RE: [Patch v3] net: mana: Batch ringing RX queue doorbell on receiving packets

[Long Li]

> Subject: Re: [Patch v3] net: mana: Batch ringing RX queue doorbell on
> receiving packets
> 
> On Fri, Jun 30, 2023 at 05:31:48PM +0000, Long Li wrote:
> > > Subject: Re: [Patch v3] net: mana: Batch ringing RX queue doorbell
> > > on receiving packets
> > >
> > > On Thu, 2023-06-29 at 18:18 +0000, Long Li wrote:
> > > > > Subject: Re: [Patch v3] net: mana: Batch ringing RX queue
> > > > > doorbell on receiving packets
> > > > >
> > > > > On Mon, 2023-06-26 at 16:57 -0700, longli@xxxxxxxxxxxxxxxxx
> wrote:
> > > > > > From: Long Li <longli@xxxxxxxxxxxxx>
> > > > > >
> > > > > > It's inefficient to ring the doorbell page every time a WQE is
> > > > > > posted to the received queue. Excessive MMIO writes result in
> > > > > > CPU spending more time waiting on LOCK instructions (atomic
> > > > > > operations), resulting in poor scaling performance.
> > > > > >
> > > > > > Move the code for ringing doorbell page to where after we have
> > > > > > posted all WQEs to the receive queue during a callback from
> > > > > > napi_poll().
> > > > > >
> > > > > > With this change, tests showed an improvement from 120G/s to
> > > > > > 160G/s on a 200G physical link, with 16 or 32 hardware queues.
> > > > > >
> > > > > > Tests showed no regression in network latency benchmarks on
> > > > > > single connection.
> > > > > >
> > > > > > While we are making changes in this code path, change the code
> > > > > > for ringing doorbell to set the WQE_COUNT to 0 for Receive
> > > > > > Queue. The hardware specification specifies that it should set to 0.
> > > > > > Although
> > > > > > currently the hardware doesn't enforce the check, in the
> > > > > > future releases it may do.
> > > > > >
> > > > > > Cc: stable@xxxxxxxxxxxxxxx
> > > > > > Fixes: ca9c54d2d6a5 ("net: mana: Add a driver for Microsoft
> > > > > > Azure Network Adapter (MANA)")
> > > > >
> > > > > Uhmmm... this looks like a performance improvement to me, more
> > > > > suitable for the net-next tree ?!? (Note that net-next is closed
> > > > > now).
> > > >
> > > > This issue is a blocker for usage on 200G physical link. I think
> > > > it can be categorized as a fix.
> > >
> > > Let me ask the question the other way around: is there any specific
> > > reason to have this fix into 6.5 and all the way back to 5.13?
> > > Especially the latest bit (CC-ing stable) looks at least debatable.
> >
> > There are many deployed Linux distributions with MANA driver on kernel
> 5.15 and kernel 6.1. (those kernels are longterm) They need this fix to achieve
> the performance target.
> 
> Why can't they be upgraded to get that performance target, and all the other
> goodness that those kernels have?  We don't normally backport new features,
> right?

I think this should be considered as a fix, not a new feature.

MANA is designed to be 200GB full duplex at the start.  Due to lack of
hardware testing capability at early stage of the project, we could only test 100GB
for the Linux driver. When hardware is fully capable of reaching designed spec,
this bug in the Linux driver shows up.

Thanks,

Long

> 
> thanks,
> 
> greg k-h