[PATCH v7 1/2] arm64/sve: Split TIF_SVE into separate execute and register state flags

[Mark Brown]

On Tue, Feb 09, 2021 at 10:16:27PM +0000, Mark Brown wrote:
> On Tue, Feb 09, 2021 at 05:59:46PM +0000, Dave Martin wrote:

> > > @@ -952,8 +1022,9 @@ void do_sve_acc(unsigned int esr, struct pt_regs *regs)
> > >  	fpsimd_flush_task_state(current);
> > >  
> > >  	fpsimd_to_sve(current);

> > Hmmm, there's a latent bug upstream here: if the WARN() fires then
> > sve_state is not safe to dereference.  But we already did.

> > So perhaps this should have been something like:

> > 	if (!WARN_ON(test_and_set_thread_flag(TIF_SVE)))
> > 		fpsimd_to_sve();

> > This might make sense as a separate Fixes patch to precede the series.

> Yes, that's definitely a separate fix I think.

Actually now I look at this properly I think that we're sufficiently
confused if this happens that it's not worth worrying about avoiding the
conversion.  We already did a sve_alloc() which will have either
allocated SVE state or dereferenced a SVE state pointer that was there
and pointing to freed memory so we *might* be OK and if we're not then
it's too late to prevent anything and we'll exit with TIF_SVE which will
have a high likelyhood of leading to future dereferences even if we stop
this one.  It's not clear to me that we're helping by potentially
leaving stale data around so I'm inclined to leave things as they are.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20210210/cae2ab2e/attachment.sig>