The material change is I_DIRTY_TIME handling without a spurious ref
acquire/release cycle.

While here a bunch of smaller changes:
1. predict there is an inode -- bpftrace suggests one is passed vast
   majority of the time
2. convert BUG_ON into VFS_BUG_ON_INODE
3. assert on ->i_count
4. assert ->i_lock is not held
5. flip the order of I_DIRTY_TIME and nlink count checks as the former
   is less likely to be true

I verified atomic_read(&inode->i_count) does not show up in asm if
debug is disabled.

Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
---

The routine kept annoying me, so here is a further revised variant.

I verified this compiles, but I still cannot runtime test. I'm sorry for
that.  My signed-off is conditional on a good samaritan making sure it
works :)

diff compared to the thing I sent "informally":
- if (unlikely(!inode))
- asserts
- slightly reworded iput_final commentary
- unlikely() on the second I_DIRTY_TIME check

Given the revamp I think it makes sense to attribute the change to me,
hence a "proper" mail.

The thing surviving from the submission by Josef is:
+       if (atomic_add_unless(&inode->i_count, -1, 1))
+               return;

And of course he is the one who brought up the spurious refcount trip in
the first place.

I'm happy with Reported-by, Co-developed-by or whatever other credit
as you guys see fit.

That aside I think it would be nice if NULL inodes passed to iput
became illegal, but that's a different story for another day.

 fs/inode.c | 46 +++++++++++++++++++++++++++++++++++-----------
 1 file changed, 35 insertions(+), 11 deletions(-)

diff --git a/fs/inode.c b/fs/inode.c
index 01ebdc40021e..01a554e11279 100644
--- a/fs/inode.c
+++ b/fs/inode.c
@@ -1908,20 +1908,44 @@ static void iput_final(struct inode *inode)
  */
 void iput(struct inode *inode)
 {
-	if (!inode)
+	if (unlikely(!inode))
 		return;
-	BUG_ON(inode->i_state & I_CLEAR);
+
 retry:
-	if (atomic_dec_and_lock(&inode->i_count, &inode->i_lock)) {
-		if (inode->i_nlink && (inode->i_state & I_DIRTY_TIME)) {
-			atomic_inc(&inode->i_count);
-			spin_unlock(&inode->i_lock);
-			trace_writeback_lazytime_iput(inode);
-			mark_inode_dirty_sync(inode);
-			goto retry;
-		}
-		iput_final(inode);
+	lockdep_assert_not_held(&inode->i_lock);
+	VFS_BUG_ON_INODE(inode->i_state & I_CLEAR, inode);
+	/*
+	 * Note this assert is technically racy as if the count is bogusly
+	 * equal to one, then two CPUs racing to further drop it can both
+	 * conclude it's fine.
+	 */
+	VFS_BUG_ON_INODE(atomic_read(&inode->i_count) < 1, inode);
+
+	if (atomic_add_unless(&inode->i_count, -1, 1))
+		return;
+
+	if ((inode->i_state & I_DIRTY_TIME) && inode->i_nlink) {
+		trace_writeback_lazytime_iput(inode);
+		mark_inode_dirty_sync(inode);
+		goto retry;
 	}
+
+	spin_lock(&inode->i_lock);
+	if (unlikely((inode->i_state & I_DIRTY_TIME) && inode->i_nlink)) {
+		spin_unlock(&inode->i_lock);
+		goto retry;
+	}
+
+	if (!atomic_dec_and_test(&inode->i_count)) {
+		spin_unlock(&inode->i_lock);
+		return;
+	}
+
+	/*
+	 * iput_final() drops ->i_lock, we can't assert on it as the inode may
+	 * be deallocated by the time the call returns.
+	 */
+	iput_final(inode);
 }
 EXPORT_SYMBOL(iput);
 
-- 
2.43.0