On 11/9/20 7:28 PM, Willem de Bruijn wrote:
> On Mon, Nov 9, 2020 at 12:37 PM Eric Dumazet <eric.duma...@gmail.com> wrote:
>>
>>
>>
>> On 11/9/20 5:56 PM, Alexander Lobakin wrote:
>>> While testing UDP GSO fraglists forwarding through driver that uses
>>> Fast GRO (via napi_gro_frags()), I was observing lots of out-of-order
>>> iperf packets:
>>>
>>> [ ID] Interval Transfer Bitrate Jitter
>>> [SUM] 0.0-40.0 sec 12106 datagrams received out-of-order
>>>
>>> Simple switch to napi_gro_receive() any other method without frag0
>>> shortcut completely resolved them.
>>>
>>> I've found that UDP GRO uses udp_hdr(skb) in its .gro_receive()
>>> callback. While it's probably OK for non-frag0 paths (when all
>>> headers or even the entire frame are already in skb->data), this
>>> inline points to junk when using Fast GRO (napi_gro_frags() or
>>> napi_gro_receive() with only Ethernet header in skb->data and all
>>> the rest in shinfo->frags) and breaks GRO packet compilation and
>>> the packet flow itself.
>>> To support both modes, skb_gro_header_fast() + skb_gro_header_slow()
>>> are typically used. UDP even has an inline helper that makes use of
>>> them, udp_gro_udphdr(). Use that instead of troublemaking udp_hdr()
>>> to get rid of the out-of-order delivers.
>>>
>>> Present since the introduction of plain UDP GRO in 5.0-rc1.
>>>
>>> Since v1 [1]:
>>> - added a NULL pointer check for "uh" as suggested by Willem.
>>>
>>> [1]
>>> https://lore.kernel.org/netdev/yazu6gezbdpyzmdmwjirxdx7b4sualpdg68adzy...@cp4-web-034.plabs.ch
>>>
>>> Fixes: e20cf8d3f1f7 ("udp: implement GRO for plain UDP sockets.")
>>> Signed-off-by: Alexander Lobakin <aloba...@pm.me>
>>> ---
>>> net/ipv4/udp_offload.c | 7 ++++++-
>>> 1 file changed, 6 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
>>> index e67a66fbf27b..7f6bd221880a 100644
>>> --- a/net/ipv4/udp_offload.c
>>> +++ b/net/ipv4/udp_offload.c
>>> @@ -366,13 +366,18 @@ static struct sk_buff *udp4_ufo_fragment(struct
>>> sk_buff *skb,
>>> static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
>>> struct sk_buff *skb)
>>> {
>>> - struct udphdr *uh = udp_hdr(skb);
>>> + struct udphdr *uh = udp_gro_udphdr(skb);
>>> struct sk_buff *pp = NULL;
>>> struct udphdr *uh2;
>>> struct sk_buff *p;
>>> unsigned int ulen;
>>> int ret = 0;
>>>
>>> + if (unlikely(!uh)) {
>>
>> How uh could be NULL here ?
>>
>> My understanding is that udp_gro_receive() is called
>> only after udp4_gro_receive() or udp6_gro_receive()
>> validated that udp_gro_udphdr(skb) was not NULL.
>
> Oh indeed. This has already been checked before.
>
>>> + NAPI_GRO_CB(skb)->flush = 1;
>>> + return NULL;
>>> + }
>>> +
>>> /* requires non zero csum, for symmetry with GSO */
>>> if (!uh->check) {
>>> NAPI_GRO_CB(skb)->flush = 1;
>>>
>>
>> Why uh2 is left unchanged ?
>>
>> uh2 = udp_hdr(p);
>
> Isn't that the same as th2 = tcp_hdr(p) in tcp_gro_receive? no frag0
> optimization to worry about for packets on the list.
My feeling was that tcp_gro_receive() is terminal in the GRO stack.
While UDP could be encapsulated in UDP :)
I guess we do not support this yet.
Years ago we made sure to propagate the current header offset into GRO stack
(when we added SIT/IPIP/GRE support to GRO)
299603e8370a93dd5d8e8d800f0dff1ce2c53d36 ("net-gro: Prepare GRO stack for the
upcoming tunneling support")
udp_hdr() is using transport header, which is unique for skb "on the list"
