Re: [PATCH v2 4/5] drm/xe: Add handler for power management unit errors which require cold-reset

[Raag Jadav <raag.jadav@intel.com>]

On Wed, Mar 18, 2026 at 12:10:21PM +0530, Mallesh Koujalagi wrote:
> This handler is designed to be called when power management unit errors are
> detected that affect device-level state persisting across warm resets. The
> cold reset recovery method signals to userspace that only a complete device
> power cycle can restore normal operation.
> 
> v2:
> - Add use case: Handling errors from power management unit,
>   which requires a complete power cycle (cold reset)
>   to recover. (Christian)
> 
> Signed-off-by: Mallesh Koujalagi <mallesh.koujalagi@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_hw_error.c | 27 +++++++++++++++++++++++++++
>  drivers/gpu/drm/xe/xe_hw_error.h |  1 +
>  drivers/gpu/drm/xe/xe_ras.c      |  3 ++-
>  3 files changed, 30 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_hw_error.c b/drivers/gpu/drm/xe/xe_hw_error.c
> index 2a31b430570e..ca965a2b092c 100644
> --- a/drivers/gpu/drm/xe/xe_hw_error.c
> +++ b/drivers/gpu/drm/xe/xe_hw_error.c
> @@ -5,6 +5,7 @@
>  
>  #include <linux/bitmap.h>
>  #include <linux/fault-inject.h>

Blank line to distinguish between linux and drm includes.

> +#include <drm/drm_drv.h>
>  
>  #include "regs/xe_gsc_regs.h"
>  #include "regs/xe_hw_error_regs.h"
> @@ -542,6 +543,32 @@ static void process_hw_errors(struct xe_device *xe)
>  	}
>  }
>  
> +/**
> + * xe_punit_error_handler - Handler for power management unit errors
> + * @xe: device instance
> + *
> + * Handles power management unit errors that affect the device and cannot
> + * be recovered through driver reload, PCIe reset, etc.
> + *
> + * Marks the device as wedged with DRM_WEDGE_RECOVERY_COLD_RESET method
> + * and notifies userspace that a complete device power cycle is required.
> + */
> +void xe_punit_error_handler(struct xe_device *xe)
> +{
> +	drm_err(&xe->drm, "CRITICAL: PMU error detected\n");
> +	drm_err(&xe->drm, "Recovery: Device cold reset required\n");

The caller is already printing these so we'd want to avoid dmesg spam.

> +	/* Set cold reset recovery method */

Comments are more useful for something that's not obvious from the code,
so we try to make the code as self documenting as possible.

> +	xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_COLD_RESET);
> +
> +	if (xe_device_wedged(xe)) {

We should not be at this point if the device is already wedged, so I
believe this needs to be handled somewhere upstream.

> +		drm_dev_wedged_event(&xe->drm, xe->wedged.method, NULL);
> +	} else {
> +		/* Declare device wedged - will trigger uevent with cold reset method */

Same as above.

Raag

> +		xe_device_declare_wedged(xe);
> +	}
> +}
> +
>  /**
>   * xe_hw_error_init - Initialize hw errors
>   * @xe: xe device instance
> diff --git a/drivers/gpu/drm/xe/xe_hw_error.h b/drivers/gpu/drm/xe/xe_hw_error.h
> index d86e28c5180c..f588320eb94d 100644
> --- a/drivers/gpu/drm/xe/xe_hw_error.h
> +++ b/drivers/gpu/drm/xe/xe_hw_error.h
> @@ -11,5 +11,6 @@ struct xe_tile;
>  struct xe_device;
>  
>  void xe_hw_error_irq_handler(struct xe_tile *tile, const u32 master_ctl);
> +void xe_punit_error_handler(struct xe_device *xe);
>  void xe_hw_error_init(struct xe_device *xe);
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_ras.c b/drivers/gpu/drm/xe/xe_ras.c
> index 777321021391..93257d0eaaa0 100644
> --- a/drivers/gpu/drm/xe/xe_ras.c
> +++ b/drivers/gpu/drm/xe/xe_ras.c
> @@ -10,6 +10,7 @@
>  #include "xe_survivability_mode.h"
>  #include "xe_sysctrl_mailbox.h"
>  #include "xe_sysctrl_mailbox_types.h"
> +#include "xe_hw_error.h"
>  
>  #define COMPUTE_ERROR_SEVERITY_MASK		GENMASK(26, 25)
>  #define GLOBAL_UNCORR_ERROR			2
> @@ -148,7 +149,7 @@ static enum xe_ras_recovery_action handle_soc_internal_errors(struct xe_device *
>  			xe_err(xe, "[RAS]: PUNIT %s error detected: 0x%x\n",
>  			       severity_to_str(xe, common_info.severity),
>  			       ieh_error->error_sources_ieh0.punit);
> -			/** TODO: Add PUNIT error handling */
> +			xe_punit_error_handler(xe);
>  			action = XE_RAS_RECOVERY_ACTION_DISCONNECT;
>  		}
>  	}
> -- 
> 2.34.1
>