[igt-dev] [RFC] IGT device scanning and selection

[Ser, Simon]

On Tue, 2019-06-04 at 08:09 +0000, Kempczynski, Zbigniew wrote:
> Hi all.
> 
> IGT device management is a little bit old. More than one GPU on the machine 
> are available on the market from years so there's time to propose new device
> management in IGT.

Nice! There is definitely room for improvement here, thanks for working
on this!

I think we already have an env variable for (some of) this:
IGT_FORCE_DRIVER. It's not great though.

> Selection using bus-device or functional specifiers.
> ---
> 1. bus-device selection
>    Current devices are mostly PCI-Express devices, but we should be able to
>    open to specify device using any bus selection. For example:
>    --device PCI:00:02.0
>    would select PCI device on 00 bus, 02 device and function 0.
> 
>    --device USB:002:003
>    would select device 003 on bus 002.

Why not specify the DRM node path, instead of this?

>    Future buses can be specified by extending --device argument parsing.

Some devices like the Raspberry Pi use platform buses. vc4 uses IGT, so
it would be nice to support that.

> 2. Functional selection
>    CI environments shouldn't rely on bus-device. Same devices (GPUs) may be 
>    inserted into different slots so functional selection is required. 
>    
>    Let's imagine machine have 5 cards:
>    00:02:0 - VendorX, product: A
>    00:03:0 - VendorY, product: B
>    00:04:0 - VendorX, product: A
>    00:05:0 - VendorY, product: C
>    00:06:0 - VendorX, product: D
> 
>    We should be able to select same device regardless order on the bus.
>    For example:
>    --device VendorX:A 
>    --device VendorX:A:0 
>    will select first A card of VendorX
> 
>    --device VendorY:A:1
>    will select second A card of VendorX

Hmm. What defines the order between the "first" and the "second" one?
This doesn't seem like it's "regardless order on the bus" to me? Am I
missing something?

>    If vendor would like to have its own parsing method for the product
>    overriding parsing will be possible. For example:
>    --device VendorX:GPU-Card:0
>    VendorX must then register its own function to handle 'GPU-Card:0'
>    string on his own. 

What is the use-case for this?

> 
> ===
> 
> Device management
> ---
> New device management should allow user to see current devices. 
> I propose IGT tool called 'lsgpu' which would discover devices in 
> any registered bus topologies scanning methods.

I believe we can do everything with libdrm's drmGetDevice.

Note that some utilities (e.g. https://github.com/ascent12/drm_info)
already expose this functionality to some extend.