drm/i915: Fix modeset handling during gpu reset, v5.
authorMaarten Lankhorst <maarten.lankhorst@linux.intel.com>
Fri, 5 Aug 2016 20:28:27 +0000 (23:28 +0300)
committerVille Syrjälä <ville.syrjala@linux.intel.com>
Fri, 5 Aug 2016 20:28:27 +0000 (23:28 +0300)
This function would call drm_modeset_lock_all, while the suspend/resume
functions already have their own locking. Fix this by factoring out
__intel_display_resume, and calling the atomic helpers for duplicating
atomic state and disabling all crtc's during suspend.

Changes since v1:
- Deal with -EDEADLK right after lock_all and clean up calls
  to hw readout.
- Always take all modeset locks so updates during gpu reset are blocked.
Changes since v2:
- Fix deadlock in intel_update_primary_planes.
- Move WARN_ON(EDEADLK) to __intel_display_resume.
- pctx -> ctx
- only call __intel_display_resume on success in intel_display_resume.
Changes since v3:
- Rebase on top of dev_priv -> dev change.
- Use drm_modeset_lock_all_ctx instead of drm_modeset_lock_all.
Changes since v4 [by vsyrjala]:
- Deal with skip_intermediate_wm
- Update comment w.r.t. mode_config.mutex vs. ->detect()
- Rebase due to INTEL_GEN() etc.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Fixes: e2c8b8701e2d ("drm/i915: Use atomic helpers for suspend, v2.")
Cc: stable@vger.kernel.org
Tested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1470428910-12125-2-git-send-email-ville.syrjala@linux.intel.com
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/intel_display.c

index c36d17659ebe4b3d483bb7311a79e9cae36578cb..54f789c75aa1aa4c933027a812b658dca5401d9e 100644 (file)
@@ -1840,6 +1840,7 @@ struct drm_i915_private {
        enum modeset_restore modeset_restore;
        struct mutex modeset_restore_lock;
        struct drm_atomic_state *modeset_restore_state;
+       struct drm_modeset_acquire_ctx reset_ctx;
 
        struct list_head vm_list; /* Global list of all address spaces */
        struct i915_ggtt ggtt; /* VM representing the global address space */
index 47ad6b3368e347be9f64c395332447f3f8d40868..a951395753240c468406eb5a55e7471eaae595d8 100644 (file)
@@ -3093,40 +3093,110 @@ static void intel_update_primary_planes(struct drm_device *dev)
 
        for_each_crtc(dev, crtc) {
                struct intel_plane *plane = to_intel_plane(crtc->primary);
-               struct intel_plane_state *plane_state;
-
-               drm_modeset_lock_crtc(crtc, &plane->base);
-               plane_state = to_intel_plane_state(plane->base.state);
+               struct intel_plane_state *plane_state =
+                       to_intel_plane_state(plane->base.state);
 
                if (plane_state->visible)
                        plane->update_plane(&plane->base,
                                            to_intel_crtc_state(crtc->state),
                                            plane_state);
+       }
+}
+
+static int
+__intel_display_resume(struct drm_device *dev,
+                      struct drm_atomic_state *state)
+{
+       struct drm_crtc_state *crtc_state;
+       struct drm_crtc *crtc;
+       int i, ret;
 
-               drm_modeset_unlock_crtc(crtc);
+       intel_modeset_setup_hw_state(dev);
+       i915_redisable_vga(dev);
+
+       if (!state)
+               return 0;
+
+       for_each_crtc_in_state(state, crtc, crtc_state, i) {
+               /*
+                * Force recalculation even if we restore
+                * current state. With fast modeset this may not result
+                * in a modeset when the state is compatible.
+                */
+               crtc_state->mode_changed = true;
        }
+
+       /* ignore any reset values/BIOS leftovers in the WM registers */
+       to_intel_atomic_state(state)->skip_intermediate_wm = true;
+
+       ret = drm_atomic_commit(state);
+
+       WARN_ON(ret == -EDEADLK);
+       return ret;
 }
 
 void intel_prepare_reset(struct drm_i915_private *dev_priv)
 {
+       struct drm_device *dev = &dev_priv->drm;
+       struct drm_modeset_acquire_ctx *ctx = &dev_priv->reset_ctx;
+       struct drm_atomic_state *state;
+       int ret;
+
        /* no reset support for gen2 */
        if (IS_GEN2(dev_priv))
                return;
 
-       /* reset doesn't touch the display */
+       /*
+        * Need mode_config.mutex so that we don't
+        * trample ongoing ->detect() and whatnot.
+        */
+       mutex_lock(&dev->mode_config.mutex);
+       drm_modeset_acquire_init(ctx, 0);
+       while (1) {
+               ret = drm_modeset_lock_all_ctx(dev, ctx);
+               if (ret != -EDEADLK)
+                       break;
+
+               drm_modeset_backoff(ctx);
+       }
+
+       /* reset doesn't touch the display, but flips might get nuked anyway, */
        if (INTEL_GEN(dev_priv) >= 5 || IS_G4X(dev_priv))
                return;
 
-       drm_modeset_lock_all(&dev_priv->drm);
        /*
         * Disabling the crtcs gracefully seems nicer. Also the
         * g33 docs say we should at least disable all the planes.
         */
-       intel_display_suspend(&dev_priv->drm);
+       state = drm_atomic_helper_duplicate_state(dev, ctx);
+       if (IS_ERR(state)) {
+               ret = PTR_ERR(state);
+               state = NULL;
+               DRM_ERROR("Duplicating state failed with %i\n", ret);
+               goto err;
+       }
+
+       ret = drm_atomic_helper_disable_all(dev, ctx);
+       if (ret) {
+               DRM_ERROR("Suspending crtc's failed with %i\n", ret);
+               goto err;
+       }
+
+       dev_priv->modeset_restore_state = state;
+       state->acquire_ctx = ctx;
+       return;
+
+err:
+       drm_atomic_state_free(state);
 }
 
 void intel_finish_reset(struct drm_i915_private *dev_priv)
 {
+       struct drm_device *dev = &dev_priv->drm;
+       struct drm_modeset_acquire_ctx *ctx = &dev_priv->reset_ctx;
+       struct drm_atomic_state *state = dev_priv->modeset_restore_state;
+       int ret;
+
        /*
         * Flips in the rings will be nuked by the reset,
         * so complete all pending flips so that user space
@@ -3138,6 +3208,8 @@ void intel_finish_reset(struct drm_i915_private *dev_priv)
        if (IS_GEN2(dev_priv))
                return;
 
+       dev_priv->modeset_restore_state = NULL;
+
        /* reset doesn't touch the display */
        if (INTEL_GEN(dev_priv) >= 5 || IS_G4X(dev_priv)) {
                /*
@@ -3149,29 +3221,32 @@ void intel_finish_reset(struct drm_i915_private *dev_priv)
                 * FIXME: Atomic will make this obsolete since we won't schedule
                 * CS-based flips (which might get lost in gpu resets) any more.
                 */
-               intel_update_primary_planes(&dev_priv->drm);
-               return;
-       }
-
-       /*
-        * The display has been reset as well,
-        * so need a full re-initialization.
-        */
-       intel_runtime_pm_disable_interrupts(dev_priv);
-       intel_runtime_pm_enable_interrupts(dev_priv);
+               intel_update_primary_planes(dev);
+       } else {
+               /*
+                * The display has been reset as well,
+                * so need a full re-initialization.
+                */
+               intel_runtime_pm_disable_interrupts(dev_priv);
+               intel_runtime_pm_enable_interrupts(dev_priv);
 
-       intel_modeset_init_hw(&dev_priv->drm);
+               intel_modeset_init_hw(dev);
 
-       spin_lock_irq(&dev_priv->irq_lock);
-       if (dev_priv->display.hpd_irq_setup)
-               dev_priv->display.hpd_irq_setup(dev_priv);
-       spin_unlock_irq(&dev_priv->irq_lock);
+               spin_lock_irq(&dev_priv->irq_lock);
+               if (dev_priv->display.hpd_irq_setup)
+                       dev_priv->display.hpd_irq_setup(dev_priv);
+               spin_unlock_irq(&dev_priv->irq_lock);
 
-       intel_display_resume(&dev_priv->drm);
+               ret = __intel_display_resume(dev, state);
+               if (ret)
+                       DRM_ERROR("Restoring old state failed with %i\n", ret);
 
-       intel_hpd_init(dev_priv);
+               intel_hpd_init(dev_priv);
+       }
 
-       drm_modeset_unlock_all(&dev_priv->drm);
+       drm_modeset_drop_locks(ctx);
+       drm_modeset_acquire_fini(ctx);
+       mutex_unlock(&dev->mode_config.mutex);
 }
 
 static bool intel_crtc_has_pending_flip(struct drm_crtc *crtc)
@@ -16180,9 +16255,10 @@ void intel_display_resume(struct drm_device *dev)
        struct drm_atomic_state *state = dev_priv->modeset_restore_state;
        struct drm_modeset_acquire_ctx ctx;
        int ret;
-       bool setup = false;
 
        dev_priv->modeset_restore_state = NULL;
+       if (state)
+               state->acquire_ctx = &ctx;
 
        /*
         * This is a cludge because with real atomic modeset mode_config.mutex
@@ -16193,43 +16269,17 @@ void intel_display_resume(struct drm_device *dev)
        mutex_lock(&dev->mode_config.mutex);
        drm_modeset_acquire_init(&ctx, 0);
 
-retry:
-       ret = drm_modeset_lock_all_ctx(dev, &ctx);
-
-       if (ret == 0 && !setup) {
-               setup = true;
-
-               intel_modeset_setup_hw_state(dev);
-               i915_redisable_vga(dev);
-       }
-
-       if (ret == 0 && state) {
-               struct drm_crtc_state *crtc_state;
-               struct drm_crtc *crtc;
-               int i;
-
-               state->acquire_ctx = &ctx;
-
-               /* ignore any reset values/BIOS leftovers in the WM registers */
-               to_intel_atomic_state(state)->skip_intermediate_wm = true;
-
-               for_each_crtc_in_state(state, crtc, crtc_state, i) {
-                       /*
-                        * Force recalculation even if we restore
-                        * current state. With fast modeset this may not result
-                        * in a modeset when the state is compatible.
-                        */
-                       crtc_state->mode_changed = true;
-               }
-
-               ret = drm_atomic_commit(state);
-       }
+       while (1) {
+               ret = drm_modeset_lock_all_ctx(dev, &ctx);
+               if (ret != -EDEADLK)
+                       break;
 
-       if (ret == -EDEADLK) {
                drm_modeset_backoff(&ctx);
-               goto retry;
        }
 
+       if (!ret)
+               ret = __intel_display_resume(dev, state);
+
        drm_modeset_drop_locks(&ctx);
        drm_modeset_acquire_fini(&ctx);
        mutex_unlock(&dev->mode_config.mutex);