[PATCH] bpf: restore the ebpf audit UNLOAD id field

[Jiri Olsa]

On Fri, Dec 23, 2022 at 10:58:37AM -0500, Paul Moore wrote:
> On Fri, Dec 23, 2022 at 10:37 AM Paul Moore <paul at paul-moore.com> wrote:
> > On Thu, Dec 22, 2022 at 6:20 PM Jiri Olsa <olsajiri at gmail.com> wrote:
> > > On Thu, Dec 22, 2022 at 02:03:41PM -0500, Paul Moore wrote:
> > > > On Thu, Dec 22, 2022 at 12:19 PM <sdf at google.com> wrote:
> > > > > On 12/21, Paul Moore wrote:
> > > > > > When changing the ebpf program put() routines to support being called
> > > > > > from within IRQ context the program ID was reset to zero prior to
> > > > > > generating the audit UNLOAD record, which obviously rendered the ID
> > > > > > field bogus (always zero).  This patch resolves this by adding a new
> > > > > > field, bpf_prog_aux::id_audit, which is set when the ebpf program is
> > > > > > allocated an ID and never reset, ensuring a valid ID field,
> > > > > > regardless of the state of the original ID field, bpf_prox_aud::id.
> > > > >
> > > > > > I also modified the bpf_audit_prog() logic used to associate the
> > > > > > AUDIT_BPF record with other associated records, e.g. @ctx != NULL.
> > > > > > Instead of keying off the operation, it now keys off the execution
> > > > > > context, e.g. '!in_irg && !irqs_disabled()', which is much more
> > > > > > appropriate and should help better connect the UNLOAD operations with
> > > > > > the associated audit state (other audit records).
> > > > >
> > > > > [..]
> > > > >
> > > > > > As an note to future bug hunters, I did briefly consider removing the
> > > > > > ID reset in bpf_prog_free_id(), as it would seem that once the
> > > > > > program is removed from the idr pool it can no longer be found by its
> > > > > > ID value, but commit ad8ad79f4f60 ("bpf: offload: free program id
> > > > > > when device disappears") seems to imply that it is beneficial to
> > > > > > reset the ID value.  Perhaps as a secondary indicator that the ebpf
> > > > > > program is unbound/orphaned.
> > > > >
> > > > > That seems like the way to go imho. Can we have some extra 'invalid_id'
> > > > > bitfield in the bpf_prog so we can set it in bpf_prog_free_id and
> > > > > check in bpf_prog_free_id (for this offloaded use-case)? Because
> > > > > having two ids and then keeping track about which one to use, depending
> > > > > on the context, seems more fragile?
> > > >
> > > > I would definitely prefer to keep just a single ID value, and that was
> > > > the first approach I took when drafting this patch, but when looking
> > > > through the git log it looked like there was some desire to reset the
> > > > ID to zero on free.  Not being an expert on the ebpf kernel code I
> > > > figured I would just write the patch up this way and make a comment
> > > > about not zero'ing out the ID in the commit description so we could
> > > > have a discussion about it.
> > > >
> > > > I'm not seeing any other comments, so I'll go ahead with putting
> > > > together a v2 that sets an invalid flag/bit and I'll post that for
> > > > further discussion/review.
> > >
> > > great, perf suffers the same issue:
> > >   https://lore.kernel.org/bpf/Y3SRWVoycV290S16@krava/
> > >
> > > any chance you could include it as well? I can send a patch
> > > later if needed
> >
> > Hi Jiri,
> >
> > I'm pretty sure the current approach recommended by Stanislav, to
> > never reset/zero the ID and instead mark it as invalid via a flag in
> > the bpf_prog struct, should resolve the perf problem as well.

ok, I misunderstood

> 
> I probably should elaborate on this a bit more, in the case of
> perf_event_bpf_event() the getter which checks the valid_id flag isn't
> used, rather a direct access of bpf_prog_aux::__id is done so that the
> ID can be retrieved even after it is free'd/marked-invalid.  Here is
> the relevant code snippet for the patch:
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index aefc1e08e015..c24e897d27f1 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -9001,7 +9001,11 @@ void perf_event_bpf_event(struct bpf_prog *prog,
>                        },
>                        .type = type,
>                        .flags = flags,
> -                       .id = prog->aux->id,
> +                       /*
> +                        * don't use bpf_prog_get_id() as the id may be marked
> +                        * invalid on PERF_BPF_EVENT_PROG_UNLOAD events
> +                        */
> +                       .id = prog->aux->__id,

looks good

>                },
>        };
> 
> > My time
> > is a little short at the moment due to the holidays, but perhaps with
> > a little luck I'll get a new revision of the patch posted soon
> > (today?) and you can take a look and give it a test.  Are you
> > subscribed to the linux-audit and/or bpf mailing lists?  If not I can
> > CC you directly on the next revision.

bpf list is fine

thanks,
jirka