On Thu, Nov 19, 2020 at 06:27:50PM +0000, Dave Martin wrote:
> On Wed, Nov 18, 2020 at 05:52:01PM +0000, Mark Brown wrote:
> > On Wed, Nov 18, 2020 at 12:51:56PM +0000, Dave Martin wrote:

> > > I think I missed this point.  It doesn't sound quite right to me: how
> > > will we ever turn SVE persistently off for a task?

> > As mentioned in the cover letter (which I inherited from Julien, it's
> > been this way since v1) we don't currently, the proposal mentioned in
> > the cover letter is to turn it off after some number of syscalls.

> I think there are two levels of this:

> 1) Basically the same behaviour as today, but optimising the fast path
> of a non-scheduling syscall to avoid dumping and reloading the regs,
> while not making other things worse (i.e., we still want SVE to get
> turned off when appropriate).

> 2) Introducing some logic to try to make an educated guess about whether
> SVE should be on or off.

> It wasn't clear that (2) would really be any better in practice than
> static logic -- after all, other arches have adopted such a thing and
> then subsequently dumped it -- but I don't have a strong argument
> against it.

> I guess I'd prefer to see this arbitrary (if straightforward) policy as
> a second patch, separate from the shovelwork in (1).

Hrm, right.  I guess part of it here is that both behaviours are equally
arbatrary and the leaving it on behaviour just falls out of the rest of
the change in the same way that the existing behaviour just falls out of
the current implementation.  I suspect given the limited availability of
SVE hardware at present and some educated guess as to what it might be
used for that current systems will struggle to notice either way but
that'll change as the feature makes it's way into a wider range of
systems.  I think the clearest thing will be to leave this initial
change with similar behaviour to what it currently has and add another
patch with a policy for turning SVE off to the series, that'll separate
things in the review if nothing else.