Date:   Tue, 1 Aug 2017 16:04:44 -0500
From:   Alan Tull <atull@...nel.org>
To:     Wu Hao <hao.wu@...el.com>
Cc:     Rob Herring <robh+dt@...nel.org>, Moritz Fischer <mdf@...nel.org>,
        linux-fpga@...r.kernel.org,
        linux-kernel <linux-kernel@...r.kernel.org>,
        "linux-api@...r.kernel.org" <linux-api@...r.kernel.org>,
        "Kang, Luwei" <luwei.kang@...el.com>,
        "Zhang, Yi Z" <yi.z.zhang@...el.com>,
        Tim Whisonant <tim.whisonant@...el.com>,
        Enno Luebbers <enno.luebbers@...el.com>,
        Shiva Rao <shiva.rao@...el.com>,
        Christopher Rauer <christopher.rauer@...el.com>,
        "devicetree@...r.kernel.org" <devicetree@...r.kernel.org>
Subject: Re: [PATCH v2 02/22] fpga: add FPGA device framework

On Tue, Aug 1, 2017 at 3:43 AM, Wu Hao <hao.wu@...el.com> wrote:
> On Mon, Jul 31, 2017 at 04:40:16PM -0500, Alan Tull wrote:
>> On Thu, Jul 27, 2017 at 2:10 PM, Rob Herring <robh+dt@...nel.org> wrote:
>> > On Thu, Jul 27, 2017 at 11:35 AM, Alan Tull <atull@...nel.org> wrote:
>> >> On Sun, Jun 25, 2017 at 8:51 PM, Wu Hao <hao.wu@...el.com> wrote:
>> >>
>> >> Hi Rob,
>> >>
>> >> I was hoping to pick your brain a bit on a DT question.
>> >>
>> >>> During FPGA device (e.g PCI-based) discovery, platform devices are
>> >>> registered for different FPGA function units. But the device node path
>> >>> isn't quite friendly to applications.
>> >>>
>> >>> Consider this case, applications want to access child device's sysfs file
>> >>> for some information.
>> >>>
>> >>> 1) Access using bus-based path (e.g PCI)
>> >>>
>> >>>   /sys/bus/pci/devices/xxxxx/fpga_func_a.0/sysfs_file
>> >>>
>> >>>   From the path, it's clear which PCI device is the parent, but not perfect
>> >>>   solution for applications. PCI device BDF is not fixed, application may
>> >>>   need to search all PCI device to find the actual FPGA Device.
>> >>>
>> >>> 2) Or access using platform device path
>> >>>
>> >>>   /sys/bus/platform/devices/fpga_func_a.0/sysfs_file
>> >>>
>> >>>   Applications find the actual function by name easily, but no information
>> >>>   about which fpga device it belongs to. It's quite confusing if multiple
>> >>>   FPGA devices are in one system.
>> >>
>> >> There's a proposal for adding sysfs nodes that correspond to each FPGA
>> >> device., with the devices located on each FPGA under them.  It makes
>> >> it easier to see which device is on which FPGA.
>> >
>> > Makes sense.
>> >
>> >>> 'FPGA Device' class is introduced to resolve this problem. Each node under
>> >>> this class represents a fpga device, which may have one or more child
>> >>> devices. Applications only need to search under this FPGA Device class
>> >>> folder to find the child device node it needs.
>> >>>
>> >>> For example, for the platform has 2 fpga devices, each fpga device has
>> >>> 3 child devices, the hierarchy looks like this.
>> >>>
>> >>> Two nodes are under /sys/class/fpga/:
>> >>> /sys/class/fpga/fpga.0
>> >>> /sys/class/fpga/fpga.1
>> >>>
>> >>> Each node has 1 function A device and 2 function B devices:
>> >>> /sys/class/fpga/fpga.0/func_a.0
>> >>> /sys/class/fpga/fpga.0/func_b.0
>> >>> /sys/class/fpga/fpga.0/func_b.1
>> >>>
>> >>> /sys/class/fpga/fpga.1/func_a.1
>> >>> /sys/class/fpga/fpga.1/func_b.2
>> >>> /sys/class/fpga/fpga.1/func_b.3
>> >
>> > A class is generally what is the function of the device, not how it is
>> > attached. Seems like what you want here is a new bus type if the
>> > existing PCI and platform bus types don't work.
>> >
>> >>
>> >> I can see the value of having sysfs nodes that correspond to fpga
>> >> devices and being able to find devices under them.  I'm thinking what
>> >> that would mean for Device Tree when fpga-dev is used on DT enabled
>> >> systems.  In Device Tree, what is a fpga-dev?
>> >
>> > Just properly setting the parent struct device on the functions should
>> > be enough to figure out which function is in which fpga. I don't see
>> > why a new class is needed.
>> >
>> >> Currently the DT would have a FPGA bridge corresponding to each FPGA's
>> >> hardware bridge and a heirarchy of bridges, regions and devices under
>> >> it.  On systems that don't support partial reconfiguration under the
>> >> OS (so not main bridge that was controlled by the OS), there would be
>> >> a FPGA region, then its child regions, bridges, and devices.
>> >
>> > The FPGA bridges could instantiate fpga bus type devices instead of
>> > platform devices.
>>
>> Yes
>>
>> Some FPGA use cases already have a base bridge per FPGA that could
>> serve as this bus.  But this use case has a static FPGA image +
>> reprogrammable child fpga regions.  There's no base bridge under Linux
>> since the FPGA was programmed and the bridge enabled before Linux
>> boots.   An added base bridge that doesn't touch hardware will be
>> required for this type of use.
>
> Hi Alan
>
> Does 'base bridge' mentioned above mean a hardware bridge just like
> PCIe or USB?

Whatever connects each FPGA to the CPU.  One base bridge per FPGA
device to create the fpga bus type devices.  Each PR region's bridge
would also be a bus.

>
> I tried to use fpga bus type device instead of fpga-dev class today,
> it works for me, e.g Intel FPGA device PCIe driver could create a
> fpga bus type dev as a child of PCIe device and its sysfs path will be
> changed to /sys/bus/fpga/devices/fpga.x/ from /sys/class/fpga/fpga.x/.
> For now, this fpga bus type device is only used as container device,
> so no driver needed for it.

That's great!  I'd like to see the code to try it out with device
tree.  Is it part of fpga-bridge or something separate for now?

>
> Do you have any concern on this? I see fpga bus type works fine, but
> I didn't see other advantages for this case, as we only use it as a
> container device to represent a FPGA device in sysfs hierarchy. :)

I could not see a way to make the fpga-dev class compatible with the
FPGA Device Tree bindings.  This was a red flag. That's why I asked
Rob's opinion.  Sysfs classes collect devices of a specific type
together; busses describe topology.  I think the goal of fpga-dev was
to describe topology.  It's more correct to define this as a bus, not
a class.  If it's done right, it can work for device tree also.

Alan

>
> Thanks
> Hao
>
>>
>> > That's really up to Linux and outside the scope of
>> > the bindings.
>>
>> Thanks for the feedback.
>>
>> Alan Tull
>>
>> >
>> > Rob

Powered by blists - more mailing lists