Re: [PATCH bpf-next v3] virtio_net: add XDP meta data support

[Jason Wang <jasowang@redhat.com>]

On 2019/7/10 上午4:03, Daniel Borkmann wrote:
> On 07/09/2019 05:04 AM, Jason Wang wrote:
>> On 2019/7/9 上午6:38, Daniel Borkmann wrote:
>>> On 07/02/2019 04:11 PM, Yuya Kusakabe wrote:
>>>> On 7/2/19 5:33 PM, Jason Wang wrote:
>>>>> On 2019/7/2 下午4:16, Yuya Kusakabe wrote:
>>>>>> This adds XDP meta data support to both receive_small() and
>>>>>> receive_mergeable().
>>>>>>
>>>>>> Fixes: de8f3a83b0a0 ("bpf: add meta pointer for direct access")
>>>>>> Signed-off-by: Yuya Kusakabe <yuya.kusakabe@gmail.com>
>>>>>> ---
>>>>>> v3:
>>>>>>     - fix preserve the vnet header in receive_small().
>>>>>> v2:
>>>>>>     - keep copy untouched in page_to_skb().
>>>>>>     - preserve the vnet header in receive_small().
>>>>>>     - fix indentation.
>>>>>> ---
>>>>>>     drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------
>>>>>>     1 file changed, 31 insertions(+), 14 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
>>>>>> index 4f3de0ac8b0b..03a1ae6fe267 100644
>>>>>> --- a/drivers/net/virtio_net.c
>>>>>> +++ b/drivers/net/virtio_net.c
>>>>>> @@ -371,7 +371,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
>>>>>>                        struct receive_queue *rq,
>>>>>>                        struct page *page, unsigned int offset,
>>>>>>                        unsigned int len, unsigned int truesize,
>>>>>> -                   bool hdr_valid)
>>>>>> +                   bool hdr_valid, unsigned int metasize)
>>>>>>     {
>>>>>>         struct sk_buff *skb;
>>>>>>         struct virtio_net_hdr_mrg_rxbuf *hdr;
>>>>>> @@ -393,7 +393,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
>>>>>>         else
>>>>>>             hdr_padded_len = sizeof(struct padded_vnet_hdr);
>>>>>>     -    if (hdr_valid)
>>>>>> +    if (hdr_valid && !metasize)
>>>>>>             memcpy(hdr, p, hdr_len);
>>>>>>           len -= hdr_len;
>>>>>> @@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
>>>>>>             copy = skb_tailroom(skb);
>>>>>>         skb_put_data(skb, p, copy);
>>>>>>     +    if (metasize) {
>>>>>> +        __skb_pull(skb, metasize);
>>>>>> +        skb_metadata_set(skb, metasize);
>>>>>> +    }
>>>>>> +
>>>>>>         len -= copy;
>>>>>>         offset += copy;
>>>>>>     @@ -644,6 +649,7 @@ static struct sk_buff *receive_small(struct net_device *dev,
>>>>>>         unsigned int delta = 0;
>>>>>>         struct page *xdp_page;
>>>>>>         int err;
>>>>>> +    unsigned int metasize = 0;
>>>>>>           len -= vi->hdr_len;
>>>>>>         stats->bytes += len;
>>>>>> @@ -683,10 +689,13 @@ static struct sk_buff *receive_small(struct net_device *dev,
>>>>>>               xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
>>>>>>             xdp.data = xdp.data_hard_start + xdp_headroom;
>>>>>> -        xdp_set_data_meta_invalid(&xdp);
>>>>>>             xdp.data_end = xdp.data + len;
>>>>>> +        xdp.data_meta = xdp.data;
>>>>>>             xdp.rxq = &rq->xdp_rxq;
>>>>>>             orig_data = xdp.data;
>>>>>> +        /* Copy the vnet header to the front of data_hard_start to avoid
>>>>>> +         * overwriting by XDP meta data */
>>>>>> +        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);
>>> I'm not fully sure if I'm following this one correctly, probably just missing
>>> something. Isn't the vnet header based on how we set up xdp.data_hard_start
>>> earlier already in front of it? Wouldn't we copy invalid data from xdp.data -
>>> vi->hdr_len into the vnet header at that point (given there can be up to 256
>>> bytes of headroom between the two)? If it's relative to xdp.data and headroom
>>> is >0, then BPF prog could otherwise mangle this; something doesn't add up to
>>> me here. Could you clarify? Thx
>> Vnet headr sits just in front of xdp.data not xdp.data_hard_start. So it could be overwrote by metadata, that's why we need a copy here.
> For the current code, you can adjust the xdp.data with a positive/negative offset
> already via bpf_xdp_adjust_head() helper. If vnet headr sits just in front of
> xdp.data, couldn't this be overridden today as well then? Anyway, just wondering
> how this is handled differently?


We will invalidate the vnet header in this case. But for the case of 
metadata adjustment without header adjustment, we want to seek a way to 
preserve that.

Thanks


>
> Thanks,
> Daniel