Re: [RFC bpf-next 01/10] uprobe: Add session callbacks to uprobe_consumer

[Andrii Nakryiko]

On Thu, Jun 6, 2024 at 9:46 AM Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
>
> On Wed, Jun 05, 2024 at 10:50:11PM +0200, Jiri Olsa wrote:
> > On Wed, Jun 05, 2024 at 07:56:19PM +0200, Oleg Nesterov wrote:
> > > On 06/05, Andrii Nakryiko wrote:
> > > >
> > > > so any such
> > > > limitations will cause problems, issue reports, investigation, etc.
> > >
> > > Agreed...
> > >
> > > > As one possible solution, what if we do
> > > >
> > > > struct return_instance {
> > > >     ...
> > > >     u64 session_cookies[];
> > > > };
> > > >
> > > > and allocate sizeof(struct return_instance) + 8 *
> > > > <num-of-session-consumers> and then at runtime pass
> > > > &session_cookies[i] as data pointer to session-aware callbacks?
> > >
> > > I too thought about this, but I guess it is not that simple.
> > >
> > > Just for example. Suppose we have 2 session-consumers C1 and C2.
> > > What if uprobe_unregister(C1) comes before the probed function
> > > returns?
> > >
> > > We need something like map_cookie_to_consumer().
> >
> > I guess we could have hash table in return_instance that gets 'consumer -> cookie' ?
>
> ok, hash table is probably too big for this.. I guess some solution that
> would iterate consumers and cookies made sure it matches would be fine
>

Yes, I was hoping to avoid hash tables for this, and in the common
case have no added overhead.

> jirka
>
> >
> > return instance is freed after the consumers' return handlers are executed,
> > so there's no leak if some consumer gets unregistered before that
> >
> > >
> > > > > +       /* The handler_session callback return value controls execution of
> > > > > +        * the return uprobe and ret_handler_session callback.
> > > > > +        *  0 on success
> > > > > +        *  1 on failure, DO NOT install/execute the return uprobe
> > > > > +        *    console warning for anything else
> > > > > +        */
> > > > > +       int (*handler_session)(struct uprobe_consumer *self, struct pt_regs *regs,
> > > > > +                              unsigned long *data);
> > > > > +       int (*ret_handler_session)(struct uprobe_consumer *self, unsigned long func,
> > > > > +                                  struct pt_regs *regs, unsigned long *data);
> > > > > +
> > > >
> > > > We should try to avoid an alternative set of callbacks, IMO. Let's
> > > > extend existing ones with `unsigned long *data`,
> > >
> > > Oh yes, agreed.
> > >
> > > And the comment about the return value looks confusing too. I mean, the
> > > logic doesn't differ from the ret-code from ->handler().
> > >
> > > "DO NOT install/execute the return uprobe" is not true if another
> > > non-session-consumer returns 0.
> >
> > well they are meant to be exclusive, so there'd be no other non-session-consumer
> >
> > jirka