Fix assertion failure and memory leak in ClockCache.

[guidotag]

Summary: This fixes two issues:

• T127355728: In the stress tests, when the ClockCache is operating close to full capacity and a burst of inserts are concurrently executed, every slot in the hash table may become occupied. This contradicts an assertion in the code, which is no longer valid in the lock-free setting. We are removing that assertion and handling the case of an insertion into a full table.
• T127427659: There was a memory leak when an insertion is performed over capacity, but no handle is provided. In that case, a handle was dynamically allocated, but the pointer wasn't stored anywhere.

Test plan:

• make -j24 check
• make -j24 USE_CLANG=1 COMPILE_WITH_ASAN=1 COMPILE_WITH_UBSAN=1 CRASH_TEST_EXT_ARGS="--duration=960 --cache_type=clock_cache" blackbox_crash_test_with_atomic_flush
• make -j24 USE_CLANG=1 COMPILE_WITH_TSAN=1 CRASH_TEST_EXT_ARGS="--duration=960 --cache_type=clock_cache" blackbox_crash_test_with_atomic_flush

[facebook-github-bot]

@guidotag has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[pdillinger]

Better title & summary for a public audience, please. Someone should be able to quickly decide whether this change is something they should care about or affects something they are wondering about. If it's marked an assertion failure in the experimental ClockCache, we can save people time.

[pdillinger]

I'm pretty sure LRUCache uses OK for this kind of case, and therefore we should also.

rocksdb/cache/lru_cache.cc

Line 327 in 8b2d429

// Don't insert the entry but still return ok, as if the entry inserted

[facebook-github-bot]

@guidotag has updated the pull request. You must reimport the pull request before landing.

[pdillinger]

LGTM thanks!

[facebook-github-bot]

@guidotag has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.