Re: [PATCH] vhost: make vhost_zerocopy_callback more efficient by poll_queue base on vhost status

[Qin Chuanyu <qinchuanyu@huawei.com>]

On 2014/2/25 15:38, Jason Wang wrote:
> On 02/25/2014 02:55 PM, Qin Chuanyu wrote:
>> guest kick vhost base on vring flag status and get perfermance improved,
>> vhost_zerocopy_callback could do this in the same way, as
>> virtqueue_enable_cb need one more check after change the status of
>> avail_ring flags, vhost also do the same thing after vhost_enable_notify
>>
>> test result list as below:
>> guest and host: suse11sp3, netperf, intel 2.4GHz
>> +-------+----------+---------+----------+---------+
>> |       |   old              |   new              |
>> +-------+----------+---------+----------+---------+
>> |  UDP  |  Gbit/s  |   PPS   |  Gbit/s  |   PPS   |
>> |  256  | 0.74805  | 321309  | 0.77933  | 334743  |
>> |  512  |   1.42   | 328475  |   1.44   | 333550  |
>> | 1024  |   2.79   | 334426  |   2.81   | 336986  |
>> | 1460  |   3.71   | 316215  |   4.02   | 342325  |
>> +-------+----------+---------+----------+---------+
>
> Looks good, do you have cpu utilization number?
+------+----------+--------+----------+--------+--------+---------+
|      |             old              |            new            |
+------+----------+--------+----------+--------+--------+---------+
| UDP  |  Gbit/s  |  PPS   |CPU idle% | Gbit/s |   PPS  |CPU idle%|
| 256  | 0.74805  | 321309 |  87.16   | 0.77933| 334743 |  90.71  |
| 512  |   1.42   | 328475 |  87.03   |  1.44  | 333550 |  90.43  |
| 1024 |   2.79   | 334426 |  89.09   |  2.81  | 336986 |  89.55  |
| 1460 |   3.71   | 316215 |  87.53   |  4.02  | 342325 |  89.58  |
+------+----------+--------+----------+--------+--------+---------+
after change, less cpu has been used.
>>
>> Signed-off-by: Chuanyu Qin <qinchuanyu@huawei.com>
>> ---
>>   drivers/vhost/net.c |   10 +++++++++-
>>   1 files changed, 9 insertions(+), 1 deletions(-)
>>
>> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
>> index a0fa5de..9bc0a15 100644
>> --- a/drivers/vhost/net.c
>> +++ b/drivers/vhost/net.c
>> @@ -322,7 +322,9 @@ static void vhost_zerocopy_callback(struct
>> ubuf_info *ubuf, bool success)
>>        * (the value 16 here is more or less arbitrary, it's tuned to
>> trigger
>>        * less than 10% of times).
>>        */
>> -    if (cnt <= 1 || !(cnt % 16))
>> +    smp_rmb();
>
> Better add a comment to explain why this is needed.
>
> Looks like what you need is a smp_mb() here to make sure the len is
> updated before testing vq->used_flags?
I wanner make sure the used_flags is updated, is smp_rmb() enough?
or a smp_mb() is needed?
>> +    if ((!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
>> +            && (cnt <= 1 || !(cnt % 16)))
>>           vhost_poll_queue(&vq->poll);
>>
>>       rcu_read_unlock_bh();
>> @@ -386,6 +388,12 @@ static void handle_tx(struct vhost_net *net)
>>                   vhost_disable_notify(&net->dev, vq);
>>                   continue;
>>               }
>> +            /* there might skb been freed between last
>> +            * vhost_zerocopy_signal_used and vhost_enable_notify,
>> +            * so one more check is needed.
>> +            */
>> +            if (zcopy)
>> +                vhost_zerocopy_signal_used(net, vq);
>>               break;
>>           }
>>           if (in) {
>
>
> .
>