drm/i915: Make request allocation caches global

As kmem_caches share the same properties (size, allocation/free behaviour)
for all potential devices, we can use global caches. While this
potential has worse fragmentation behaviour (one can argue that
different devices would have different activity lifetimes, but you can
also argue that activity is temporal across the system) it is the
default behaviour of the system at large to amalgamate matching caches.

The benefit for us is much reduced pointer dancing along the frequent
allocation paths.

v2: Defer shrinking until after a global grace period for futureproofing
multiple consumers of the slab caches, similar to the current strategy
for avoiding shrinking too early.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190228102035.5857-1-chris@chris-wilson.co.uk

1 files changed, 113 insertions, 0 deletions

diff --git a/drivers/gpu/drm/i915/i915_globals.c b/drivers/gpu/drm/i915/i915_globals.c
new file mode 100644
index 000000000000..7fd1b3945a04
--- /dev/null
+++ b/drivers/gpu/drm/i915/i915_globals.c
@@ -0,0 +1,113 @@
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#include <linux/slab.h>
+#include <linux/workqueue.h>
+
+#include "i915_active.h"
+#include "i915_globals.h"
+#include "i915_request.h"
+#include "i915_scheduler.h"
+
+int __init i915_globals_init(void)
+{
+	int err;
+
+	err = i915_global_active_init();
+	if (err)
+		return err;
+
+	err = i915_global_request_init();
+	if (err)
+		goto err_active;
+
+	err = i915_global_scheduler_init();
+	if (err)
+		goto err_request;
+
+	return 0;
+
+err_request:
+	i915_global_request_exit();
+err_active:
+	i915_global_active_exit();
+	return err;
+}
+
+static void i915_globals_shrink(void)
+{
+	/*
+	 * 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.
+	 */
+	i915_global_active_shrink();
+	i915_global_request_shrink();
+	i915_global_scheduler_shrink();
+}
+
+static atomic_t active;
+static atomic_t epoch;
+struct park_work {
+	struct rcu_work work;
+	int epoch;
+};
+
+static void __i915_globals_park(struct work_struct *work)
+{
+	struct park_work *wrk = container_of(work, typeof(*wrk), work.work);
+
+	/* Confirm nothing woke up in the last grace period */
+	if (wrk->epoch == atomic_read(&epoch))
+		i915_globals_shrink();
+
+	kfree(wrk);
+}
+
+void i915_globals_park(void)
+{
+	struct park_work *wrk;
+
+	/*
+	 * Defer shrinking the global slab caches (and other work) until
+	 * after a RCU grace period has completed with no activity. This
+	 * is to try and reduce the latency impact on the consumers caused
+	 * by us shrinking the caches the same time as they are trying to
+	 * allocate, with the assumption being that if we idle long enough
+	 * for an RCU grace period to elapse since the last use, it is likely
+	 * to be longer until we need the caches again.
+	 */
+	if (!atomic_dec_and_test(&active))
+		return;
+
+	wrk = kmalloc(sizeof(*wrk), GFP_KERNEL);
+	if (!wrk)
+		return;
+
+	wrk->epoch = atomic_inc_return(&epoch);
+	INIT_RCU_WORK(&wrk->work, __i915_globals_park);
+	queue_rcu_work(system_wq, &wrk->work);
+}
+
+void i915_globals_unpark(void)
+{
+	atomic_inc(&epoch);
+	atomic_inc(&active);
+}
+
+void __exit i915_globals_exit(void)
+{
+	/* Flush any residual park_work */
+	rcu_barrier();
+	flush_scheduled_work();
+
+	i915_global_scheduler_exit();
+	i915_global_request_exit();
+	i915_global_active_exit();
+
+	/* And ensure that our DESTROY_BY_RCU slabs are truly destroyed */
+	rcu_barrier();
+}