Re: [PATCH v5] pidfd: add ioctl to retrieve pid info

[Christian Brauner]

On Sun, Oct 06, 2024 at 10:33:08PM GMT, Luca Boccassi wrote:
> On Sun, 6 Oct 2024 at 19:56, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> >
> > On 10/06, Luca Boccassi wrote:
> > >
> > > I see, so what should I do here then? Check both? Or none?
> >
> > I don't know, because I don't know how are you going to use this API.
> >
> > > The caller
> > > needs to verify that the data is still valid at the point they use it
> > > anyway,
> >
> > So "none" should work fine? Just it should be documented that, say,
> > kinfo.pid can be 0 if we race with the exiting task.
> >
> > Just in case, you can also use lock_task_sighand() || return -ESRCH,
> > this way kinfo.*pid can't be zero. But I don't think this will buy too
> > much, the task can exit right after pidfd_info() returns.
> >
> > Oleg.
> 
> I think what we should do is check if any of those fields was set to
> 0, and return ESRCH in that case. This way either we return a full set
> of metadata that was correct at the time it was taken, or we provide
> nothing and a clear error. We cannot solve the race as you mentioned,
> but I think it is important to avoid providing incomplete information,
> so that either the data is complete or nothing is given back. If the
> information is complete but becomes stale later, that can happen and
> it's ok.

Agree.