netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Marcelo Ricardo Leitner <mleitner@redhat.com>
To: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>,
	netdev <netdev@vger.kernel.org>,
	Eric Dumazet <edumazet@google.com>
Subject: Re: TCP NewReno and single retransmit
Date: Mon, 03 Nov 2014 14:38:05 -0200	[thread overview]
Message-ID: <5457AF6D.6010105@redhat.com> (raw)
In-Reply-To: <CAK6E8=dYfJw16Q0D40QD7RLr=wq=y+5W59zHmZ24L49OPS9O5A@mail.gmail.com>

On 31-10-2014 01:51, Yuchung Cheng wrote:
> On Thu, Oct 30, 2014 at 7:24 PM, Marcelo Ricardo Leitner
> <mleitner@redhat.com> wrote:
>> On 30-10-2014 00:03, Neal Cardwell wrote:
>>>
>>> On Mon, Oct 27, 2014 at 2:49 PM, Marcelo Ricardo Leitner
>>> <mleitner@redhat.com> wrote:
>>>>
>>>> Hi,
>>>>
>>>> We have a report from a customer saying that on a very calm connection,
>>>> like
>>>> having only a single data packet within some minutes, if this packet gets
>>>> to
>>>> be re-transmitted, retrans_stamp is only cleared when the next acked
>>>> packet
>>>> is received. But this may make we abort the connection too soon if this
>>>> next
>>>> packet also gets lost, because the reference for the initial loss is
>>>> still
>>>> for a big while ago..
>>>
>>> ...
>>>>
>>>> @@ -2382,31 +2382,32 @@ static inline bool tcp_may_undo(const struct
>>>> tcp_sock *tp)
>>>>    static bool tcp_try_undo_recovery(struct sock *sk)
>>>
>>> ...
>>>>
>>>>           if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
>>>>                   /* Hold old state until something *above* high_seq
>>>>                    * is ACKed. For Reno it is MUST to prevent false
>>>>                    * fast retransmits (RFC2582). SACK TCP is safe. */
> Or we can just remove this strange state-holding logic?
>
> I couldn't find such a "MUST" statement in RFC2582. RFC2582 section 3
> step 5 suggests exiting the recovery procedure when an ACK acknowledges
> the "recover" variable (== tp->high_seq - 1).
>
> Since we've called tcp_reset_reno_sack() before tcp_try_undo_recovery(),
> I couldn't see how false fast retransmits can be triggered without
> this state-holding.
>
> Any insights?

Nice one, me neither. Neal?

 From RFC2582, Section 5, Avoiding Multiple Fast Retransmits:

    Nevertheless, unnecessary Fast Retransmits can occur with Reno or
    NewReno TCP, particularly if a Retransmit Timeout occurs during Fast
    Recovery.  (This is illustrated for Reno on page 6 of [F98], and for
    NewReno on page 8 of [F98].)  With NewReno, the data sender remains
    in Fast Recovery until either a Retransmit Timeout, or *until all of
    the data outstanding when Fast Retransmit was entered has been
    acknowledged*.  Thus with NewReno, the problem of multiple Fast
    Retransmits from a single window of data can only occur after a
    Retransmit Timeout.

Bolding mark is mine. If I didn't miss anything, as that condition was met, we 
should be good to keep that cwnd reduction (required by section 3 step 5) and 
but get back to Open state right away.

Marcelo

>>>>                   tcp_moderate_cwnd(tp);
>>>> +               tp->retrans_stamp = 0;
>>>>                   return true;
>>>>           }
>>>>           tcp_set_ca_state(sk, TCP_CA_Open);
>>>>           return false;
>>>>    }
>>>>
>>>> We would still hold state, at least part of it.. WDYT?
>>>
>>>
>>> This approach sounds OK to me as long as we include a check of
>>> tcp_any_retrans_done(), as we do in the similar code paths (for
>>> motivation, see the comment above tcp_any_retrans_done()).
>>
>>
>> Yes, okay. I thought that this would be taken care of already by then but
>> reading the code again now after your comment, I can see what you're saying.
>> Thanks.
>>
>>> So it sounds fine to me if you change that one new line to the following
>>> 2:
>>>
>>> +  if (!tcp_any_retrans_done(sk))
>>> +    tp->retrans_stamp = 0;
>>
>>
>> Will do.
>>
>>> Nice catch!
>>
>>
>> A good part of it (including the diagram) was done by customer. :)
>> I'll post the patch as soon as we sync with them (credits).
>>
>> Marcelo
>>
>

  reply	other threads:[~2014-11-03 16:38 UTC|newest]

Thread overview: 12+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2014-10-27 18:49 TCP NewReno and single retransmit Marcelo Ricardo Leitner
2014-10-30  2:03 ` Neal Cardwell
2014-10-30 11:24   ` Marcelo Ricardo Leitner
2014-10-31  3:51     ` Yuchung Cheng
2014-11-03 16:38       ` Marcelo Ricardo Leitner [this message]
2014-11-03 20:08         ` Neal Cardwell
2014-11-03 21:35           ` Marcelo Ricardo Leitner
2014-11-03 23:17             ` Neal Cardwell
2014-11-04  7:59               ` Yuchung Cheng
2014-11-04 13:12                 ` Marcelo Ricardo Leitner
2014-11-04 14:38                 ` Neal Cardwell
2014-11-04  9:56           ` David Laight

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=5457AF6D.6010105@redhat.com \
    --to=mleitner@redhat.com \
    --cc=edumazet@google.com \
    --cc=ncardwell@google.com \
    --cc=netdev@vger.kernel.org \
    --cc=ycheng@google.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).