On Wed, Sep 23, 2020 at 7:05 PM Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote: > > On Wed, Sep 23, 2020 at 6:49 PM Alexei Starovoitov > <alexei.starovoitov@xxxxxxxxx> wrote: > > > > Hi Stanislav, > > > > looks like flow_dissector selftest got quite unstable recently. > > test_link_update_invalid_opts:FAIL:340 > > bpf_link_create(prog1): Argument list too long > > #33/25 flow dissector link update invalid opts:FAIL > > test_link_update_invalid_prog:FAIL:400 > > bpf_link_create(prog1): Argument list too long > > #33/26 flow dissector link update invalid prog:FAIL > > #33/27 flow dissector link update netns gone:OK > > > > I've seen similar flakiness for cgroup_link selftest that used to be > rock solid. And it just clicked when I saw this, that this patch might > be a culprit: > > https://patchwork.ozlabs.org/project/netdev/patch/20200917074453.20621-1-songmuchun@xxxxxxxxxxxxx/ > > It makes bpf_link detachment delayed, so now anything that relies on > the fact that bpf_link gets auto-detached immediately after the last > link FD was closed is flaky. But that is a pretty reasonable and > convenient assumption. So can we please revert that patch? It's a > really nice guarantee to have, while the benefits of the fix in that > patch is a bit ephemeral. Indeed. Reverted. The usage of in_atomic() is fine there.
