Re: [PATCH] drm/xe: Check for dead config space on reading all 1s

[Michal Wajdeczko <michal.wajdeczko@intel.com>]

On 5/12/2026 8:01 AM, Arun R Murthy wrote:
> Reading the VF_CAP returns all 1s when the config space is dead leading
> to missdetection of VF and confusing the GuC.
> Check for all 1s and pci device detect can act as a sanity.
> 
> Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/work_items/7941
> Signed-off-by: Arun R Murthy <arun.r.murthy@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_sriov.c | 14 ++++++++++++++
>  1 file changed, 14 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_sriov.c b/drivers/gpu/drm/xe/xe_sriov.c
> index f3835867fce5..40e142d78924 100644
> --- a/drivers/gpu/drm/xe/xe_sriov.c
> +++ b/drivers/gpu/drm/xe/xe_sriov.c
> @@ -4,6 +4,7 @@
>   */
>  
>  #include <linux/fault-inject.h>
> +#include <linux/pci.h>
>  
>  #include <drm/drm_managed.h>
>  
> @@ -39,8 +40,21 @@ const char *xe_sriov_mode_to_string(enum xe_sriov_mode mode)
>  
>  static bool test_is_vf(struct xe_device *xe)
>  {
> +	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
>  	u32 value = xe_mmio_read32(xe_root_tile_mmio(xe), VF_CAP_REG);
>  
> +	/*
> +	 * If the device is inaccessible (e.g. parent bridge stuck in D3cold,
> +	 * fatal AER Errors) MMIO reads returns all-ones, 0xffff.
> +	 * VF_CAP would appear set and would misdetect as VF mode.
> +	 * Sanity check PCI presence before trusting the read.
> +	 */

can we have this sanity check elsewhere please?
this is not a SR-IOV specific problem and we will not be here when info.has_sriov is not set

some attempts to show the problem were already done on both i915 and xe, see:

[1] https://elixir.bootlin.com/linux/v7.1-rc1/source/drivers/gpu/drm/i915/intel_uncore.c#L559
[2] https://elixir.bootlin.com/linux/v7.1-rc1/source/drivers/gpu/drm/xe/xe_force_wake.c#L121

so maybe we should just extend our xe_mmio_probe_early() and do some sanity checks there?
then any error will be correctly propagated to abort the probe sooner than today.

also likely similar sanity checks shall be done on the resume paths

Michal


> +	if (value == U32_MAX && !pci_device_is_present(pdev)) {
> +		drm_err(&xe->drm,

nit: we can use xe_err(xe, ...)

> +			"VF_CAP_REG returned all 1s, config space looks to be dead, skipping SR-IOV mode detection\n");

nit: this is still misleading, we will continue as PF/native
and likely we will see just different set of random errors
without realizing that PCI/MMIO is dead ...


> +		return false;
> +	}
> +
>  	return value & VF_CAP;
>  }
>