> bool called_prctl = false;
> prctl(PR_SET_TAGGED_ADDR_CTRL, ...); // effect is to change
> GCR_EL1.Excl from 0xffff to 1
> called_prctl = true;
> }
>
> This may be compiled as something like (well, a real compiler wouldn't
> compile it like this but rather use sp-relative stores or eliminate
> the dead stores entirely, but imagine that the stores to called_prctl
> are obfuscated somehow, e.g. in another translation unit):
>
> sub x19, sp, #16
> irg x19, x19 // compute a tag base for the function
> addg x0, x19, #0, #1 // add tag offset for "called_prctl"
> stzg x0, [x0]
> bl prctl
> addg x0, x19, #0, #1 // rematerialize "called_prctl" address
> mov w1, #1
> strb w1, [x0]
> ret
>
> The first addg will materialize a tag of 0 due to the default Excl
> value, so the stzg will set the memory tag to 0. However, the second
> addg will materialize a tag of 1 because of the new Excl value, which
> will result in a tag fault in the strb instruction.
>
> This problem is less likely to be detected if we transition Excl from
> 0 to 1. It will only be detected in the case where the irg instruction
> produces a tag of 0xf, which would be incremented to 0 by the first
> addg but to 1 by the second one.
Thanks for the explanation. For some reason I thought ADDG would only be
used once (per variable or frame). I now agree that a default exclude
mask of 0xffff would catch such issues early.
> > > If we change the default in this way, maybe it would be worth
> > > considering flipping the meaning of the tag mask and have it be a mask
> > > of tags to allow. That would be consistent with the existing behaviour
> > > where userspace sets bits in tagged_addr_ctrl in order to enable
> > > tagging features.
> >
> > The first question is whether the C runtime requires a default
> > GCR_EL1.Excl mask of 0xffff (or 0xfffe) so that IRG, ADDG, SUBG always
> > generate tag 0. If the runtime is fine with a default exclude mask of 0,
> > I'm tempted to go back to an exclude mask for prctl().
> >
> > (to me it feels more natural to use an exclude mask as it matches the
> > ARM ARM definition but maybe I stare too much at the hardware specs ;))
>
> I think that would be fine with me. With the transition from 0 to 1
> the above problem would still be detected, but only 1/16 of the time.
> But if the problem exists in the early startup code which will be
> executed many times during a typical system boot, it makes it likely
> that the problem will be detected eventually.
I'm not a big fan of hitting a problem 1/16 times, it makes debugging
harder. So I'll stick to a default exclude mask of 0xffff, in which case
it makes sense to invert the polarity for prctl() and make it an include
mask (as in v4 of the patchset).
Thanks.
--
Catalin
