Re: [RFC PATCHv5 4/5] [TCP]: new LOST marker optimizations

[David Miller <davem@davemloft.net>]

From: "Ilpo_Järvinen" <ilpo.jarvinen@helsinki.fi>
Date: Fri, 16 Mar 2007 20:41:05 +0200

> 1) Couple of skb states are mutually exclusive
> 
> Skb cannot be in both S and L states at the same time, adding
> non-S nor L skb count (below highest sack) to that can be
> compared against fackets_out to see if anything below the
> current skb can still be marked with L or not. If they're
> equal (or fackets_out is smaller), the next skb will have L
> set for sure.
> 
> 2) Create a fastpath for the new LOST marker
> 
> The fastpath takes advantage of the fact that the latest ACK
> very likely contains the globally highest SACK block. By
> verifying that is larger than tp->reordering, whole SACK block
> can be skipped while counting for not-to-be marked skbs until
> tp->reordering of them is encountered. Since TCP now can know,
> that should block exists as highest, the marking can begin right
> below it. If the latest ACK does not contain a SACK block that
> reaches all the way to the highest_sack (e.g., due to
> reordering), a slow path is used.

Excellent observation. :)

> Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>

Patch applied, thanks.