On Tue, 2 Mar 2021 10:38:46 +0100 Eric Dumazet wrote:
> On Tue, Mar 2, 2021 at 7:08 AM Jakub Kicinski <k...@kernel.org> wrote:
> >
> > When receiver does not accept TCP Fast Open it will only ack
> > the SYN, and not the data. We detect this and immediately queue
> > the data for (re)transmission in tcp_rcv_fastopen_synack().
> >
> > In DC networks with very low RTT and without RFS the SYN-ACK
> > may arrive before NIC driver reported Tx completion on
> > the original SYN. In which case skb_still_in_host_queue()
> > returns true and sender will need to wait for the retransmission
> > timer to fire milliseconds later.
> >
> > Revert back to non-fast clone skbs, this way
> > skb_still_in_host_queue() won't prevent the recovery flow
> > from completing.
> >
> > Suggested-by: Eric Dumazet <eduma...@google.com>
> > Fixes: 355a901e6cf1 ("tcp: make connect() mem charging friendly")
>
> Hmmm, not sure if this Fixes: tag makes sense.
>
> Really, if we delay TX completions by say 10 ms, other parts of the
> stack will misbehave anyway.
>
> Also, backporting this patch up to linux-3.19 is going to be tricky.
Indeed, the problem is minor in practical terms. Maybe it's enough if I
spell that out more in the description? Are you thinking net-next or
net without a Fixes tag?
> The real issue here is that skb_still_in_host_queue() can give a false
> positive.
>
> I have mixed feelings here, as you can read my answer :/
>
> Maybe skb_still_in_host_queue() signal should not be used when a part
> of the SKB has been received/acknowledged by the remote peer
> (in this case the SYN part).
FWIW I was pondering this, when the rtx is requested by the receiver
we are relatively sure we can ignore skb_still_in_host_queue() because
we know our system should Tx in order so if receiver saw N + 1, N can't
be in our queues.
But AFAICT generalizing the test doesn't matter much. In cases other
than TFO worst case a loss probe will chase the rtx out. And I don't
grasp enough of TCP to implement the general optimization :)
> Alternative is that drivers unable to TX complete their skbs in a
> reasonable time should call skb_orphan()
> to avoid skb_unclone() penalties (and this skb_still_in_host_queue() issue)
>
> If you really want to play and delay TX completions, maybe provide a
> way to disable skb_still_in_host_queue() globally,
> using a static key ?
I see the TFO issue with rx and tx completions set to 33us both,
with two different NIC vendors, so the timing just influences the
likelihood.
> My personal WIP/hack was something like :
LGTM, are you happy with that being the fix?
> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> index
> 69a545db80d2ead47ffcf2f3819a6d066e95f35d..666f6f0a6a06fece204199e07a79e21d1faf8f92
> 100644
> --- a/net/ipv4/tcp_input.c
> +++ b/net/ipv4/tcp_input.c
> @@ -5995,7 +5995,8 @@ static bool tcp_rcv_fastopen_synack(struct sock
> *sk, struct sk_buff *synack,
> else
> tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
> skb_rbtree_walk_from(data) {
> - if (__tcp_retransmit_skb(sk, data, 1))
> + /* segs = -1 to bypass
> skb_still_in_host_queue() check */
> + if (__tcp_retransmit_skb(sk, data, -1))
> break;
> }
> tcp_rearm_rto(sk);
> diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
> index
> fbf140a770d8e21b936369b79abbe9857537acd8..1d1489e596976e352fe7d5ccee7a6eae55fdbcce
> 100644
> --- a/net/ipv4/tcp_output.c
> +++ b/net/ipv4/tcp_output.c
> @@ -3155,8 +3155,12 @@ int __tcp_retransmit_skb(struct sock *sk,
> struct sk_buff *skb, int segs)
> sk->sk_sndbuf))
> return -EAGAIN;
>
> - if (skb_still_in_host_queue(sk, skb))
> - return -EBUSY;
> + if (segs > 0) {
> + if (skb_still_in_host_queue(sk, skb))
> + return -EBUSY;
> + } else {
> + segs = -segs;
> + }
>
> if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
> if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
