Re: [PATCH 1/1] aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock()

[Oleg Nesterov <oleg@redhat.com>]

On 04/30, Mateusz Guzik wrote:
>
> On Tue, Apr 29, 2014 at 05:22:22PM -0400, Benjamin LaHaise wrote:
> > On Tue, Apr 29, 2014 at 04:42:17PM -0400, Benjamin LaHaise wrote:
> > > > Signed-off-by: Oleg Nesterov <oleg@redhat.com>
> > >
> > > Your patch does not apply because it is whitespace damaged.  Please resend
> > > and verify that it applies with 'git am'.
> >
> > Whoops, it's not whitespace damange, but rather that it doesn't apply with
> > the other changes that are queued up in the aio-next tree.  You can find a
> > copy of that tree at git://git.kvack.org/~bcrl/aio-next.git .  The change
> > that conflicts is an additional parameter to kill_ioctx().
>
> While here is there any reason for:
> rcu_assign_pointer(mm->ioctx_table, NULL);
>
> Nothing looks at this pointer afterwards and mm is about to be freed.

Yees, and initially I was going to remove it, but this needs "3" in the
changelog and I am lazy ;)

> I thought it would be used to sanity check that everything was cleared
> before freeing,

Actually, ->mm_count can be > 1, we do not know if we are going to free
this mm_struct or not.

We do not really need to nullify this ptr anyway, but perhaps it makes
sense to keep this initialization anyway, I dunno. But I'll change it to
use RCU_INIT_POINTER(NULL) in v2.

Oleg.