[Intel-gfx] [PATCH] drm/i915: Wait for the GPU whilst shrinking, if truly desperate.

[Chris Wilson]

By idling the GPU and discarding everything we can when under extreme
memory pressure, the number of OOM-killer events is dramatically
reduced. For instance, this makes it possible to run
firefox-planet-gnome.trace again on my swapless 512MiB i915.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_gem.c |   45 +++++++++++++++++++++++++++++++++++++++
 1 files changed, 45 insertions(+), 0 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index a3fd19c..244d4ff 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -5250,6 +5250,20 @@ void i915_gem_release(struct drm_device * dev, struct drm_file *file_priv)
 }
 
 static int
+i915_gpu_is_active(struct drm_device *dev)
+{
+	drm_i915_private_t *dev_priv = dev->dev_private;
+	int lists_empty;
+
+	spin_lock(&dev_priv->mm.active_list_lock);
+	lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
+		      list_empty(&dev_priv->mm.active_list);
+	spin_unlock(&dev_priv->mm.active_list_lock);
+
+	return !lists_empty;
+}
+
+static int
 i915_gem_shrink(int nr_to_scan, gfp_t gfp_mask)
 {
 	drm_i915_private_t *dev_priv, *next_dev;
@@ -5278,6 +5292,7 @@ i915_gem_shrink(int nr_to_scan, gfp_t gfp_mask)
 
 	spin_lock(&shrink_list_lock);
 
+rescan:
 	/* first scan for clean buffers */
 	list_for_each_entry_safe(dev_priv, next_dev,
 				 &shrink_list, mm.shrink_list) {
@@ -5335,6 +5350,36 @@ i915_gem_shrink(int nr_to_scan, gfp_t gfp_mask)
 		would_deadlock = 0;
 	}
 
+	if (nr_to_scan) {
+		int active = 0;
+
+		/*
+		 * We are desperate for pages, so as a last resort, wait
+		 * for the GPU to finish and discard whatever we can.
+		 * This has a dramatic impact to reduce the number of
+		 * OOM-killer events whilst running the GPU aggressively.
+		 */
+		list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
+			struct drm_device *dev = dev_priv->dev;
+
+			if (!mutex_trylock(&dev->struct_mutex))
+				continue;
+
+			spin_unlock(&shrink_list_lock);
+
+			if (i915_gpu_is_active(dev)) {
+				i915_gpu_idle(dev);
+				active++;
+			}
+
+			spin_lock(&shrink_list_lock);
+			mutex_unlock(&dev->struct_mutex);
+		}
+
+		if (active)
+			goto rescan;
+	}
+
 	spin_unlock(&shrink_list_lock);
 
 	if (would_deadlock)
-- 
1.7.0.4