Here's how the deadlock can happen:
* The vm-service background compiler asks all threads to go to a
safepoint before it finalizes the code. Then it finalizes the
code and notifies code observers. A code observer posts a
message to the vm-service. Posting this message requires holding the
monitor of the receiving isolate's message handler (see
MessageHandler::PostMessage).
* The vm-service mutator holds it's message handlers monitor and
enters the isolate (see MessageHandler::HandleMessages). The
StartIsolateScope schedules the mutator thread, which will in return
see that BG compiler wants to safepoint, so it waits for BG compiler
to finish.
At this point the mutator holds the MessageHandler's monitor and waits
for BG compiler to finish safepoint. The BG compiler tries to post a
message to the message handler (and the monitor is locked).
-> This CL ensures the BG compiler only arranges a safepoint before it
starts creating the instructions object. It fills it in and exits the
safepoint again. Only then do we notify code observers.
Furthermore this CL ensures we transition out of VM to native state
before calling the Dart_PostCMessage() function.
Change-Id: I5320440d54ca303a1ccc7556600a42d37cd6e1de
Reviewed-on: https://dart-review.googlesource.com/c/93421
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Siva Annamalai <asiva@google.com>
Martin Kustermann
d3fde891bf