[dpdk-dev] [PATCH] vhost: fix virtio_net cache sharing of broadcast_rarp

[Maxime Coquelin]

On 03/17/2017 06:47 AM, Yuanhan Liu wrote:
> On Thu, Mar 16, 2017 at 10:10:05AM +0000, Kevin Traynor wrote:
>> On 03/16/2017 06:21 AM, Yuanhan Liu wrote:
>>> On Wed, Mar 15, 2017 at 07:10:49PM +0000, Kevin Traynor wrote:
>>>> The virtio_net structure is used in both enqueue and dequeue datapaths.
>>>> broadcast_rarp is checked with cmpset in the dequeue datapath regardless
>>>> of whether descriptors are available or not.
>>>>
>>>> It is observed in some cases where dequeue and enqueue are performed by
>>>> different cores and no packets are available on the dequeue datapath
>>>> (i.e. uni-directional traffic), the frequent checking of broadcast_rarp
>>>> in dequeue causes performance degradation for the enqueue datapath.
>>>>
>>>> In OVS the issue can cause a uni-directional performance drop of up to 15%.
>>>>
>>>> Fix that by moving broadcast_rarp to a different cache line in
>>>> virtio_net struct.
>>>
>>> Thanks, but I'm a bit confused. The drop looks like being caused by
>>> cache false sharing, but I don't see anything would lead to a false
>>> sharing. I mean, there is no write in the same cache line where the
>>> broadcast_rarp belongs. Or, the "volatile" type is the culprit here?
>>>
>>
>> Yes, the cmpset code uses cmpxchg and that performs a write regardless
>> of the result - it either writes the new value or back the old value.
>
> Oh, right, I missed this part!
>
>>> Talking about that, I had actually considered to turn "broadcast_rarp"
>>> to a simple "int" or "uint16_t" type, to make it more light weight.
>>> The reason I used atomic type is to exactly send one broadcast RARP
>>> packet once SEND_RARP request is recieved. Otherwise, we may send more
>>> than one RARP packet when MQ is invovled. But I think we don't have
>>> to be that accurate: it's tolerable when more RARP are sent. I saw 4
>>> SEND_RARP requests (aka 4 RARP packets) in the last time I tried
>>> vhost-user live migration after all. I don't quite remember why
>>> it was 4 though.
>>>
>>> That said, I think it also would resolve the performance issue if you
>>> change "rte_atomic16_t" to "uint16_t", without moving the place?
>>>
>>
>> Yes, that should work fine, with the side effect you mentioned of
>> possibly some more rarps - no big deal.
>>
>> I tested another solution also - as it is unlikely we would need to send
>> the broadcast_rarp, you can first read and only do the cmpset if it is
>> likely to succeed. This resolved the issue too.
>>
>> --- a/lib/librte_vhost/virtio_net.c
>> +++ b/lib/librte_vhost/virtio_net.c
>> @@ -1057,7 +1057,8 @@ static inline bool __attribute__((always_inline))
>>          *
>>          * Check user_send_rarp() for more information.
>>          */
>> -       if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
>> +       if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
>> +                       rte_atomic16_cmpset((volatile uint16_t *)
>>                                          &dev->broadcast_rarp.cnt, 1, 0))) {
>>                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
>>                 if (rarp_mbuf == NULL) {
>
> I'm okay with this one. It's simple and clean enough, that it could
> be picked to a stable release. Later, I'd like to send another patch
> to turn it to "uint16_t". Since it changes the behaviour a bit, it
> is not a good candidate for stable release.
>
> BTW, would you please include the root cause (false sharing) into
> your commit log?
And maybe also adds the info to the comment just above?
I will help people wondering why we read before cmpset.

Maxime