On Wed, Nov 23, 2022 at 3:31 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Wed, Nov 23, 2022 at 2:39 PM Stanislav Fomichev <sdf@xxxxxxxxxx> wrote:
> >
> > On Wed, Nov 23, 2022 at 12:39 PM Alexei Starovoitov
> > <alexei.starovoitov@xxxxxxxxx> wrote:
> > >
> > > On Wed, Nov 23, 2022 at 12:08 PM Stanislav Fomichev <sdf@xxxxxxxxxx> wrote:
> > > >
> > > > Jiri reports broken test_progs after recent commit 68f8e3d4b916
> > > > ("selftests/bpf: Make sure zero-len skbs aren't redirectable").
> > > > Apparently we don't remount debugfs when we switch back networking namespace.
> > > > Let's explicitly mount /sys/kernel/debug.
> > > >
> > > > 0: https://lore.kernel.org/bpf/63b85917-a2ea-8e35-620c-808560910819@xxxxxxxx/T/#ma66ca9c92e99eee0a25e40f422489b26ee0171c1
> > > >
> > > > Fixes: a30338840fa5 ("selftests/bpf: Move open_netns() and close_netns() into network_helpers.c")
> > > > Reported-by: Jiri Olsa <olsajiri@xxxxxxxxx>
> > > > Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> > > > ---
> > > > tools/testing/selftests/bpf/network_helpers.c | 4 ++++
> > > > 1 file changed, 4 insertions(+)
> > > >
> > > > diff --git a/tools/testing/selftests/bpf/network_helpers.c b/tools/testing/selftests/bpf/network_helpers.c
> > > > index bec15558fd93..1f37adff7632 100644
> > > > --- a/tools/testing/selftests/bpf/network_helpers.c
> > > > +++ b/tools/testing/selftests/bpf/network_helpers.c
> > > > @@ -426,6 +426,10 @@ static int setns_by_fd(int nsfd)
> > > > if (!ASSERT_OK(err, "mount /sys/fs/bpf"))
> > > > return err;
> > > >
> > > > + err = mount("debugfs", "/sys/kernel/debug", "debugfs", 0, NULL);
> > > > + if (!ASSERT_OK(err, "mount /sys/kernel/debug"))
> > > > + return err;
> > > > +
> > > > return 0;
> > > > }
> > >
> > > Thanks.
> > > It fixes part of it but it's still racy.
> > > I see:
> > > do_read:FAIL:open open /sys/fs/bpf/bpf_iter_test1 failed: No such file
> > > or directory
> > >
> > > I suspect it happens when iter tests are running while test_empty_skb
> > > is cleaning the netns.
> > >
> > > So I've added:
> > > -void test_empty_skb(void)
> > > +void serial_test_empty_skb(void)
> > > -void test_xdp_do_redirect(void)
> > > +void serial_test_xdp_do_redirect(void)
> > > -void test_xdp_synproxy(void)
> > > +void serial_test_xdp_synproxy(void)
> > >
> > > to stop the bleeding and applied.
> >
> > Not sure I understand where the race is coming from and no luck
> > reproducing locally :-(
> > Looks like we run the tests in the forked workers, so that
> > unshare(mountns) shouldn't theoretically affect the rest, but
> > obviously I'm missing something..
>
> I'm equally confused.
[..]
> If what you're describing was the case than the bug of
> not mounting debugfs wouldn't have caused issues in parallel run.
The workers are reused and aren't respawned for every test.
If some of them runs empty_skb test, its debugfs is not mounted
anymore and any other test that requires debugfs (and lands on this
worker) will be broken.
> That close_netns is somehow messing with other forked processes.
Yeah, agreed, looks like /sys/fs/bpf disappears while bpf_iter test is running.
I can try to stick open/close_netns in a bunch of places in test_progs
to see if I can trigger the issue. (or maybe have a test in parallel
that does while(1) {openns(); closens(); }
