On Sat, Jun 29, 2019 at 05:08:44AM +0100, Al Viro wrote: > > The reason we don't hit that problem with regular memory shrinker is > > this: > > unregister_shrinker(&s->s_shrink); > > fs->kill_sb(s); > > in deactivate_locked_super(). IOW, shrinker for this fs is gone > > before we get around to shutdown. And so are all normal sources > > of dentry eviction for that fs. > > > > Your earlier variants all suffer the same problem - picking a page > > shared by dentries from several superblocks can run into trouble > > if it overlaps with umount of one of those. PS: the problem is not gone in the next iteration of the patchset in question. The patch I'm proposing (including dput_to_list() and _ONLY_ compile-tested) follows. Comments? diff --git a/fs/dcache.c b/fs/dcache.c index 8136bda27a1f..dfe21a649c96 100644 --- a/fs/dcache.c +++ b/fs/dcache.c @@ -860,6 +860,32 @@ void dput(struct dentry *dentry) } EXPORT_SYMBOL(dput); +static void __dput_to_list(struct dentry *dentry, struct list_head *list) +__must_hold(&dentry->d_lock) +{ + if (dentry->d_flags & DCACHE_SHRINK_LIST) { + /* let the owner of the list it's on deal with it */ + --dentry->d_lockref.count; + } else { + if (dentry->d_flags & DCACHE_LRU_LIST) + d_lru_del(dentry); + if (!--dentry->d_lockref.count) + d_shrink_add(dentry, list); + } +} + +void dput_to_list(struct dentry *dentry, struct list_head *list) +{ + rcu_read_lock(); + if (likely(fast_dput(dentry))) { + rcu_read_unlock(); + return; + } + rcu_read_unlock(); + if (!retain_dentry(dentry)) + __dput_to_list(dentry, list); + spin_unlock(&dentry->d_lock); +} /* This must be called with d_lock held */ static inline void __dget_dlock(struct dentry *dentry) @@ -1088,18 +1114,9 @@ static void shrink_dentry_list(struct list_head *list) rcu_read_unlock(); d_shrink_del(dentry); parent = dentry->d_parent; + if (parent != dentry) + __dput_to_list(parent, list); __dentry_kill(dentry); - if (parent == dentry) - continue; - /* - * We need to prune ancestors too. This is necessary to prevent - * quadratic behavior of shrink_dcache_parent(), but is also - * expected to be beneficial in reducing dentry cache - * fragmentation. - */ - dentry = parent; - while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) - dentry = dentry_kill(dentry); } }