"kfd kfd: error getting iommu info. is the iommu enabled?"

40 views
Skip to first unread message

Claudia

unread,
Dec 30, 2019, 1:57:43 PM12/30/19
to qubes...@googlegroups.com
I'm getting this message in my logs, about an IOMMU error, on both 4.0.2 and the F31-based 4.1 build. I'm as certain as I can be that the IOMMU is enabled in BIOS. I'm having issues with passthru and I'm wondering if this might be the cause.

In dom0 kernel logs:
AMD IOMMUv2 driver by Joerg Roedel <jro...@suse.de
AMD IOMMUv2 functionality not available on this system
...
kfd kfd: Initialized module
...
kfd kfd: Allocated 3969056 bytes on gart
Virtual CRAT table created for GPU
Parsing CRAT table with 1 nodes
Creating topology SYSFS entries
Topology: Add dGPU node [0x15dd:0x1002]
kfd kfd: error getting iommu info. is the iommu enabled? <-- red
kfd kfd: Error initializing iommuv2 <-- red
Creating topology SYSFS entries
kfd kfd: device 1002:15dd NOT added due to errors <-- red

1002:15dd is the integrated GPU on my Ryzen 2500U.

In xl dmesg
(XEN) AMD-Vi: Disabled HAP memory map sharing with IOMMU
(XEN) AMD-Vi: IOMMU Extended Features:
(XEN) - Peripheral Page Service Request
(XEN) - NX bit Supported
(XEN) - Guest Translation
(XEN) - Invalidate All Command
(XEN) - Guest APIC supported
(XEN) - Performance Counters
(XEN) AMD-Vi: IOMMU 0 Enabled.


https://community.amd.com/thread/170292
This thread recommends creating /etc/udev/rules.d/kfd.rules with MODE="0666". Qubes 4.1 has file /usr/lib/udev/rules.d/50-udev-default.rules which contains:
SUBSYSTEM=="kfd", GROUP="render", MODE="0666"
although 4.0.2 appears to have no such file. However I get the same error even on 4.1 so I don't think that's the fix.

https://unix.stackexchange.com/questions/263901/kfd-error-getting-iommu-info
This thread says add kernel parameter "iommu=pt" but I don't know if that's effective or safe in Qubes, because Xen handles the IOMMU instead of the kernel I think. I'm not really sure what that parameter does.

https://bugzilla.redhat.com/show_bug.cgi?id=1404139
This thread says it's not a bug, it's just because the system doesn't support IOMMUv2 and amdkfd (whatever that is) requires IOMMUv2 (whatever that is) for HSA (whatever that is). I don't really know what any of that means.

Is it an error? Should I be worried? Could it be causing my passthru problems? Or does it just mean that my GPU is not protected by the IOMMU (same as iommu=no-igfx on Intel) and I don't have to worry about it?

awokd

unread,
Jan 4, 2020, 4:35:28 AM1/4/20
to qubes...@googlegroups.com
Claudia:
> I'm getting this message in my logs, about an IOMMU error, on both 4.0.2 and the F31-based 4.1 build. I'm as certain as I can be that the IOMMU is enabled in BIOS. I'm having issues with passthru and I'm wondering if this might be the cause.

> https://bugzilla.redhat.com/show_bug.cgi?id=1404139
> This thread says it's not a bug, it's just because the system doesn't support IOMMUv2 and amdkfd (whatever that is) requires IOMMUv2 (whatever that is) for HSA (whatever that is). I don't really know what any of that means.

> Or does it just mean that my GPU is not protected by the IOMMU (same as iommu=no-igfx on Intel) and I don't have to worry about it?

I saw in a different thread you made progress, so this is probably
outdated. However, I think your conclusions are correct. If your IOMMU
was not working correctly, you would have problems starting HVMs period.
The Xen logs look similar to mine. I don't have kfd on my system.

--
- don't top post
Mailing list etiquette:
- trim quoted reply to only relevant portions
- when possible, copy and paste text instead of screenshots

Claudia

unread,
Jan 4, 2020, 8:41:29 AM1/4/20
to awokd, qubes...@googlegroups.com
January 4, 2020 9:35 AM, "awokd' via qubes-users" <qubes...@googlegroups.com> wrote:

> Claudia:
>
>> I'm getting this message in my logs, about an IOMMU error, on both 4.0.2 and the F31-based 4.1
>> build. I'm as certain as I can be that the IOMMU is enabled in BIOS. I'm having issues with
>> passthru and I'm wondering if this might be the cause.
>>
>> https://bugzilla.redhat.com/show_bug.cgi?id=1404139
>> This thread says it's not a bug, it's just because the system doesn't support IOMMUv2 and amdkfd
>> (whatever that is) requires IOMMUv2 (whatever that is) for HSA (whatever that is). I don't really
>> know what any of that means.
>>
>> Or does it just mean that my GPU is not protected by the IOMMU (same as iommu=no-igfx on Intel) and
>> I don't have to worry about it?
>
> I saw in a different thread you made progress, so this is probably
> outdated. However, I think your conclusions are correct. If your IOMMU
> was not working correctly, you would have problems starting HVMs period.
>

I made progress on a couple of things, but not USB Qube. I was just wondering if might have something to do with USB controller passthru. However it may be that USB Qube is simply not supported on this system due to IOMMU grouping. Other passthru HVMs such as sys-net work fine, though, so I guess that means the IOMMU is working despite the error.


> The Xen logs look similar to mine. I don't have kfd on my system.

From what I saw, I think kfd is an AMD thing. Do you have AMD?

awokd

unread,
Jan 4, 2020, 1:49:52 PM1/4/20
to Claudia, qubes...@googlegroups.com
Claudia:
> January 4, 2020 9:35 AM, "awokd' via qubes-users" <qubes...@googlegroups.com> wrote:

>> The Xen logs look similar to mine. I don't have kfd on my system.
>
> From what I saw, I think kfd is an AMD thing. Do you have AMD?

Yes. Pre-Ryzen, but the AMD IOMMU init in Xen logs looks the same.
Reply all
Reply to author
Forward
0 new messages