GitLab

Two independent but related fixes here:

1. IPCSupportInitializer could previously outlive its IO
   runner, resulting in Mojo not being shut down in time.
   The EDK posts tasks directly to a MessageLoopForIO
   pointer because it's assumed to remain valid until
   EDK shutdown. This CL limits IPCSupportInitializer's
   lifetime to that of the IO thread, regardless of
   ScopedIPCSupport lifetime.

2. In single-process mode, the child IO thread cannot be
   used to create the client ChannelMojo, because all channels
   in-process must be created on the EDK IO thread. This problem only
   surfaced in practice because #1 above changed some shutdown
   ordering and tickled a DCHECK in EDK shutdown.

   This CL has ChannelMojo hop to its IO runner (only if
   necessary) before creating underlying messaging pipes.

BUG=None
R=morrita@chromium.org

Review URL: https://codereview.chromium.org/1130413002

Cr-Commit-Position: refs/heads/master@{#329262}