Re: [PATCH net-next 2/2] tcp: factorize logic into tcp_epollin_ready()

On Sat, Feb 13, 2021 at 12:05 AM Eric Dumazet [off-list ref] wrote:

On Sat, Feb 13, 2021 at 8:50 AM Eric Dumazet [off-list ref] wrote:

quoted

On Sat, Feb 13, 2021 at 1:30 AM Wei Wang [off-list ref] wrote:

quoted

quoted

quoted

 void tcp_data_ready(struct sock *sk)
 {
-       const struct tcp_sock *tp = tcp_sk(sk);
-       int avail = tp->rcv_nxt - tp->copied_seq;
-
-       if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
-           !sock_flag(sk, SOCK_DONE) &&

Seems "!sock_flag(sk, SOCK_DONE)" is not checked in
tcp_epollin_read(). Does it matter?

Yes, probably, good catch.

Not sure where tcp_poll() gets this, I have to double check.

It gets the info from sk->sk_hutdown & RCV_SHUTDOWN

tcp_find() sets both sk->sk_shutdown |= RCV_SHUTDOWN and
sock_set_flag(sk, SOCK_DONE);

This seems to suggest tcp_fin() could call sk->sk_data_ready() so that
we do not have to test for this unlikely condition in tcp_data_ready()

When a thread is subsequently then woken up due to sk_data_ready(),
and it calls tcp_stream_is_readable() but we had lowat > 1 set, is
there a chance of that thread then thinking that the stream is not
readable, despite SOCK_DONE being set? This is assuming that the check
is not added to the refactored logic.

Note that on a related note if the tcp memory pressure check (for
system-wide pressure) is added just to the original code in
tcp_data_ready() but not added to tcp_stream_is_readable() we had this
kind of issue (sk_data_ready() was called but tcp_stream_is_readable()
returned false).

-Arjun

-Arjun