Re: [PATCH net] net: enforce xmit_recursion for devices with a queue

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

On 03/14/2019 10:40 AM, Sabrina Dubroca wrote:
> 2019-03-14, 07:56:10 -0700, Eric Dumazet wrote:
>>
>>
>> On 03/14/2019 07:15 AM, Sabrina Dubroca wrote:
>>> 2019-03-14, 05:58:03 -0700, Eric Dumazet wrote:
>>>>
>>>>
>>>> On 03/14/2019 03:15 AM, Sabrina Dubroca wrote:
>>>>> Commit 745e20f1b626 ("net: add a recursion limit in xmit path")
>>>>> introduced a recursion limit, but it only applies to devices without a
>>>>> queue. Virtual devices with a queue (either because they don't have
>>>>> the IFF_NO_QUEUE flag, or because the administrator added one) can
>>>>> still cause an unbounded recursion, via __dev_queue_xmit ->
>>>>> __dev_xmit_skb -> qdisc_run -> __qdisc_run -> qdisc_restart ->
>>>>> sch_direct_xmit -> dev_hard_start_xmit . Jianlin reported this in a
>>>>> setup with 16 gretap devices stacked on top of one another.
>>>>>
>>>>> This patch prevents the stack overflow by incrementing xmit_recursion in
>>>>> code paths that can call dev_hard_start_xmit() (like commit 745e20f1b626
>>>>> did). If the recursion limit is exceeded, the packet is enqueued and the
>>>>> qdisc is scheduled.
>>>>>
>>>>> Reported-by: Jianlin Shi <jishi@redhat.com>
>>>>> Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
>>>>> Reviewed-by: Stefano Brivio <sbrivio@redhat.com>
>>>>
>>>> Hi Sabrina, thanks for the patch.
>>>>
>>>> Can't we detect this in the control path instead ?
>>>
>>> I don't see how. You could have a perfectly reasonable set of gretap
>>> devices that trigger this situation from simply reshuffling the IP
>>> addresses:
>>>
>>> gretap$x remote 1.1.$((x-1)).{1,2}
>>> (all those addresses set on a single veth device)
>>>
>>> Then you move those addresses to the corresponding device
>>> (1.1.${x}.{1,2} on gretap$x), and your machine crashes.
>>>
>>
>> If this only can be done with gretap, why gretap cant implement the protection,
>> outside of the fast path ?
> 
> It's not just gretap. VXLAN will do the same as long as you add a
> qdisc. I expect other types of tunnels to behave like that.
> 

It might make sense to add a helper using dev_queue_xmit()
for tunnel users.

Then remove the xmit recursion stuff out of the dev_queue_xmit()

Lets make the fast path fast again.