[COLLECTIONS-885] PatriciaTrie submap and range view operations silently mutate trie structure resulting in ConcurrentModificationException for any iterating threads

[griffinjm]

Some methods in AbstractPatriciaTrie use a pattern where they silently mutate the trie when searching for a key which is not an exact match for an existing key in the Trie. Said methods are invoked by subMap/tailMap/headMap/prefixMap under the hood. This results in CMEs for even simple shared read access.

ceilingEntry() floorEntry() higherEntry() lowerEntry() operations:
1. insert a phantom node into the trie with the search key
2. increment the modCount
3. find the neighbor of that phantom node via nextEntry()/previousEntry()
4. remove the phantom node
5. increment the modCount again
6. then decrement modCount by 2 to hide the 2 modifications (add + remove)

The referenced modCount is used as a fail-fast mechanism in the iterators.

This fix prevents the issue by instead:

1. Finding the nearest entry via getNearestEntryForKey().
2. Determining whether the search key is less than or greater than the found entry using isBitSet() at the first differing bit.
3. Walking forward (nextEntry()) or backward (previousEntry()) to locate the correct ceiling/floor neighbor.

See https://issues.apache.org/jira/browse/COLLECTIONS-885 for all details.

[garydgregory]

Hello Hello @griffinjm
If I only apply the test side of the patch, all the tests still pass, the new tests should fail. This means a regression could undo whatever the main changes claim to make.

[griffinjm]

Hi @garydgregory
The only test which can expose this bug, and prevent regressions would be a multithreaded test which triggers the phantom insertion causing the original iterator to fail. I can add one if necessary. I'll also attach some sequence diagrams below to help visualize the issue. Thanks.

[griffinjm]

[griffinjm]

[griffinjm]

@garydgregory I added a threaded test to verify behaviour, that test should fail for the existing main branch and pass for this bugfix branch.

[garydgregory]

...

[garydgregory]

Hello @griffinjm

Please see my one comment.

[garydgregory]

@griffinjm
I mered the PR. Thank you!

[griffinjm]

Great! Thank you @garydgregory