[Intel-gfx] [PATCH 1/2] drm/i915: Ratelimit i915_global_park

[Chris Wilson]

When doing our global park, we like to be a good citizen and shrink our
slab caches (of which we have quite a few now), but each
kmem_cache_shrink() incurs a stop_machine() and so ends up being quite
expensive, causing machine-wide stalls. While ideally we would like to
throw away unused pages in our slab caches whenever it appears that we
are idling, doing so will require a much cheaper mechanism. In the
meantime use a delayed worked to impose a rate-limit that means we have
to have been idle for more than 2 seconds before we start shrinking.

References: https://gitlab.freedesktop.org/drm/intel/issues/848
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 drivers/gpu/drm/i915/i915_globals.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_globals.c b/drivers/gpu/drm/i915/i915_globals.c
index be127cd28931..e6984b897a61 100644
--- a/drivers/gpu/drm/i915/i915_globals.c
+++ b/drivers/gpu/drm/i915/i915_globals.c
@@ -20,7 +20,8 @@ static LIST_HEAD(globals);
 static atomic_t active;
 static atomic_t epoch;
 static struct park_work {
-	struct rcu_work work;
+	struct delayed_work work;
+	struct rcu_head rcu;
 	int epoch;
 } park;
 
@@ -44,6 +45,13 @@ static void __i915_globals_park(struct work_struct *work)
 		i915_globals_shrink();
 }
 
+static void __i915_globals_grace(struct rcu_head *rcu)
+{
+	/* Ratelimit parking as shrinking is quite slow */
+	mod_delayed_work(system_wq, &park.work,
+			 round_jiffies_up_relative(2 * HZ));
+}
+
 void __init i915_global_register(struct i915_global *global)
 {
 	GEM_BUG_ON(!global->shrink);
@@ -85,7 +93,7 @@ int __init i915_globals_init(void)
 		}
 	}
 
-	INIT_RCU_WORK(&park.work, __i915_globals_park);
+	INIT_DELAYED_WORK(&park.work, __i915_globals_park);
 	return 0;
 }
 
@@ -103,8 +111,9 @@ void i915_globals_park(void)
 	if (!atomic_dec_and_test(&active))
 		return;
 
+	/* Queue cleanup after the next RCU grace period has freed slabs */
 	park.epoch = atomic_inc_return(&epoch);
-	queue_rcu_work(system_wq, &park.work);
+	call_rcu(&park.rcu, __i915_globals_grace);
 }
 
 void i915_globals_unpark(void)
@@ -116,8 +125,9 @@ void i915_globals_unpark(void)
 void __exit i915_globals_exit(void)
 {
 	/* Flush any residual park_work */
-	atomic_inc(&epoch);
-	flush_rcu_work(&park.work);
+	atomic_inc(&epoch); /* skip shrinking */
+	rcu_barrier(); /* wait for the work to be queued */
+	flush_delayed_work(&park.work);
 
 	__i915_globals_cleanup();
 
-- 
2.24.1