[RFC PATCH 1/3] signal: Ensure every siginfo we send has all bits initialized
Dave Martin
Dave.Martin at arm.com
Thu Apr 19 18:26:57 AEST 2018
On Wed, Apr 18, 2018 at 09:22:09AM -0500, Eric W. Biederman wrote:
> Dave Martin <Dave.Martin at arm.com> writes:
>
> > On Tue, Apr 17, 2018 at 02:37:38PM -0500, Eric W. Biederman wrote:
[...]
> >> My intention is to leave 0 instances of clear_siginfo in the
> >> architecture specific code. Ideally struct siginfo will be limited to
> >> kernel/signal.c but I am not certain I can quite get that far.
> >> The function do_coredump appears to have a legit need for siginfo.
> >
> > So, you mean we can't detect that the caller didn't initialise all the
> > members, or initialised the wrong union member?
>
> Correct. Even when we smuggled the the union member in the upper bits
> of si_code we got it wrong. So an interface that helps out and does
> more and is harder to misues looks desirable.
>
> > What would be the alternative? Have a separate interface for each SIL_
> > type, with only kernel/signal.c translating that into the siginfo_t that
> > userspace sees?
>
> Yes. It really isn't bad as architecture specific code only generates
> faults. In general faults only take a pointer. I have already merged
> the needed helpers into kernel/signal.c
Good point. I hadn't considered that only one class of signal comes
from the arch code, but now that you point it out, it sounds right.
> > Either way, I don't see how we force the caller to initilise the whole
> > structure.
>
> In general the plan is to convert the callers to call force_sig_fault,
> and then there is no need to have siginfo in the architecture specific
> code. I have all of the necessary helpers are already merged into
> kernel/signal.c
Makes sense.
I wonder if all the relevant siginfo data could be passed to
force_sig_fault() (or whatever) as arguments. Then the problem of
uninitialised fields goes away. Perhaps that's what you had in mind.
[...]
> >> Unless gcc has changed it's stance on type-punning through unions
> >> or it's semantics with -fno-strict_aliasing we should be good.
> >
> > In practice you're probably right.
> >
> > Today, gcc is pretty conservative in this area, and I haven't been able
> > to convince clang to optimise away memset in this way either.
> >
> > My concern is that is this assumption turns out to be wrong it may be
> > some time before anybody notices, because the leakage of kernel stack may
> > be the only symptom.
> >
> > I'll try to nail down a compiler guy to see if we can get a promise on
> > this at least with -fno-strict-aliasing.
> >
> >
> > I wonder whether it's worth protecting ourselves with something like:
> >
> >
> > static void clear_siginfo(siginfo_t *si)
> > {
> > asm ("" : "=m" (*si));
> > memset(si, 0, sizeof(*si));
> > asm ("" : "+m" (*si));
> > }
> >
> > Probably needs to be thought about more widely though. I guess it's out
> > of scope for this series.
>
> It is definitely a question worth asking.
I may follow it up later if I find myself at a loose end...
Cheers
---Dave
More information about the Linuxppc-dev
mailing list