From bug 1065128 SettingsManager has been changed to listen the
dom-window-destroyed event for its cleanup. However, when running Gaia
in Mulet, a race condition is exposed. For B2G, when loading a page,
about:blank is first used. This means that window destroyed events will
be triggered. However, from the dom-window-destroyed event we cannot
distinguish whether this is about:blank or a legit application being
closed. SettingsManager gets initialized (i.e., init() called) when the
application makes use of navigator.mozSettings. So the chain of event is
that we have a SettingsManager living because System app did some
request. At this time, about:blank is being unloaded and triggers a
dom-window-destroyed event. This makes SettingsManager doing its
cleanup, especially freeing the window reference. Then in the meantime,
we have the navigator.mozSettings use that is progressing. At some
point, SettingsManager has no more window to send messages to, and Gaia
is not able to even start.
SettingsRequestManager lives on the parent process and SettingsManager
lives on the child side. Part of the cleanup performed by
SettingsManager was to ensure pending locks on the parent process would
be forced to finalize to make sure those are being properly committed.
We move this cleanup to SettingsRequestManager and we augment the lock
informations with the proper inner window id. This way we can track
which lock is attached to which inner window when the lock gets created.
And thus we can listen on inner-window-destroyed from
SettingsRequestManager to be able to force finalize on any pending lock.
Impacted code path are those were we are not running out of process.
When we are running out of process, SettingsRequestManager already
listens on the child-process-shutdown event to perform the lock
finalization.
When child message manager dies, it sends a child-process-shutdown
message to the SettingsRequestManager. This would just close the locks
and tasks of this message manager. The race happens when some
applications can shutdown quickly: settings requests will never be
committed to the database. One example is the callscreen. The fix,
provided by qDot, simply put those tasks in a finalize state to make
sure they are properly executed and committed.
