[PATCH] drm/i915: Unshare the idle-barrier from other kernel requests

* [PATCH] drm/i915: Unshare the idle-barrier from other kernel requests
@ 2019-07-24 12:43 Chris Wilson
  2019-07-24 12:50 ` Chris Wilson
  2019-07-24 16:35 ` ✗ Fi.CI.BAT: failure for drm/i915: Unshare the idle-barrier from other kernel requests (rev2) Patchwork
  0 siblings, 2 replies; 8+ messages in thread
From: Chris Wilson @ 2019-07-24 12:43 UTC (permalink / raw)
  To: intel-gfx

Under some circumstances (see intel_context_prepare_remote_request), we
may use a requst along a kernel context to modify the logical state of
another. To keep the target context in place while the request executes,
we take an active reference on it using the kernel timeline. This is the
same timeline as we use for the idle-barrier, and so we end up reusing
the same active node. Except that the idle barrier is special and cannot
be reused in this manner! Give the idle-barrier a reserved timeline
index (0) so that is will always be unique (give or take we may issue
multiple idle barriers across multiple engines).

Reported-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: ce476c80b8bf ("drm/i915: Keep contexts pinned until after the next kernel context switch")
Fixes: a9877da2d629 ("drm/i915/oa: Reconfigure contexts on the fly")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
---
 drivers/gpu/drm/i915/i915_active.c | 63 ++++++++++++++++++++++--------
 1 file changed, 47 insertions(+), 16 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c
index 13f304a29fc8..4f7f698bff15 100644
--- a/drivers/gpu/drm/i915/i915_active.c
+++ b/drivers/gpu/drm/i915/i915_active.c
@@ -184,6 +184,7 @@ active_instance(struct i915_active *ref, u64 idx)
 	ref->cache = node;
 	mutex_unlock(&ref->mutex);
 
+	BUILD_BUG_ON(offsetof(typeof(*node), base));
 	return &node->base;
 }
 
@@ -212,6 +213,8 @@ int i915_active_ref(struct i915_active *ref,
 	struct i915_active_request *active;
 	int err;
 
+	GEM_BUG_ON(!timeline); /* reserved for idle-barrier */
+
 	/* Prevent reaping in case we malloc/wait while building the tree */
 	err = i915_active_acquire(ref);
 	if (err)
@@ -342,6 +345,31 @@ void i915_active_fini(struct i915_active *ref)
 }
 #endif
 
+static struct active_node *idle_barrier(struct i915_active *ref)
+{
+	struct active_node *node = NULL;
+	struct rb_node *rb;
+
+	mutex_lock(&ref->mutex);
+
+	rb = rb_first(&ref->tree);
+	if (!rb)
+		goto unlock;
+
+	node = rb_entry(rb, typeof(*node), node);
+	if (node->timeline || i915_active_request_isset(&node->base)) {
+		node = NULL;
+		goto unlock;
+	}
+
+	GEM_BUG_ON(!list_empty(&node->base.link));
+	rb_erase(rb, &ref->tree);
+
+unlock:
+	mutex_unlock(&ref->mutex);
+	return node;
+}
+
 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 					    struct intel_engine_cs *engine)
 {
@@ -352,22 +380,29 @@ int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 
 	GEM_BUG_ON(!engine->mask);
 	for_each_engine_masked(engine, i915, engine->mask, tmp) {
-		struct intel_context *kctx = engine->kernel_context;
 		struct active_node *node;
 
-		node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
-		if (unlikely(!node)) {
-			err = -ENOMEM;
-			goto unwind;
+		node = idle_barrier(ref);
+		if (!node) {
+			node = kmem_cache_alloc(global.slab_cache,
+						GFP_KERNEL |
+						__GFP_RETRY_MAYFAIL |
+						__GFP_NOWARN);
+			if (unlikely(!node)) {
+				err = -ENOMEM;
+				goto unwind;
+			}
+
+			node->ref = ref;
+			node->timeline = 0;
+			node->base.retire = node_retire;
 		}
 
-		i915_active_request_init(&node->base,
-					 (void *)engine, node_retire);
-		node->timeline = kctx->ring->timeline->fence_context;
-		node->ref = ref;
+		intel_engine_pm_get(engine);
+
+		RCU_INIT_POINTER(node->base.request, (void *)engine);
 		atomic_inc(&ref->count);
 
-		intel_engine_pm_get(engine);
 		llist_add((struct llist_node *)&node->base.link,
 			  &ref->barriers);
 	}
@@ -402,6 +437,7 @@ void i915_active_acquire_barrier(struct i915_active *ref)
 
 		node = container_of((struct list_head *)pos,
 				    typeof(*node), base.link);
+		GEM_BUG_ON(node->timeline);
 
 		engine = (void *)rcu_access_pointer(node->base.request);
 		RCU_INIT_POINTER(node->base.request, ERR_PTR(-EAGAIN));
@@ -410,12 +446,7 @@ void i915_active_acquire_barrier(struct i915_active *ref)
 		p = &ref->tree.rb_node;
 		while (*p) {
 			parent = *p;
-			if (rb_entry(parent,
-				     struct active_node,
-				     node)->timeline < node->timeline)
-				p = &parent->rb_right;
-			else
-				p = &parent->rb_left;
+			p = &parent->rb_left;
 		}
 		rb_link_node(&node->node, parent, p);
 		rb_insert_color(&node->node, &ref->tree);
-- 
2.22.0

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

^ permalink raw reply related	[flat|nested] 8+ messages in thread