[ovs-dev] [PATCH] openvswitch: Orphan frags before sending to userspace via Netlink to avoid guest stall

[Pravin Shelar]

On Fri, Mar 7, 2014 at 10:05 AM, Thomas Graf <tgraf at redhat.com> wrote:
> On 03/07/2014 06:19 PM, Pravin Shelar wrote:
>>
>> On Fri, Mar 7, 2014 at 7:58 AM, Thomas Graf <tgraf at redhat.com> wrote:
>>>
>>> On 03/07/2014 05:46 AM, Pravin Shelar wrote:
>>>>
>>>>
>>>> But I found bug in datapath user-space queue code. I am not sure how
>>>> this can work with skb fragments and MMAP-netlink socket.
>>>> Here is what happens, OVS allocates netlink skb and adds fragments to
>>>> skb using skb_zero_copy(), then calls genlmsg_unicast().
>>>> But if netlink sock is mmped then netlink-send queues netlink
>>>> allocated skb->head (linear data of skb) and ignore skb frags.
>>>>
>>>> Currently this is not problem with OVS vswitchd since it does not use
>>>> netlink MMAP sockets. But if vswitchd stats using MMAP-netlink socket,
>>>> it can break it.
>>>
>>>
>>>
>>> The secret is out ;-)
>>>
>>> I was very surprised too when I noticed that it worked. It's not just
>>> OVS, it's nfqueue as well. The reason is that an netlink mmaped skb is
>>> setup with a giant tailroom in netlink_ring_setup_skb():
>>>
>>>          skb->end        = skb->tail + size;
>>>
>> For OVS use-case, the size is linear part of skb. so I think for
>> mmap-netlink socket it will fail.
>
>
> Could you rephrase? I'm not sure I understand correctly.
>
I meant OVS (queue_userspace_packet) passes skb_zerocopy_headlen() to
genlmsg_new_unicast().

> The tailroom size equals to the configured frame payload size of
> the ring buffer. So as long as the frame size chosen is large
> enough to hold whatever pieces comes out of skb_gso_segment() we are
> fine. That said, I agree that we should fix this properly before we
> enable mmap on the OVS user space side.

I see. So this work assuming that userspace sets
nl_mmap_req.nm_frame_size large enough.