Re: Do you know about skb_header_pointer() ?

Re: Do you know about skb_header_pointer() ?

[Harald Welte]

[-- Attachment #1: Type: text/plain, Size: 1770 bytes --]

[Cc'ing davem]

On Thu, Aug 19, 2004 at 04:21:19PM +0900, Yasuyuki Kozakai wrote:

> According to a message in "netdev" mailing list, skb_header_pointer() will be
> introduced in mainline kernel. (see attached message)

yes, David Miller did so.  

If I remember correctly, his reasoning for skb_header_pointer() was that
even though we are not guaranteed a completely linear transport layer
header, it doesn't really happen at any point in the network stack at
this time.

I personally still like Yasuyuki's implementation[1] better, especially as
long as I'm not convinced that even in the event of IPv6 extension
headers, tunnels, ipsec, whaetever we're can still assume that
everything up to (and including) tcp option headers is linear.

And I still owe Dave some profiles on whether skb_header_pointer() as
opoosed to skb_copy_bits() actually still shows up in the profiles or not.

I'll do this _right now_ before some customer starts to annoy me again
with some stupid problem ;)

> If I remember correctly, this approach was rejected in the result of our
> discussions. And I sent a patch to this mailing list at 2004/08/08.

Yes, I thought your approach is better ;)  But maybe I'm just
overgeneralizing or trying to be too perfect.  Anyway, in the end it's
Dave's call.

[1]
http://lists.netfilter.org/pipermail/netfilter-devel/2004-August/016277.html

-- 
- Harald Welte <laforge@netfilter.org>             http://www.netfilter.org/
============================================================================
  "Fragmentation is like classful addressing -- an interesting early
   architectural error that shows how much experimentation was going
   on while IP was being designed."                    -- Paul Vixie

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: Do you know about skb_header_pointer() ?

[David S. Miller]

On Thu, 19 Aug 2004 11:44:01 +0200
Harald Welte <laforge@netfilter.org> wrote:

> I personally still like Yasuyuki's implementation[1] better,

Thanks for letting me know about this after we agreed upon
skb_header_pointer() (or somthing like it) at the networking
conference _AND_ I did the full implementation for you Harald.

Or did you just learn about Ysauyuki's work recently?

Re: Do you know about skb_header_pointer() ?

[David S. Miller]

On Thu, 19 Aug 2004 11:44:01 +0200
Harald Welte <laforge@netfilter.org> wrote:

> http://lists.netfilter.org/pipermail/netfilter-devel/2004-August/016277.html

I totally disagree with this change.

The goal was to eliminate any and all linearization done by any
part of the networking, especially on output paths.

Re: Do you know about skb_header_pointer() ?

[Harald Welte]

[-- Attachment #1: Type: text/plain, Size: 1107 bytes --]

On Thu, Aug 19, 2004 at 07:16:18AM -0700, David S. Miller wrote:
> On Thu, 19 Aug 2004 11:44:01 +0200
> Harald Welte <laforge@netfilter.org> wrote:
> 
> > I personally still like Yasuyuki's implementation[1] better,
> 
> Thanks for letting me know about this after we agreed upon
> skb_header_pointer() (or somthing like it) at the networking
> conference _AND_ I did the full implementation for you Harald.

No offence meant, Dave :)  I am grateful that you did this work, please
don't get me wrong.

> Or did you just learn about Ysauyuki's work recently?

Yasuyiki's first implementation dates July 29th.  
The one that I'm referring to was a patch from August 8th.

The netconf discussion predates any of this.

-- 
- Harald Welte <laforge@netfilter.org>             http://www.netfilter.org/
============================================================================
  "Fragmentation is like classful addressing -- an interesting early
   architectural error that shows how much experimentation was going
   on while IP was being designed."                    -- Paul Vixie

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: Do you know about skb_header_pointer() ?

[Harald Welte]

[-- Attachment #1: Type: text/plain, Size: 1237 bytes --]

On Thu, Aug 19, 2004 at 07:17:50AM -0700, David S. Miller wrote:
> On Thu, 19 Aug 2004 11:44:01 +0200
> Harald Welte <laforge@netfilter.org> wrote:
> 
> > http://lists.netfilter.org/pipermail/netfilter-devel/2004-August/016277.html
> 
> I totally disagree with this change.

Ok, that's fine with me.   Let's stay with the skb_header_pointer(),
which by the way [as expected] is not visible on my profile runs.

Still, I am curious to know if anyone on one of the Cc'ed mailinglists
can imagine a case where parts of the header up and including the tcp
options could be non-linear.

And no, I do not want to know this in order to say that
skb_heder_pointer() is bad.  I just want to know, because then we might
still need some magic for those cases in order to avoid a case where we
end up with per-rule copies or similar stuff again.

Thanks!

-- 
- Harald Welte <laforge@netfilter.org>             http://www.netfilter.org/
============================================================================
  "Fragmentation is like classful addressing -- an interesting early
   architectural error that shows how much experimentation was going
   on while IP was being designed."                    -- Paul Vixie

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: Do you know about skb_header_pointer() ?

[David S. Miller]

On Thu, 19 Aug 2004 19:00:40 +0200
Harald Welte <laforge@netfilter.org> wrote:

> can imagine a case where parts of the header up and including the tcp
> options could be non-linear.

A driver could do this, by putting all receive SKB data into
pages.

I am not aware of any that currently do this.

Such a driver would have to be very careful, as things like
the ethernet header parsing code on receive is not ready
for such things.  Things like pskb_may_pull() calls would
need to be added to that code for this to work.