Re: NULL pointer dereference panic in stable (2.6.33.2), amd64

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 12 avril 2010 à 13:24 +0530, Krishna Kumar2 a écrit :

> If the dst got changed between call to vlan_dev_hwaccel_hard_start_xmit
> and it's call to dev_queue_xmit, that change to dst should have reset
> sk_tx_queue_mapping to -1 by calling sk_tx_queue_clear (assuming that I
> have changed in all paths, eg __sk_dst_reset), and thus result in a new
> mapping in dev_pick_tx. Would the patch hide the actual bug where we do
> not clear sk_tx_queue_mapping, eg __sk_dst_set does it? I agree the
> patch will fix the panic, but this check could be removed if the code
> which changes the dst is fixed to clear the mapping. I could check that
> if you think this assumption is correct.
> 

I believe you focus on another problem. I am not saying we dont have
another bug (forgetting to reset sk_dst_cache somewhere).

I am only saying that when we want to cache the queue number on a given
socket, we have to make sure current packet routing decision was taken
on same dst_entries than current and future ones. Denys hit the problem
because of long delays caused by traffic shaping.

So the cache renew must be safe, which I tried to fix.

You are saying that cache invalidation might be missing from some paths.
I dont think so because I took an extensive look at these spots when
working on yet another RCU conversion two days ago (sk_dst_lock becomes
a spinlock). This was fresh in my mind, this is why I probably found
Denys problem origin so quickly ;)