Linux Virtualization - Mar 2018 - [PATCH] vhost-net: add time limitation for tx polling(Internet mail)

On 2018?03?28? 12:01, haibinzhang(???) wrote:
> On 2018?03?27? 19:26, Jason wrote
> On 2018?03?27? 17:12, haibinzhang wrote:
>>> handle_tx() will delay rx for a long time when busy tx polling udp
packets
>>> with short length(ie: 1byte udp payload), because setting
VHOST_NET_WEIGHT
>>> takes into account only sent-bytes but no time.
>> Interesting.
>>
>> Looking at vhost_can_busy_poll() it tries to poke pending vhost work
and
>> exit the busy loop if it found one. So I believe something block the
>> work queuing. E.g did reverting
8241a1e466cd56e6c10472cac9c1ad4e54bc65db
>> fix the issue?
> "busy tx polling" means using netperf send udp packets with 1
bytes payload(total 47bytes frame lenght),
> and handle_tx() will be busy sending packets continuously.
>
>>>    It's not fair for handle_rx(),
>>> so needs to limit max time of tx polling.
>>>
>>> ---
>>>    drivers/vhost/net.c | 3 ++-
>>>    1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
>>> index 8139bc70ad7d..dc9218a3a75b 100644
>>> --- a/drivers/vhost/net.c
>>> +++ b/drivers/vhost/net.c
>>> @@ -473,6 +473,7 @@ static void handle_tx(struct vhost_net *net)
>>>    	struct socket *sock;
>>>    	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
>>>    	bool zcopy, zcopy_used;
>>> +	unsigned long start = jiffies;
>> Checking jiffies is tricky, need to convert it to ms or whatever
others.
>>
>>>    
>>>    	mutex_lock(&vq->mutex);
>>>    	sock = vq->private_data;
>>> @@ -580,7 +581,7 @@ static void handle_tx(struct vhost_net *net)
>>>    		else
>>>    			vhost_zerocopy_signal_used(net, vq);
>>>    		vhost_net_tx_packet(net);
>>> -		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
>>> +		if (unlikely(total_len >= VHOST_NET_WEIGHT) ||
unlikely(jiffies - start >= 1)) {
>> How value 1 is determined here? And we need a complete test to make
sure
>> this won't affect other use cases.
> We just want <1ms ping latency, but actually we are not sure what value
is reasonable.
> We have some test results using netperf before this patch as follow,
>
>      Udp payload    1byte    100bytes    1000bytes    1400bytes
>    Ping avg latency    25ms     10ms       2ms         1.5ms
>
> What is other testcases?
Something like https://patchwork.kernel.org/patch/10151645/.

Btw, you need use time_before() to properly handle jiffies overflow and 
I would also suggest you to try something like #packets limit (e.g 64).

For long term, we definitely need more worker threads.

Thanks
>
>> Another thought is introduce another limit of #packets, but this need
>> benchmark too.
>>
>> Thanks
>>
>>>    			vhost_poll_queue(&vq->poll);
>>>    			break;
>>>    		}
>>

On Wed, Mar 28, 2018 at 02:37:04PM +0800, Jason Wang
wrote:
> 
> 
> On 2018?03?28? 12:01, haibinzhang(???) wrote:
> > On 2018?03?27? 19:26, Jason wrote
> > On 2018?03?27? 17:12, haibinzhang wrote:
> > > > handle_tx() will delay rx for a long time when busy tx
polling udp packets
> > > > with short length(ie: 1byte udp payload), because setting
VHOST_NET_WEIGHT
> > > > takes into account only sent-bytes but no time.
> > > Interesting.
> > > 
> > > Looking at vhost_can_busy_poll() it tries to poke pending vhost
work and
> > > exit the busy loop if it found one. So I believe something block
the
> > > work queuing. E.g did reverting
8241a1e466cd56e6c10472cac9c1ad4e54bc65db
> > > fix the issue?
> > "busy tx polling" means using netperf send udp packets with
1 bytes payload(total 47bytes frame lenght),
> > and handle_tx() will be busy sending packets continuously.
> > 
> > > >    It's not fair for handle_rx(),
> > > > so needs to limit max time of tx polling.
> > > > 
> > > > ---
> > > >    drivers/vhost/net.c | 3 ++-
> > > >    1 file changed, 2 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> > > > index 8139bc70ad7d..dc9218a3a75b 100644
> > > > --- a/drivers/vhost/net.c
> > > > +++ b/drivers/vhost/net.c
> > > > @@ -473,6 +473,7 @@ static void handle_tx(struct vhost_net
*net)
> > > >    	struct socket *sock;
> > > >    	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
> > > >    	bool zcopy, zcopy_used;
> > > > +	unsigned long start = jiffies;
> > > Checking jiffies is tricky, need to convert it to ms or whatever
others.
> > > 
> > > >    	mutex_lock(&vq->mutex);
> > > >    	sock = vq->private_data;
> > > > @@ -580,7 +581,7 @@ static void handle_tx(struct vhost_net
*net)
> > > >    		else
> > > >    			vhost_zerocopy_signal_used(net, vq);
> > > >    		vhost_net_tx_packet(net);
> > > > -		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
> > > > +		if (unlikely(total_len >= VHOST_NET_WEIGHT) ||
unlikely(jiffies - start >= 1)) {
> > > How value 1 is determined here? And we need a complete test to
make sure
> > > this won't affect other use cases.
> > We just want <1ms ping latency, but actually we are not sure what
value is reasonable.
> > We have some test results using netperf before this patch as follow,
> > 
> >      Udp payload    1byte    100bytes    1000bytes    1400bytes
> >    Ping avg latency    25ms     10ms       2ms         1.5ms
> > 
> > What is other testcases?
> 
> Something like https://patchwork.kernel.org/patch/10151645/.
> 
> Btw, you need use time_before() to properly handle jiffies overflow and I
> would also suggest you to try something like #packets limit (e.g 64).
Maybe a ring size?
> For long term, we definitely need more worker threads.
> 
> Thanks
Only helps when you have spare CPUs.
> > 
> > > Another thought is introduce another limit of #packets, but this
need
> > > benchmark too.
> > > 
> > > Thanks
> > > 
> > > >    			vhost_poll_queue(&vq->poll);
> > > >    			break;
> > > >    		}
> > >

On 2018?03?28? 23:31, Michael S. Tsirkin wrote:
> On Wed, Mar 28, 2018 at 02:37:04PM +0800, Jason Wang wrote:
>>
>> On 2018?03?28? 12:01, haibinzhang(???) wrote:
>>> On 2018?03?27? 19:26, Jason wrote
>>> On 2018?03?27? 17:12, haibinzhang wrote:
>>>>> handle_tx() will delay rx for a long time when busy tx
polling udp packets
>>>>> with short length(ie: 1byte udp payload), because setting
VHOST_NET_WEIGHT
>>>>> takes into account only sent-bytes but no time.
>>>> Interesting.
>>>>
>>>> Looking at vhost_can_busy_poll() it tries to poke pending vhost
work and
>>>> exit the busy loop if it found one. So I believe something
block the
>>>> work queuing. E.g did reverting
8241a1e466cd56e6c10472cac9c1ad4e54bc65db
>>>> fix the issue?
>>> "busy tx polling" means using netperf send udp packets
with 1 bytes payload(total 47bytes frame lenght),
>>> and handle_tx() will be busy sending packets continuously.
>>>
>>>>>     It's not fair for handle_rx(),
>>>>> so needs to limit max time of tx polling.
>>>>>
>>>>> ---
>>>>>     drivers/vhost/net.c | 3 ++-
>>>>>     1 file changed, 2 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
>>>>> index 8139bc70ad7d..dc9218a3a75b 100644
>>>>> --- a/drivers/vhost/net.c
>>>>> +++ b/drivers/vhost/net.c
>>>>> @@ -473,6 +473,7 @@ static void handle_tx(struct vhost_net
*net)
>>>>>     	struct socket *sock;
>>>>>     	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
>>>>>     	bool zcopy, zcopy_used;
>>>>> +	unsigned long start = jiffies;
>>>> Checking jiffies is tricky, need to convert it to ms or
whatever others.
>>>>
>>>>>     	mutex_lock(&vq->mutex);
>>>>>     	sock = vq->private_data;
>>>>> @@ -580,7 +581,7 @@ static void handle_tx(struct vhost_net
*net)
>>>>>     		else
>>>>>     			vhost_zerocopy_signal_used(net, vq);
>>>>>     		vhost_net_tx_packet(net);
>>>>> -		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
>>>>> +		if (unlikely(total_len >= VHOST_NET_WEIGHT) ||
unlikely(jiffies - start >= 1)) {
>>>> How value 1 is determined here? And we need a complete test to
make sure
>>>> this won't affect other use cases.
>>> We just want <1ms ping latency, but actually we are not sure
what value is reasonable.
>>> We have some test results using netperf before this patch as
follow,
>>>
>>>       Udp payload    1byte    100bytes    1000bytes    1400bytes
>>>     Ping avg latency    25ms     10ms       2ms         1.5ms
>>>
>>> What is other testcases?
>> Something like https://patchwork.kernel.org/patch/10151645/.
>>
>> Btw, you need use time_before() to properly handle jiffies overflow and
I
>> would also suggest you to try something like #packets limit (e.g 64).
> Maybe a ring size?
Yes or a factor of ring size.
>
>> For long term, we definitely need more worker threads.
>>
>> Thanks
> Only helps when you have spare CPUs.
Right.

Thanks
>>>> Another thought is introduce another limit of #packets, but
this need
>>>> benchmark too.
>>>>
>>>> Thanks
>>>>
>>>>>     			vhost_poll_queue(&vq->poll);
>>>>>     			break;
>>>>>     		}