Open unman opened 1 year ago
Hm, I get the idea, but also: people will 100% forget this warning. Nobody remembers warnings from - possibly - months before.
@marmarek @HW42 Can assignment be based on PCIe path rather than bus-slot-function?
Yes a warning probably not the way to go about this. A technical solution is much better than documentation or notes through the user interface.
And I see that sometimes PCIe devices numbering (within the VM) is randomized at boot of the host. This makes some of my HVMs sometimes boot and other times not boot (the passed-through graphics card is assigned to one slot).
@neowutran provides a way to circumvent that (changing the Xorg config on boot) here, but that's cumbersome. In my case I just connect the qube to a netVM (from none), which adds a virtual PCIe network card before, making the GPU one be in the right place.
A permanent solution to this (having some sort of PCIe numbering attribution) would make it easier to use. Maybe the slots could be user-assignable? Here's a (terrible) mockup of what I mean:
(for those who can't see the picture, it's basically a list of devices on the left and numbered slots on the right, where the user can drag and drop the devices).
I was hit by this issue.
I am trying to resolve cooling issues with my SSD. It was under GPU, so I cannot add a heatsink. So, I decided to move GPU.
The QubesOS started booting as usual. However, after entering the password, the system always rebooted after few seconds.
After some trials and errors (including reinserting the GPU), I've added nomodeset
parameter to the kernel in grub. Qubes OS was able to boot then. Also, it has helped me to identify the root issue: devices were renumbered and GPU was assignet to my NetVM instead of one of my network cards. (I am not sure how could nomodeset
have helped.)
I've looked at qvm-pci
output in order to find out whether there are any other unwanted assignments.
qvm-pci
whether all devices are assigned correctly.Note that this workaround might not be applicable if some PCI device dies, as user cannot do step no. 1. (Assuming that the user cannot predict the device death.) EDIT: It seems that you can disable autostarts even if you haven't done step no. 1 in advance. (I haven't tested that, though.)
How to file a helpful issue
The problem you're addressing (if any)
PCI numbering may change if devices are added or altered.(Cf #7792) This means that device allocation to qubes may change.
The solution you'd like
Absent a full solution at least provide warning when device allocation is made. A warning could also be made in the docs.
The value to a user, and who that user might be
Any user who may make hardware changes will have been warned of possible issues when they first make device allocations. Whether they remember that warning is a separate issue.
Cf Forum discussion - https://forum.qubes-os.org/t/usability-issues-with-hardware-changes-on-system-qubes/17676