There appears to be a race condition in the destruction of the ExtensionsGuestViewMessageFilter if the ProcessIdToFilterMap is modified concurrently.
See the comment in the code:
// This map is created and accessed on the UI thread. Remove the reference to
// |this| here so that it will not be accessed again; but leave erasing the
// key from the global map to UI thread to avoid races when accessing the
// underlying data structure (https:/crbug.com/869791).
(*GetProcessIdToFilterMap())[render_process_id_] = nullptr;
This comment doesn't describe behaviour that appears to be safe to me - there's no explicit mention of the safety of concurrent modification of base::flat_map; but it is noted that iterators are invalidated on insertion/erase, so as there is no further synchronisation, it doesn't appear that it is safe to read or write from this map if another thread may be concurrently modifying it.
This issue was detected by TSAN during fuzzing; but I don't think the behaviour is related to the fuzz case - it's just a dormant issue that was caught. I've attached the TSAN splat that resulted, but I don't have a reliable testcase to reproduce this issue.
This bug is subject to a 90 day disclosure deadline. After 90 days elapse
or a patch has been made broadly available (whichever is earlier), the bug
report will become visible to the public.
Ah, you can reproduce this quite easily without even having a compromised renderer. Before I submitted the report to Chrome I thought I should try just spawning lots of renderers, and that repros fairly quickly - still the same (benign, I think) race that's occuring, but that's unsurprising since the renderers in this case aren't doing anything. The attached testcase triggers in < 5 min with a TSAN build.
The testcase assumes that the domains test0.com - test63.com all resolve to your local server.
Proof of Concept: