MatroskaExtractor crashes with NPE on nested ChapterAtom elements (regression in 1.11.0) #3438

Version

Media3 1.11.1

More version details

Also reproduces on 1.11.0. Works on 1.10.1 (before Matroska chapter parsing was added to MatroskaExtractor).

Devices that reproduce the issue

• Robolectric (SDK 35) unit tests on JVM
• Not device-specific: the crash is in pure extractor parsing logic

Devices that do not reproduce the issue

None known on 1.11.x.

Reproducible in the demo app?

Not tested

Reproduction steps

1. Take a Matroska file whose chapters are nested, i.e. a ChapterAtom contains child ChapterAtoms. This is allowed by the spec: ChapterAtom is recursive (\Segment\Chapters\EditionEntry\+ChapterAtom, RFC 9559).
2. Extract it with MatroskaExtractor, e.g. via MetadataRetriever.retrieveTrackGroups() / retrieveDurationUs() or by playing it.

The file below was muxed with mkvmerge from this chapter XML using this script. mkvinfo (v102) parses it without warnings, and ffprobe reads its chapters without problems.

Structure:

EditionEntry
  ChapterAtom "Introduction"
    ChapterAtom "Getting Started"
    ChapterAtom "Basic Concepts"
  ChapterAtom "Main Content"
    ChapterAtom "Advanced Techniques"
    ChapterAtom "Practical Examples"
  ChapterAtom "Conclusion"

Expected result

The file is extracted successfully: tracks and duration are reported and chapters are exposed (at least the top-level ones, ideally the nested ones too).

Actual result

MatroskaExtractor throws an NPE before any track is output:

java.lang.NullPointerException
    at com.google.common.base.Preconditions.checkNotNull(Preconditions.java:902)
    at androidx.media3.extractor.mkv.MatroskaExtractor.endMasterElement(MatroskaExtractor.java:978)
    at androidx.media3.extractor.mkv.MatroskaExtractor$InnerEbmlProcessor.endMasterElement(MatroskaExtractor.java:2356)
    at androidx.media3.extractor.mkv.DefaultEbmlReader.read(DefaultEbmlReader.java:88)
    at androidx.media3.extractor.mkv.MatroskaExtractor.read(MatroskaExtractor.java:659)

Claude Analysis:

Cause: the extractor tracks a single currentChapter field instead of a stack. startMasterElement(ID_CHAPTER_ATOM) overwrites it with the nested atom, the nested atom's endMasterElement sets it to null, and the parent atom's endMasterElement then hits checkNotNull(currentChapter). A ChapterDisplay/ChapterTimeStart etc. following a nested atom inside the parent would similarly fail in getCurrentChapter().

Because the error is thrown while parsing, the whole file becomes unusable: MetadataRetriever fails for both track groups and duration, and the file can't be played, even though it is valid and worked in 1.10.1.

A side note on the symptom: under Robolectric, the MetadataRetriever future never completes (instead of failing), because MetadataRetrieverInternal only surfaces the prepare error through a sendEmptyMessageDelayed poll, and the paused looper's clock doesn't advance while the test thread is blocked waiting on the future.

Context: this blocks the Voice audiobook player from moving to 1.11 and adopting the new chapters API: PaulWoitaschek/Voice#3779

You can find the exact reproducer also here: PaulWoitaschek/Voice#3703

Media

https://github.com/PaulWoitaschek/Voice/raw/f827b30dc6d8db88d206332b70331dd31e38563c/core/scanner/src/test/resources/mka_nested_chapters.mka (~1 MB, silent AAC audio with nested chapters)