__gen_oldest() linearly scans the entire session array (conn->session_cnt slots) on every call.
Generation checks run on every page/memory free, so discarding a large tree from cache - e.g. __wt_evict_file(WT_SYNC_DISCARD) during a non-forced drop - costs O(in-cache pages * active sessions).

Observed in cetrain cases: perf flame graphs show __gen_oldest dominating CPU during ident drops.
The stall caused connection pile-up, raising open session counts to 3-7k which further slowed the discard - a feedback loop.
Single-ident discards held the schema and dhandle-list write locks for 16-70s.

The discard path amplifies the cost but any generation check under high session counts pays the same linear scan.

Possible directions (to be evaluated):

• Cache/track the oldest generation instead of recomputing by full scan
• Maintain a per-generation minimum incrementally as sessions enter/leave
• Batch frees during file discard so one generation check covers many pages