netdev
[Top] [All Lists]

Re: [PATCH] pkt_cls.h incompatiables

To: YOSHIFUJI Hideaki / 吉藤英明 <yoshfuji@xxxxxxxxxxxxxx>
Subject: Re: [PATCH] pkt_cls.h incompatiables
From: Jamal Hadi Salim <hadi@xxxxxxxx>
Date: 23 Jul 2004 15:57:49 -0400
Cc: shemminger@xxxxxxxx, "David S. Miller" <davem@xxxxxxxxxx>, arekm@xxxxxxxxxxxxx, netdev@xxxxxxxxxxx
In-reply-to: <20040723.110007.27520072.yoshfuji@xxxxxxxxxxxxxx>
Organization: ZNYX Networks
References: <20040722134522.4e7e0b07@xxxxxxxxxxxxxxxxxxxxx> <20040722.200426.99255296.yoshfuji@xxxxxxxxxxxxxx> <1090593676.1128.25.camel@xxxxxxxxxxxxxxxx> <20040723.110007.27520072.yoshfuji@xxxxxxxxxxxxxx>
Reply-to: hadi@xxxxxxxx
Sender: netdev-bounce@xxxxxxxxxxx
On Fri, 2004-07-23 at 11:00, YOSHIFUJI Hideaki / 吉藤英明 wrote:

> 
> BTW, what's for rcnt etc.? I don't see the point. 
> 
> They're not (effectively) used in kernel.

It is transported to user space for analysis of how to reorganize
hash table optimally.

> I'd suggest to remove these things and to maintain the original ABI.
> 
> If rcnt etc. are for other purposes, such as statistics for userspace,
> please allocate another structure / interface for it.

I could send a separate TLV, i dont think this will resolve the problem
of ABI breakage but it will be more of a cleans. Part of the counters i need 
for analysis is in the
key structure.

> (And... cheking size is too strict;
> we need to relax it to accept old binaries if we add something
> at the tail of structure.)

A lot of size checks for every structure at both kernel/userspace is
needed. Not much is done today as is - which means old binaries may
break when modified structures are transported to user space.
I think one of the main challenges with the u32 structures is the keys[]
table. In the future if we could make that a TLV it would help. 
We need not just be backward compatible but also forward compatible.

I will do some testing when i get back next week to see if current patch
breaks anything.

cheers,
jamal



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