linux-kernel - Re: Get amount of fast retransmissions from TCP info

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAF0XkCAZJd7adZf6VHdDc2yesyKQ8DqQnDOv_uk8o8mOVGMRyA@mail.gmail.com>
Date:   Mon, 24 Apr 2017 23:31:25 +0200
From:   Lars Erik Storbukås <storbukas.dev@...il.com>
To:     Neal Cardwell <ncardwell@...gle.com>
Cc:     LKML <linux-kernel@...r.kernel.org>,
        Netdev <netdev@...r.kernel.org>
Subject: Re: Get amount of fast retransmissions from TCP info

2017-04-24 23:00 GMT+02:00 Neal Cardwell <ncardwell@...gle.com>:
> On Mon, Apr 24, 2017 at 4:20 PM, Lars Erik Storbukås
> <storbukas.dev@...il.com> wrote:
>> 2017-04-24 21:42 GMT+02:00 Neal Cardwell <ncardwell@...gle.com>:
>>> On Mon, Apr 24, 2017 at 3:11 PM, Lars Erik Storbukås
>>> <storbukas.dev@...il.com> wrote:
>>>> I'm trying to get amount of congestion events in TCP caused by
>>>> DUPACK's (fast retransmissions), and can't seem to find any variable
>>>> in the TCP info struct which hold that value. There are three
>>>> variables in the TCP info struct that seem to hold similar congestion
>>>> values: __u8 tcpi_retransmits;__u32 tcpi_retrans; __u32
>>>> tcpi_total_retrans;
>>>>
>>>> Does anyone have any pointers on how to find this value in the TCP code?
>>>>
>>>> Please CC me personally if answering this question. Any help is
>>>> greatly appreciated.
>>>
>>> [I'm cc-ing the netdev list.]
>>>
>>> Do you need this per-socket? On a per-socket basis, I do not think
>>> there are separate totals for fast retransmits and timeout
>>> retransmits.
>>>
>>> If a global number is good enough, then you can get that number from
>>> the global network statistics. In "nstat" output they look like:
>>>
>>>   TcpExtTCPFastRetrans = packets sent in fast retransmit / fast recovery
>>>
>>>   TcpExtTCPSlowStartRetrans = packets sent in timeout recovery
>>>
>>> It sounds like TcpExtTCPFastRetrans is what you are after.
>>>
>>> Hope that helps,
>>> neal
>>
>> Thanks for your answer Neal.
>>
>> Yes, I need this information per-socket. What would be the most
>> appropriate place to update this value?
>
> Is this for a custom kernel you are building? Or are you proposing
> this for upstream?

This is currently for a custom kernel.

> IMHO the best place to add this for your custom kernel would be in
> _tcp_retransmit_skb() around the spot with the comment "Update global
> and local TCP statistics". Something like:
>
>   /* Update global and local TCP statistics. */
> ...
>   tp->total_retrans += segs;
>   if (icsk->icsk_ca_state == TCP_CA_Loss)
>     tp->slow_retrans += segs;
>   else
>     tp->fast_retrans += segs;
>

Excellent. That seems like a logical place.

>> If none of the variables (mentioned above) contain any value in
>> regards to fast retransmits, what does the different values represent?
>
> tcpi_retransmits: consecutive retransmits of lowest-sequence outstanding packet
>
> tcpi_retrans: retransmitted packets estimated to be in-flight in the network now
>
> tcpi_total_retrans: total number of retransmitted packets over the
> life of the connection
>
> Can you sketch out why you need to have separate counts for fast
> retransmits and timeout/slow-start retransmits?
>
> neal

I'm working on the implementation of a Deadline Aware, Less than Best
Effort framework proposed by David A. Hayes, David Ros, Andreas
Petlund. A framework for adding both LBE behaviour and awareness of
“soft” delivery deadlines to any congestion control (CC) algorithm,
whether loss-based, delay- based or explicit signaling-based. This
effectively allows it to turn an arbitrary CC protocol into a
scavenger protocol that dynamically adapts its sending rate to network
conditions and remaining time before the deadline, to balance
timeliness and transmission aggressiveness.

/Lars Erik