On Wed, 2006-08-09 at 15:19 +0200, Thomas Graf wrote:

* Daniel Phillips [off-list ref] 2006-08-08 22:47

quoted

David Miller wrote:

quoted

From: Daniel Phillips <redacted>

 >>Can you please characterize the conditions under which skb->dev changes

quoted

quoted

after the alloc?  Are there writings on this subtlety?

The packet scheduler and classifier can redirect packets to different
devices, and can the netfilter layer.

The setting of skb->dev is wholly transient and you cannot rely upon
it to be the same as when you set it on allocation.

Even simple things like the bonding device change skb->dev on every
receive.

Thankyou, this is easily fixed.

It's not that simple, in order to just fix the most obvious case
being packet forwarding when skb->dev changes its meaning from
device the packet is coming from to device the packet will be leaving
on is difficult.

You can't unreserve at that point so you need to keep the original
skb->dev. Since the packet is mostly likely queued before freeing
you will lose the refcnt on the original skb->dev. Keeping a
refcnt just for this memalloc stuff is out of question. Even keeping
the ifindex on a best effort basis is unlikely an option, sk_buff is
way overweight already.

I think Daniel was thinking of adding struct net_device *
sk_buff::alloc_dev,
I know I was after reading the first few mails. However if adding a
field 
there is strict no-no....

/me takes a look at struct sk_buff

Hmm, what does sk_buff::input_dev do? That seems to store the initial
device?