Re: [regression, bisected, pci/iommu] Bug 216865 - Black screen when amdgpu started during 6.2-rc1 boot with AMD IOMMU enabled

["Christian König" <christian.koenig@amd.com>]

Am 10.01.23 um 14:25 schrieb Jason Gunthorpe:
> On Tue, Jan 10, 2023 at 01:48:39PM +0800, Baolu Lu wrote:
>> On 2023/1/6 22:14, Jason Gunthorpe wrote:
>>> On Thu, Jan 05, 2023 at 03:57:28PM +0530, Vasant Hegde wrote:
>>>> Matt,
>>>>
>>>> On 1/5/2023 6:39 AM, Matt Fagnani wrote:
>>>>> I built 6.2-rc2 with the patch applied. The same black screen problem happened
>>>>> with 6.2-rc2 with the patch. I tried to use early kdump with 6.2-rc2 with the
>>>>> patch twice by panicking the kernel with sysrq+alt+c after the black screen
>>>>> happened. The system rebooted after about 10-20 seconds both times, but no kdump
>>>>> and dmesg files were saved in /var/crash. I'm attaching the lspci -vvv output as
>>>>> requested.
>>>>>
>>>> Thanks for testing. As mentioned earlier I was not expecting this patch to fix
>>>> the black screen issue. It should fix kernel warnings and IOMMU page fault
>>>> related call traces. By any chance do you have the kernel boot logs?
>>>>
>>>>
>>>> @Baolu,
>>>>     Looking into lspci output, it doesn't list ACS feature for Graphics card. So
>>>> with your fix it didn't enable PASID and hence it failed to boot.
>>> The ACS checks being done are feature of the path not the end point or
>>> root port.
>>>
>>> If we are expecting ACS on the end port then it is just a bug in how
>>> the test was written.. The test should be a NOP because there are no
>>> switches in this topology.
>>>
>>> Looking at it, this seems to just be because pci_enable_pasid is
>>> calling pci_acs_path_enabled wrong, the only other user is here:
>>>
>>> 	for (bus = pdev->bus; !pci_is_root_bus(bus); bus = bus->parent) {
>>> 		if (!bus->self)
>>> 			continue;
>>>
>>> 		if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
>>> 			break;
>>>
>>> 		pdev = bus->self;
>>>
>>> 		group = iommu_group_get(&pdev->dev);
>>> 		if (group)
>>> 			return group;
>>> 	}
>>>
>>> And notice it is calling it on pdev->bus not on pdev itself which
>>> naturally excludes the end point from the ACS validation.
>>>
>>> So try something like:
>>>
>>> 	if (!pci_acs_path_enabled(pdev->bus->self, NULL, PCI_ACS_RR | PCI_ACS_UF))
>>>
>>> (and probably need to check for null ?)
>> Hi Matt,
>>
>> Do you mind helping to test below change? No other change needed.
>>
>> diff --git a/drivers/pci/ats.c b/drivers/pci/ats.c
>> index f9cc2e10b676..48f34cc996e4 100644
>> --- a/drivers/pci/ats.c
>> +++ b/drivers/pci/ats.c
>> @@ -382,8 +382,15 @@ int pci_enable_pasid(struct pci_dev *pdev, int
>> features)
>>          if (!pasid)
>>                  return -EINVAL;
>>
>> -       if (!pci_acs_path_enabled(pdev, NULL, PCI_ACS_RR | PCI_ACS_UF))
>> -               return -EINVAL;
>> +       if (pdev->multifunction) {
>> +               if (!pci_acs_path_enabled(pdev, NULL, PCI_ACS_RR |
>> PCI_ACS_UF))
>> +                       return -EINVAL;
> The AMD device is multi-function according to the lspci, and we
> already know that 'pci_acs_path_enabled' will fail on it because that
> is the problem..
>
> Actually, I remember it is supposed to be like this:
>
>   https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Flinux-iommu%2FYgpb6CxmTdUHiN50%408bytes.org%2F&data=05%7C01%7Cchristian.koenig%40amd.com%7Cb45e8c5a24394d66ae2908daf30e3802%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638089539666187724%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=vf9QsDFqp9s1NUxuP5iMsQJn1R0K9tVRTImTR6uZWAE%3D&reserved=0
>
> The GPU and sound device are considered non-isolated by the group
> code, presumably because of the missing ACS caps.
>
> So, if I remember the issue, PCIe says that MemWr/Rd are routed
> according to their address and ignore the PASID header.
>
> A multifunction device is permitted to loop back DMAs one function
> issues that match a MMIO BAR of another function. eg the GPU could DMA
> to an MMIO address that overlaps the sound device and the function
> will deliver the MMIO to the sound device not the host bridge even
> though it is PASID tagged.
>
> This is what get_pci_function_alias_group() is looking for.
>
> Multifunction devices that do not do that are supposed to set the ACS
> RR|UF bits and get_pci_function_alias_group()/etc are supposed to
> succeed.
>
> Thus - the PCI information is telling us that the AMD GPU device does
> not support PASID because it may be looping back the MMIO to the other
> functions on the device and thus creating an unacceptable hole in the
> PASID address space.
>
> So - we need AMD to comment on which of these describes their GPU device:
>
>   1) Is the issue that the PCI Caps are incorrect on this device and
>   there is no loopback? Thus we should fix it with a quirk to correct
>   the caps which will naturally split the iommu group too.
>
>   2) Is the device broken and loops back PASID DMAs and we are
>   legimiate and correct in blocking PASID? So far AMD just got lucky
>   that no user had a SVA that overlaps with MMIO? Seems unlikely
>
>   3) Is the device odd in that it doesn't loop back PASID tagged DMAs,
>   but does loop untagged? I would say this is non-compliant and PCI
>   provides no way to describe this. But we should again quirk it to
>   allow the PASID to be enabled but keep the group separated.

Mhm, I don't have a Kabini at hand but I have a Raven and there I see on 
the GPU:

     Capabilities: [2a0 v1] Access Control Services
         ACSCap:    SrcValid- TransBlk- ReqRedir- CmpltRedir- 
UpstreamFwd- EgressCtrl- DirectTrans-
         ACSCtl:    SrcValid- TransBlk- ReqRedir- CmpltRedir- 
UpstreamFwd- EgressCtrl- DirectTrans-

     Capabilities: [2b0 v1] Address Translation Service (ATS)
         ATSCap:    Invalidate Queue Depth: 00
         ATSCtl:    Enable+, Smallest Translation Unit: 00

On the bridge:

     Capabilities: [2a0 v1] Access Control Services
         ACSCap:    SrcValid+ TransBlk+ ReqRedir- CmpltRedir- 
UpstreamFwd- EgressCtrl- DirectTrans-
         ACSCtl:    SrcValid+ TransBlk- ReqRedir- CmpltRedir- 
UpstreamFwd- EgressCtrl- DirectTrans-

And I'm like 99% sure that Kabini/Wani should be identical to that.

Since this is a device integrated in the CPU it could be that the 
ACS/ATS functionalities are controlled by the BIOS and can be 
enabled/disabled there. But this should always enable/disable both.

Regards,
Christian.

>
> Alex/Christian/Pan - can you please find out? The HW is:
>
> 00:01.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] Wani [Radeon R5/R6/R7 Graphics] (rev ca) (prog-if 00 [VGA controller])
> 	DeviceName: ATI EG BROADWAY
> 	Subsystem: Hewlett-Packard Company Device 8332
> 00:01.1 Audio device: Advanced Micro Devices, Inc. [AMD/ATI] Kabini HDMI/DP Audio
> 	Subsystem: Hewlett-Packard Company Device 8332
>
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fall%2F223ee6d6-70ea-1d53-8bc2-2d22201d8dde%40bell.net%2F&data=05%7C01%7Cchristian.koenig%40amd.com%7Cb45e8c5a24394d66ae2908daf30e3802%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638089539666187724%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=TMB3pXS0eUKPZhcRCvUIxzvJvPYosvv3ofrFKZx7b%2FI%3D&reserved=0
>
>> +       } else {
>> +               if (!pdev->bus->self ||
>> +                   !pci_acs_path_enabled(pdev->bus->self, NULL,
>> +                                         PCI_ACS_RR | PCI_ACS_UF))
>> +                       return -EINVAL;
>> +       }
> Why would these be exclusive? Both the path and endpoint needs to be
> checked
>
> Jason