net/sched: act_police: more accurate MTU policing

net/sched: act_police: more accurate MTU policing

[Davide Caratti]

hello,

Ilya reports bad TCP throughput when GSO packets hit an OVS rule that does
tc MTU policing. According to his observations [1], the problem is fixed
by upstream commit 4ddc844eb81d ("net/sched: act_police: more accurate MTU
policing"). Can we queue this commit for inclusion in stable trees?

thanks!
-- 
davide


[1] https://mail.openvswitch.org/pipermail/ovs-dev/2022-June/394802.html

Re: net/sched: act_police: more accurate MTU policing

[Greg KH]

On Fri, Jun 10, 2022 at 04:58:37PM +0200, Davide Caratti wrote:
> hello,
> 
> Ilya reports bad TCP throughput when GSO packets hit an OVS rule that does
> tc MTU policing. According to his observations [1], the problem is fixed
> by upstream commit 4ddc844eb81d ("net/sched: act_police: more accurate MTU
> policing"). Can we queue this commit for inclusion in stable trees?

Did you test this, and what kernel(s) do you want it applied to?

thanks,

greg k-h

Re: net/sched: act_police: more accurate MTU policing

[Davide Caratti]

hello Greg,

thanks for looking at this!

On Fri, Jun 10, 2022 at 05:09:21PM +0200, Greg KH wrote:
> On Fri, Jun 10, 2022 at 04:58:37PM +0200, Davide Caratti wrote:
> > hello,
> > 
> > Ilya reports bad TCP throughput when GSO packets hit an OVS rule that does
> > tc MTU policing. According to his observations [1], the problem is fixed
> > by upstream commit 4ddc844eb81d ("net/sched: act_police: more accurate MTU
> > policing"). Can we queue this commit for inclusion in stable trees?
> 
> Did you test this,

I tested it on upstream, RHEL8 and RHEL9 kernels. BTW, the kselftest I included
in the commit only verifies the correct setting for the MTU threshold, not
the GSO problem (to test GSO, we should use netperf / iperf3 rather than
mausezahn to generate traffic).

> and what kernel(s) do you want it applied to?

the reported bug is in act_police since the very beginning; however, the
patch should apply cleanly at least on 5.x kernels. On older ones, there
might be a small conflict due to lack of RCU-ification of struct
tcf_police_params.
A conflict that gets fixed easily, but in case we need it I volunteer to
write a patch for kernels older than 4.20. @Ilya, what is the
minimum kernel usable for openvswitch with MTU policing?

-- 
davide

Re: net/sched: act_police: more accurate MTU policing

[Ilya Maximets]

On 6/10/22 19:08, Davide Caratti wrote:
> hello Greg,
> 
> thanks for looking at this!
> 
> On Fri, Jun 10, 2022 at 05:09:21PM +0200, Greg KH wrote:
>> On Fri, Jun 10, 2022 at 04:58:37PM +0200, Davide Caratti wrote:
>>> hello,
>>>
>>> Ilya reports bad TCP throughput when GSO packets hit an OVS rule that does
>>> tc MTU policing. According to his observations [1], the problem is fixed
>>> by upstream commit 4ddc844eb81d ("net/sched: act_police: more accurate MTU
>>> policing"). Can we queue this commit for inclusion in stable trees?
>>
>> Did you test this,
> 
> I tested it on upstream, RHEL8 and RHEL9 kernels. BTW, the kselftest I included
> in the commit only verifies the correct setting for the MTU threshold, not
> the GSO problem (to test GSO, we should use netperf / iperf3 rather than
> mausezahn to generate traffic).
> 
>> and what kernel(s) do you want it applied to?
> 
> the reported bug is in act_police since the very beginning; however, the
> patch should apply cleanly at least on 5.x kernels. On older ones, there
> might be a small conflict due to lack of RCU-ification of struct
> tcf_police_params.
> A conflict that gets fixed easily, but in case we need it I volunteer to
> write a patch for kernels older than 4.20. @Ilya, what is the
> minimum kernel usable for openvswitch with MTU policing?

act_police MTU check is supposed to be an alternative for OVS' check_pkt_len()
action which was introduced in 5.2.  So, backport to 5.2+ should be enough.

Best regards, Ilya Maximets.

Re: net/sched: act_police: more accurate MTU policing

[Greg KH]

On Fri, Jun 10, 2022 at 07:08:48PM +0200, Davide Caratti wrote:
> hello Greg,
> 
> thanks for looking at this!
> 
> On Fri, Jun 10, 2022 at 05:09:21PM +0200, Greg KH wrote:
> > On Fri, Jun 10, 2022 at 04:58:37PM +0200, Davide Caratti wrote:
> > > hello,
> > > 
> > > Ilya reports bad TCP throughput when GSO packets hit an OVS rule that does
> > > tc MTU policing. According to his observations [1], the problem is fixed
> > > by upstream commit 4ddc844eb81d ("net/sched: act_police: more accurate MTU
> > > policing"). Can we queue this commit for inclusion in stable trees?
> > 
> > Did you test this,
> 
> I tested it on upstream, RHEL8 and RHEL9 kernels. BTW, the kselftest I included
> in the commit only verifies the correct setting for the MTU threshold, not
> the GSO problem (to test GSO, we should use netperf / iperf3 rather than
> mausezahn to generate traffic).
> 
> > and what kernel(s) do you want it applied to?
> 
> the reported bug is in act_police since the very beginning; however, the
> patch should apply cleanly at least on 5.x kernels. On older ones, there
> might be a small conflict due to lack of RCU-ification of struct
> tcf_police_params.
> A conflict that gets fixed easily, but in case we need it I volunteer to
> write a patch for kernels older than 4.20. @Ilya, what is the
> minimum kernel usable for openvswitch with MTU policing?
> 

It does not apply to 5.10 or earlier, so please provide a working
backport for those kernels if you wish it to be applied there.

thanks,

greg k-h

Re: net/sched: act_police: more accurate MTU policing

[Davide Caratti]

hello Greg,

On Mon, Jun 13, 2022 at 11:11:18AM +0200, Greg KH wrote:
> On Fri, Jun 10, 2022 at 07:08:48PM +0200, Davide Caratti wrote:
> > hello Greg,

[...]

> > > and what kernel(s) do you want it applied to?
> > 
> > the reported bug is in act_police since the very beginning; however, the
> > patch should apply cleanly at least on 5.x kernels. On older ones, there
> > might be a small conflict due to lack of RCU-ification of struct
> > tcf_police_params.
> > A conflict that gets fixed easily, but in case we need it I volunteer to
> > write a patch for kernels older than 4.20. @Ilya, what is the
> > minimum kernel usable for openvswitch with MTU policing?
> > 
> 
> It does not apply to 5.10 or earlier, so please provide a working
> backport for those kernels if you wish it to be applied there.

sure, I will do that.
thanks!
-- 
davide