Re: [Patch bpf v2] skmsg: check sk_rcvbuf limit before queuing to ingress_skb

On Sat, Jul 03, 2021 at 07:52 PM CEST, Jakub Sitnicki wrote:

[...]

Then there is the case when a parser prog is attached. In this case the
skb is really gone if we drop it on redirect.

In sk_psock_strp_read, we ignore the -EIO error from
sk_psock_verdict_apply, and return to tcp_read_sock how many bytes have
been parsed.

  sk->sk_data_ready
    sk_psock_verdict_data_ready
      ->read_sock(..., sk_psock_verdict_recv)
        tcp_read_sock (used = copied = eaten)
          strp_recv -> ret = eaten
            __strp_recv -> ret = eaten
              strp->cb.rcv_msg -> -EIO
                sk_psock_verdict_apply -> -EIO
                  sk_psock_redirect -> -EIO

Copy-paste error. The call chain for the parser case goes as so:

  sk->sk_data_ready
    sk_psock_strp_data_ready
      strp_data_ready
        strp_read_sock
          ->read_sock(..., strp_recv)
            tcp_read_sock (used = copied = skb->len)
              strp_recv -> ret = skb->len
                __strp_recv -> ret = skb->len (dummy parser case)
                  strp->cb.rcv_msg -> -EIO
                    sk_psock_verdict_apply -> -EIO
                      sk_psock_redirect -> -EIO