Re: [PATCH] tests/intel/xe_configfs: Skip ctx-restore subtests on SR-IOV VF

["Bernatowicz, Marcin" <marcin.bernatowicz@linux.intel.com>]

On 3/10/2026 5:37 AM, Gote, Nitin R wrote:
> Hi Marcin,
>
>> -----Original Message-----
>> From: Bernatowicz, Marcin <marcin.bernatowicz@linux.intel.com>
>> Sent: Thursday, March 5, 2026 9:51 PM
>> To: Kamil Konieczny <kamil.konieczny@linux.intel.com>; Gote, Nitin R
>> <nitin.r.gote@intel.com>; igt-dev@lists.freedesktop.org; Wajdeczko, Michal
>> <Michal.Wajdeczko@intel.com>; Laguna, Lukasz <lukasz.laguna@intel.com>
>> Subject: Re: [PATCH] tests/intel/xe_configfs: Skip ctx-restore subtests on SR-IOV
>> VF
>>
>>
>> On 2/24/2026 2:20 PM, Kamil Konieczny wrote:
>>> Hi Nitin,
>>> On 2026-02-24 at 16:50:42 +0530, Nitin Gote wrote:
>>>> The ctx-restore subtests use intel_register_read() via direct MMIO
>>>> (PCI BAR) access to verify that the batch buffer wrote the expected
>>>> register values. On SR-IOV Virtual Functions (VF), only a very
>>>> limited subset of registers is accessible. Registers like 0x4F100 are
>>>> not exposed to VFs, so the read always returns 0, causing a false
>>>> test failure.
>>>>
>>>> Skip all ctx-restore subtests (post-bb, mid-bb, and their invalid
>>>> variants) when the device is running in SR-IOV VF mode.
>> Does it mean the feature is not applicable to VFs or just the verification is a bit
>> harder ?
> The ctx-restore feature itself works on VFs -- the GuC handles context
> save/restore for VFs just like the PF. The issue is strictly with the
> test's verification method.
>
> The test uses intel_register_read() via direct PCI BAR MMIO to read
> back register 0x4F100, which is blocked by the HW MMIO filter on VFs.
> I looked into alternatives but couldn't find a suitable VF-accessible
> register -- the VF MMIO filter only exposes IRQ registers
> (0x190010-0x1900f4) and VF_SW_FLAG registers (0x190240), which are
> used by GuC communication and not suitable as test scratch.
>
> Given there's no VF-safe register available, I think the skip with an
> accurate message seems like the correct approach:
>
> igt_require_f(!is_vf_device,
> "MMIO register readback not possible on VF\n");

I do not see a better way, so let skip

NIT: igt_skip_on(is_vf_device, ...)  reads a bit easier

Reviewed-by: Marcin Bernatowicz <marcin.bernatowicz@linux.intel.com>

>
> Please let me know if you'd prefer a different approach.
> Thank you.
>
> - Nitin
>
>>> +cc Lukasz and Marcin
>>>
>>>> Signed-off-by: Nitin Gote <nitin.r.gote@intel.com>
>>>> ---
>>>>    tests/intel/xe_configfs.c | 4 ++++
>>>>    1 file changed, 4 insertions(+)
>>>>