On Tue, 2026-09-22 at 16:37 +1000, NeilBrown wrote: > On Tue, 22 Sep 2026, Jeff Layton wrote: > > nfsd_global_mutex still serializes file cache flushes against server > > start/stop in unrelated namespaces. nfsd_file_cache_purge() is reached from > > expkey_flush() on every "exportfs -f", which makes it the only frequent > > taker of the global lock -- and it has nothing to do with nfsd_users or the > > NFSv4 global tables it would be waiting on. > > > > Move NFSD_FILE_CACHE_UP and the objects it covers (rhltable, LRU, slabs, > > shrinker, fsnotify groups) onto a nfsd_file_cache_mutex private to > > filecache.c. nfsd_file_cache_init() and nfsd_file_cache_shutdown() take it > > themselves rather than asserting the caller holds the global one, so > > nfssvc.c no longer needs to know how the cache locks itself. > > > > nfsd_global_mutex is left holding only nfsd_users, the notifier count and > > user_recovery_dirname, all of which are touched solely on server > > start/stop. > > > > Ordering is nn->nfsd_mutex outside nfsd_global_mutex outside > > nfsd_file_cache_mutex. The cache mutex is reached three ways, all > > consistent with that: from nfsd_startup_generic()/nfsd_shutdown_generic() > > under the global mutex, from nfsd_shutdown_net() under nn->nfsd_mutex, and > > bare from expkey_flush() and the filecache stats file. Nothing called with > > the cache mutex held takes either of the others. > > > > Assisted-by: LLM > > Signed-off-by: Jeff Layton <[email protected]> > > --- > > fs/nfsd/filecache.c | 39 ++++++++++++++++++++++++++++----------- > > fs/nfsd/nfssvc.c | 9 +++++---- > > 2 files changed, 33 insertions(+), 15 deletions(-) > > > > diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c > > index 501772bc8f8d..179362e5394b 100644 > > --- a/fs/nfsd/filecache.c > > +++ b/fs/nfsd/filecache.c > > @@ -67,6 +67,17 @@ > > */ > > static DEFINE_SPINLOCK(nfsd_gc_lock); > > > > +/* > > + * Guards NFSD_FILE_CACHE_UP and the host-wide objects it covers: the > > + * rhltable, the LRU, the slabs, the shrinker and the fsnotify groups. > > + * Held across cache bring-up and teardown, and by readers and walkers > > + * that need the cache to stay up for the duration (->cache_purge, the > > + * stats file). > > + * > > + * Nests inside nfsd_global_mutex and inside nn->nfsd_mutex. > > + */ > > +static DEFINE_MUTEX(nfsd_file_cache_mutex); > > + > > static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits); > > static DEFINE_PER_CPU(unsigned long, nfsd_file_acquisitions); > > static DEFINE_PER_CPU(unsigned long, nfsd_file_allocations); > > @@ -869,9 +880,11 @@ nfsd_file_cache_init(void) > > { > > int ret; > > > > - lockdep_assert_held(&nfsd_global_mutex); > > - if (test_and_set_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1) > > + mutex_lock(&nfsd_file_cache_mutex); > > What would you think of making this > > guard(mutex)(&nfsd_file_cache_mutex); > > so you don't need to make other changes in this function?? > >
Sounds good. I'll do that. Thanks! -- Jeff Layton <[email protected]>
