On Tue, Aug 05, 2025 at 02:07:20PM -0700, Yonghong Song wrote:
>
>
> On 8/1/25 5:10 AM, Florian Lehner wrote:
> > 73b11c2a introduced LINK_DETACH and implemented it for some link types,
> > like xdp, netns and others.
> >
> > This patch implements LINK_DETACH for perf and iter links, re-using
> > existing link release handling code.
[..]
> > static void bpf_iter_link_dealloc(struct bpf_link *link)
> > {
> > struct bpf_iter_link *iter_link =
> > @@ -490,6 +496,7 @@ static int bpf_iter_link_fill_link_info(const struct bpf_link *link,
> > static const struct bpf_link_ops bpf_iter_link_lops = {
> > .release = bpf_iter_link_release,
> > + .detach = bpf_iter_link_detach,
>
> Not sure how useful for this one. For bpf_iter programs,
> the loaded prog will expect certain bpt_iter (e.g., bpf_map_elem, bpf_map, ...).
> So even if you have detach, you won't be able to attach to a different
> bpf_iter flavor.
>
> Do you have a use case for this one?
>
A key reason for adding this was to enable the temporary disabling and re-enabling of
an attached BPF program while keeping the same bpf_iter flavor. If you don't think
this is a strong enough use case, I'm open to removing this from the patch.
> > static void bpf_perf_link_dealloc(struct bpf_link *link)
> > {
> > struct bpf_perf_link *perf_link = container_of(link, struct bpf_perf_link, link);
> > @@ -4027,6 +4033,7 @@ static void bpf_perf_link_show_fdinfo(const struct bpf_link *link,
> > static const struct bpf_link_ops bpf_perf_link_lops = {
> > .release = bpf_perf_link_release,
> > + .detach = bpf_perf_link_detach,
>
> This one may be possible. You might be able to e.g., try a different bpf_cookie, or
> different perf event.
>
The primary use case for this feature is to allow for the temporary disabling of
uprobes that are attached using bpf_perf_links.
