The real locking/fencing/waiting semantics we want are Acquire and Release for both Producer and Consumer. We need this architecture for using `SharedMutex`s with D3D11.

I had a super quick look at this and it looks like what we want.

Comment on attachment 8492478 [details] [diff] [review] recycle-fence Review of attachment 8492478 [details] [diff] [review]: ----------------------------------------------------------------- ::: gfx/layers/composite/TextureHost.cpp @@ +1166,5 @@ > + MOZ_ASSERT(!mIsLocked); > + > + EnsureTexSource(); > + > + mSurf->ConsumerAcquire(); It would work better if we did ConsumerAcquire before EnsureTexSource. I want to make it so that we only have to make it so that we only have to OpenShareHandle once per draw.

This patch seems to be hitting the assertion in ProducerRelease() called from GLScreenBuffer::Swap()

This fixes the assertion by calling Acquire in ::Resize and fixes another assertion by calling Acquire before EnsureTexSource.

Comment on attachment 8505651 [details] [diff] [review] acquire-release.patch Review of attachment 8505651 [details] [diff] [review]: ----------------------------------------------------------------- Cool. It looks like we currently call Fence twice. This might explain some of the talos regressions. ::: gfx/gl/SharedSurface.h @@ +80,5 @@ > virtual void UnlockProdImpl() = 0; > > + virtual void ProducerAcquireImpl() {} > + virtual void ProducerReleaseImpl() { > + Fence(); This file is still 4-space.

Looks like we're still failing an assertion: https://tbpl.mozilla.org/?tree=Try&rev=002446aa01a2

This was without the double readback fix though. Trying again.

Still failing against current inbound: https://tbpl.mozilla.org/?tree=Try&rev=8a543c535e56

This adds a ProducerRelease() in Resize() that's called when needed.