PPAPI: Make TrackedCallback more threadsafe

ppapi/shared_impl/callback_tracker.cc

Index: ppapi/shared_impl/callback_tracker.cc
diff --git a/ppapi/shared_impl/callback_tracker.cc b/ppapi/shared_impl/callback_tracker.cc
index feefb0e117d70aa862553838656da09763927824..6d489d5b32d73f79dafc79d94b4b496de8b37cfa 100644
--- a/ppapi/shared_impl/callback_tracker.cc
+++ b/ppapi/shared_impl/callback_tracker.cc
@@ -11,18 +11,25 @@
 #include "base/logging.h"
 #include "base/message_loop/message_loop.h"
 #include "ppapi/c/pp_completion_callback.h"
+#include "ppapi/shared_impl/proxy_lock.h"
 #include "ppapi/shared_impl/tracked_callback.h"
 
 namespace ppapi {
 
 // CallbackTracker -------------------------------------------------------------
 
-CallbackTracker::CallbackTracker() {}
+CallbackTracker::CallbackTracker() : abort_all_called_(false) {}
 
 void CallbackTracker::AbortAll() {
-  // Iterate over a copy since |Abort()| calls |Remove()| (indirectly).
-  // TODO(viettrungluu): This obviously isn't so efficient.
-  CallbackSetMap pending_callbacks_copy = pending_callbacks_;
+  // Iterate over a copy:
+  // 1) because |Abort()| calls |Remove()| (indirectly).
+  // 2) So we can drop the lock before calling in to TrackedCallback.
+  CallbackSetMap pending_callbacks_copy;
+  {
+    base::AutoLock acquire(lock_);
+    pending_callbacks_copy = pending_callbacks_;
+    abort_all_called_ = true;
+  }
   for (CallbackSetMap::iterator it1 = pending_callbacks_copy.begin();
        it1 != pending_callbacks_copy.end();
        ++it1) {
@@ -35,15 +42,20 @@ void CallbackTracker::AbortAll() {
 }
 
 void CallbackTracker::PostAbortForResource(PP_Resource resource_id) {
-  CHECK(resource_id != 0);
-  CallbackSetMap::iterator it1 = pending_callbacks_.find(resource_id);
-  if (it1 == pending_callbacks_.end())
-    return;
-  for (CallbackSet::iterator it2 = it1->second.begin();
-       it2 != it1->second.end();
-       ++it2) {
-    // Post the abort.
-    (*it2)->PostAbort();
+  CHECK_NE(resource_id, 0);

[bbudge, 2015/03/27 23:46:20]

I think I asked already, but why not just DCHECK here? It's an internal error
and should silently return on line 52 in this case. Or is there a reason to
terminate the process?

[dmichael (off chromium), 2015/03/31 19:38:04]

Sure, I made it a DCHECK and added a small comment of explanation.

+  CallbackSet callbacks_for_resource;
+  {
+    base::AutoLock acquire(lock_);
+    CallbackSetMap::iterator iter = pending_callbacks_.find(resource_id);
+    // The resource may have no callbacks, so it won't be found, and we're done.
+    if (iter == pending_callbacks_.end())
+      return;
+    // Copy the set so we can drop the lock before calling in to
+    // TrackedCallback.
+    callbacks_for_resource = iter->second;
+  }
+  for (const auto& iter : callbacks_for_resource) {
+    iter->PostAbort();
   }
 }
 
@@ -54,6 +66,8 @@ CallbackTracker::~CallbackTracker() {
 
 void CallbackTracker::Add(
     const scoped_refptr<TrackedCallback>& tracked_callback) {
+  base::AutoLock acquire(lock_);
+  DCHECK(!abort_all_called_);
   PP_Resource resource_id = tracked_callback->resource_id();

[bbudge, 2015/03/27 23:46:20]

A related question is why not DCHECK_NE(resource_id, 0) here? tracked_callback
might not be associated with a resource.

[dmichael (off chromium), 2015/03/31 19:38:04]

Done.

   DCHECK(pending_callbacks_[resource_id].find(tracked_callback) ==
          pending_callbacks_[resource_id].end());
@@ -62,6 +76,7 @@ void CallbackTracker::Add(
 
 void CallbackTracker::Remove(
     const scoped_refptr<TrackedCallback>& tracked_callback) {
+  base::AutoLock acquire(lock_);
   CallbackSetMap::iterator map_it =
       pending_callbacks_.find(tracked_callback->resource_id());
   DCHECK(map_it != pending_callbacks_.end());