Re: [RFC net-next 0/8] Introducing subdev bus and devlink extension

[Jakub Kicinski]

Parav, please wrap your responses to at most 80 characters.
This is hard to read.

On Mon, 4 Mar 2019 04:41:01 +0000, Parav Pandit wrote:
> > -----Original Message-----
> > From: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>
> > Sent: Friday, March 1, 2019 2:04 PM
> > To: Parav Pandit <parav@xxxxxxxxxxxx>; Or Gerlitz <gerlitz.or@xxxxxxxxx>
> > Cc: netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > michal.lkml@xxxxxxxxxxx; davem@xxxxxxxxxxxxx;
> > gregkh@xxxxxxxxxxxxxxxxxxx; Jiri Pirko <jiri@xxxxxxxxxxxx>
> > Subject: Re: [RFC net-next 0/8] Introducing subdev bus and devlink extension
> > 
> > On Thu, 28 Feb 2019 23:37:44 -0600, Parav Pandit wrote:  
> > > Requirements for above use cases:
> > > --------------------------------
> > > 1. We need a generic user interface & core APIs to create sub devices
> > > from a parent pci device but should be generic enough for other parent
> > > devices 2. Interface should be vendor agnostic 3. User should be able
> > > to set device params at creation time 4. In future if needed, tool
> > > should be able to create passthrough device to map to a virtual
> > > machine  
> > 
> > Like a mediated device?
> 
> Yes.
>  
> > https://www.kernel.org/doc/Documentation/vfio-mediated-device.txt
> > https://www.dpdk.org/wp-content/uploads/sites/35/2018/06/Mediated-
> > Devices-Better-Userland-IO.pdf
> > 
> > Other than pass-through it is entirely unclear to me why you'd need a bus.
> > (Or should I say VM pass through or DPDK?)  Could you clarify why the need
> > for a bus?
> >   
> A bus follow standard linux kernel device driver model to attach a
> driver to specific device. Platform device with my limited
> understanding looks a hack/abuse of it based on documentation [1],
> but it can possibly be an alternative to bus if it looks fine to Greg
> and others.

I grok from this text that the main advantage you see is the ability to
choose a driver for the subdevice.

> > My thinking is that we should allow spawning subports in devlink
> > and if user specifies "passthrough" the device spawned would be an
> > mdev. 
> 
> devlink device is much more comprehensive way to create sub-devices
> than sub-ports for at least below reasons.
>
> 1. devlink device already defines device->port relation which enables
> to create multiport device.

I presume that by devlink device you mean devlink instance?  Yes, this
part I'm following.

> subport breaks that.

Breaks what?  The ability to create a devlink instance with multiple
ports?

> 2. With bus model, it enables us to load driver of same vendor or
> generic one such a vfio in future.

Yes, sorry, I'm not an expert on mdevs, but isn't that the goal of
those?  Could you go into more detail why not just use mdevs?

> 3. Devices live on the bus, mapping a subport to 'struct device' is
> not intuitive. 

Are you saying that the main devlink instance would not have any port
information for the subdevices?

Devices live on a bus.  Software constructs - depend on how one wants
to model them - don't have to.

> 4. sub-device allows to use existing devlink port,
> registers, health infrastructure to sub devices, which otherwise need
> to be duplicated for ports. 

Health stuff is not tied to a port, I'm not following you.  You can
create a reporter per port, per ACL rule or per SB or per whatever your
heart desires..

> 5. Even though current devlink devices are networking devices, there
> is nothing restricts it to be that way. So subport is a restricted
> view. 
> 6. devlink device already covers
> port sub-object, hence creating devlink device is desired.
> 
> > > 5. A device can have multiple ports  
> > 
> > What does this mean, in practice?  You want to spawn a subdev which
> > can access both ports?  That'd be for RDMA use cases, more than
> > Ethernet, right?  (Just clarifying :))
> >  
> Yep, you got it right. :-)
>  
> > > So how is it done?
> > > ------------------
> > > (a) user in control
> > > To address above requirements, a generic tool iproute2/devlink is
> > > extended for sub device's life cycle.
> > > However a devlink tool and its kernel counter part is not
> > > sufficient to create protocol agnostic devices on a existing PCI
> > > bus.  
> > 
> > "Protocol agnostic"?...  What does that mean?
> >   
> Devlink works on bus,device model. It doesn't matter what class of
> device is. For example, for pci class can be anything. So newly
> created sub-devices are not limited to netdev/rdma devices. Its
> agnostic to protocol. More importantly, we don't want to create these
> sub-devices who bus type is 'pci'. Because as described below, PCI
> has its addressing scheme and pci bus must not have mix-n match
> devices.
> 
> So probably better wording should be,
> 'a devlink tool and its kernel counterpart is not sufficient to
> create sub-devices of same class as that of PCI device.

Let me clarify - for networking devices the partition will most likely
end up as a subport, but its not a requirement that each partition must
be a subport..  The question was about the necessity to invent a new
bus, and have every resource have a struct device..

> > > (b) subdev bus
> > > A given bus defines well defined addressing scheme. Creating sub
> > > devices on existing PCI bus with a different naming scheme is
> > > just weird. So, creating well named devices on appropriate bus is
> > > desired.  
> > 
> > What's that address scheme you're referring to, you seem to assign
> > IDs in sequence?
> >  
> Yes. a device on subdev bus follows standard linux driver model based
> id assignment scheme = u32. And devices are well named as 'subdev0'.
> Prefix + id as the default scheme of core driver model. 

I thought "well defined addressing scheme" means I can address
subdevice X of device Y with your scheme.  I can't, it's just an
global ID.  Thanks for clarifying.

> > > Given that, these are user created devices for a given hardware
> > > and in absence of a central entity like PCISIG to assign vendor
> > > and device ids, A unique vendor and device id are maintained as
> > > enum in include/linux/subdev_ids.h.  
> > 
> > Why do we need IDs?  The sysfs hierarchy isn't sufficient?    
> 
> > Do we need a driver to match on those again?  Is it going to be a
> > different driver? 
> IDs are used to match driver against the created device.
> It can be same or different driver.
> Even in same driver case, it provides a clear code separation for
> creating sub-devices and their respective one or more protocol
> devices (netdev, rep-netdev, rdma ..)
> 
> > > subdev bus device names follow default device naming scheme of
> > > Linux kernel. It is done as 'subdev<instance_id>' such as,
> > > subdev0, subdev3.
> > >
> > > System example view:
> > > --------------------
> > >
> > > $ devlink dev show
> > > pci/0000:05:00.0
> > >
> > > $ devlink dev add pci/0000:05:00.0  
> > 
> > That does not look great.
> >  
> Yes, It must return bus+device attributes in user output too
> Code in existing patchset returns it, it is not shown here.
> I will fix the cover-letter.
> 
> > Also you have to return the id of the spawned device, otherwise
> > this is very racy.
> >   
> Yes, that is correct. It must return an devlink device id =
> {bus+device} attr. I will update the example in v2.
> 
> > > $ devlink dev show
> > > pci/0000:05:00.0
> > > subdev/subdev0  
>
> > Please don't spawn devlink instances.  Devlink instance is supposed
> > to represent an ASIC.  If we start spawning them willy nilly for
> > whatever software construct we want to model the clarity of the
> > ontology will suffer a lot.  
> Devlink devices not restricted to ASIC even though today it is
> representing ASIC for one vendor. Today for one ASIC, it already
> presents multiple devlink devices (128 or more) for PF and VFs, two
> PFs on same ASIC etc. VF is just a sub-device which is well defined
> by PCISIG, whereas sub-device is not. Sub-device do consume actual
> ASIC resources (just like PFs and VFs), Hence point-(6) of
> cover-letter indicate that the devlink capability to tell how many
> such sub-devices can be created.
> 
> In above example, they are created for a given bus-device following
> existing devlink construct.
>
> > Please see the discussion on my recent patchset.  I think Jiri CCed
> > you. 
> I will review the discussion in short while after this reply, and
> provide comments.
> 
> > > Alternatives considered:
> > > ------------------------
> > > Will discuss separately if needed to keep this RFC short.  
> > 
> > Please do discuss.
> >  
> (a) subports instead of subdevices.
> We dropped this option because its two restrictive; I explained above
> the benefits of devlink device.
> 
> (b) extending iproute2/ip link and iproute2/rdma tools to creating
> sub-devices. But that is too limiting which doesn't provide all the
> features we get using devlink. It also doesn't address the
> passthrough needs and its just ugly to create and manage PCI level
> devices using high level tools like 'ip' and 'rdma'.
> 
> (c) creating platform device and platform driver instead of subdev bus
> Our understanding is that - platform device for this purpose would be
> an abuse/misuse, but our view is limited based on kernel
> documentation in [2]. [1] says "platform devices typically appear as
> autonomous entities" Sub-devices are well managed, created,
> configurable by user. Most things of [1] -> "Platform devices"
> section do not match with subdev.
> 
> Greg suggested to use mfd framework (wrapper to platform), which also
> needs extension. mfd_remove_devices() removes all the devices, while
> here based on user request, we want to add/remove individual device.
> Will wait if he is ok with subdev bus or he prefers to extend the
> platform documentation and mfd for removing individual devices.
> 
> (d) drivers/visorbus
> This bus is limited to UUID/GUID based naming scheme and very
> specific to s-Par standard and vendor. Additionally its guest drivers
> are living in staging for more than year. So it doesn't appear the
> right direction.
> 
> (e) creating subdev as child objects of devlink device (such as port,
> registers, health, etc). In this mode, a given devlink device has
> multiport child device which is anchored using 'struct device' and
> life cycled through devlink. Only difference with current proposal is
> it doesn't follow standard driver model to bind to other driver. It
> also doesn't show in unified way using devlink dev show.
> 
> So instead of these alternatives, devlink device that matches PF, VF,
> sub-device, + subdev bus seems better design. This follows all
> standard constructs of 1. Devlink, 2. Linux driver model. It is not
> limited to ports and generic enough for networking and not networking
> devices. 
> > The things key thing for me on the netdev side is what is the
> > forwarding model to this new entity.  Is this basically VMDQ?
> > Should we just go ahead and mandate "switchdev mode" here?
> >   
> It will follow the switchdev mode, but it not limited to it.
> Switchdev mode is for the eswitch functionality. There isn't a need
> to combine this. rdma Infiniband will be able to use this without
> switchdev mode.

It's the devlink instance that's in "switchdev mode", regardless of
type of any of its ports.

> > Thanks for working on a common architecture and suffering through
> > people's reviews rather than adding a debugfs interface that does
> > this like a different vendor did :)  
> Oh yes, lets not do debugfs.
> Thanks a lot Jakub for the review.
> This common architecture should be able to address such common needs.
> Please let me know if this needs more refinement, if I missed
> something.
> 
> [1] https://www.kernel.org/doc/Documentation/driver-model/platform.txt
>