Fixed
Status Update
Comments
ap...@google.com <ap...@google.com> #2
A couple of questions:
1. Have you saw crash in real device or only in simulators?
2. Do you use dynamic feature for language ID?
1. Have you saw crash in real device or only in simulators?
2. Do you use dynamic feature for language ID?
ap...@google.com <ap...@google.com> #3
Tested on Android 12 Emulator with custom executor, but cannot repro this issue.
na...@google.com <na...@google.com> #4
-
Second crash in the description is from a real device. Experienced it myself on two different Xiaomi phones, plus lots of crashes from users in the Google Play console.
-
Dynamic features are not used in the application.
As a wild guess, I have downgraded build tools from 31.0.0 to 30.0.3, compileSdk from 31 to 30, and moved all work with Language ID to the service in a separate process (just to be sure that crash can kill secondary process instead of main). This combination is in beta for 2 days by now and I don't see any SIGSEGV crashes.
Description
Some more details in b/388353336
Start of trace:
This happens because when we draw we use a shared ContentDrawScope implementation that we render all DrawModifierNodes in. We do this by mutating a property as we draw successive nodes. However, this means that if anyone captures a reference to this draw scope and tries to draw it later on, this will cause a NPE error as we don't know what the 'current node' that needs to be drawn in.
For the general case this is an error and scopes shouldn't be captured and stored outside the scope in this way, but for GraphicsLayer this is needed when recording drawContent to support some edge cases, such as when we need to draw a layer implementation that can't be drawn in software, within software rendering (such as drawing to a Bitmap). For this we just end up re-invoking the draw block we captured.
We considered some solutions where we would allocate a new draw scope instance per modifier, but this adds some overhead and 'storing' these draw scopes for future use is challenging. (we considered putting them on the node object, in a map on the node coordinator, or some other chain structure on the node coordinator, but none of these are really great)
Instead we can explicitly support the GraphicsLayer.record case by overriding the DrawScope#GraphicsLayer.record function and making it handle this case properly. This can still cause issues if using the non-DrawScope scoped record function, but developers should move to the DrawScope one instead.