Re: [PATCH v2 bpf-next 1/2] bpf: Reject bpf_skb_output() from return-side tracing

[Feng Yang]

On Tue, 22 Sep 2026 04:15:13 +0200, Kumar Kartikeya Dwivedi wrote:

> On Tue Sep 22, 2026 at 3:54 AM CEST, Feng Yang wrote:
> > From: Feng Yang <yangfeng@xxxxxxxxxx>
> >
> > BPF fexit programs run after the traced function returns, while their
> > context still contains the original function argument values. A traced
> > function is free to consume an skb argument before returning, so the
> > pointer seen by fexit can already be stale.
> >
> > The verifier checks that the first argument to bpf_skb_output() has the
> > BTF type of struct sk_buff, but that does not establish its lifetime.
> > bpf_skb_event_output() then dereferences skb->len and can trigger a
> > use-after-free.
> >
> > Do not expose bpf_skb_output() to tracing programs which can run after
> > the target: fexit, fexit.multi, fsession and fsession.multi. Keep it
> > available to fentry and other tracing attach types where it is already
> > supported. fsession must be rejected because the same program runs on
> > both entry and return and the verifier cannot prove that a helper call
> > is entry-only.
> >
> > Fixes: fec56f5890d9 ("bpf: Introduce BPF trampoline")
> > Reported-by: Quan Sun <2022090917019@xxxxxxxxxxxxxxxx>
> > Reported-by: Yinhao Hu <dddddd@xxxxxxxxxxx>
> > Reported-by: Kaiyan Mei <M202472210@xxxxxxxxxxx>
> > Closes: https://lore.kernel.org/all/9d61b891-2d52-42b9-bc1a-ad963ccb675d@xxxxxxxxxxxxxxxx/
> > Signed-off-by: Yun Lu <luyun@xxxxxxxxxx>
> > Signed-off-by: Feng Yang <yangfeng@xxxxxxxxxx>
> > ---
> 
> Sorry, this is not an acceptable fix. This negates usage of the function in all
> fexit and fession attach points. Unless there is a simpler way to enumerate in
> which attach points this helper should be disabled, it might not be worth doing
> and leaving this be as is.
> 

Thanks for your reply. That is indeed correct.

Moreover, the comment for `bpf_skb_output` in `include/uapi/linux/bpf.h`
reads: "This helper is similar to bpf_perf_event_output() but restricted to
raw_tracepoint bpf programs.", yet this comment is actually outdated.

> pw-bot: cr