On Mon, Sep 25, 2023 at 08:52 PM -07, John Fastabend wrote:

With e5c6de5fa0258 ("bpf, sockmap: Incorrectly handling copied_seq") we
started fixing the available bytes accounting by moving copied_seq to
where the user actually reads the bytes.

However we missed handling MSG_PEEK correctly and we need to ensure
that we don't kfree_skb() a skb off the receive_queue when the
copied_seq number is not incremented by user reads for some time.

v2: drop seq var in tcp_read_skb its no longer necessary per Jakub's
    suggestion

Credit goes to Simon Horman.

v3: drop tcp_sock as well its also not used anymore. sorry for the extra
    noise there.

John Fastabend (3):
  bpf: tcp_read_skb needs to pop skb regardless of seq
  bpf: sockmap, do not inc copied_seq when PEEK flag set
  bpf: sockmap, add tests for MSG_F_PEEK

 net/ipv4/tcp.c                                | 10 +---
 net/ipv4/tcp_bpf.c                            |  4 +-
 .../selftests/bpf/prog_tests/sockmap_basic.c  | 52 +++++++++++++++++++
 3 files changed, 57 insertions(+), 9 deletions(-)

For the series:

Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>