Fabric: Resetting *all* ShadowNode pointers in MountingCoordinator::revoke

Summary:
With this diff (and maybe other coming in the future), we are enforcing a model where Fabric Core actively/architecturally prevents `jsi::Object` to overlive `UIManager`. That never was a promise that but I think it's achievable; it should help to eliminate the whole class of possible crashes (application layer still must ensure that Scheduler is deallocated before `jsi::Runtime`).

Clearing all pointers to `ShadowNode`s in `MountingCoordinator` prevents from extending the lifetime of ShadowNodes stored inside revoked `MountingCoordinator` (the lifetime of `MountingCoordinator` is up to application layer).

That should help prevent crashes (not all of them!) caused by `jsi::Object`s overliving `jsi::Runtime`.

Changelog: [Internal]

Reviewed By: sammy-SC

Differential Revision: D18333708

fbshipit-source-id: 4391998b8f805563cfaff834fc56e85b98bc4d68
This commit is contained in:
Valentin Shergin authored and Facebook Github Bot committed 2019-11-08 22:20:39 -08:00
1 parent f7385ac1b8
commit 0eb7c7bc43
1 file changed
+5
@@ -52,6 +52,11 @@ void MountingCoordinator::push(ShadowTreeRevision &&revision) const {
void MountingCoordinator::revoke() const {
std::lock_guard<std::mutex> lock(mutex_);
// We have two goals here.
// 1. We need to stop retaining `ShadowNode`s to not prolong their lifetime
// to prevent them from overliving `ComponentDescriptor`s.
// 2. A possible call to `pullTransaction()` should return empty optional.
baseRevision_.rootShadowNode_.reset();
lastRevision_.reset();
}