Re: [PATCH v12 4/4] drm/xe/ras: Use fault-inject to trigger punit error handler

[Mallesh, Koujalagi]

On 30-07-2026 12:56 pm, Mallesh, Koujalagi wrote:

On 29-07-2026 12:55 am, Rodrigo Vivi wrote:
On Fri, Jul 24, 2026 at 03:33:07PM +0530, Mallesh Koujalagi wrote:
Use fault-inject framework to trigger punit_error_handler()
for testing.

Usage:
   echo 100 > .../inject_punit_error/probability
   echo 1   > .../inject_punit_error/times

Signed-off-by: Mallesh Koujalagi <mallesh.koujal...@intel.com>
---
  drivers/gpu/drm/xe/xe_debugfs.c |  4 ++++
  drivers/gpu/drm/xe/xe_debugfs.h |  2 ++
  drivers/gpu/drm/xe/xe_ras.c     | 17 +++++++++++++++++
  3 files changed, 23 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_debugfs.c 
b/drivers/gpu/drm/xe/xe_debugfs.c
index 5a3877fcb0f0..5599d493feca 100644
--- a/drivers/gpu/drm/xe/xe_debugfs.c
+++ b/drivers/gpu/drm/xe/xe_debugfs.c
@@ -42,6 +42,7 @@
    DECLARE_FAULT_ATTR(gt_reset_failure);
  DECLARE_FAULT_ATTR(inject_csc_hw_error);
+DECLARE_FAULT_ATTR(inject_punit_error);
    static bool csc_hw_error_available(struct xe_device *xe)
  {
@@ -62,6 +63,8 @@ static struct {
      { .name = "inject_csc_hw_error",
        .attr = &inject_csc_hw_error,
        .is_visible = csc_hw_error_available },
+    { .name = "inject_punit_error",
+      .attr = &inject_punit_error },
  };
    /*
@@ -76,6 +79,7 @@ bool xe_fault_##name(void)                \
    FAULT_ACTION(gt_reset, gt_reset_failure)
  FAULT_ACTION(csc_hw_error, inject_csc_hw_error)
+FAULT_ACTION(punit_error, inject_punit_error)
    static void xe_fault_inject_debugfs_register(struct xe_device *xe,
                           struct dentry *root)
diff --git a/drivers/gpu/drm/xe/xe_debugfs.h 
b/drivers/gpu/drm/xe/xe_debugfs.h
index cd56f7442b99..dd57914dd4f2 100644
--- a/drivers/gpu/drm/xe/xe_debugfs.h
+++ b/drivers/gpu/drm/xe/xe_debugfs.h
@@ -13,10 +13,12 @@ struct xe_device;
  #ifdef CONFIG_DEBUG_FS
  bool xe_fault_gt_reset(void);
  bool xe_fault_csc_hw_error(void);
+bool xe_fault_punit_error(void);
  void xe_debugfs_register(struct xe_device *xe);
  #else
  static inline bool xe_fault_gt_reset(void) { return false; }
  static inline bool xe_fault_csc_hw_error(void) { return false; }
+static inline bool xe_fault_punit_error(void) { return false; }
  static inline void xe_debugfs_register(struct xe_device *xe) { }
  #endif
  diff --git a/drivers/gpu/drm/xe/xe_ras.c 
b/drivers/gpu/drm/xe/xe_ras.c
index 92b4181026cb..da06c16cf7ed 100644
--- a/drivers/gpu/drm/xe/xe_ras.c
+++ b/drivers/gpu/drm/xe/xe_ras.c
@@ -3,6 +3,7 @@
   * Copyright © 2026 Intel Corporation
   */
  +#include "xe_debugfs.h"
  #include "xe_device.h"
  #include "xe_drm_ras.h"
  #include "xe_pm.h"
@@ -391,6 +392,22 @@ enum xe_ras_recovery_action 
xe_ras_process_errors(struct xe_device *xe)
      size_t rlen;
      int ret;
  +    /*
+     * Only allow the injected PUNIT error once the DRM device is 
registered.
That entirely defeats the purpose! :)

Sashiko identified this corner case so I asked you to ensure that 
this is covered
and to prove that you should provide a fault-inject with test case.

But then you are skipping exactly the corner case. That makes 
absolutely no sense.

You are absolutely right!! Guarding the injection with 
xe->drm.registered was the wrong approach.

The correct fix is structural, we are splitting xe_ras_init function 
into xe_ras_init_early() (keep all bookkeeping) and

xe_ras_init() (called after  drm_dev_register in which 
xe_ras_process_errors now lives).


Thanks,

-/Mallesh

Sashiko corner case issue is addressed with patch: 
https://gitlab.freedesktop.org/drm/tip/-/commit/20bc4883c7c0e28c3ba6c76ccc279486c349dd3e

Do we still need this patch? Please let me know anything else I need to 
address here.

Thanks,

-/Mallesh

+     * xe_ras_process_errors() also runs during probe (via 
xe_ras_init()), before
+     * drm_dev_register() calls device_add() on the DRM minor's 
kdev. At that
+     * point kdev->kobj.parent is not yet linked into the sysfs 
hierarchy, so
+     * kobject_get_path() returns "/card0" instead of the real 
sysfs path.
+     * drm_dev_wedged_event() would then emit a KOBJ_CHANGE uevent 
with a wrong
+     * DEVPATH that udev cannot resolve, silently dropping the 
event and leaving
+     * the cold-reset recovery broken.
+     */
+    if (xe->drm.registered && xe_fault_punit_error()) {
+        xe_err(xe, "[RAS]: PUNIT error injected\n");
+        punit_error_handler(xe);
+        return XE_RAS_RECOVERY_ACTION_DISCONNECT;
+    }
+
      if (!xe->info.has_sysctrl)
          return XE_RAS_RECOVERY_ACTION_RESET;
  --
2.48.1