Re: [PATCHv2 bpf-next 1/9] uprobe: Add support for session consumer

[Jiri Olsa <olsajiri@gmail.com>]

On Tue, Jul 02, 2024 at 01:52:38PM -0700, Andrii Nakryiko wrote:
> On Tue, Jul 2, 2024 at 9:11 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> >
> > On Tue, Jul 02, 2024 at 03:04:08PM +0200, Peter Zijlstra wrote:
> > > On Mon, Jul 01, 2024 at 06:41:07PM +0200, Jiri Olsa wrote:
> > >
> > > > +static void
> > > > +uprobe_consumer_account(struct uprobe *uprobe, struct uprobe_consumer *uc)
> > > > +{
> > > > +   static unsigned int session_id;
> > > > +
> > > > +   if (uc->session) {
> > > > +           uprobe->sessions_cnt++;
> > > > +           uc->session_id = ++session_id ?: ++session_id;
> > > > +   }
> > > > +}
> > >
> > > The way I understand this code, you create a consumer every time you do
> > > uprobe_register() and unregister makes it go away.
> > >
> > > Now, register one, then 4g-1 times register+unregister, then register
> > > again.
> > >
> > > The above seems to then result in two consumers with the same
> > > session_id, which leads to trouble.
> > >
> > > Hmm?
> >
> > ugh true.. will make it u64 :)
> >
> > I think we could store uprobe_consumer pointer+ref in session_consumer,
> > and that would make the unregister path more interesting.. will check
> 
> More interesting how? It's actually a great idea, uprobe_consumer

nah, got confused ;-)

> pointer itself is a unique ID and 64-bit. We can still use lowest bit
> for RC (see my other reply).

I used pointers in the previous version, but then I thought what if the
consumer gets free-ed and new one created (with same address.. maybe not
likely but possible, right?) before the return probe is hit

jirka