Fixed null dereference in GPU watchdog termination code....

Switched GPU watchdog timeout to be based on main thread's user + kernel time rather than wall clock time on Windows.

Only on Windows because it is the only platform 58396 appears to be happening on. My latest hairbrained theory is that it is timing out and aborting while Windows boxes are resuming from hypernation. In that case wall clock time might incorporate a lot of I/O time for paging in data from swapfile. This is an attempt to hide that I/O time, counting only active CPU time.

It catches hangs like this:

for (;;) {
}

and this:

for (;;) {
  Sleep(0);
}

but not this:

for (;;) {
  Sleep(1000);
}

because that just makes the thread largely idle. It also does not catch deadlocks.

Also fixed null dereference in GPU watchdog termination code.

BUG=64648, 58396
TEST=test GPU watchdog locally, try


Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=67891

[apatrick_chromium, 2010-11-29 22:00:14]

http://codereview.chromium.org/5301007/diff/1/chrome/gpu/gpu_watchdog_thread.cc
File chrome/gpu/gpu_watchdog_thread.cc (left):

http://codereview.chromium.org/5301007/diff/1/chrome/gpu/gpu_watchdog_thread....
chrome/gpu/gpu_watchdog_thread.cc:64: watched_message_loop_ = NULL;
I don't think this was actually needed.

[Ken Russell (switch to Gerrit), 2010-11-29 22:45:55]

Did you run with --disable-gpu-watchdog and verify that there are no assertion
failures upon shutdown? That was the original motivation for the tests against
the watched_message_loop, though I think the code structure has changed since
then.

[apatrick_chromium, 2010-11-30 23:05:18]

Some more changes...

[Ken Russell (switch to Gerrit), 2010-11-30 23:23:44]

This looks like it'll work great on Windows. It would be nice if we had the same
guards on other platforms to prevent the GPU process from being unexpectedly
terminated if the machine is running out of resources...

LGTM