Re: [PATCH 2/2 net-next] tcp: sk_add_backlog() is too agressive for TCP

[Eric Dumazet <eric.dumazet@gmail.com>]

On Tue, 2012-04-24 at 11:01 +0300, Ilpo Järvinen wrote:
> On Tue, 24 Apr 2012, Eric Dumazet wrote:
> 
> > On Mon, 2012-04-23 at 22:37 +0200, Eric Dumazet wrote:
> > 
> > > We could try to coalesce ACKs before backlogging them. I'll work on
> > > this.
> > 
> > I did an experiment, and found a basic coalescing was not working in
> > case of packet loss and SACK storm.
> 
> ...That case might also have some performance issues at the receiver end 
> when a hole is filled and TCP pushes stuff higher up.
> 
> > Doing a smart coalescing in this case sounds really complex.
> 
> Why's that? ...We'd compare options 32-bit at a time (like you already do 
> anyway) and if we find difference we check the previous bits to validate 
> it's a SACK option (the changing one should be in the first start-end 
> pair). ...As long as there's no hole in every other segment we'd be 
> winners I think.
> 
> > Should we really continue this way ? 
> 
> Why not, but wouldn't it be nicer to coalesce them already in GRO below 
> with an assumption that GRO is likely to find some "mss" equivivalent 
> which tells the gap between consecutive ACK (or even SACK) seqnos?
> 
> I've been long thinking that it would be nice to run offloading for ACKs 
> too, and possibly even for SACKs, in both ends, although that might not be 
> possible with other than GSO/GRO, at least atm.
> 
> 


GRO doesnt coalesce pure acks, thats part of GRO contract.

By the way, I find GRO less and less attractive, if we have fragged skbs
provided by drivers, we can do the GRO almost for free in tcp stack,
instead of very complex/duplicated logic before tcp stack.

http://git.kernel.org/?p=linux/kernel/git/davem/net-next.git;a=commit;h=1402d366019fedaa2b024f2bac06b7cc9a8782e1

Doing this coalescing in tcp stack solves many problems GRO is unable to
address, like reordering...