Re: [PATCH] PCI: Mark AMD Raven iGPU ATS as broken

From: Huang Rui
Date: Tue Nov 24 2020 - 01:43:05 EST


On Tue, Nov 24, 2020 at 06:51:11AM +0800, Kuehling, Felix wrote:
> On 2020-11-23 5:33 p.m., Will Deacon wrote:
> > On Mon, Nov 23, 2020 at 09:04:14PM +0000, Deucher, Alexander wrote:
> >> [AMD Public Use]
> >>
> >>> -----Original Message-----
> >>> From: Will Deacon <will@xxxxxxxxxx>
> >>> Sent: Monday, November 23, 2020 8:44 AM
> >>> To: linux-kernel@xxxxxxxxxxxxxxx
> >>> Cc: linux-pci@xxxxxxxxxxxxxxx; iommu@xxxxxxxxxxxxxxxxxxxxxxxxxx; Will
> >>> Deacon <will@xxxxxxxxxx>; Bjorn Helgaas <bhelgaas@xxxxxxxxxx>;
> >>> Deucher, Alexander <Alexander.Deucher@xxxxxxx>; Edgar Merger
> >>> <Edgar.Merger@xxxxxxxxxxx>; Joerg Roedel <jroedel@xxxxxxx>
> >>> Subject: [PATCH] PCI: Mark AMD Raven iGPU ATS as broken
> >>>
> >>> Edgar Merger reports that the AMD Raven GPU does not work reliably on his
> >>> system when the IOMMU is enabled:
> >>>
> >>> | [drm:amdgpu_job_timedout [amdgpu]] *ERROR* ring gfx timeout,
> >>> signaled seq=1, emitted seq=3
> >>> | [...]
> >>> | amdgpu 0000:0b:00.0: GPU reset begin!
> >>> | AMD-Vi: Completion-Wait loop timed out
> >>> | iommu ivhd0: AMD-Vi: Event logged [IOTLB_INV_TIMEOUT
> >>> device=0b:00.0 address=0x38edc0970]
> >>>
> >>> This is indicative of a hardware/platform configuration issue so, since
> >>> disabling ATS has been shown to resolve the problem, add a quirk to match
> >>> this particular device while Edgar follows-up with AMD for more information.
> >>>
> >>> Cc: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> >>> Cc: Alex Deucher <alexander.deucher@xxxxxxx>
> >>> Reported-by: Edgar Merger <Edgar.Merger@xxxxxxxxxxx>
> >>> Suggested-by: Joerg Roedel <jroedel@xxxxxxx>
> >>> Link:
> >>> https://lore.
> >>> kernel.org/linux-
> >>> iommu/MWHPR10MB1310F042A30661D4158520B589FC0@MWHPR10M
> >>> B1310.namprd10.prod.outlook.com
> >>> her%40amd.com%7C1a883fe14d0c408e7d9508d88fb5df4e%7C3dd8961fe488
> >>> 4e608e11a82d994e183d%7C0%7C0%7C637417358593629699%7CUnknown%7
> >>> CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwi
> >>> LCJXVCI6Mn0%3D%7C1000&amp;sdata=TMgKldWzsX8XZ0l7q3%2BszDWXQJJ
> >>> LOUfX5oGaoLN8n%2B8%3D&amp;reserved=0
> >>> Signed-off-by: Will Deacon <will@xxxxxxxxxx>
> >>> ---
> >>>
> >>> Hi all,
> >>>
> >>> Since Joerg is away at the moment, I'm posting this to try to make some
> >>> progress with the thread in the Link: tag.
> >> + Felix
> >>
> >> What system is this? Can you provide more details? Does a sbios update
> >> fix this? Disabling ATS for all Ravens will break GPU compute for a lot
> >> of people. I'd prefer to just black list this particular system (e.g.,
> >> just SSIDs or revision) if possible.
>
> +Ray
>
> There are already many systems where the IOMMU is disabled in the BIOS,
> or the CRAT table reporting the APU compute capabilities is broken. Ray
> has been working on a fallback to make APUs behave like dGPUs on such
> systems. That should also cover this case where ATS is blacklisted. That
> said, it affects the programming model, because we don't support the
> unified and coherent memory model on dGPUs like we do on APUs with
> IOMMUv2. So it would be good to make the conditions for this workaround
> as narrow as possible.

Yes, besides the comments from Alex and Felix, may we get your firmware
version (SMC firmware which is from SBIOS) and device id?

> >>> | [drm:amdgpu_job_timedout [amdgpu]] *ERROR* ring gfx timeout,
> >>> signaled seq=1, emitted seq=3

It looks only gfx ib test passed, and fails to lanuch desktop, am I right?

We would like to see whether it is Raven, Raven kicker (new Raven), or
Picasso. In our side, per the internal test result, we didn't see the
similiar issue on Raven kicker and Picasso platform.

Thanks,
Ray

>
> These are the relevant changes in KFD and Thunk for reference:
>
> ### KFD ###
>
> commit 914913ab04dfbcd0226ecb6bc99d276832ea2908
> Author: Huang Rui <ray.huang@xxxxxxx>
> Date:   Tue Aug 18 14:54:23 2020 +0800
>
>     drm/amdkfd: implement the dGPU fallback path for apu (v6)
>
>     We still have a few iommu issues which need to address, so force raven
>     as "dgpu" path for the moment.
>
>     This is to add the fallback path to bypass IOMMU if IOMMU v2 is
> disabled
>     or ACPI CRAT table not correct.
>
>     v2: Use ignore_crat parameter to decide whether it will go with
> IOMMUv2.
>     v3: Align with existed thunk, don't change the way of raven, only
> renoir
>         will use "dgpu" path by default.
>     v4: don't update global ignore_crat in the driver, and revise fallback
>         function if CRAT is broken.
>     v5: refine acpi crat good but no iommu support case, and rename the
>         title.
>     v6: fix the issue of dGPU initialized firstly, just modify the report
>         value in the node_show().
>
>     Signed-off-by: Huang Rui <ray.huang@xxxxxxx>
>     Reviewed-by: Felix Kuehling <Felix.Kuehling@xxxxxxx>
>     Signed-off-by: Alex Deucher <alexander.deucher@xxxxxxx>
>
> ### Thunk ###
>
> commit e32482fa4b9ca398c8bdc303920abfd672592764
> Author: Huang Rui <ray.huang@xxxxxxx>
> Date:   Tue Aug 18 18:54:05 2020 +0800
>
>     libhsakmt: remove is_dgpu flag in the hsa_gfxip_table
>
>     Whether use dgpu path will check the props which exposed from kernel.
>     We won't need hard code in the ASIC table.
>
>     Signed-off-by: Huang Rui <ray.huang@xxxxxxx>
>     Change-Id: I0c018a26b219914a41197ff36dbec7a75945d452
>
> commit 7c60f6d912034aa67ed27b47a29221422423f5cc
> Author: Huang Rui <ray.huang@xxxxxxx>
> Date:   Thu Jul 30 10:22:23 2020 +0800
>
>     libhsakmt: implement the method that using flag which exposed by
> kfd to configure is_dgpu
>
>     KFD already implemented the fallback path for APU. Thunk will use flag
>     which exposed by kfd to configure is_dgpu instead of hardcode before.
>
>     Signed-off-by: Huang Rui <ray.huang@xxxxxxx>
>     Change-Id: I445f6cf668f9484dd06cd9ae1bb3cfe7428ec7eb
>
> Regards,
>   Felix
>
>
> > Cheers, Alex. I'll have to defer to Edgar for the details, as my
> > understanding from the original thread over at:
> >
> > https://lore.kernel.org/linux-iommu/MWHPR10MB1310CDB6829DDCF5EA84A14689150@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
> >
> > is that this is a board developed by his company.
> >
> > Edgar -- please can you answer Alex's questions?
> >
> > Will