On Tue, Oct 07, 2025 at 12:32:19AM +0200, Stefano Brivio wrote: > On Fri, 3 Oct 2025 13:19:17 +1000 > David Gibson wrote: > > > On Thu, Oct 02, 2025 at 01:58:41PM +0200, Stefano Brivio wrote: > > > On Thu, 2 Oct 2025 12:41:08 +1000 > > > David Gibson wrote: > > > > > > > On Thu, Oct 02, 2025 at 02:06:43AM +0200, Stefano Brivio wrote: > > > > > If we reach end-of-file on a socket (or get EPOLLRDHUP / EPOLLHUP) and > > > > > send a FIN segment to the guest / container acknowledging a sequence > > > > > number that's behind what we received so far, we won't have any > > > > > further trigger to send an updated ACK segment, as we are now > > > > > switching the epoll socket monitoring to edge-triggered mode. > > > > > > > > > > To avoid this situation, in tcp_update_seqack_wnd(), we set the next > > > > > acknowledgement sequence to the current observed sequence, regardless > > > > > of what was acknowledged socket-side. > > > > > > > > To double check my understanding: things should work if we always > > > > acknowledged everything we've received. Acknowledging only what the > > > > peer has acked is a refinement to give the guest a view that's closer > > > > to what it would be end-to-end with the peer (which might improve the > > > > operation of flow control). > > > > > > Right. > > > > > > > We can't use that refined mechanism when the socket is closing > > > > (amongst other cases), because while we can get the peer acked bytes > > > > from TCP_INFO, we can't get events when that changes, so we have no > > > > mechanism to provide updates to the guest at the right time. So we > > > > fall back to the simpler method. > > > > > > > > Is that correct? > > > > > > Also correct, yes. If you have a better idea to summarise this in the > > > comment in tcp_buf_data_from_sock() let me know. > > > > Hm, I might. Or actually a way to reorganise the code that I think > > will be a bit clearer and probably allow a clearer comment too. > > I would keep reworks for a later moment. Right now, it's already taking > me long enough to find a moment to investigate these issues, write these > fixes, and test them. I mean... the change I'm proposing reduces lines of code (excepting the big new comment), makes it easier to reason about and is localised to the immediately surrounding code. But whatever, I don't particularly care about the order we do things. > > > Maybe I could mention > > > EPOLLET explicitly there? > > > > I don't think EPOLLET is actually relevant. Even if we had level > > triggered events, a change in bytes_acked doesn't count as an event > > (AFAIK). > > It doesn't count, but with level-triggered events, we would be busy > polling bytes_acked, as you noted. I was mentioning EPOLLET because it > could be taken, intuitively, as a "stop listening for events" (almost) > step. I'll leave that out then. I mean, if busy polling were acceptable we could accomplish that easily enough by doing it in tcp_defer_handler() regardless of EPOLLET. > > So either some other event is on, in which case we'd > > effectively be busy polling bytes_acked, or it's not in which case we > > don't get updates, just like now. > > > > I principle we could implement some sort of timer based polling, but > > that sounds like way more trouble than it's worth. > > We already have something similar, based on ACK_INTERVAL and > ACK_TO_TAP_DUE, and it shouldn't be overly complicated to extend that > to a new FIN_TO_TAP_DUE flag. But indeed beyond the scope of this > series. Certainly. > > > > > However, we don't necessarily call tcp_update_seqack_wnd() before > > > > > sending the FIN segment, which might potentially lead to a situation, > > > > > not observed in practice, where we unnecessarily cause a > > > > > retransmission at some point after our FIN segment. > > > > > > > > > > Avoid that by setting the ACK sequence to whatever we received from > > > > > the container / guest, before sending a FIN segment and switching to > > > > > EPOLLET. > > > > > > > > > > Signed-off-by: Stefano Brivio > > > > > > > > Based on my understanding above, this looks correct to me, so, > > > > > > > > Reviewed-by: David Gibson > > > > > > > > My only concern is whether we could instead insert an extra call to > > > > tcp_update_seqack_wnd() to reduce duplicated logic. > > > > > > Hmm, maybe, but on the other hand we're closing the connection. Should > > > we really spend time querying TCP_INFO to recalculate the window at > > > this point? I wouldn't. > > > > Good point. I mean tcp_update_seqack_wnd() could skip the TCP_INFO in > > that case, but that does look a bit fiddly. > > > > On the other hand, in favour of not duplicating logic... > > > > [snip] > > > > > @@ -368,7 +368,19 @@ int tcp_buf_data_from_sock(const struct ctx *c, struct tcp_tap_conn *conn) > > > > > conn_flag(c, conn, STALLED); > > > > > } else if ((conn->events & (SOCK_FIN_RCVD | TAP_FIN_SENT)) == > > > > > SOCK_FIN_RCVD) { > > > > > - int ret = tcp_buf_send_flag(c, conn, FIN | ACK); > > > > > + int ret; > > > > > + > > > > > + /* On TAP_FIN_SENT, we won't get further data events > > > > > + * from the socket, and this might be the last ACK > > > > > + * segment we send to the tap, so update its sequence to > > > > > + * include everything we received until now. > > > > > + * > > > > > + * See also the special handling on CONN_IS_CLOSING() in > > > > > + * tcp_update_seqack_wnd(). > > > > > + */ > > > > > + conn->seq_ack_to_tap = conn->seq_from_tap; > > > > ... the equivalent bits in tcp_update_seqack_wnd() have after them: > > if (SEQ_LT(conn->seq_ack_to_tap, prev_ack_to_tap)) > > conn->seq_ack_to_tap = prev_ack_to_tap; > > > > Don't we need that here as well, in case the guest is retransmitting > > when we get the sock side FIN? > > Not really, because we don't rewind conn->seq_from_tap, so we don't > risk jumping back here. Ah, because this one's on the sock->tap data path, whereas the other calls are on the tap->sock data path. Good point. > In tcp_update_seqack_wnd(), we might jump back (that should be double > checked eventually, I'm not sure it's still the case) if we happened to > acknowledge more than acknowledged socket-side while handling some > particular condition, and then we switch back to acknowledging only > bytes_acked. > > It should happen if the destination is/was a low RTT one, but we run > out of slots in low_rtt_dst in favour of other entries. I don't > remember any other case. -- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson