Re: [PATCH 2/8] uprobes: revamp uprobe refcounting and lifetime management

[Oleg Nesterov <oleg@redhat.com>]

On 08/02, Andrii Nakryiko wrote:
>
> On Fri, Aug 2, 2024 at 1:50 AM Oleg Nesterov <oleg@redhat.com> wrote:
> >
> > On 08/01, Andrii Nakryiko wrote:
> > >
> > > > +               /* TODO : cant unregister? schedule a worker thread */
> > > > +               WARN(err, "leaking uprobe due to failed unregistration");
> >
> > > Ok, so now that I added this very loud warning if
> > > register_for_each_vma(uprobe, NULL) returns error, it turns out it's
> > > not that unusual for this unregistration to fail.
> >
> > ...
> >
> > > So, is there something smarter we can do in this case besides leaking
> > > an uprobe (and note, my changes don't change this behavior)?
> >
> > Something like schedule_work() which retries register_for_each_vma()...
>
> And if that fails again, what do we do?

try again after some timeout ;)

> Because I don't think we even
> need schedule_work(), we can just keep some list of "pending to be
> retried" items and check them after each
> uprobe_register()/uprobe_unregister() call.

Agreed, we need a list of "pending to be retried", but rightly or not
I think this should be done from work_struct->func.

Lets discuss this later? We seem to agree this is a known problem, and
this has nothing to do with your optimizations.

> I'm just not clear how we
> should handle stubborn cases (but honestly I haven't even tried to
> understand all the details about this just yet).

Same here.

Oleg.