Re: [RFC 7/8] fpga-region: add sysfs interface

From: Alan Tull
Date: Thu Feb 16 2017 - 12:47:14 EST


On Wed, Feb 15, 2017 at 6:16 PM, Moritz Fischer
<moritz.fischer@xxxxxxxxx> wrote:
> On Wed, Feb 15, 2017 at 2:42 PM, Alan Tull <delicious.quinoa@xxxxxxxxx> wrote:
>> On Wed, Feb 15, 2017 at 3:36 PM, Moritz Fischer
>> <moritz.fischer@xxxxxxxxx> wrote:
>>> Jason,
>>>
>>> On Wed, Feb 15, 2017 at 1:15 PM, Jason Gunthorpe
>>> <jgunthorpe@xxxxxxxxxxxxxxxxxxxx> wrote:
>>>> On Wed, Feb 15, 2017 at 12:54:27PM -0800, Moritz Fischer wrote:
>>>>
>>>>> Well I don't know ;-) With something fdt based we already have
>>>>> parsers there,
>>>>
>>>> Not sure.. How does incbin work in DTB?
>>>>
>>>> We have the FPGA in a s/g list so we cannot pass the entire file to
>>>> libfdt - is that consistent with incbin?
>>>
>>> Well you could attach the (for lack of better word) blob to the beginning,
>>> instead of doing incbin
>>>
>>>> Can we force a specific alignment for the included data?
>>>
>>> I'd say probably, but haven't checked.
>>>
>>>> How complex will the userspace tool be to make the image?
>>>
>>> Userspace can be as complex as it needs to be, imho, if it makes
>>> kernel space easier & safer.
>>>
>>> I'll need to do some more reading over the weekend before I can make
>>> more sensible comments :)
>>>
>>> Thanks,
>>>
>>> Moritz
>>
>> Another thought I have about this is that adding the header to
>> bitstreams can be a piece of independence from DT for systems that
>> aren't already using DT. This includes x86 in Linux. It also
>> includes other OS's that aren't using DT, they can reuse the same
>> image files without having to add dtc. As much as I like DT, it is
>> something I'm having to think about.
>
> Just to clarify:
> I was proposing using the binary format of dts, not actually requiring
> devicetree
> for it to work. There's plenty of people running u-boot on x86 using FIT images
> to boot.

The FPGA images should not be required to have OS specific parts.
Some ahem non-Linux OS's that use FPGAs don't use device tree, so that
adds an extra complication for them unnecessarily.

>
> W.r.t to Jason's script, it's there. Almost any company dealing with
> Xilinx FPGAs
> will have one of those. We have one, too. I recall having seen another one made
> and shared by Mike @ topic.
>
> While it's a good starting point ,I *really* don't like the idea
> parsing user-land
> provided strings in kernel space in a parser that we open-code.

Why do you not like about it? Jason posted some very clear practices
on how to do that properly and safely.

>
> Good discussion ;-)

Yes, I like it. :)

Alan

>
> Cheers,
> Moritz