1462411 - Try to FlushAsyncPaints in ClientLayerManager::EndTransaction

Status: RESOLVED FIXED

(Core :: Graphics: Layers, enhancement, P3)

Description

[Ryan Hunt [:rhunt]]

Currently we block entering a layer transaction while previous async paints are still finishing. [1] The reason we do this is because we are double buffered and cannot have two active frames being painted into a single backbuffer.

But, the painting only happens in EndTransaction [2]. It seems to me that we can delay blocking for an async paint until we get to EndTransaction. This would let us do display list building and frame layer building at the same time we are painting.

In the future we might determine cases where we don't need to block at all, like an empty transaction or if there are no invalidated painted regions.

The concern for race conditions lies in two areas, IPDL and content clients/texture clients.

When we have async paints still active, we postpone all IPDL messages on the CompositorBridge protocol until they complete [3]. This is important so that the Update message is not sent until the paints are complete and so that other messages are in order as they are sent. This may have some effects here, and I will need to think about it.

The other thing is we need to make sure that no content clients are manipulated or destroyed that have async paints happening. We currently protect against this by flushing async paints before we call root->RenderLayer(), and that would continue with this change.

Here's a try run with a prototype of this patch:
https://treeherder.mozilla.org/#/jobs?repo=try&revision=a7731a2f09f6132566c30e171cacd7771af63b27

It looks green to me.

[1] https://searchfox.org/mozilla-central/rev/d4b9e50875ad7e5d20f2fee6a53418315f6dfcc0/gfx/layers/client/ClientLayerManager.cpp#228
[2] https://searchfox.org/mozilla-central/rev/d4b9e50875ad7e5d20f2fee6a53418315f6dfcc0/gfx/layers/client/ClientLayerManager.cpp#356
[3] https://searchfox.org/mozilla-central/rev/d4b9e50875ad7e5d20f2fee6a53418315f6dfcc0/gfx/layers/ipc/ShadowLayers.cpp#771

Comment 1

[Ryan Hunt [:rhunt]]

Matt, would you have any concerns about this change? Are there any things we need to watch out here during DL building or FLB?

Comment 2

[Matt Woodrow (:mattwoodrow)]

I can't think of anything potentially dangerous during DL building.

Might be worth double checking (or finding a way to assert) that we never try to clear buffers or copy data during FLB.

FLB might also delete the Layers used in the previous paint, and that would probably drop a reference to the pixel buffer. We'd want to make sure we're holding the right refs, and that nobody unlocks/unmaps memory that we still want to write to.

Comment 3

[Ryan Hunt [:rhunt]]

(In reply to Matt Woodrow (:mattwoodrow) from comment #2)
> I can't think of anything potentially dangerous during DL building.
> 
> Might be worth double checking (or finding a way to assert) that we never
> try to clear buffers or copy data during FLB.
> 

I haven't seen anything in the public interface of ClientPaintedLayer that would
clear buffers or copy data.

There is one exception, when gfxEnv::DumpPaint() is enabled GetCompositableClient()
is used with Dump()[1] to read the current buffer. This happens in FLB so with this
change we could be reading bogus values. Not sure what the best solution here is,
but detecting this and falling back to the old behavior seems reasonable.

> FLB might also delete the Layers used in the previous paint, and that would
> probably drop a reference to the pixel buffer. We'd want to make sure we're
> holding the right refs, and that nobody unlocks/unmaps memory that we still
> want to write to.

We currently hold on to texture clients for this reason [2]. I don't see a reason
this would become worse now. The IPDL/TextureClient destruction stuff is a bit
of a mystery to me, but I believe it will behave the same as before.

I sent an email to David Anderson as he originally wrote this code, and he
wasn't sure exactly the reason. He thinks it was most likely a safety
precaution around FLB mutating the layer tree, but also that he's not
sure of anything in the paint thread code that would depend on that so
that this could be safe.

I think this is worth landing and monitoring for regressions.

[1] https://searchfox.org/mozilla-central/rev/da499aac682d0bbda5829327b60a865cbc491611/gfx/layers/Layers.cpp#1608
[2] https://searchfox.org/mozilla-central/rev/da499aac682d0bbda5829327b60a865cbc491611/gfx/layers/ipc/CompositorBridgeChild.h#246

Comment 4

[Ryan Hunt [:rhunt]]

Attachment: flush-later.patch

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/c0b2977cab3c
Move FlushAsyncPaints to ClientLayerManager::EndTransactionInternal. r=mattwoodrow

Comment 6

[Cristian Brindusan [:cbrindusan]]

https://hg.mozilla.org/mozilla-central/rev/c0b2977cab3c

Comment 7

[Ionuț Goldan [:igoldan]]

Perf wins!

== Change summary for alert #13421 (as of Thu, 24 May 2018 11:36:11 GMT) ==

Improvements:

  9%  tsvgx windows7-32 opt e10s stylo     261.64 -> 236.82
  8%  tsvgx osx-10-10 opt e10s stylo       320.37 -> 294.97
  8%  tsvgx windows7-32 pgo e10s stylo     232.51 -> 214.63

For up to date results, see: https://treeherder.mozilla.org/perf.html#/alerts?id=13421

Comment 8

[Ryan Hunt [:rhunt]]

This also had a major impact on the SCALARS_GFX.OMTP.PAINT_WAIT_RATIO telemetry probe. [1]

Comparing the probe from 5/25 to 5/24, we see the median drop by half from 161.2 to 86.7. That is a change of 1.6% of blocking paints waiting over 200us to 0.87% of paints.

[1] http://alerts.telemetry.mozilla.org/index.html#/detectors/1/metrics/2530/alerts/?from=2018-05-25&to=2018-05-25
[2] https://telemetry.mozilla.org/new-pipeline/dist.html#!cumulative=0&end_date=2018-05-25&keys=__none__!__none__!__none__&max_channel_version=nightly%252F62&measure=SCALARS_GFX.OMTP.PAINT_WAIT_RATIO&min_channel_version=null&processType=*&product=Firefox&sanitize=1&sort_keys=submissions&start_date=2018-05-25&table=0&trim=1&use_submission_date=0
[3] https://telemetry.mozilla.org/new-pipeline/dist.html#!cumulative=0&end_date=2018-05-24&keys=__none__!__none__!__none__&max_channel_version=nightly%252F62&measure=SCALARS_GFX.OMTP.PAINT_WAIT_RATIO&min_channel_version=null&processType=*&product=Firefox&sanitize=1&sort_keys=submissions&start_date=2018-05-24&table=0&trim=1&use_submission_date=0