Re: [PATCH v13 3/4] drm/xe: Handle PUNIT errors by requesting cold-reset recovery

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Unprotected concurrent access to `xe->wedged.method` can cause the `COLD_RESET` recovery method to be overwritten by the `DEFAULT` fallback method during simultaneous wedge declarations.
--

--- Patch [3]: [PATCH v13 3/4] drm/xe: Handle PUNIT errors by requesting cold-reset recovery ---
commit 5edfcf29ea367e4df826b556199c8154b0704d81
Author: Mallesh Koujalagi <[email protected]>

drm/xe: Handle PUNIT errors by requesting cold-reset recovery

This commit handles PUNIT errors by marking the device as wedged with
DRM_WEDGE_RECOVERY_COLD_RESET, indicating that a full device power cycle
is required to restore normal operation.

> diff --git a/drivers/gpu/drm/xe/xe_ras.c b/drivers/gpu/drm/xe/xe_ras.c
> index 352f056fd9ef4..3033a377bace9 100644
> --- a/drivers/gpu/drm/xe/xe_ras.c
> +++ b/drivers/gpu/drm/xe/xe_ras.c
[ ... ]
> @@ -254,6 +254,12 @@ static u8 handle_core_compute_errors(struct xe_ras_error_array *arr)
>  	return XE_RAS_RECOVERY_ACTION_RECOVERED;
>  }
>  
> +static void punit_error_handler(struct xe_device *xe)
> +{
> +	xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_COLD_RESET);
> +	xe_device_declare_wedged(xe);
> +}
        ^^^^

[Severity: High]
This is a pre-existing issue, but can the COLD_RESET recovery method be lost
here due to unprotected concurrent access to xe->wedged.method?

If a GT hang occurs concurrently with a PUNIT error, a race condition can
overwrite the COLD_RESET escalation with the fallback method:

CPU1 (GT hang)
xe_device_declare_wedged()
// Evaluates as true (sees 0)
    if (!xe->wedged.method)

CPU2 (PUNIT error)
punit_error_handler()
    xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_COLD_RESET);

CPU1 (GT hang)
    // Overwrites COLD_RESET
    xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_REBIND | 
                                    DRM_WEDGE_RECOVERY_BUS_RESET);

Because xe->wedged.method is assigned directly and read in a check-then-act
pattern without locking, both threads might pass the incorrect fallback method
to drm_dev_wedged_event(). 

Could this cause userspace to perform a generic bus reset instead of the 
required full power cycle?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3