Re: [PATCH v2] ipv4: dst_entry leak in ip_append_data()

[Vasily Averin <vvs@parallels.com>]

On 15.10.2014 08:46, Eric Dumazet wrote:
> On Tue, 2014-10-14 at 08:57 +0400, Vasily Averin wrote:
>> v2: adjust the indentation of the arguments __ip_append_data() call
>>
>> Fixes: 2e77d89b2fa8 ("net: avoid a pair of dst_hold()/dst_release() in ip_append_data()")
>>
>> If sk_write_queue is empty ip_append_data() executes ip_setup_cork()
>> that "steals" dst entry from rt to cork. Later it calls __ip_append_data()
>> that creates skb and adds it to sk_write_queue.
>>
>> If skb was added successfully following ip_push_pending_frames() call
>> reassign dst entries from cork to skb, and kfree_skb frees dst_entry.
>>
>> However nobody frees stolen dst_entry if skb was not added into sk_write_queue.
> 
> I thought this was done by ip_flush_pending_frames() ?

Take look at ip_send_unicast_reply():

ip_flush_pending_frames() is not called if skb was not added to sk_write_queue.
And ip_rt_put() does not work, because dst entry was stolen in ip_setup_cork().

Probably it can happen in raw_sendmsg() and udp_sendmsg() too.