mirror of
https://github.com/versity/scoutfs.git
synced 2026-04-16 03:46:59 +00:00
5bea29a168
The use of pages in the item cache got us pretty far but it fundmanetally couldn't escape the contention around the global or per-page read locks. Some loads became bottlenecked in contention in the item cache. Worse, we were seeing inconsistency in the per-cpu cached mappings of key ranges to pages. All the users of items in the cache are transitioned from searching for items in locked pages to searching for items in the cwskip list. It's fundamentally built around a seqlock-like begin/retry pattern so most of the item work gets wrapped around search and retry helpers. Without pages we no longer have a global list of dirty pages. Instead we have per-cpu lists of dirty items that are later sorted and handed to the btree insertion iterator. We take the opportunity to clean up that interface now that it's very easy for us to iterate through the stable list of dirty items. Rather than a global lru of pages we have an algorithm for maintaining items in rough groups of ages. Shrinking randomly walks the cwskip list looking for regions of sufficiently old items rather than walking a precise global lru list of pages. Signed-off-by: Zach Brown <zab@versity.com>