On Wed, Sep 03, 2025 at 11:05:57AM +1000, Dave Chinner wrote: > I think that the buftarg rhashtable needs to initialised before the > shrinker is registered, then freed from xfs_destroy_buftarg() after > the shrinker has been shut down as it must live longer than the > buftarg shrinker instance. i.e. the buftarg rhashtable needs to have > the same life cycle as the buftarg LRU list.... My RFC patch to switch back to a per-buftarg hashtable would do that: https://git.infradead.org/?p=users/hch/xfs.git;a=commitdiff;h=e3cc537864a7ab980abfa18a3efe01a111aad1d7 I still haven't gotten around doing serious performance testing on it, and I'm busy right now. But in about two weeks I'll probably have both a bit of time and access to a big enough system to do serious scalability testing on it.
