Re: sockmap test is broken

[John Fastabend <john.fastabend@gmail.com>]

John Fastabend wrote:
> Andrii Nakryiko wrote:
> > On Wed, Nov 17, 2021 at 10:00 AM John Fastabend
> > <john.fastabend@gmail.com> wrote:
> > >
> > > Andrii Nakryiko wrote:
> > > > On Sun, Nov 14, 2021 at 9:21 PM John Fastabend <john.fastabend@gmail.com> wrote:
> > > > >
> > > > > Alexei Starovoitov wrote:
> > > > > > test_maps is failing in bpf tree:
> > > > > >
> > > > > > $ ./test_maps
> > > > > > Failed sockmap recv
> > > > > >
> > > > > > and causing BPF CI to stay red.
> > > > > >
> > > > > > Since bpf-next is fine, I suspect it is one of John's or Jussi's patches.
> > > > > >
> > > > > > Please take a look.
> > > > >
> > > > > I'll look into it thanks.
> > > >
> > > > Any updates, John? Should we just disable test_maps in CI to make it
> > > > useful again?
> > >
> > > I'm debugging this now. Hopefully I'll have a fix shortly (today I hope).
> > > Maybe, it makes sense to wait for EOD and if I still don't have the fix
> > > disable it then. Anyways fixing it is top of list now.
> > 
> > Sounds good, let's hope you find it and fix it today.
> 
> OK got the fix, but its fairly subtle. Whats happening is when socks are
> removed from a map their programs are not actually being removed. They
> continue to live with the sock for the lifetime of the socket or until
> the last reference held from BPF side is lost. At which point all progs
> are dropped and socket returns to normal/preBPF state. We never noticed
> it on our real use cases because once we move sockets into BPF we never
> release them until the socket is free. The fix is to null the set progs
> and then do the update_sk_prot call which will decide based on the
> configured programs what proto ops need to be set to.

Fix posted here,

https://lore.kernel.org/bpf/20211119181418.353932-1-john.fastabend@gmail.com/

Thanks,
John