tcp: make smp_rmb() conditional in tcp_poll()

Commit a4d258036ed9 ("tcp: Fix race in tcp_poll") added smp_rmb() in
tcp_poll() and smp_wmb() in tcp_reset() (now tcp_done_with_error())
to ensure that if tcp_poll() observed socket closure, it would also
observe sk->sk_err.

Currently, tcp_poll() unconditionally executes smp_rmb() at the end
of every invocation, which on weakly-ordered architectures such as ARM64
emits a memory barrier instruction (dmb ishld) on the poll fast path,
even for healthy, active sockets.

However, tcp_poll() only needs this barrier if socket closure has been
observed, to ensure that the error code set by tcp_done_with_error()
before socket closure is visible before returning EPOLLERR.

Move smp_rmb() inside the conditional block handling socket closure
(shutdown == SHUTDOWN_MASK || state == TCP_CLOSE). For healthy
connected sockets in epoll, tcp_poll() avoids the barrier entirely.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260913123224.762935-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>