> On Jul 10, 2023, at 9:03 AM, Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > On Sun, 2023-07-09 at 11:45 -0400, Chuck Lever wrote: >> This series optimizes DRC scalability by freeing cache objects only >> once the hash bucket lock is no longer held. There are a couple of >> related clean-ups to go along with this optimization. >> > > > The conventional wisdom that I've always heard is that a kfree under > spinlock is generally no big deal. It can't block and is usually quite > fast. Are you able to measure any performance delta from this set? Yes, a couple of percent better throughput with a tmpfs export and a fast transport, which is not a common use case, granted. I think the difference is that, after this change, we're holding each bucket lock for a shorter period of time. That's enough to reduce some lock contention. But I also like the shrinker improvements. >> --- >> >> Chuck Lever (6): >> NFSD: Refactor nfsd_reply_cache_free_locked() >> NFSD: Rename nfsd_reply_cache_alloc() >> NFSD: Replace nfsd_prune_bucket() >> NFSD: Refactor the duplicate reply cache shrinker >> NFSD: Remove svc_rqst::rq_cacherep >> NFSD: Rename struct svc_cacherep >> >> >> fs/nfsd/cache.h | 8 +- >> fs/nfsd/nfscache.c | 203 ++++++++++++++++++++++++------------- >> fs/nfsd/nfssvc.c | 10 +- >> fs/nfsd/trace.h | 26 ++++- >> include/linux/sunrpc/svc.h | 1 - >> 5 files changed, 165 insertions(+), 83 deletions(-) >> >> -- >> Chuck Lever >> > > -- > Jeff Layton <jlayton@xxxxxxxxxx> -- Chuck Lever
