drm/i915: Unwind i915_gem_init() failure
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 13 Dec 2017 13:43:47 +0000 (13:43 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 13 Dec 2017 18:53:35 +0000 (18:53 +0000)
Since Michal introduced new user controllable errors other than -EIO
during i915_gem_init(), we need to actually unwind on the error path as
we have to abort the module load (and we expect to do so cleanly!).

As we now teardown key state and then mark the driver as wedged (on
EIO), we have to be careful to not allow ourselves to resume and
unwedge, thus attempting to use the uninitialised driver.

v2: Try not to free driver state for the suppressed EIO
v3: Use load-fault-injection to test both error/recovery paths.

References: 8620eb1dbbf2 ("drm/i915/uc: Don't use -EIO to report missing firmware")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Sagar Arun Kamble <sagar.a.kamble@intel.com>
Reviewed-by: MichaƂ Winiarski <michal.winiarski@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171213134347.4608-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem.c

index 8c3d801..13fa262 100644 (file)
@@ -4865,7 +4865,8 @@ void i915_gem_resume(struct drm_i915_private *i915)
        i915_gem_restore_gtt_mappings(i915);
        i915_gem_restore_fences(i915);
 
-       /* As we didn't flush the kernel context before suspend, we cannot
+       /*
+        * As we didn't flush the kernel context before suspend, we cannot
         * guarantee that the context image is complete. So let's just reset
         * it and start again.
         */
@@ -4886,8 +4887,10 @@ out_unlock:
        return;
 
 err_wedged:
-       DRM_ERROR("failed to re-initialize GPU, declaring wedged!\n");
-       i915_gem_set_wedged(i915);
+       if (!i915_terminally_wedged(&i915->gpu_error)) {
+               DRM_ERROR("failed to re-initialize GPU, declaring wedged!\n");
+               i915_gem_set_wedged(i915);
+       }
        goto out_unlock;
 }
 
@@ -5170,22 +5173,28 @@ int i915_gem_init(struct drm_i915_private *dev_priv)
        intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
 
        ret = i915_gem_init_ggtt(dev_priv);
-       if (ret)
-               goto out_unlock;
+       if (ret) {
+               GEM_BUG_ON(ret == -EIO);
+               goto err_unlock;
+       }
 
        ret = i915_gem_contexts_init(dev_priv);
-       if (ret)
-               goto out_unlock;
+       if (ret) {
+               GEM_BUG_ON(ret == -EIO);
+               goto err_ggtt;
+       }
 
        ret = intel_engines_init(dev_priv);
-       if (ret)
-               goto out_unlock;
+       if (ret) {
+               GEM_BUG_ON(ret == -EIO);
+               goto err_context;
+       }
 
        intel_init_gt_powersave(dev_priv);
 
        ret = i915_gem_init_hw(dev_priv);
        if (ret)
-               goto out_unlock;
+               goto err_pm;
 
        /*
         * Despite its name intel_init_clock_gating applies both display
@@ -5199,9 +5208,53 @@ int i915_gem_init(struct drm_i915_private *dev_priv)
        intel_init_clock_gating(dev_priv);
 
        ret = __intel_engines_record_defaults(dev_priv);
-out_unlock:
+       if (ret)
+               goto err_init_hw;
+
+       if (i915_inject_load_failure()) {
+               ret = -ENODEV;
+               goto err_init_hw;
+       }
+
+       if (i915_inject_load_failure()) {
+               ret = -EIO;
+               goto err_init_hw;
+       }
+
+       intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
+       mutex_unlock(&dev_priv->drm.struct_mutex);
+
+       return 0;
+
+       /*
+        * Unwinding is complicated by that we want to handle -EIO to mean
+        * disable GPU submission but keep KMS alive. We want to mark the
+        * HW as irrevisibly wedged, but keep enough state around that the
+        * driver doesn't explode during runtime.
+        */
+err_init_hw:
+       i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
+       i915_gem_contexts_lost(dev_priv);
+       intel_uc_fini_hw(dev_priv);
+err_pm:
+       if (ret != -EIO) {
+               intel_cleanup_gt_powersave(dev_priv);
+               i915_gem_cleanup_engines(dev_priv);
+       }
+err_context:
+       if (ret != -EIO)
+               i915_gem_contexts_fini(dev_priv);
+err_ggtt:
+err_unlock:
+       intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
+       mutex_unlock(&dev_priv->drm.struct_mutex);
+
+       if (ret != -EIO)
+               i915_gem_cleanup_userptr(dev_priv);
+
        if (ret == -EIO) {
-               /* Allow engine initialisation to fail by marking the GPU as
+               /*
+                * Allow engine initialisation to fail by marking the GPU as
                 * wedged. But we only want to do this where the GPU is angry,
                 * for all other failure, such as an allocation failure, bail.
                 */
@@ -5211,9 +5264,8 @@ out_unlock:
                }
                ret = 0;
        }
-       intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
-       mutex_unlock(&dev_priv->drm.struct_mutex);
 
+       i915_gem_drain_freed_objects(dev_priv);
        return ret;
 }