bind specific Vendor:Device about vfio-pci-bind HOT 10 CLOSED

Interesting use case. One question: If you are unsure if the Vendor:Device device lives at the supplied domain address, why not parse the whole PCIe tree first to find it, and use the result to bind it with the current algorithm?

from vfio-pci-bind.

The difficulty is if you have two pieces of the same hardware, they will share the Vendor:Device code.

For instance, my motherboard has two Intel nics:

#lspci -n -d 8086:1533
02:00.0 0200: 8086:1533 (rev 03)
03:00.0 0200: 8086:1533 (rev 03)

(each with their own IOMMU group) so I need the pci address to distinguish them.

I thought about switching it and making the Vendor:Device be the required field and Domain:Bus:Device.Function be optional, but that would be a pretty invasive change and I didn't want to break backward compatibility.

Hmm... I suppose we don't have to break backward compatibility since we have regexes that can distinguish which parameter is which. It wouldn't be that invasive either... if you pass in just a Vendor:Device code we could use lcpci to determine the Domain:Bus:Device.Function and then the script can continue.

I'll take a crack at that.

The question is what should happen if you just pass a Vendor:Device code and there are multiple results:

1. bind the first one
2. loop through and bind them all
3. throw an error and require you to specify a pci address too

I think I'll go with option 3.

from vfio-pci-bind.

Hmm... I suppose we don't have to break backward compatibility since we have regexes that can distinguish which parameter is which. It wouldn't be that invasive either... if you pass in just a Vendor:Device code we could use lcpci to determine the Domain:Bus:Device.Function and then the script can continue.

Should work. B:D.F and Vendor:Device can be distinguished fast through the "." character.

I'll take a crack at that.

The question is what should happen if you just pass a Vendor:Device code and there are multiple results:

1. bind the first one

2. loop through and bind them all

3. throw an error and require you to specify a pci address too

I think I'll go with option 3.

Agree with 3.

By the way, are you planning on keeping the Ruby script in sync too?

from vfio-pci-bind.

I like this!

• If the user has a single Vendor:Device, they can provide that and it will be bound regardless of how the PCI address might change
• If they have multiple Vendor:Devices, they can provide the Domain:Bus:Device.Function for the one they want to bind
• For backwards compatibility, they can also provide a Domain:Bus:Device.Function to bind

What do you think, do you see any issues?

I'm sorry that I have no experience with Ruby, are you up for syncing the changes there?

from vfio-pci-bind.

Actually, I think the ruby user base for this script would be very minimal. We can just delete it for now and see if someone complains.

Would be great if you can add that to your PR please?

from vfio-pci-bind.

Sure thing. I also added more detail to the readme

from vfio-pci-bind.

Awesome, thanks, is this complete and tested from your side now? I won’t have time to test it myself the next few days. But if you say it’s good to go I’ll just merge it 👍

from vfio-pci-bind.

I believe this is good to go. But I have limited hardware to test on, I don't mind waiting until you've had a chance to smoke test it.

from vfio-pci-bind.

Thanks for all your work! Merged it 👍

from vfio-pci-bind.

Great! Thanks!

from vfio-pci-bind.