RE: [PATCH net] tcp: Fix TLP implementation in case receive window limits send

[Itzcak Pechtalt <itzcak@flashnetworks.com>]

> -----Original Message-----
> From: Neal Cardwell [mailto:ncardwell@google.com]
> Sent: Wednesday, September 24, 2014 10:25 PM
> To: Nandita Dukkipati
> Cc: Yuchung Cheng; Itzcak Pechtalt; netdev@vger.kernel.org
> Subject: Re: [PATCH net] tcp: Fix TLP implementation in case receive
> window limits send
> 
> On Wed, Sep 24, 2014 at 3:00 PM, Nandita Dukkipati
> <nanditad@google.com> wrote:
> >> On Tue, Sep 23, 2014 at 9:28 AM, Itzcak Pechtalt
> >> <itzcak@flashnetworks.com> wrote:
> >>> From: Itzcak Pechtalt <itzcak@flashnetworks.com>
> >>>
> >>> TCP Tail loss probe (TLP) algorithm implementation has some problem.
> >>> According to paper (draft-dukkipati-tcpm-tcp-loss-probe-0 ):
> >>> In case recive window of receiver limits send of new packet in probe
> >>> time than a retransmit of last packet send should be done.
> >>>
> >>> Actually, return code from tcp_write_xmit is not checked and only
> >>> RTO is scheduled, So, it will take more time for reovery in this
> >>> case than without TLP.
> >>>
> >>> Signed-off-by: Itzcak Pechtalt <itzcak@flashnetworks.com>
> >> Acked-by: Yuchung Cheng <ycheng@google.com>
> > Acked-by: Nandita Dukkipati <nanditad@google.com>
> 
> Acked-by: Neal Cardwell <ncardwell@google.com>
> 
> BTW, Itzcak, have you been able to construct a test case where this patch
> now allows a TLP in a case where the sender is limited by the receive
> window?
> 
> Often we will be prevented from doing any kind of TLP if we are limited by
> the receive window, due to these lines in
> tcp_schedule_loss_probe():
> 
>   if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
>        tcp_send_head(sk))
>           return false;
> 
> But there is something else going on as well, since I haven't been able to
> force a TLP even when cwnd == rwin == tcp_packets_in_flight() == 10. But I
> haven't had much time to spend on it.
> 
> neal

The scenario will be  in case the sender process calls "send" per packet size data, 
so when tcp_schedule_loss_probe  is called there is no additional data 
(tcp_send_head(sk)) returns false),  but further calls to send fail due to receive 
window limit  (The same is right if data bulk size ended exactly with receive 
window limit).

But really I can't understand what is the goal of the above 3 lines of code.
If send was stopped from any reason, why not schedule TLP timer?
Anyway next send will clear the timer by tcp_event_new_data_sent function.

What is your opinion about another patch to remove these 3 lines of code?

Itzcak