[vfio-users] "No signal" on dual Nvidia setup

[Nicolas Roy-Renaud]

Alright, using VNC helped immensely, but as it turns out removing 
PCI:01:00.0 from the list of passed devices in order to investigate a 
blue screen and adding it back in while the VM was live thinking it 
would queue the operation for the next reboot did the trick.
Attempting to boot the VM with the GPU already bridged causes the boot 
process to hang whether or not I pass a vBIOS file to qemu and vfio then 
complains that my ROM contents are invalid, but booting the VM with the 
sound subsystem alone (I figured it ensures that the IOMMU group is 
grabbed by kvm, but I'm not aware of the technical details) and then 
using virsh or virtmanager to attach the actual GPU device to the guest 
works where everything I've been doing for the last week has failed. 
While that is some counter-intuitive behavior, and I should probably 
file a bug on this somewhere, I'm glad I at least got this working.

Thank you Will and Jonathan, and thank you vfio-users.

On 2016-01-19 11:24, Nicolas Roy-Renaud wrote:
> On 2016-01-19 03:19, Jonathan Scruggs wrote:
>>
>> Hi,
>>
>> > Since the 970 is set at my primary GPU, it is responsible for 
>> displaying my bios and bootloader until linux boots, where I have its 
>> framebuffer disabled and vfio-pci latch onto it. The 210 GT, however, 
>> is still managed by the nouveau driver.
>>
>> Out of curiosity, why do you have the 970 set as the primary boot 
>> display? I know on some motherboards the first slot will be the 
>> fastest and the other ones have reduced PCI-e lanes when multiple 
>> cards are used. However, my BIOS has a setting to override which slot 
>> is the default. You may want to check if you have this AMD set the 
>> 210 as primary, or for testing purposes, swap the cards around to 
>> take the 970 out of the primary slot. I have a 970 in a secondary 
>> slot and it passes through correctly.
>>
> I know it's going to sound silly, but it's because my motherboard is 
> made in a way that if I put my (massive) 970 in the secondary port, I 
> end up blocking access to all my SATA ports and aligning one of the 
> intake fans with the PSU's outtake. Those are all things I could 
> technically work my way around, but unless it's not possible to get 
> the passthrough to work otherwise, it just seems like more trouble 
> than it's worth.
>
> Also, I did check in my mobo settings for a way to change which card 
> should act as the boot device, but the only option that ressembles 
> that is an integrated/dedicated graphics switch.
>>
>> In your win8 XML file, did you pass through the audio part of your 
>> card or did I miss it. Even if you don't use HDMI audio, you can try 
>> passing it in.
>>
> Right, I removed that line shortly before I decided come asking for 
> help. Whether it's there or not doesn't seem to change much.
>
> On 2016-01-18 16:23, Will Marler wrote:
>> Well it looks to me like your GeForce GTX 970 is correctly being 
>> claimed by vfio-pci, so I would expect that if you passed it to a VM, 
>> the VM should be able to see it. I'd suggest removing <timer 
>> name='hypervclock' present='yes'/> from your XML file, and accessing 
>> it via VNC. You should be able to go into the Windows device manager 
>> and see the video card there (where I actually think you'll see an 
>> Error 43 currently, because of the hypervclock line).
>
> Good thinking, I'll try to give the VNC server a shot. Also, from what 
> I've read, using hv_vendor_id= with recent (git) versions of qemu 
> fools the nvidia driver without requiring you to disable the hyperv 
> clock and other performence enchancing features, which is why I've 
> left them there.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20160119/d73646df/attachment.htm>