drm/i915: Avoid GPU hang when coming out of s3 or s4
authorPeter Antoine <peter.antoine@intel.com>
Mon, 11 May 2015 07:50:45 +0000 (08:50 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 29 Jun 2015 19:29:08 +0000 (12:29 -0700)
commit 364aece01a2dd748fc36a1e8bf52ef639b0857bd upstream.

This patch fixes a timing issue that causes a GPU hang when the system
comes out of power saving.

During pm_resume, We are submitting batchbuffers before enabling
Interrupts this is causing us to miss the context switch interrupt,
and in consequence intel_execlists_handle_ctx_events is not triggered.

This patch is based on a patch from Deepak S <deepak.s@intel.com>
from another platform.

The patch fixes an issue introduced by:
  commit e7778be1eab918274f79603d7c17b3ec8be77386
  drm/i915: Fix startup failure in LRC mode after recent init changes

The above patch added a call to init_context() to fix an issue introduced
by a previous patch. But, it then opened up a small timing window for the
batches being added by the init_context (basically setting up the context)
to complete before the interrupts have been turned on, thus hanging the
GPU.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89600
Cc: stable@vger.kernel.org # 4.0+
Signed-off-by: Peter Antoine <peter.antoine@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
[Jani: fixed typo in subject, massaged the comments a bit]
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/gpu/drm/i915/i915_drv.c

index ec4d932f8be483e0fe4aa4edd837e0c595330273..169123a6ad0ee3c37d9d49c0c6c9ba3065399633 100644 (file)
@@ -693,6 +693,16 @@ static int i915_drm_resume(struct drm_device *dev)
                intel_init_pch_refclk(dev);
                drm_mode_config_reset(dev);
 
+               /*
+                * Interrupts have to be enabled before any batches are run.
+                * If not the GPU will hang. i915_gem_init_hw() will initiate
+                * batches to update/restore the context.
+                *
+                * Modeset enabling in intel_modeset_init_hw() also needs
+                * working interrupts.
+                */
+               intel_runtime_pm_enable_interrupts(dev_priv);
+
                mutex_lock(&dev->struct_mutex);
                if (i915_gem_init_hw(dev)) {
                        DRM_ERROR("failed to re-initialize GPU, declaring wedged!\n");
@@ -700,9 +710,6 @@ static int i915_drm_resume(struct drm_device *dev)
                }
                mutex_unlock(&dev->struct_mutex);
 
-               /* We need working interrupts for modeset enabling ... */
-               intel_runtime_pm_enable_interrupts(dev_priv);
-
                intel_modeset_init_hw(dev);
 
                spin_lock_irq(&dev_priv->irq_lock);