Fix unitialized memory access in workers.

Fix unitialized memory access in workers.

The primary issue was that OnDestroy didn't change the entangled port
to have its entangled port be none.

A secondary issues that came up is that in very rare circumstances (like a crash
happening early in a worker process), it seemed like it may be possible that one
of the message ports may think it is entangled and the other half may not, so
the Erase method guards against this.

Also, some code was added to verify the internal structure before running code
and after.

BUG=27839
TEST=valgrind on linux running ui tests, specifically WorkerTest.WorkerFastLayoutTests.

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=32586

[jam, 2009-11-20 00:44:01]

lgtm, thanks for tracking this down

http://codereview.chromium.org/402106/diff/1/3
File chrome/browser/worker_host/message_port_dispatcher.h (right):

http://codereview.chromium.org/402106/diff/1/3#newcode75
Line 75: void CheckInvariants(bool check_entanglements) { }
nit: CheckInvariants's name isn't clear to me about what it's checking.  what
about CheckMessagePortMap or something similar?

http://codereview.chromium.org/402106/diff/1/3#newcode77
Line 77: void CheckInvariants(bool check_entanglements);
nit: the common way we do these checks is to have a function only defined in
ifdef NDEBUG, then we call out to it as such:

DCHECK(CheckBlah)

The function can even return true all the time since you have DCHECKs inside of
it.

This makes it obvious to the reader that the check only happens in debug builds,
and that there's no overhead for adding that code.