Recreate the WebData database on a catastrophic SQL error

components/webdata/common/web_database_service.cc

Index: components/webdata/common/web_database_service.cc
diff --git a/components/webdata/common/web_database_service.cc b/components/webdata/common/web_database_service.cc
index a5ed8633b2f705ba2a83c7c2f73b87591a3a2b4f..203bb7a3b4a8b425fb45821dfbcaed167e09d997 100644
--- a/components/webdata/common/web_database_service.cc
+++ b/components/webdata/common/web_database_service.cc
@@ -137,19 +137,24 @@ void WebDatabaseService::OnDatabaseLoadDone(sql::InitStatus status,
   if (status == sql::INIT_OK) {
     db_loaded_ = true;
 
-    for (const auto& loaded_callback : loaded_callbacks_) {
+    while (!loaded_callbacks_.empty()) {
+      DBLoadedCallback loaded_callback = loaded_callbacks_.back();
+      loaded_callbacks_.pop_back();
       if (!loaded_callback.is_null())
         loaded_callback.Run();
     }
-
-    loaded_callbacks_.clear();
   } else {
     // Notify that the database load failed.
-    for (const auto& error_callback : error_callbacks_) {
+    while (!error_callbacks_.empty()) {
+      // The profile error callback is a message box that runs in a nested run
+      // loop. While it's being displayed, other OnDatabaseLoadDone() will run
+      // (posted from WebDatabaseBackend::Delegate::DBLoaded()). We need to make
+      // sure that after the callback running the message box returns, it checks
+      // |error_callbacks_| before it accesses it.

[Scott Hess - ex-Googler, 2016/08/10 17:16:19]

If I understand this correctly, another option might be something like:
  ErrorCallbacks callbacks;
  std::swap(callbacks, error_callbacks_);
  for (const auto& error_callback : callbacks) {
    ...
  }

[afakhry, 2016/08/11 17:32:54]

This works too, but it exposes a problem that was already there. Now we get the
profile error dialog 4 times, once after every "OK! The nested runloop of the
message box blocks the execution until it's dismissed, and then the next
callback is executed, which shows the profile error dialog again!

This pop_back() implementation *hides* the problem, since all other callbacks
are executed fast while the message box is still being shown. This is good but
that's not a robust solution! I'm not sure why the code is written that way. Is
it expected that we get one profile error dialog for each WebData DB type
(autofill, keyword, token, password ...)? I'd need more time to think about how
to fix this cleanly (if that's important).

+      DBLoadErrorCallback error_callback = error_callbacks_.back();
+      error_callbacks_.pop_back();
       if (!error_callback.is_null())
         error_callback.Run(status, diagnostics);
     }
-
-    error_callbacks_.clear();
   }
 }