netdev - Re: TCP reordering and Multiple DL path

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 for Android: free password hash cracker in your pocket

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

Date:	Mon, 10 Mar 2014 09:15:08 -0700
From:	Eric Dumazet <eric.dumazet@...il.com>
To:	"Soum, Redouane" <redouane.soum@...el.com>
Cc:	"netdev@...r.kernel.org" <netdev@...r.kernel.org>
Subject: Re: TCP reordering and Multiple DL path

On Mon, 2014-03-10 at 15:44 +0000, Soum, Redouane wrote:
> Hi,
> 
> I have several path in DL between an Application Processor and a Modem
> using different network interfaces (USB CDC NCMs) with the same ip
> address (connected to the same network)
> 
> Having several network interfaces used in DL implies to have out of
> order packets, even if the ooo packets will most likely come close
> together.
> 
> Here is my understanding on the impact on TCP throughput :
> 
> -        OOO packets will lead to Dup acks.
> -        When sender receives 3 Dup acks it'll start fast recovery
> which will reduce TCP congestion window.
> 
> First question is my understanding correct ?
> 
> Assuming it is : 
> In order to solve that issue I am thinking to add a small delay before
> sending dup acks for OoO packets.
> The objective is to be sure that dup acks for OoO packets will only be
> sent for OoO packet due to different path in the network, and not
> because of the different network interfaces.
> 
> However I am not sure how complex it would be to implement and which
> part of the TCP stack need to be modified.
> Any advice to solve that issue ?

This is quite complex.

Note that we are working on a more generic solution, involving a high resolution timer,
for not only this case, but more generally :

Delayed ACK (application could tell TCP stack its max latency for RPC)

TSO autodefer bugs (currently we rely on future ACKS, while we should
not)

TCP_CORK_US (extend TCP auto cork with a usec timer)

Trick is to not add uncontrolled jitter on ACKs, otherwise we screw up
delay based CC.



--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html