Re: [RFC PATCH for 4.18] rseq: use __u64 for rseq_cs fields, validate user inputs

[Peter Zijlstra]

On Tue, Jul 03, 2018 at 10:29:55AM +0200, Heiko Carstens wrote:
> On Tue, Jul 03, 2018 at 10:14:49AM +0200, Peter Zijlstra wrote:
> > On Mon, Jul 02, 2018 at 10:30:09PM -0400, Mathieu Desnoyers wrote:
> > > > Use "get_user()". It works for 64-bit objects too, and it will be
> > > > atomic in the 32-bit sub-parts on a 32-bit architecture.
> > > 
> > > Is it really ? Last time we had this discussion, not all architectures
> > > guaranteed that reading a 64-bit integer would happen in two atomic
> > > 32-bit sub-parts. This was the main motivation for the LINUX_FIELD_u32_u64()
> > > macro as it stands today (rather than using a union).
> > 
> > Just state, as a requirement for supporting rseq, that the arch
> > {get,put}_user(u64) on 32bit targets must be exactly 2 u32 loads/stores.
> > 
> > We're piece-wise enabling rseq across architectures anyway, and when the
> > relevant maintains do this, they can have a look at their
> > {get,put}_user() implementations and fix them.
> > 
> > If you rely on get_user(u64) working, that means microblaze is already
> > broken, but I suppose it already was, since their rseq enablement patch
> > is extremely dodgy. Michal?
> 
> s390 uses the mvcos instruction to implement get_user(). That instruction
> is not defined to be atomic, but may copy bytes piecemeal.. I had the
> impression that the rseq fields are supposed to be updated within the
> context of a single thread (user + kernel space).
> 
> However if another user space thread is allowed to do this as well, then
> the get_user() approach won't fly on s390.
> 
> That leaves the question: does it even make sense for a thread to update
> the rseq structure of a different thread?

The problem is interrupts; we need interrupts on the CPU doing the store
to observe either the old or the new value, not a mix.

If mvcos does not guarantee that, we're having problems. Is there a
reason get_user() cannot use a 'regular' load?