Pause and unpause queue

Pause and unpause queue

[team lnx]

Hello everyone,

1. Is there a way to pause/unpause queue ?
2.  Is QoS (work in progress for XDP) ? Can we use work in progress
series if  there exists any ?

Thanks !

Re: Pause and unpause queue

[Toke Høiland-Jørgensen]

team lnx <teamlnxi8@gmail.com> writes:

> Hello everyone,
>
> 1. Is there a way to pause/unpause queue ?
> 2.  Is QoS (work in progress for XDP) ? Can we use work in progress
> series if  there exists any ?

XDP has no pushback from the driver at all, nor any way to queue
packets. I am planning to add this, see this presentation at LPC last
year:

https://lpc.events/event/16/contributions/1351/

Could you elaborate a bit on what your use case is, specifically? :)

-Toke

Re: Pause and unpause queue

[team lnx]

sure, during xdp redirect i found that sometimes the number of packets
arrived are too many for an interface
to handle in which case. Hence thought to experiment with tx pause to
make room for successive packets and then unpause !

Thanks !

On Fri, Jan 6, 2023 at 5:00 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
>
> team lnx <teamlnxi8@gmail.com> writes:
>
> > Hello everyone,
> >
> > 1. Is there a way to pause/unpause queue ?
> > 2.  Is QoS (work in progress for XDP) ? Can we use work in progress
> > series if  there exists any ?
>
> XDP has no pushback from the driver at all, nor any way to queue
> packets. I am planning to add this, see this presentation at LPC last
> year:
>
> https://lpc.events/event/16/contributions/1351/
>
> Could you elaborate a bit on what your use case is, specifically? :)
>
> -Toke
>

Re: Pause and unpause queue

[Dave Taht]

On Sat, Jan 7, 2023 at 10:08 AM team lnx <teamlnxi8@gmail.com> wrote:
>
> sure, during xdp redirect i found that sometimes the number of packets
> arrived are too many for an interface

That is often the case.

> to handle in which case. Hence thought to experiment with tx pause to
> make room for successive packets and then unpause !

"pauses" are really not how the internet works, there needs to be end
to end signalling to "slow down", either via packet loss, or marking.
RFC970 is a good read here.

I like to think that some of the most advanced work for xdp is our
ongoing work in libreqos, which
leverages XDP for the input path, and then a complicated setup to
manage 10s of thousands of ISP bandwidth plans.

Demo here: https://payne.taht.net/ (hit bandwidth test, click on things)
Code here: https://github.com/LibreQoE/

I think it would be good for that project to attempt a simpler setup
someday merely to load balance XDP connections at some point, rather
than emulating plans.

>
> Thanks !
>
> On Fri, Jan 6, 2023 at 5:00 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
> >
> > team lnx <teamlnxi8@gmail.com> writes:
> >
> > > Hello everyone,
> > >
> > > 1. Is there a way to pause/unpause queue ?
> > > 2.  Is QoS (work in progress for XDP) ? Can we use work in progress
> > > series if  there exists any ?
> >
> > XDP has no pushback from the driver at all, nor any way to queue
> > packets. I am planning to add this, see this presentation at LPC last
> > year:
> >
> > https://lpc.events/event/16/contributions/1351/
> >
> > Could you elaborate a bit on what your use case is, specifically? :)
> >
> > -Toke
> >



-- 
This song goes out to all the folk that thought Stadia would work:
https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
Dave Täht CEO, TekLibre, LLC

Re: Pause and unpause queue

[Toke Høiland-Jørgensen]

team lnx <teamlnxi8@gmail.com> writes:

> sure, during xdp redirect i found that sometimes the number of packets
> arrived are too many for an interface
> to handle in which case. Hence thought to experiment with tx pause to
> make room for successive packets and then unpause !

Right, so this is using XDP as a straight-forward forwarding
accelerator, redirecting from one interface to another? That will indeed
only work well if both interfaces are exactly the same link speed; this
is one of the motivating use cases for the queueing work :)

-Toke

Re: Pause and unpause queue

[Toke Høiland-Jørgensen]

Dave Taht <dave.taht@gmail.com> writes:

> On Sat, Jan 7, 2023 at 10:08 AM team lnx <teamlnxi8@gmail.com> wrote:
>>
>> sure, during xdp redirect i found that sometimes the number of packets
>> arrived are too many for an interface
>
> That is often the case.
>
>> to handle in which case. Hence thought to experiment with tx pause to
>> make room for successive packets and then unpause !
>
> "pauses" are really not how the internet works, there needs to be end
> to end signalling to "slow down", either via packet loss, or marking.
> RFC970 is a good read here.

Actually it is, kinda, at this level: Pausing the TX interface (when the
HWQ is full) is how the regular network stack creates backpressure
against the qdisc, which is what allows fq_codel and friends to
function. This is missing entirely from the XDP redirect path, which is
what we're trying to fix...

-Toke

Re: Pause and unpause queue

[team lnx]

Thanks Toke and Dave for the response with explanation and knowledge
which explains that its work in progress !!

Thanks !

On Sat, Jan 7, 2023 at 10:49 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
>
> Dave Taht <dave.taht@gmail.com> writes:
>
> > On Sat, Jan 7, 2023 at 10:08 AM team lnx <teamlnxi8@gmail.com> wrote:
> >>
> >> sure, during xdp redirect i found that sometimes the number of packets
> >> arrived are too many for an interface
> >
> > That is often the case.
> >
> >> to handle in which case. Hence thought to experiment with tx pause to
> >> make room for successive packets and then unpause !
> >
> > "pauses" are really not how the internet works, there needs to be end
> > to end signalling to "slow down", either via packet loss, or marking.
> > RFC970 is a good read here.
>
> Actually it is, kinda, at this level: Pausing the TX interface (when the
> HWQ is full) is how the regular network stack creates backpressure
> against the qdisc, which is what allows fq_codel and friends to
> function. This is missing entirely from the XDP redirect path, which is
> what we're trying to fix...
>
> -Toke
>

Re: Pause and unpause queue

[Toke Høiland-Jørgensen]

team lnx <teamlnxi8@gmail.com> writes:

> Thanks Toke and Dave for the response with explanation and knowledge
> which explains that its work in progress !!
>
> Thanks !

You're welcome! :)

-Toke

Re: Pause and unpause queue

[Marcus Wichelmann]

Am 07.01.23 um 19:47 schrieb Toke Høiland-Jørgensen:
> team lnx <teamlnxi8@gmail.com> writes:
> 
>> sure, during xdp redirect i found that sometimes the number of packets
>> arrived are too many for an interface
>> to handle in which case. Hence thought to experiment with tx pause to
>> make room for successive packets and then unpause !
> 
> Right, so this is using XDP as a straight-forward forwarding
> accelerator, redirecting from one interface to another? That will indeed
> only work well if both interfaces are exactly the same link speed; this
> is one of the motivating use cases for the queueing work :)
> 
> -Toke
> 

Hi,

I saw your talk and the patch set from july about this. What is blocking 
things right now from getting into the kernel?

XDP queuing support would be really nice to have. ;)

Thanks for your work!
Marcus

Re: Pause and unpause queue

[Toke Høiland-Jørgensen]

Marcus Wichelmann <marcus.wichelmann@hetzner-cloud.de> writes:

> Am 07.01.23 um 19:47 schrieb Toke Høiland-Jørgensen:
>> team lnx <teamlnxi8@gmail.com> writes:
>> 
>>> sure, during xdp redirect i found that sometimes the number of packets
>>> arrived are too many for an interface
>>> to handle in which case. Hence thought to experiment with tx pause to
>>> make room for successive packets and then unpause !
>> 
>> Right, so this is using XDP as a straight-forward forwarding
>> accelerator, redirecting from one interface to another? That will indeed
>> only work well if both interfaces are exactly the same link speed; this
>> is one of the motivating use cases for the queueing work :)
>> 
>> -Toke
>> 
>
> Hi,
>
> I saw your talk and the patch set from july about this. What is blocking 
> things right now from getting into the kernel?

Well, the immediate blocker is me getting around to revising the patch
set and incorporating feedback + some other ideas for how to do it
properly. I managed to get awfully sidetracked before the holidays, but
it's on my list of things to pick back up this year :)

> XDP queuing support would be really nice to have. ;)

Knowing that there's interest in this certainly helps motivation-wise,
so thanks for that!

-Toke