Here’s a treat for Oracle performance professionals and geeks who are looking for more systematic ways for cache buffers chains (CBC) latch contention troubleshooting. Cache buffers chains latches are taken when a process wants to walk through a cache buffer hash chain, looking if the block with required DBA (data block address) is in buffer cache. If the block happens to be in cache, then in most cases it has to be pinned first before use and unpinned after use, to make sure no-one else can perform an incompatible operation on that block at the same time. The modification of pin structures (pinning/unpinning) is also protected by CBC latches.
CBC latch contention can happen for multiple reasons, but one reason is that there is some really hot block in a SMP system with high number of CPUs (or CMT system with high number of threads like Sun T-series servers). Sometimes there happen to be multiple moderately hot blocks “under” the same CBC latch, which can result in latch contention again.
Traditionally DBAs used to look up the child latch address from V$SESSION_WAIT, sql_trace output or ASH and then look up all buffers protected by that latch from X$BH, using HLADDR column (HLADDR stands for Hash Latch Address). I also have a script for that, bhla.sql (Buffer Headers by Latch Address), which reports me all blocks currently in buffer cache, “under” that particular latch and the corresponding data block addresses and object names:



