Implemented new fence syncs which replaces the old sync points.

gpu/command_buffer/service/sync_point_manager.h

Index: gpu/command_buffer/service/sync_point_manager.h
diff --git a/gpu/command_buffer/service/sync_point_manager.h b/gpu/command_buffer/service/sync_point_manager.h
index 3f11e05dabbe407ea9c27e018c759701e3fedbb3..26930602e3df57a8ca9489c4de8ffa8f3bf30934 100644
--- a/gpu/command_buffer/service/sync_point_manager.h
+++ b/gpu/command_buffer/service/sync_point_manager.h
@@ -5,6 +5,8 @@
 #ifndef GPU_COMMAND_BUFFER_SERVICE_SYNC_POINT_MANAGER_H_
 #define GPU_COMMAND_BUFFER_SERVICE_SYNC_POINT_MANAGER_H_
 
+#include <functional>
+#include <queue>
 #include <vector>
 
 #include "base/atomic_sequence_num.h"
@@ -19,29 +21,25 @@
 #include "gpu/command_buffer/common/constants.h"
 #include "gpu/gpu_export.h"
 
+namespace base {
+class SingleThreadTaskRunner;
+}  // namespace base
+
 namespace gpu {
 
 class SyncPointClient;
+class SyncPointClientState;
 class SyncPointManager;
 
-class GPU_EXPORT SyncPointClientState
-    : public base::RefCountedThreadSafe<SyncPointClientState> {
+class GPU_EXPORT SyncPointOrderData
+    : public base::RefCountedThreadSafe<SyncPointOrderData> {
  public:
-  static scoped_refptr<SyncPointClientState> Create();
-  uint32_t GenerateUnprocessedOrderNumber(SyncPointManager* sync_point_manager);
-
-  void BeginProcessingOrderNumber(uint32_t order_num) {
-    DCHECK(processing_thread_checker_.CalledOnValidThread());
-    DCHECK_GE(order_num, current_order_num_);
-    current_order_num_ = order_num;
-  }
+  static scoped_refptr<SyncPointOrderData> Create();
+  void Destroy();
 
-  void FinishProcessingOrderNumber(uint32_t order_num) {
-    DCHECK(processing_thread_checker_.CalledOnValidThread());
-    DCHECK_EQ(current_order_num_, order_num);
-    DCHECK_GT(order_num, processed_order_num());
-    base::subtle::Release_Store(&processed_order_num_, order_num);
-  }
+  uint32_t GenerateUnprocessedOrderNumber(SyncPointManager* sync_point_manager);
+  void BeginProcessingOrderNumber(uint32_t order_num);
+  void FinishProcessingOrderNumber(uint32_t order_num);
 
   uint32_t processed_order_num() const {
     return base::subtle::Acquire_Load(&processed_order_num_);
@@ -57,11 +55,33 @@ class GPU_EXPORT SyncPointClientState
   }
 
  protected:
-  friend class base::RefCountedThreadSafe<SyncPointClientState>;
-  friend class SyncPointClient;
+  friend class base::RefCountedThreadSafe<SyncPointOrderData>;
+  friend class SyncPointClientState;
 
-  SyncPointClientState();
-  virtual ~SyncPointClientState();
+  struct OrderFence {
+    uint32_t order_num;
+    uint64_t fence_release;
+    scoped_refptr<SyncPointClientState> client_state;
+
+    OrderFence(uint32_t order, uint64_t release,
+               scoped_refptr<SyncPointClientState> state);
+    ~OrderFence();
+
+    bool operator>(const OrderFence& rhs) const {
+      return (order_num > rhs.order_num) ||
+             ((order_num == rhs.order_num) &&
+              (fence_release > rhs.fence_release));
+    }
+  };
+  typedef std::priority_queue<OrderFence,
+                              std::vector<OrderFence>,
+                              std::greater<OrderFence>> OrderFenceQueue;
+
+  SyncPointOrderData();
+  virtual ~SyncPointOrderData();

[piman, 2015/09/30 22:50:23]

This doesn't need to be virtual, because nothing derives from it, and it has no
other virtual member.

[David Yen, 2015/09/30 23:55:18]

Done.

+
+  void AddOrderFenceGuard(uint32_t order_num, uint64_t fence_release,
+                          scoped_refptr<SyncPointClientState> client_state);
 
   // Last finished IPC order number.
   base::subtle::Atomic32 processed_order_num_;
@@ -75,6 +95,79 @@ class GPU_EXPORT SyncPointClientState
   // Current IPC order number being processed (only used on processing thread).
   uint32_t current_order_num_;
 
+  // In situations where we are waiting on fence syncs that do not exist, we
+  // validate by making sure the order number does not pass the order number
+  // which the wait command was issued. If the order number reaches the
+  // wait command's, we should automatically release up to the expected
+  // release count. Note that this also releases other lower release counts,
+  // so a single misbehaved fence sync is enough to invalidate/signal all
+  // previous fence syncs.
+  base::Lock order_fence_queue_lock_;
+  OrderFenceQueue order_fence_queue_;
+
+  DISALLOW_COPY_AND_ASSIGN(SyncPointOrderData);
+};
+
+class GPU_EXPORT SyncPointClientState
+    : public base::RefCountedThreadSafe<SyncPointClientState> {
+ public:
+  scoped_refptr<SyncPointOrderData> order_data() { return order_data_; }
+
+  bool IsFenceSyncReleased(uint64_t release) {
+    return release <= fence_sync_release();
+  }
+
+  uint64_t fence_sync_release() {
+    base::AutoLock auto_lock(fence_sync_lock_);
+    return fence_sync_release_;
+  }
+
+ protected:
+  friend class base::RefCountedThreadSafe<SyncPointClientState>;
+  friend class SyncPointClient;
+  friend class SyncPointOrderData;
+
+  struct ReleaseCallback {
+    uint64_t release_count;
+    base::Closure callback_closure;
+
+    ReleaseCallback(uint64_t release, const base::Closure& callback);
+    ~ReleaseCallback();
+
+    bool operator>(const ReleaseCallback& rhs) const {
+      return release_count > rhs.release_count;
+    }
+  };
+  typedef std::priority_queue<
+      ReleaseCallback,
+      std::vector<ReleaseCallback>,
+      std::greater<ReleaseCallback>> ReleaseCallbackQueue;
+
+  SyncPointClientState(scoped_refptr<SyncPointOrderData> order_data);
+  virtual ~SyncPointClientState();

[piman, 2015/09/30 22:50:23]

same here, no need to be virtual.

[David Yen, 2015/09/30 23:55:18]

Done.

+
+  bool WaitForRelease(uint32_t wait_order_num,
+                      uint64_t release,
+                      const base::Closure& callback);
+
+  void ReleaseFenceSync(uint64_t release);
+  void EnsureReleased(uint64_t release);
+  void ReleaseFenceSyncLocked(uint64_t release,
+                              std::vector<base::Closure>* callback_list);
+
+  // Global order data where releases will originate from.
+  scoped_refptr<SyncPointOrderData> order_data_;
+
+  // Protects fence_sync_release_, fence_callback_queue_.
+  base::Lock fence_sync_lock_;
+
+  // Current fence sync release that has been signaled.
+  uint64_t fence_sync_release_;
+
+  // In well defined fence sync operations, fence syncs are released in order
+  // so simply having a priority queue for callbacks is enough.
+  ReleaseCallbackQueue release_callback_queue_;
+
   DISALLOW_COPY_AND_ASSIGN(SyncPointClientState);
 };
 
@@ -84,11 +177,29 @@ class GPU_EXPORT SyncPointClient {
 
   scoped_refptr<SyncPointClientState> client_state() { return client_state_; }
 
+  // Wait for a release count to be reached on a SyncPointClientState. If this
+  // function returns false, that means the wait was invalid. Otherwise if it
+  // returns True it means wait_complete_callback will be called eventually.
+  // In the case where the release happened already, it will call the callback
+  // immediately before returning. The callback function may be called on
+  // on another thread so it should be thread-safe. For convenience, another
+  // non-threadsafe version is defined below where you can supply a task runner.
+  bool Wait(scoped_refptr<SyncPointClientState> release_state,
+            uint64_t release_count,
+            const base::Closure& wait_complete_callback);
+
+  bool WaitNonThreadSafe(scoped_refptr<SyncPointClientState> release_state,
+                         uint64_t release_count,
+                         scoped_refptr<base::SingleThreadTaskRunner> runner,
+                         const base::Closure& wait_complete_callback);
+
+  void ReleaseFenceSync(uint64_t release);
+
  private:
   friend class SyncPointManager;
 
   SyncPointClient(SyncPointManager* sync_point_manager,
-                  scoped_refptr<SyncPointClientState> state,
+                  scoped_refptr<SyncPointOrderData> order_data,
                   CommandBufferNamespace namespace_id, uint64_t client_id);
 
   // Sync point manager is guaranteed to exist in the lifetime of the client.
@@ -98,8 +209,8 @@ class GPU_EXPORT SyncPointClient {
   scoped_refptr<SyncPointClientState> client_state_;
 
   // Unique namespace/client id pair for this sync point client.
-  CommandBufferNamespace namespace_id_;
-  uint64_t client_id_;
+  const CommandBufferNamespace namespace_id_;
+  const uint64_t client_id_;
 
   DISALLOW_COPY_AND_ASSIGN(SyncPointClient);
 };
@@ -113,7 +224,7 @@ class GPU_EXPORT SyncPointManager {
 
   // Creates/Destroy a sync point client which message processors should hold.
   scoped_ptr<SyncPointClient> CreateSyncPointClient(
-      scoped_refptr<SyncPointClientState> client_state,
+      scoped_refptr<SyncPointOrderData> order_data,
       CommandBufferNamespace namespace_id, uint64_t client_id);
 
   // Finds the state of an already created sync point client.
@@ -142,7 +253,7 @@ class GPU_EXPORT SyncPointManager {
 
  private:
   friend class SyncPointClient;
-  friend class SyncPointClientState;
+  friend class SyncPointOrderData;
 
   typedef std::vector<base::Closure> ClosureList;
   typedef base::hash_map<uint32, ClosureList> SyncPointMap;