There's a scenario that we're currently not allowing for, but I'd really like to use in DM:
1) client calls add(SomeRunnable*) several times
2) client calls wait()
3) any of the runnables added by the client _themselves_ call add(SomeOtherRunnable*)
4-inf) maybe those SomeOtherRunnables too call add(SomeCrazyThirdRunnable*), etc.
Right now in this scenario we'll assert in debug mode in step 3) when we call
add() and we're waiting to stop, and do strange unspecified things in release
mode.
The old threadpool had basically two states: running, and waiting to stop. If
a thread saw we were waiting to stop and the queue was empty, that thread shut
down. This wasn't accounting for any work that other threads might be doing;
potentially they were about to add to the queue.
So now we have three states: running, waiting, and halting. When the client
calls wait() (or the destructor triggers), we move into waiting. When a thread
notices we're _really_ done, that is, have an empty queue and there are no
active threads, we move into halting. The halting state actually triggers the
threads to stop, which wait() is patiently join()ing on.
BUG=
R=bungeman@google.com, bsalomon@google.com
Author: mtklein@google.com
Review URL: https://codereview.chromium.org/26389005
git-svn-id: http://skia.googlecode.com/svn/trunk@11852 2bbb7eff-a529-9590-31e7-b0007b416f81
6ee68583f8