Re: [PATCH 2/2] uswsusp: build only when configured

[Domenico Andreoli <domenico.andreoli@linux.com>]

On Wed, Apr 29, 2020 at 01:20:53PM +0200, Rafael J. Wysocki wrote:
> On Mon, Apr 27, 2020 at 11:48 AM Domenico Andreoli <domenico.andreoli@linux.com> wrote:
> >
> > On Sun, Apr 26, 2020 at 06:16:29PM +0200, Rafael J. Wysocki wrote:
> > >

[...]

> > >
> > > It is possible in theory that two processes write "disk" to /sys/power/state
> > > concurrently.
> > >
> > > Is there enough mutual exclusion in place to handle this gracefully after the
> > > above change?
> >
> > No, indeed.
> >
> > It looks like hibernate.c needs the mutual exclusion and user.c could
> > just use it. Should I move snapshot_device_available to hibernate.c
> > and rename it hibernate_available?
> 
> There is hibernation_available() already.
> 
> Maybe switch over to the refcount_t API, call the variable
> hibernate_refcount and use refcount_add_not_zero() on it for the
> mutual exclusion.

I'm doing as you ask but I'm not understanding what we actually gain
from using the refcount_t API.

I'm reading about relaxation of memory ordering and there is no mention
on what this implies for the add_not_zero operation that we use.

Honestly I would stick with the atomic_t, we just need the mutual
exclusion and no saturation semantics.

Dom

-- 
rsa4096: 3B10 0CA1 8674 ACBA B4FE  FCD2 CE5B CF17 9960 DE13
ed25519: FFB4 0CC3 7F2E 091D F7DA  356E CC79 2832 ED38 CB05