Re: [PATCH net] bpf: sockmap: Remove preempt_disable in sock_map_sk_acquire

[Alexei Starovoitov]

On Tue, Aug 1, 2023 at 12:44 AM Paolo Abeni <pabeni@xxxxxxxxxx> wrote:
>
> On Mon, 2023-07-31 at 20:58 -0700, John Fastabend wrote:
> > Jakub Sitnicki wrote:
> > >
> > > On Fri, Jul 28, 2023 at 08:44 AM +02, tglozar@xxxxxxxxxx wrote:
> > > > From: Tomas Glozar <tglozar@xxxxxxxxxx>
> > > >
> > > > Disabling preemption in sock_map_sk_acquire conflicts with GFP_ATOMIC
> > > > allocation later in sk_psock_init_link on PREEMPT_RT kernels, since
> > > > GFP_ATOMIC might sleep on RT (see bpf: Make BPF and PREEMPT_RT co-exist
> > > > patchset notes for details).
> > > >
> > > > This causes calling bpf_map_update_elem on BPF_MAP_TYPE_SOCKMAP maps to
> > > > BUG (sleeping function called from invalid context) on RT kernels.
> > > >
> > > > preempt_disable was introduced together with lock_sk and rcu_read_lock
> > > > in commit 99ba2b5aba24e ("bpf: sockhash, disallow bpf_tcp_close and update
> > > > in parallel"), probably to match disabled migration of BPF programs, and
> > > > is no longer necessary.
> > > >
> > > > Remove preempt_disable to fix BUG in sock_map_update_common on RT.
> > > >
> > > > Signed-off-by: Tomas Glozar <tglozar@xxxxxxxxxx>
> > > > ---
> > >
> > > We disable softirq and hold a spin lock when modifying the map/hash in
> > > sock_{map,hash}_update_common so this LGTM:
> > >
> >
> > Agree.
> >
> > > Reviewed-by: Jakub Sitnicki <jakub@xxxxxxxxxxxxxx>
> >
> > Reviewed-by: John Fastabend <john.fastabend@xxxxxxxxx>
> >
> > >
> > > You might want some extra tags:
> > >
> > > Link: https://lore.kernel.org/all/20200224140131.461979697@xxxxxxxxxxxxx/
> > > Fixes: 99ba2b5aba24 ("bpf: sockhash, disallow bpf_tcp_close and update in parallel")
>
> ENOCOFFEE here (which is never an excuse, but at least today is really
> true), but I considered you may want this patch via the ebpf tree only
> after applying it to net.
>
> Hopefully should not be tragic, but please let me know if you prefer
> the change reverted from net and going via the other path.

It's fine. Probably it shouldn't conflict with other sockmap fixes.