netdev
[Top] [All Lists]

Re: New net features for added performance

To: jgarzik@xxxxxxxxxxxxxxxx (Jeff Garzik)
Subject: Re: New net features for added performance
From: kuznet@xxxxxxxxxxxxx
Date: Tue, 27 Feb 2001 22:14:55 +0300 (MSK)
Cc: netdev@xxxxxxxxxxx
In-reply-to: <3A9AEFAF.1DC89A8A@xxxxxxxxxxxxxxxx> from "Jeff Garzik" at Feb 27, 1 03:15:07 am
Sender: owner-netdev@xxxxxxxxxxx
Hello!

> I only want to know if more are coming, not actually pass multiples..

BTW trying to strain brains how could driver to use this...

TX irq mitigation is maximally aggressive even without this,
such information is simply useless for mitigation.

Merging index updates? Well, it is too easy to do again
(for drivers using tx spinlock).
Just do not tell driver about new index until at least one of two conditions
happen:
1. >N frames are queued at driver (in hard_start_xmit)
2. <M frames are pushed to device (in tx interrupt).

That's all, you cannot win more using only "more" flag
and not pushing real vector of skbs to hard_start_xmit.

Alexey

<Prev in Thread] Current Thread [Next in Thread>