Android: Fix GPU recovery issues

Android: Fix GPU recovery issues

    a)
    The swap accounting is a bit unreliable when the GPU process
    crashes or the context is lost. We might end up thinking that
    too many swap acks are pending (which will never arrive) and
    bail out of CompositorImpl::Composite() early.
    Ignore the swap limit when deciding to composite if we know
    the context was lost. Then calling Composite() will cause the
    output surface to get recreated.

    b)
    Fix infinite retry loops of which there are two:
    One is that we could indefinitely early-out of Composite()
    and repost if the GPU process keeps crashing at startup before
    LTH ever gets a chance to try create a context.

    Fix regression with https://codereview.chromium.org/738983002/
    which both removed the LOG(FATAL) for too many successive retry
    attempts (used to happen inside LTH), and added a similar retry
    loop where we might never call SetOutputSurface and keep
    relaunching the GPU process.

    Also essentially revert https://codereview.chromium.org/750643003/
    to simplify things. We don't need to early-out in Composite()
    anymore if the GPU channel is lost. This is handled better in
    cc now and we can just attempt to Composite(). It will schedule
    OutputSurface creation. Even if we call Composite() again in the
    meantime, that should be harmless. And after we eventually call
    SetOutputSurface(), this will trigger ScheduleComposite().

BUG=453511, 453671, 453819, 453649, 448549
NOTRY=True

Committed: https://crrev.com/bce21923514e1b1d92d5dd743785830fe9c10ad3
Cr-Commit-Position: refs/heads/master@{#315390}

[no sievers, 2015-02-06 22:25:24]

sievers@chromium.org changed reviewers:
+ epenner@chromium.org

[no sievers, 2015-02-06 22:27:27]

ptal, the conservative fix, but adds yet more state to CompositorImpl :(

With https://codereview.chromium.org/344743002/ this would get simpler since
it'd only be handled within cc which probably already has a lot of unit tests
for handling context lost wrt scheduler and OutputSurface creation.

[no sievers, 2015-02-07 00:00:43]

sievers@chromium.org changed reviewers:
+ enne@chromium.org

[no sievers, 2015-02-07 00:15:05]

removed tracking output surface validity in CompositorImpl, and just ask LTH.

[epennerAtGoogle, 2015-02-07 00:16:17]

On 2015/02/07 00:15:05, sievers wrote:
> removed tracking output surface validity in CompositorImpl, and just ask LTH.

Cool. LGTM. I'll try to think of any other cases, but seems good.

[enne (OOO), 2015-02-07 00:46:13]

lgtm too.  Getting rid of defer_composite_for_gpu_channel_ is nice.

[no sievers, 2015-02-09 20:08:25]

New patchsets have been uploaded after l-g-t-m from
epenner@google.com,enne@chromium.org

[no sievers, 2015-02-09 20:24:04]

The CQ bit was checked by sievers@chromium.org

[commit-bot: I haz the power, 2015-02-09 20:25:12]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/903933005/80001

[no sievers, 2015-02-09 21:00:35]

The CQ bit was unchecked by sievers@chromium.org

[no sievers, 2015-02-09 21:00:40]

The CQ bit was checked by sievers@chromium.org

[commit-bot: I haz the power, 2015-02-09 21:00:54]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/903933005/80001

[commit-bot: I haz the power, 2015-02-09 21:01:59]

Committed patchset #5 (id:80001)

[commit-bot: I haz the power, 2015-02-09 21:02:40]

Patchset 5 (id:??) landed as
https://crrev.com/bce21923514e1b1d92d5dd743785830fe9c10ad3
Cr-Commit-Position: refs/heads/master@{#315390}