Re: [PATCH 3/8] arm: Fix possible race on task->mm

[Oleg Nesterov <oleg@redhat.com>]

On 02/09, Anton Vorontsov wrote:
>
> On Thu, Feb 09, 2012 at 07:33:46PM +0400, Anton Vorontsov wrote:
> > On Wed, Feb 08, 2012 at 05:08:08PM +0100, Oleg Nesterov wrote:
> > > On 02/07, Anton Vorontsov wrote:
> > > >
> > > > Checking for task->mm is dangerous as ->mm might disappear (exit_mm()
> > > > assigns NULL under task_lock(), so tasklist lock is not enough).
> > > >
> > > > We can't use get_task_mm()/mmput() pair as mmput() might sleep,
> > >
> > > Yes, but
> > >
> > > > so let's take the task lock while we care about its mm.
> > >
> > > it seems that this needs find_lock_task_mm() too ?
> > >
> > > the same for the rest patches in this series.
> >
> > Yep, I think you're right, will add this change.
>
> Thinking about it more... making the code use find_lock_task_mm
> would be a behaviour change. Sure, in trivial cases like ARM this
> looks like a 100% safe thing to do, but in e.g. UML case, I
> wouldn't bet much money on that 'mm->context.id.u.pid' would be
> still meaningful.

OK, perhaps UML differs. I don't know what context.id.u.pid means.
Although at first glance it would be meaningful anyway...

> So, I'd rather do it in a separate change, so this can be easily
> reverted.

In the !UML case find_lock_task_mm() "obviously looks like the right
thing...

But I won't argue, up to you.

Oleg.