Re: [RFC PATCH v9 for 4.15 01/14] Restartable sequences system call

[Florian Weimer]

On 10/13/2017 07:24 PM, Andy Lutomirski wrote:
> On Fri, Oct 13, 2017 at 7:27 AM, Mathieu Desnoyers
> <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
> > ----- On Oct 13, 2017, at 9:56 AM, Florian Weimer fweimer@xxxxxxxxxx wrote:
> >
> > > On 10/13/2017 03:40 PM, Mathieu Desnoyers wrote:
> > > > The proposed ABI does not require to store any function pointer. For a given
> > > > rseq_finish() critical section, pointers to specific instructions (within a
> > > > function) are emitted at link-time into a struct rseq_cs:
> > > >
> > > > struct rseq_cs {
> > > >           RSEQ_FIELD_u32_u64(start_ip);
> > > >           RSEQ_FIELD_u32_u64(post_commit_ip);
> > > >           RSEQ_FIELD_u32_u64(abort_ip);
> > > >           uint32_t flags;
> > > > } __attribute__((aligned(4 * sizeof(uint64_t))));
> > > >
> > > > Then, at runtime, the fast-path stores the address of that struct rseq_cs
> > > > into the TLS struct rseq "rseq_cs" field.
> > > >
> > > > So all we store at runtime is a pointer to data, not a pointer to functions.
> > > >
> > > > But you seem to hint that having a pointer to data containing pointers to code
> > > > may still be making it easier for exploit writers. Can you elaborate on the
> > > > scenario ?
> > >
> > > I'm concerned that the exploit writer writes a totally made up struct
> > > rseq_cs object into writable memory, along with function pointers, and
> > > puts the address of that in to the rseq_cs field.
> > >
> > > This would be comparable to how C++ vtable pointers are targeted
> > > (including those in the glibc libio implementation of stdio streams).
> > >
> > > Does this answer your questions?
> >
> > Yes, it does. How about we add a "canary" field to the TLS struct rseq, e.g.:
> >
> > struct rseq {
> >          union rseq_cpu_event u;
> >          RSEQ_FIELD_u32_u64(rseq_cs);  -> pointer to struct rseq_cs
> >          uint32_t flags;
> >          uint32_t canary;   -> 32 low bits of rseq_cs ^ canary_mask
> > };
> >
> > We could then add a "uint32_t canary_mask" argument to sys_rseq, e.g.:
> >
> > SYSCALL_DEFINE3(rseq, struct rseq __user *, rseq, uint32_t, canary_mask, int, flags);
> >
> > So a thread which does not care about hardening would simply register its
> > struct rseq TLS with a canary mask of "0". Nothing changes on the fast-path.
> >
> > A thread belonging to a process that cares about hardening could use a random
> > value as canary, and pass it as canary_mask argument to the syscall. The
> > fast-path could then set the struct rseq "canary" value to
> > (32-low-bits of rseq_cs) ^ canary_mask just surrounding the critical section,
> > and set it back to 0 afterward.
> >
> > In the kernel, whenever the rseq_cs pointer would be loaded, its 32 low bits
> > would be checked to match (canary ^ canary_mask). If it differs, then the
> > kernel kills the process with SIGSEGV.
> >
> > Would that take care of your concern ?
>
> I would propose a slightly different solution: have the kernel verify
> that it jumps to a code sequence that occurs just after some
> highly-unlikely magic bytes in the text *and* that those bytes have
> some signature that matches a signature in the struct rseq that's
> passed in.

And the signature is fixed at the time of the rseq syscall?

Yes, that would be far more reliable.

Thanks,
Florian