With this change only 0->1 and 1->0 transitions need the lock.

I verified all places which look at the refcount either only care about
it staying 0 (and have the lock enforce it) or don't hold the inode lock
to begin with (making the above change irrelevant to their correcness or
lack thereof).

I also confirmed nfs and btrfs like to call into these a lot and now
avoid the lock in the common case, shaving off some atomics.

Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
---
 fs/dcache.c        |  4 +++
 fs/inode.c         | 65 ++++++++++++++++++++++++++++++++++++++++++++++
 include/linux/fs.h |  4 +--
 3 files changed, 71 insertions(+), 2 deletions(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index 9ceab142896f..b63450ebb85c 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -2033,6 +2033,10 @@ void d_instantiate_new(struct dentry *entry, struct inode *inode)
 	__d_instantiate(entry, inode);
 	spin_unlock(&entry->d_lock);
 	WARN_ON(!(inode_state_read(inode) & I_NEW));
+	/*
+	 * Paired with igrab_try_lockless()
+	 */
+	smp_wmb();
 	inode_state_clear(inode, I_NEW | I_CREATING);
 	inode_wake_up_bit(inode, __I_NEW);
 	spin_unlock(&inode->i_lock);
diff --git a/fs/inode.c b/fs/inode.c
index 013470e6d144..03472be4e1a9 100644
--- a/fs/inode.c
+++ b/fs/inode.c
@@ -1029,6 +1029,7 @@ long prune_icache_sb(struct super_block *sb, struct shrink_control *sc)
 }
 
 static void __wait_on_freeing_inode(struct inode *inode, bool hash_locked, bool rcu_locked);
+static bool igrab_try_lockless(struct inode *inode);
 
 /*
  * Called with the inode lock held.
@@ -1053,6 +1054,11 @@ static struct inode *find_inode(struct super_block *sb,
 			continue;
 		if (!test(inode, data))
 			continue;
+		if (igrab_try_lockless(inode)) {
+			rcu_read_unlock();
+			*isnew = false;
+			return inode;
+		}
 		spin_lock(&inode->i_lock);
 		if (inode_state_read(inode) & (I_FREEING | I_WILL_FREE)) {
 			__wait_on_freeing_inode(inode, hash_locked, true);
@@ -1095,6 +1101,11 @@ static struct inode *find_inode_fast(struct super_block *sb,
 			continue;
 		if (inode->i_sb != sb)
 			continue;
+		if (igrab_try_lockless(inode)) {
+			rcu_read_unlock();
+			*isnew = false;
+			return inode;
+		}
 		spin_lock(&inode->i_lock);
 		if (inode_state_read(inode) & (I_FREEING | I_WILL_FREE)) {
 			__wait_on_freeing_inode(inode, hash_locked, true);
@@ -1212,6 +1223,10 @@ void unlock_new_inode(struct inode *inode)
 	lockdep_annotate_inode_mutex_key(inode);
 	spin_lock(&inode->i_lock);
 	WARN_ON(!(inode_state_read(inode) & I_NEW));
+	/*
+	 * Paired with igrab_try_lockless()
+	 */
+	smp_wmb();
 	inode_state_clear(inode, I_NEW | I_CREATING);
 	inode_wake_up_bit(inode, __I_NEW);
 	spin_unlock(&inode->i_lock);
@@ -1223,6 +1238,10 @@ void discard_new_inode(struct inode *inode)
 	lockdep_annotate_inode_mutex_key(inode);
 	spin_lock(&inode->i_lock);
 	WARN_ON(!(inode_state_read(inode) & I_NEW));
+	/*
+	 * Paired with igrab_try_lockless()
+	 */
+	smp_wmb();
 	inode_state_clear(inode, I_NEW);
 	inode_wake_up_bit(inode, __I_NEW);
 	spin_unlock(&inode->i_lock);
@@ -1582,6 +1601,14 @@ EXPORT_SYMBOL(ihold);
 
 struct inode *igrab(struct inode *inode)
 {
+	/*
+	 * Read commentary above igrab_try_lockless() for an explanation why this works.
+	 */
+	if (atomic_add_unless(&inode->i_count, 1, 0)) {
+		VFS_BUG_ON_INODE(inode_state_read_once(inode) & (I_FREEING | I_WILL_FREE), inode);
+		return inode;
+	}
+
 	spin_lock(&inode->i_lock);
 	if (!(inode_state_read(inode) & (I_FREEING | I_WILL_FREE))) {
 		__iget(inode);
@@ -1599,6 +1626,44 @@ struct inode *igrab(struct inode *inode)
 }
 EXPORT_SYMBOL(igrab);
 
+/*
+ * igrab_try_lockless - special inode refcount acquire primitive for the inode hash
+ * (don't use elsewhere!)
+ *
+ * It provides lockless refcount acquire in the common case of no problematic
+ * flags being set and the count being > 0.
+ *
+ * There are 4 state flags to worry about and the routine makes sure to not bump the
+ * ref if any of them is present.
+ *
+ * I_NEW and I_CREATING can only legally get set *before* the inode becomes visible
+ * during lookup. Thus if the flags are not spotted, they are guaranteed to not be
+ * a factor. However, we need an acquire fence before returning the inode just
+ * in case we raced against clearing the state to make sure our consumer picks up
+ * any other changes made prior. atomic_add_unless provides a full fence, which
+ * takes care of it.
+ *
+ * I_FREEING and I_WILL_FREE can only legally get set if ->i_count == 0 and it is
+ * illegal to bump the ref if either is present. Consequently if atomic_add_unless
+ * managed to replace a non-0 value with a bigger one, we have a guarantee neither
+ * of these flags is set. Note this means explicitly checking of these flags below
+ * is not necessary, it is only done because it does not cost anything on top of the
+ * load which already needs to be done to handle the other flags.
+ */
+static bool igrab_try_lockless(struct inode *inode)
+{
+	if (inode_state_read_once(inode) & (I_NEW | I_CREATING | I_FREEING | I_WILL_FREE))
+		return false;
+	/*
+	 * Paired with routines clearing I_NEW
+	 */
+	if (atomic_add_unless(&inode->i_count, 1, 0)) {
+		VFS_BUG_ON_INODE(inode_state_read_once(inode) & (I_FREEING | I_WILL_FREE), inode);
+		return true;
+	}
+	return false;
+}
+
 /**
  * ilookup5_nowait - search for an inode in the inode cache
  * @sb:		super block of file system to search
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 07363fce4406..119e0a3d2f42 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -2234,8 +2234,8 @@ static inline int icount_read_once(const struct inode *inode)
 }
 
 /*
- * returns the refcount on the inode. The lock guarantees no new references
- * are added, but references can be dropped as long as the result is > 0.
+ * returns the refcount on the inode. The lock guarantees no 0->1 or 1->0 transitions
+ * of the count are going to take place, otherwise it changes arbitrarily.
  */
 static inline int icount_read(const struct inode *inode)
 {
-- 
2.48.1