GitLab

When the UI compositor (or any compositor with a Display attached to it
in process - such as layout tests) does SwapBuffers, it submits the
CompositorFrame to the SurfaceFactory.

The SurfaceFactory pokes the DisplayScheduler, which can lead to it
doing the Display::DrawAndSwap immediately, in the same callstack. In
that case, the callback to the SurfaceDisplayOutputSurface (or the
TestDelegatingOutputSurface) that it drew happens inline also, which
was previously calling out to the OutputSurfaceClient.

The problem is, the compositor expects SwapBuffers() to return before
it hears that DidSwapBuffersComplete(). To avoid this, and similar
problems, we prevent DisplayScheduler from running a new BeginFrame
inside an external call to DisplayScheduler::SurfaceDamaged() which
prevents re-entrancy for its clients.

We do this by posting the missed BeginFrame to respond to it on a new
stack (and cancelling it if another BeginFrame comes in the meantime
since it was already posted).

For TestDelegatedOutputSurface synchronous composites, we simply
post the SwapBuffersComplete message to get it on a new call stack.

R=jbauman
BUG=495650
CQ_INCLUDE_TRYBOTS=master.tryserver.blink:linux_precise_blink_rel

Review-Url: https://codereview.chromium.org/2183333003
Cr-Commit-Position: refs/heads/master@{#408289}