From: Chris Wilson <chris@chris-wilson.co.uk>
Date: Wed, 24 Jan 2018 11:36:08 +0000
Subject: drm/i915: Shrink the GEM kmem_caches upon idling
Git-commit: 84a1074920523430f9dc30ff907f4801b4820072
Patch-mainline: v4.17-rc1
References: FATE#326289 FATE#326079 FATE#326049 FATE#322398 FATE#326166

When we finally decide the gpu is idle, that is a good time to shrink
our kmem_caches.

v3: Defer until an rcu grace period after we idle.
v4: Think about epoch wraparound and how likely that is.
v5: Use I915_EPOCH_INVALID magic.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180124113608.14909-2-chris@chris-wilson.co.uk

Acked-by: Petr Tesarik <ptesarik@suse.com>
---
 drivers/gpu/drm/i915/i915_gem.c |   79 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 79 insertions(+)

--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -3345,6 +3345,65 @@ i915_gem_retire_work_handler(struct work
 				   round_jiffies_up_relative(HZ));
 }
 
+static void shrink_caches(struct drm_i915_private *i915)
+{
+	/*
+	 * kmem_cache_shrink() discards empty slabs and reorders partially
+	 * filled slabs to prioritise allocating from the mostly full slabs,
+	 * with the aim of reducing fragmentation.
+	 */
+	kmem_cache_shrink(i915->priorities);
+	kmem_cache_shrink(i915->dependencies);
+	kmem_cache_shrink(i915->requests);
+	kmem_cache_shrink(i915->luts);
+	kmem_cache_shrink(i915->vmas);
+	kmem_cache_shrink(i915->objects);
+}
+
+struct sleep_rcu_work {
+	union {
+		struct rcu_head rcu;
+		struct work_struct work;
+	};
+	struct drm_i915_private *i915;
+	unsigned int epoch;
+};
+
+static inline bool
+same_epoch(struct drm_i915_private *i915, unsigned int epoch)
+{
+	/*
+	 * There is a small chance that the epoch wrapped since we started
+	 * sleeping. If we assume that epoch is at least a u32, then it will
+	 * take at least 2^32 * 100ms for it to wrap, or about 326 years.
+	 */
+	return epoch == READ_ONCE(i915->gt.epoch);
+}
+
+static void __sleep_work(struct work_struct *work)
+{
+	struct sleep_rcu_work *s = container_of(work, typeof(*s), work);
+	struct drm_i915_private *i915 = s->i915;
+	unsigned int epoch = s->epoch;
+
+	kfree(s);
+	if (same_epoch(i915, epoch))
+		shrink_caches(i915);
+}
+
+static void __sleep_rcu(struct rcu_head *rcu)
+{
+	struct sleep_rcu_work *s = container_of(rcu, typeof(*s), rcu);
+	struct drm_i915_private *i915 = s->i915;
+
+	if (same_epoch(i915, s->epoch)) {
+		INIT_WORK(&s->work, __sleep_work);
+		queue_work(i915->wq, &s->work);
+	} else {
+		kfree(s);
+	}
+}
+
 static inline bool
 new_requests_since_last_retire(const struct drm_i915_private *i915)
 {
@@ -3357,6 +3416,7 @@ i915_gem_idle_work_handler(struct work_s
 {
 	struct drm_i915_private *dev_priv =
 		container_of(work, typeof(*dev_priv), gt.idle_work.work);
+	unsigned int epoch = I915_EPOCH_INVALID;
 	bool rearm_hangcheck;
 	ktime_t end;
 
@@ -3416,6 +3476,8 @@ i915_gem_idle_work_handler(struct work_s
 
 	GEM_BUG_ON(!dev_priv->gt.awake);
 	dev_priv->gt.awake = false;
+	epoch = dev_priv->gt.epoch;
+	GEM_BUG_ON(epoch == I915_EPOCH_INVALID);
 	rearm_hangcheck = false;
 
 	if (INTEL_GEN(dev_priv) >= 6)
@@ -3432,6 +3494,23 @@ out_rearm:
 		GEM_BUG_ON(!dev_priv->gt.awake);
 		i915_queue_hangcheck(dev_priv);
 	}
+
+	/*
+	 * When we are idle, it is an opportune time to reap our caches.
+	 * However, we have many objects that utilise RCU and the ordered
+	 * i915->wq that this work is executing on. To try and flush any
+	 * pending frees now we are idle, we first wait for an RCU grace
+	 * period, and then queue a task (that will run last on the wq) to
+	 * shrink and re-optimize the caches.
+	 */
+	if (same_epoch(dev_priv, epoch)) {
+		struct sleep_rcu_work *s = kmalloc(sizeof(*s), GFP_KERNEL);
+		if (s) {
+			s->i915 = dev_priv;
+			s->epoch = epoch;
+			call_rcu(&s->rcu, __sleep_rcu);
+		}
+	}
 }
 
 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)