On 17-02-16 09:10 PM, Jason Wang wrote:
>
>
> On 2017年02月17日 12:53, John Fastabend wrote:
>> On 17-02-15 01:08 AM, Jason Wang wrote:
>>> We set queues before reset which will cause a crash[1]. This is
>>> because is_xdp_raw_buffer_queue() depends on the old xdp queue pairs
>>> number to do the correct detection. So fix this by:
>>>
>>> - set queues after reset, to keep the old vi->curr_queue_pairs. (in
>>> fact setting queues before reset does not works since after feature
>>> set, all queue pairs were enabled by default during reset).
>>> - change xdp_queue_pairs only after virtnet_reset() is succeed.
>>>
>>> [1]
>> I'm guessing this occurs when enabling XDP while receiving lots of traffic?
>
> I hit this then disabling XDP while receiving lots of traffic.
>
[...]
>>> + vi->xdp_queue_pairs = xdp_qp;
>> But xdp_queue_pairs is being used to detect if we should allocate the XDP
>> headroom. If we move it here do we have a set of buffers in the ring without
>> the proper headroom when we assign the xdp program below?
>
> Right, so how about passing xdp_queue_pairs as a parameter to virtnet_reset().
> Then virtnet_reset() can set it after _remove_vq_common() but before
> virtnet_restore_up()?
>
> Thanks
>
Sounds like a reasonable fix to me.
>>
>>> + }
>>> +
>>> + err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
>>> + if (err) {
>>> + dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
>>> + goto virtio_queue_err;
>>> }
>>> netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
>>> @@ -1844,17 +1844,18 @@ static int virtnet_xdp_set(struct net_device *dev,
>>> struct bpf_prog *prog)
>>> return 0;
>> Thanks,
>> John
>>
>
_______________________________________________
Virtualization mailing list
Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
