[PATCH] Convert pppoe to new style protocol

[PATCH] Convert pppoe to new style protocol

[Andi Kleen]

Convert pppoe to a new style protocol, otherwise it is unusable 
on SMP compiled kernels because of the printk in deliver*old_ones.
It works fine here, but only tested on a UP box.

I checked with paulus and he confirmed that ppp_input is safe
to be called with the new softirq semantics. The PPP path also 
appears to work with shared skbs (never modify skb data, 
not 100% audited however) but not with non linear skbs.

The pppoe code itself is safe because it never uses timers
or does anything complicated.

If someone knows a place in ppp_input that modifies the data 
area then please contain now so that a skb_copy can be added.
 
-Andi

diff -u linux-2.5.72-work/drivers/net/pppoe.c-o linux-2.5.72-work/drivers/net/pppoe.c
--- linux-2.5.72-work/drivers/net/pppoe.c-o	2003-06-14 23:42:51.000000000 +0200
+++ linux-2.5.72-work/drivers/net/pppoe.c	2003-06-17 23:15:11.000000000 +0200
@@ -348,6 +348,9 @@
 	struct pppox_opt *po = pppox_sk(sk);
 	struct pppox_opt *relay_po = NULL;
 
+	if (skb_is_nonlinear(skb) && skb_linearize(skb, GFP_ATOMIC))
+		goto abort_kfree;
+
 	if (sk->sk_state & PPPOX_BOUND) {
 		skb_pull(skb, sizeof(struct pppoe_hdr));
 		ppp_input(&po->chan, skb);
@@ -463,11 +466,13 @@
 struct packet_type pppoes_ptype = {
 	.type	= __constant_htons(ETH_P_PPP_SES),
 	.func	= pppoe_rcv,
+	.data =	(void*) 1,
 };
 
 struct packet_type pppoed_ptype = {
 	.type	= __constant_htons(ETH_P_PPP_DISC),
 	.func	= pppoe_disc_rcv,
+	.data =	(void*) 1,
 };
 
 /***********************************************************************

Re: [PATCH] Convert pppoe to new style protocol

[David S. Miller]

   From: Andi Kleen <ak@muc.de>
   Date: Wed, 18 Jun 2003 00:04:20 +0200

   Convert pppoe to a new style protocol, otherwise it is unusable 
   on SMP compiled kernels because of the printk in deliver*old_ones.
   It works fine here, but only tested on a UP box.

Please don't add new skb_linearize() users, I'm trying
to make that only local to net/core/dev.c

Otherwise I'm fine with your patch.

Re: [PATCH] Convert pppoe to new style protocol

[Andi Kleen]

On Tue, Jun 17, 2003 at 03:07:51PM -0700, David S. Miller wrote:
>    From: Andi Kleen <ak@muc.de>
>    Date: Wed, 18 Jun 2003 00:04:20 +0200
> 
>    Convert pppoe to a new style protocol, otherwise it is unusable 
>    on SMP compiled kernels because of the printk in deliver*old_ones.
>    It works fine here, but only tested on a UP box.
> 
> Please don't add new skb_linearize() users, I'm trying
> to make that only local to net/core/dev.c

Then offer a flag or something for protocols that also doesn't trigger
printks ?


> 
> Otherwise I'm fine with your patch.

How else should it be done? 

I'm not going through the whole PPP layer now to audit it for non linear
skb cleanliness...

-Andi

Re: [PATCH] Convert pppoe to new style protocol

[Werner Almesberger]

On Tue, Jun 17, 2003 at 03:07:51PM -0700, David S. Miller wrote:
> Please don't add new skb_linearize() users, I'm trying
> to make that only local to net/core/dev.c

I know you find documentation unmanly, but maybe ...

- Werner

--- skbuff.h.orig	Wed Jun 18 09:13:57 2003
+++ skbuff.h	Wed Jun 18 09:15:05 2003
@@ -1126,6 +1126,9 @@
  *
  *	If there is no free memory -ENOMEM is returned, otherwise zero
  *	is returned and the old skb data released.
+ *
+ *	DO NOT USE THIS IN NEW CODE ! skb_linearize will be for internal
+ *	use by net/core/dev.c only.
  */
 int skb_linearize(struct sk_buff *skb, int gfp);
 
-- 
  _________________________________________________________________________
 / Werner Almesberger, Buenos Aires, Argentina         wa@almesberger.net /
/_http://www.almesberger.net/____________________________________________/

Re: [PATCH] Convert pppoe to new style protocol

[Andi Kleen]

On Wed, Jun 18, 2003 at 09:21:26AM -0300, Werner Almesberger wrote:
> On Tue, Jun 17, 2003 at 03:07:51PM -0700, David S. Miller wrote:
> > Please don't add new skb_linearize() users, I'm trying
> > to make that only local to net/core/dev.c
> 
> I know you find documentation unmanly, but maybe ...

But what is the replacement? For moving over whole subsystems to 
non linear skbs it's a bit late in the release...

And the "legacy protocol" setting cannot be used anymore because
it triggers a printk for every packet on a SMP kernel.

-Andi

Re: [PATCH] Convert pppoe to new style protocol

[David S. Miller]

   From: Andi Kleen <ak@colin2.muc.de>
   Date: 18 Jun 2003 14:51:59 +0200,Wed, 18 Jun 2003 14:51:59 +0200

   On Wed, Jun 18, 2003 at 09:21:26AM -0300, Werner Almesberger wrote:
   > I know you find documentation unmanly, but maybe ...
   
   But what is the replacement? For moving over whole subsystems to 
   non linear skbs it's a bit late in the release...
   
We did it for all of netfilter just the other week, networking
development still continues, it's not too late.

Stop looking for excuses and just do the work Andi.

Re: [PATCH] Convert pppoe to new style protocol

[David S. Miller]

   From: Werner Almesberger <wa@almesberger.net>
   Date: Wed, 18 Jun 2003 09:21:26 -0300

   + *
   + *	DO NOT USE THIS IN NEW CODE ! skb_linearize will be for internal
   + *	use by net/core/dev.c only.
     */
    int skb_linearize(struct sk_buff *skb, int gfp);

I have a better idea, I just added the __deprecated tag
to this function declaration :-)

Re: [PATCH] Convert pppoe to new style protocol

[David S. Miller]

   From: Andi Kleen <ak@suse.de>
   Date: Wed, 18 Jun 2003 10:01:18 +0200

   I'm not going through the whole PPP layer now to audit it for non
   linear skb cleanliness...
   
Then you're not making it a compatible protocol.
And the printk will remain until someone fixes it.

If I could get Rusty to fixup _ALL_ of netfilter, I can get
you to fix just _ONE_ thing.