Date:   Sat, 9 Nov 2019 09:27:47 -0800
From:   Jakub Kicinski <jakub.kicinski@...ronome.com>
To:     Jason Gunthorpe <jgg@...pe.ca>
Cc:     Parav Pandit <parav@...lanox.com>, Jiri Pirko <jiri@...nulli.us>,
        David M <david.m.ertman@...el.com>,
        "gregkh@...uxfoundation.org" <gregkh@...uxfoundation.org>,
        "alex.williamson@...hat.com" <alex.williamson@...hat.com>,
        "davem@...emloft.net" <davem@...emloft.net>,
        "kvm@...r.kernel.org" <kvm@...r.kernel.org>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        Saeed Mahameed <saeedm@...lanox.com>,
        "kwankhede@...dia.com" <kwankhede@...dia.com>,
        "leon@...nel.org" <leon@...nel.org>,
        "cohuck@...hat.com" <cohuck@...hat.com>,
        Jiri Pirko <jiri@...lanox.com>,
        "linux-rdma@...r.kernel.org" <linux-rdma@...r.kernel.org>,
        Or Gerlitz <gerlitz.or@...il.com>
Subject: Re: [PATCH net-next 00/19] Mellanox, mlx5 sub function support

On Fri, 8 Nov 2019 20:44:26 -0400, Jason Gunthorpe wrote:
> On Fri, Nov 08, 2019 at 01:45:59PM -0800, Jakub Kicinski wrote:
> > Yes, my suggestion to use mdev was entirely based on the premise that
> > the purpose of this work is to get vfio working.. otherwise I'm unclear
> > as to why we'd need a bus in the first place. If this is just for
> > containers - we have macvlan offload for years now, with no need for a
> > separate device.  
> 
> This SF thing is a full fledged VF function, it is not at all like
> macvlan. This is perhaps less important for the netdev part of the
> world, but the difference is very big for the RDMA side, and should
> enable VFIO too..

Well, macvlan used VMDq so it was pretty much a "legacy SR-IOV" VF.
I'd perhaps need to learn more about RDMA to appreciate the difference.

> > On the RDMA/Intel front, would you mind explaining what the main
> > motivation for the special buses is? I'm a little confurious.  
> 
> Well, the issue is driver binding. For years we have had these
> multi-function netdev drivers that have a single PCI device which must
> bind into multiple subsystems, ie mlx5 does netdev and RDMA, the cxgb
> drivers do netdev, RDMA, SCSI initiator, SCSI target, etc. [And I
> expect when NVMe over TCP rolls out we will have drivers like cxgb4
> binding to 6 subsytems in total!]

What I'm missing is why is it so bad to have a driver register to
multiple subsystems.

I've seen no end of hacks caused people trying to split their driver
too deeply by functionality. Separate sub-drivers, buses and modules.

The nfp driver was split up before I upstreamed it, I merged it into
one monolithic driver/module. Code is still split up cleanly internally,
the architecture doesn't change in any major way. Sure 5% of developers
were upset they can't do some partial reloads they were used to, but
they got used to the new ways, and 100% of users were happy about the
simplicity.

For the nfp I think the _real_ reason to have a bus was that it
was expected to have some out-of-tree modules bind to it. Something 
I would not encourage :)

Maybe RDMA and storage have some requirements where the reload of the
part of the driver is important, IDK..

> > My understanding is MFD was created to help with cases where single
> > device has multiple pieces of common IP in it.   
> 
> MFD really seems to be good at splitting a device when the HW is
> orthogonal at the register level. Ie you might have regs 100-200 for
> ethernet and 200-300 for RDMA.
> 
> But this is not how modern HW works, the functional division is more
> subtle and more software based. ie on most devices a netdev and rdma
> queue are nearly the same, just a few settings make them function
> differently.
> 
> So what is needed isn't a split of register set like MFD specializes
> in, but a unique per-driver API between the 'core' and 'subsystem'
> parts of the multi-subsystem device.

Exactly, because the device is one. For my simplistic brain one device
means one driver, which can register to as many subsystems as it wants.

> > Do modern RDMA cards really share IP across generations?   
> 
> What is a generation? Mellanox has had a stable RDMA driver across
> many sillicon generations. Intel looks like their new driver will
> support at least the last two or more sillicon generations..
> 
> RDMA drivers are monstrous complex things, there is a big incentive to
> not respin them every time a new chip comes out.

Ack, but then again none of the drivers gets rewritten from scratch,
right? It's not that some "sub-drivers" get reused and some not, no?

> > Is there a need to reload the drivers for the separate pieces (I
> > wonder if the devlink reload doesn't belong to the device model :().  
> 
> Yes, it is already done, but without driver model support the only way
> to reload the rdma driver is to unload the entire module as there is
> no 'unbind'

The reload is the only thing that I can think of (other than
out-of-tree code), but with devlink no I believe it can be solved
differently.

Thanks a lot for the explanation Jason, much appreciated!

The practicality of this is still a little elusive to me, but since 
Greg seems on board I guess it's just me :)

Powered by blists - more mailing lists