[5.15.y] build issue when building the queue

[5.15.y] build issue when building the queue

[Matthieu Baerts]

Hi Sasha, Greg,

(+cc Kuniyuki)

Our CI at MPTCP cannot build the patches being queued for v5.15:


  net/unix/af_unix.c: In function 'bpf_iter_unix_hold_batch':
  net/unix/af_unix.c:3352:22: error: 'unix_table_locks' undeclared
(first use in this function); did you mean 'unix_table_lock'?
   3352 |         spin_unlock(&unix_table_locks[start_sk->sk_hash]);
        |                      ^~~~~~~~~~~~~~~~
        |                      unix_table_lock

  net/unix/af_unix.c: In function 'bpf_iter_unix_batch':
  net/unix/af_unix.c:3397:14: error: implicit declaration of function
'unix_get_first'; did you mean 'unix_get_socket'?
[-Werror=implicit-function-declaration]
   3397 |         sk = unix_get_first(seq, pos);
        |              ^~~~~~~~~~~~~~
        |              unix_get_socket


https://github.com/multipath-tcp/mptcp_net-next/actions/runs/26156680393/job/76937788263#step:9:2175

For the first one, unix_table_locks has been introduced in afd20b9290e1
("af_unix: Replace the big lock with small locks."). Before, the version
without 's' (unix_table_lock) was used.

For the second one, unix_get_first has been introduced in 4408d55a6467
("af_unix: Refactor unix_next_socket().").

It looks like these errors are due to:

  bpf-af_unix-use-batching-algorithm-in-bpf-unix-iter.patch

This backport of 855d8e77ffb0 ("bpf: af_unix: Use batching algorithm in
bpf unix iter.") is needed for 4d328dd69538 ("bpf, sockmap: Fix af_unix
iter deadlock") apparently.

I'm not sure how you want to fix this in v5.15: backporting new
improvements, or adapting. But maybe easier to drop the faulty patch for
the moment, and bpf-sockmap-fix-af_unix-iter-deadlock.patch I suppose.

Cheers,
Matt
-- 
Sponsored by the NGI0 Core fund.

Re: [5.15.y] build issue when building the queue

[Sasha Levin]

On Wed, May 20, 2026 at 11:28:28PM +1000, Matthieu Baerts wrote:
>Hi Sasha, Greg,
>
>(+cc Kuniyuki)
>
>Our CI at MPTCP cannot build the patches being queued for v5.15:
>
>
>  net/unix/af_unix.c: In function 'bpf_iter_unix_hold_batch':
>  net/unix/af_unix.c:3352:22: error: 'unix_table_locks' undeclared
>(first use in this function); did you mean 'unix_table_lock'?
>   3352 |         spin_unlock(&unix_table_locks[start_sk->sk_hash]);
>        |                      ^~~~~~~~~~~~~~~~
>        |                      unix_table_lock
>
>  net/unix/af_unix.c: In function 'bpf_iter_unix_batch':
>  net/unix/af_unix.c:3397:14: error: implicit declaration of function
>'unix_get_first'; did you mean 'unix_get_socket'?
>[-Werror=implicit-function-declaration]
>   3397 |         sk = unix_get_first(seq, pos);
>        |              ^~~~~~~~~~~~~~
>        |              unix_get_socket
>
>
>https://github.com/multipath-tcp/mptcp_net-next/actions/runs/26156680393/job/76937788263#step:9:2175
>
>For the first one, unix_table_locks has been introduced in afd20b9290e1
>("af_unix: Replace the big lock with small locks."). Before, the version
>without 's' (unix_table_lock) was used.
>
>For the second one, unix_get_first has been introduced in 4408d55a6467
>("af_unix: Refactor unix_next_socket().").
>
>It looks like these errors are due to:
>
>  bpf-af_unix-use-batching-algorithm-in-bpf-unix-iter.patch
>
>This backport of 855d8e77ffb0 ("bpf: af_unix: Use batching algorithm in
>bpf unix iter.") is needed for 4d328dd69538 ("bpf, sockmap: Fix af_unix
>iter deadlock") apparently.
>
>I'm not sure how you want to fix this in v5.15: backporting new
>improvements, or adapting. But maybe easier to drop the faulty patch for
>the moment, and bpf-sockmap-fix-af_unix-iter-deadlock.patch I suppose.

Thanks for the report. I've just dropped this (and few other follow-up
commits).

-- 
Thanks,
Sasha