Re: [RFC] net: release dst entry in dev_queue_xmit()

[Eric Dumazet <dada1@cosmosbay.com>]

Jarek Poplawski a écrit :
> On Wed, Mar 25, 2009 at 07:41:27PM +0100, Eric Dumazet wrote:
>> Jarek Poplawski a écrit :
>>> David Miller wrote, On 03/25/2009 08:17 AM:
>>>
>>>> From: Eric Dumazet <dada1@cosmosbay.com>
>>>> Date: Wed, 25 Mar 2009 08:13:30 +0100
>>>>
>>>>> If done in dev_hard_start_xmit(), skb could be requeued (because of
>>>>> NETDEV_TX_BUSY).  Then if requeued, maybe at this time, dst being
>>>>> NULL is not a problem ?
>>>> Usually it should be OK because the packet schedulers have
>>>> a sort-of one-behind state that allows them to reinsert
>>>> the SKB into their queue datastructures without reclassifying.
>>>
>>> Actually, since David has dumped requeuing there is no reinserting;
>>> this last one "requeued" skb is buffered at the top in q->gso_skb
>>> and waiting for better times.
>> Yes indeed, this is what I thought too, thanks Jarek.
> 
> Alas I'm a bit concerned with virtual devs, e.g. now I'm looking at
> xmits in macvlan and pppoe. Maybe this patch should exclude them?
> 

Yes, MACVLAN :) its macvlan_start_xmit() function calls
dev_queue_xmit(skb) again, so we go back to packet schedulers and
classifiers, they might need dst again :(

Only other potential problem I found was in 
drivers/net/appletalk/ipddp.c

static int ipddp_xmit(struct sk_buff *skb, struct net_device *dev)
{
__be32 paddr = ((struct rtable*)skb->dst)->rt_gateway;

Not sure this driver is still supported, or if this paddr can be found elsewhere...
 __sk_dst_get(skb->sk) ???