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

From: Jiri Olsa
Date: Wed Jun 05 2024 - 16:19:17 EST


On Wed, Jun 05, 2024 at 06:36:25PM +0200, Oleg Nesterov wrote:
> On 06/05, Oleg Nesterov wrote:
> >
> > On 06/05, Oleg Nesterov wrote:
> > >
> > > > +/*
> > > > + * Make sure all the uprobe consumers have only one type of entry
> > > > + * callback registered (either handler or handler_session) due to
> > > > + * different return value actions.
> > > > + */
> > > > +static int consumer_check(struct uprobe_consumer *curr, struct uprobe_consumer *uc)
> > > > +{
> > > > + if (!curr)
> > > > + return 0;
> > > > + if (curr->handler_session || uc->handler_session)
> > > > + return -EBUSY;
> > > > + return 0;
> > > > +}
> > >
> > > Hmm, I don't understand this code, it doesn't match the comment...
> > >
> > > The comment says "all the uprobe consumers have only one type" but
> > > consumer_check() will always fail if the the 1st or 2nd consumer has
> > > ->handler_session != NULL ?
> > >
> > > Perhaps you meant
> > >
> > > if (!!curr->handler != !!uc->handler)
> > > return -EBUSY;
> > >
> > > ?
> >
> > OK, the changelog says
> >
> > Which means that there can be only single user of a uprobe (inode +
> > offset) when session consumer is registered to it.
> >
> > so the code is correct. But I still think the comment is misleading.
>
> Cough... perhaps it is correct but I am still confused even we forget about
> the comment ;)
>
> OK, uprobe can have a single consumer with ->handler_session != NULL. I guess
> this is because return_instance->data is "global".
>
> So uprobe can have multiple handler_session == NULL consumers before
> handler_session != NULL, but not after ?

ah yea it should have done what's in the comment, so it's missing
the check for handler.. session handlers are meant to be exclusive

thanks,
jirka