[PATCH 3/3] common-tl-sync

[Chris Wilson]

---
 drivers/gpu/drm/i915/i915_gem_request.c  | 71 ++++++++++++++++++++------------
 drivers/gpu/drm/i915/i915_gem_timeline.c |  6 +++
 drivers/gpu/drm/i915/i915_gem_timeline.h |  3 ++
 3 files changed, 54 insertions(+), 26 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c
index fc7bdbe3bf94..3d96568a8152 100644
--- a/drivers/gpu/drm/i915/i915_gem_request.c
+++ b/drivers/gpu/drm/i915/i915_gem_request.c
@@ -717,9 +717,7 @@ int
 i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
 				 struct dma_fence *fence)
 {
-	struct dma_fence_array *array;
 	int ret;
-	int i;
 
 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 		return 0;
@@ -727,6 +725,11 @@ i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
 	/* Squash repeated waits to the same timelines, picking the latest */
 	if (fence->context != req->i915->mm.unordered_timeline) {
 		void __rcu **slot;
+		void *p_seqno;
+
+		p_seqno = radix_tree_lookup(&req->timeline->sync, fence->context);
+		if ((int)((u32)(uintptr_t)p_seqno - fence->seqno) >= 0)
+			return 0;
 
 		slot = radix_tree_lookup_slot(&req->waits, fence->context);
 		if (!slot) {
@@ -745,37 +748,53 @@ i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
 		}
 	}
 
-	if (dma_fence_is_i915(fence))
-		return i915_gem_request_await_request(req, to_request(fence));
-
-	if (!dma_fence_is_array(fence)) {
+	if (dma_fence_is_i915(fence)) {
+		ret = i915_gem_request_await_request(req, to_request(fence));
+		if (ret < 0)
+			return ret;
+	} else if (!dma_fence_is_array(fence)) {
 		ret = i915_sw_fence_await_dma_fence(&req->submit,
 						    fence, I915_FENCE_TIMEOUT,
 						    GFP_KERNEL);
-		return ret < 0 ? ret : 0;
-	}
+		if (ret < 0)
+			return ret;
+	} else {
+		struct dma_fence_array *array = to_dma_fence_array(fence);
+		int i;
+
+		/* Note that if the fence-array was created in signal-on-any mode,
+		 * we should *not* decompose it into its individual fences. However,
+		 * we don't currently store which mode the fence-array is operating
+		 * in. Fortunately, the only user of signal-on-any is private to
+		 * amdgpu and we should not see any incoming fence-array from
+		 * sync-file being in signal-on-any mode.
+		 */
 
-	/* Note that if the fence-array was created in signal-on-any mode,
-	 * we should *not* decompose it into its individual fences. However,
-	 * we don't currently store which mode the fence-array is operating
-	 * in. Fortunately, the only user of signal-on-any is private to
-	 * amdgpu and we should not see any incoming fence-array from
-	 * sync-file being in signal-on-any mode.
-	 */
+		for (i = 0; i < array->num_fences; i++) {
+			struct dma_fence *child = array->fences[i];
 
-	array = to_dma_fence_array(fence);
-	for (i = 0; i < array->num_fences; i++) {
-		struct dma_fence *child = array->fences[i];
+			if (dma_fence_is_i915(child))
+				ret = i915_gem_request_await_request(req,
+								     to_request(child));
+			else
+				ret = i915_sw_fence_await_dma_fence(&req->submit,
+								    child, I915_FENCE_TIMEOUT,
+								    GFP_KERNEL);
+			if (ret < 0)
+				return ret;
+		}
+	}
+
+	if (fence->context != req->i915->mm.unordered_timeline) {
+		struct radix_tree_root *root = &req->timeline->sync;
+		void *p_seqno = (void *)(uintptr_t)fence->seqno;
+		void __rcu **slot;
 
-		if (dma_fence_is_i915(child))
-			ret = i915_gem_request_await_request(req,
-							     to_request(child));
+		slot = radix_tree_lookup_slot(root, fence->context);
+		if (slot)
+			radix_tree_replace_slot(root, slot, p_seqno);
 		else
-			ret = i915_sw_fence_await_dma_fence(&req->submit,
-							    child, I915_FENCE_TIMEOUT,
-							    GFP_KERNEL);
-		if (ret < 0)
-			return ret;
+			radix_tree_insert(root, fence->context, p_seqno);
 	}
 
 	return 0;
diff --git a/drivers/gpu/drm/i915/i915_gem_timeline.c b/drivers/gpu/drm/i915/i915_gem_timeline.c
index b596ca7ee058..e295aaebb8ea 100644
--- a/drivers/gpu/drm/i915/i915_gem_timeline.c
+++ b/drivers/gpu/drm/i915/i915_gem_timeline.c
@@ -56,6 +56,7 @@ static int __i915_gem_timeline_init(struct drm_i915_private *i915,
 #endif
 		init_request_active(&tl->last_request, NULL);
 		INIT_LIST_HEAD(&tl->requests);
+		INIT_RADIX_TREE(&tl->sync, GFP_KERNEL);
 	}
 
 	return 0;
@@ -89,6 +90,11 @@ void i915_gem_timeline_fini(struct i915_gem_timeline *timeline)
 
 	for (i = 0; i < ARRAY_SIZE(timeline->engine); i++) {
 		struct intel_timeline *tl = &timeline->engine[i];
+		struct radix_tree_iter iter;
+		void __rcu **slot;
+
+		radix_tree_for_each_slot(slot, &tl->sync, &iter, 0)
+			radix_tree_iter_delete(&tl->sync, &iter, slot);
 
 		GEM_BUG_ON(!list_empty(&tl->requests));
 	}
diff --git a/drivers/gpu/drm/i915/i915_gem_timeline.h b/drivers/gpu/drm/i915/i915_gem_timeline.h
index 6c53e14cab2a..db0443e3ff22 100644
--- a/drivers/gpu/drm/i915/i915_gem_timeline.h
+++ b/drivers/gpu/drm/i915/i915_gem_timeline.h
@@ -26,6 +26,7 @@
 #define I915_GEM_TIMELINE_H
 
 #include <linux/list.h>
+#include <linux/radix-tree.h>
 
 #include "i915_gem_request.h"
 
@@ -55,6 +56,8 @@ struct intel_timeline {
 	 * struct_mutex.
 	 */
 	struct i915_gem_active last_request;
+	struct radix_tree_root sync;
+
 	u32 sync_seqno[I915_NUM_ENGINES];
 
 	struct i915_gem_timeline *common;
-- 
2.11.0