drm/i915: Generalise GPU activity tracking

We currently track GPU memory usage inside VMA, such that we never
release memory used by the GPU until after it has finished accessing it.
However, we may want to track other resources aside from VMA, or we may
want to split a VMA into multiple independent regions and track each
separately. For this purpose, generalise our request tracking (akin to
struct reservation_object) so that we can embed it into other objects.

v2: Tweak error handling during selftest setup.

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/20190205130005.2807-2-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/Makefile

@@ -57,7 +57,9 @@ i915-$(CONFIG_DEBUG_FS) += i915_debugfs.o intel_pipe_crc.o
 i915-$(CONFIG_PERF_EVENTS) += i915_pmu.o
 
 # GEM code
-i915-y += i915_cmd_parser.o \
+i915-y += \
+	  i915_active.o \
+	  i915_cmd_parser.o \
 	  i915_gem_batch_pool.o \
 	  i915_gem_clflush.o \
 	  i915_gem_context.o \

drivers/gpu/drm/i915/i915_active.c

@@ -0,0 +1,228 @@
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#include "i915_drv.h"
+#include "i915_active.h"
+
+#define BKL(ref) (&(ref)->i915->drm.struct_mutex)
+
+struct active_node {
+	struct i915_gem_active base;
+	struct i915_active *ref;
+	struct rb_node node;
+	u64 timeline;
+};
+
+static void
+__active_retire(struct i915_active *ref)
+{
+	GEM_BUG_ON(!ref->count);
+	if (!--ref->count)
+		ref->retire(ref);
+}
+
+static void
+node_retire(struct i915_gem_active *base, struct i915_request *rq)
+{
+	__active_retire(container_of(base, struct active_node, base)->ref);
+}
+
+static void
+last_retire(struct i915_gem_active *base, struct i915_request *rq)
+{
+	__active_retire(container_of(base, struct i915_active, last));
+}
+
+static struct i915_gem_active *
+active_instance(struct i915_active *ref, u64 idx)
+{
+	struct active_node *node;
+	struct rb_node **p, *parent;
+	struct i915_request *old;
+
+	/*
+	 * We track the most recently used timeline to skip a rbtree search
+	 * for the common case, under typical loads we never need the rbtree
+	 * at all. We can reuse the last slot if it is empty, that is
+	 * after the previous activity has been retired, or if it matches the
+	 * current timeline.
+	 *
+	 * Note that we allow the timeline to be active simultaneously in
+	 * the rbtree and the last cache. We do this to avoid having
+	 * to search and replace the rbtree element for a new timeline, with
+	 * the cost being that we must be aware that the ref may be retired
+	 * twice for the same timeline (as the older rbtree element will be
+	 * retired before the new request added to last).
+	 */
+	old = i915_gem_active_raw(&ref->last, BKL(ref));
+	if (!old || old->fence.context == idx)
+		goto out;
+
+	/* Move the currently active fence into the rbtree */
+	idx = old->fence.context;
+
+	parent = NULL;
+	p = &ref->tree.rb_node;
+	while (*p) {
+		parent = *p;
+
+		node = rb_entry(parent, struct active_node, node);
+		if (node->timeline == idx)
+			goto replace;
+
+		if (node->timeline < idx)
+			p = &parent->rb_right;
+		else
+			p = &parent->rb_left;
+	}
+
+	node = kmalloc(sizeof(*node), GFP_KERNEL);
+
+	/* kmalloc may retire the ref->last (thanks shrinker)! */
+	if (unlikely(!i915_gem_active_raw(&ref->last, BKL(ref)))) {
+		kfree(node);
+		goto out;
+	}
+
+	if (unlikely(!node))
+		return ERR_PTR(-ENOMEM);
+
+	init_request_active(&node->base, node_retire);
+	node->ref = ref;
+	node->timeline = idx;
+
+	rb_link_node(&node->node, parent, p);
+	rb_insert_color(&node->node, &ref->tree);
+
+replace:
+	/*
+	 * Overwrite the previous active slot in the rbtree with last,
+	 * leaving last zeroed. If the previous slot is still active,
+	 * we must be careful as we now only expect to receive one retire
+	 * callback not two, and so much undo the active counting for the
+	 * overwritten slot.
+	 */
+	if (i915_gem_active_isset(&node->base)) {
+		/* Retire ourselves from the old rq->active_list */
+		__list_del_entry(&node->base.link);
+		ref->count--;
+		GEM_BUG_ON(!ref->count);
+	}
+	GEM_BUG_ON(list_empty(&ref->last.link));
+	list_replace_init(&ref->last.link, &node->base.link);
+	node->base.request = fetch_and_zero(&ref->last.request);
+
+out:
+	return &ref->last;
+}
+
+void i915_active_init(struct drm_i915_private *i915,
+		      struct i915_active *ref,
+		      void (*retire)(struct i915_active *ref))
+{
+	ref->i915 = i915;
+	ref->retire = retire;
+	ref->tree = RB_ROOT;
+	init_request_active(&ref->last, last_retire);
+	ref->count = 0;
+}
+
+int i915_active_ref(struct i915_active *ref,
+		    u64 timeline,
+		    struct i915_request *rq)
+{
+	struct i915_gem_active *active;
+
+	active = active_instance(ref, timeline);
+	if (IS_ERR(active))
+		return PTR_ERR(active);
+
+	if (!i915_gem_active_isset(active))
+		ref->count++;
+	i915_gem_active_set(active, rq);
+
+	GEM_BUG_ON(!ref->count);
+	return 0;
+}
+
+bool i915_active_acquire(struct i915_active *ref)
+{
+	lockdep_assert_held(BKL(ref));
+	return !ref->count++;
+}
+
+void i915_active_release(struct i915_active *ref)
+{
+	lockdep_assert_held(BKL(ref));
+	__active_retire(ref);
+}
+
+int i915_active_wait(struct i915_active *ref)
+{
+	struct active_node *it, *n;
+	int ret = 0;
+
+	if (i915_active_acquire(ref))
+		goto out_release;
+
+	ret = i915_gem_active_retire(&ref->last, BKL(ref));
+	if (ret)
+		goto out_release;
+
+	rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
+		ret = i915_gem_active_retire(&it->base, BKL(ref));
+		if (ret)
+			break;
+	}
+
+out_release:
+	i915_active_release(ref);
+	return ret;
+}
+
+static int __i915_request_await_active(struct i915_request *rq,
+				       struct i915_gem_active *active)
+{
+	struct i915_request *barrier =
+		i915_gem_active_raw(active, &rq->i915->drm.struct_mutex);
+
+	return barrier ? i915_request_await_dma_fence(rq, &barrier->fence) : 0;
+}
+
+int i915_request_await_active(struct i915_request *rq, struct i915_active *ref)
+{
+	struct active_node *it, *n;
+	int ret;
+
+	ret = __i915_request_await_active(rq, &ref->last);
+	if (ret)
+		return ret;
+
+	rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
+		ret = __i915_request_await_active(rq, &it->base);
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
+void i915_active_fini(struct i915_active *ref)
+{
+	struct active_node *it, *n;
+
+	GEM_BUG_ON(i915_gem_active_isset(&ref->last));
+
+	rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
+		GEM_BUG_ON(i915_gem_active_isset(&it->base));
+		kfree(it);
+	}
+	ref->tree = RB_ROOT;
+}
+
+#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
+#include "selftests/i915_active.c"
+#endif

drivers/gpu/drm/i915/i915_active.h

@@ -0,0 +1,69 @@
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#ifndef _I915_ACTIVE_H_
+#define _I915_ACTIVE_H_
+
+#include "i915_active_types.h"
+
+/*
+ * GPU activity tracking
+ *
+ * Each set of commands submitted to the GPU compromises a single request that
+ * signals a fence upon completion. struct i915_request combines the
+ * command submission, scheduling and fence signaling roles. If we want to see
+ * if a particular task is complete, we need to grab the fence (struct
+ * i915_request) for that task and check or wait for it to be signaled. More
+ * often though we want to track the status of a bunch of tasks, for example
+ * to wait for the GPU to finish accessing some memory across a variety of
+ * different command pipelines from different clients. We could choose to
+ * track every single request associated with the task, but knowing that
+ * each request belongs to an ordered timeline (later requests within a
+ * timeline must wait for earlier requests), we need only track the
+ * latest request in each timeline to determine the overall status of the
+ * task.
+ *
+ * struct i915_active provides this tracking across timelines. It builds a
+ * composite shared-fence, and is updated as new work is submitted to the task,
+ * forming a snapshot of the current status. It should be embedded into the
+ * different resources that need to track their associated GPU activity to
+ * provide a callback when that GPU activity has ceased, or otherwise to
+ * provide a serialisation point either for request submission or for CPU
+ * synchronisation.
+ */
+
+void i915_active_init(struct drm_i915_private *i915,
+		      struct i915_active *ref,
+		      void (*retire)(struct i915_active *ref));
+
+int i915_active_ref(struct i915_active *ref,
+		    u64 timeline,
+		    struct i915_request *rq);
+
+int i915_active_wait(struct i915_active *ref);
+
+int i915_request_await_active(struct i915_request *rq,
+			      struct i915_active *ref);
+
+bool i915_active_acquire(struct i915_active *ref);
+
+static inline void i915_active_cancel(struct i915_active *ref)
+{
+	GEM_BUG_ON(ref->count != 1);
+	ref->count = 0;
+}
+
+void i915_active_release(struct i915_active *ref);
+
+static inline bool
+i915_active_is_idle(const struct i915_active *ref)
+{
+	return !ref->count;
+}
+
+void i915_active_fini(struct i915_active *ref);
+
+#endif /* _I915_ACTIVE_H_ */

drivers/gpu/drm/i915/i915_active_types.h

@@ -0,0 +1,26 @@
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#ifndef _I915_ACTIVE_TYPES_H_
+#define _I915_ACTIVE_TYPES_H_
+
+#include <linux/rbtree.h>
+
+#include "i915_request.h"
+
+struct drm_i915_private;
+
+struct i915_active {
+	struct drm_i915_private *i915;
+
+	struct rb_root tree;
+	struct i915_gem_active last;
+	unsigned int count;
+
+	void (*retire)(struct i915_active *ref);
+};
+
+#endif /* _I915_ACTIVE_TYPES_H_ */

drivers/gpu/drm/i915/i915_gem_gtt.c

@@ -1917,14 +1917,13 @@ static struct i915_vma *pd_vma_create(struct gen6_hw_ppgtt *ppgtt, int size)
 	if (!vma)
 		return ERR_PTR(-ENOMEM);
 
+	i915_active_init(i915, &vma->active, NULL);
 	init_request_active(&vma->last_fence, NULL);
 
 	vma->vm = &ggtt->vm;
 	vma->ops = &pd_vma_ops;
 	vma->private = ppgtt;
 
-	vma->active = RB_ROOT;
-
 	vma->size = size;
 	vma->fence_size = size;
 	vma->flags = I915_VMA_GGTT;