There's a race condition in the destruction of the BindingState for bindings to the StoragePartitionService.
It looks like the root cause of the issue is that since we can get two concurrent calls to callbacks returned from mojo::BindingSet::GetBadMessageCallback() from the same BindingSet, which results in a data race destroying the same BindingState.
One case where this can be called is in a synchronous context when validating a received message, or when tearing down the connection. The other context is the callback passed to OpenSessionStorage here
This is then posted to the mojo task runner thread here:
And can be called on this thread directly here:
Since this is a tight race, the repro is not terribly reliable. Opening multiple tabs at once will increase your chances of reproducing to the point where the issue triggers around 50% of the time (on my machine, mileage will vary depending on system load etc...). I found the most convenient way to reproduce was to use a ThreadSanitizer build, but since the most common result is a double-free, this will often crash a normal release build when the freelist corruption is detected.
$ python ./copy_mojo_js_bindings.py /path/to/chrome/.../out/Asan/gen
$ python -m SimpleHTTPServer&
$ /ssd/chrome_trunk/src/out/Tsan/chrome --enable-blink-features=MojoJS
--no-sandbox --user-data-dir=/tmp/aa 'http://localhost:8000/index.html' 'http://localhost:8000/index.html' 'http://localhost:8000/index.html' 'http://localhost:8000/index.html' 'http://localhost:8000/index.html' 'http://localhost:8000/index.html' 'http://localhost:8000/index.html' 'http://localhost:8000/index.html'
Proof of Concept: