During async unlink, we drop the `i_nlink` counter before we receive
the completion (that will eventually update the `i_nlink`) because "we
assume that the unlink will succeed".  That is not a bad idea, but it
races against deletions by other clients (or against the completion of
our own unlink) and can lead to an underrun which emits a WARNING like
this one:

 WARNING: CPU: 85 PID: 25093 at fs/inode.c:407 drop_nlink+0x50/0x68
 Modules linked in:
 CPU: 85 UID: 3221252029 PID: 25093 Comm: php-cgi8.1 Not tainted 6.14.11-cm4all1-ampere #655
 Hardware name: Supermicro ARS-110M-NR/R12SPD-A, BIOS 1.1b 10/17/2023
 pstate: 60400009 (nZCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
 pc : drop_nlink+0x50/0x68
 lr : ceph_unlink+0x6c4/0x720
 sp : ffff80012173bc90
 x29: ffff80012173bc90 x28: ffff086d0a45aaf8 x27: ffff0871d0eb5680
 x26: ffff087f2a64a718 x25: 0000020000000180 x24: 0000000061c88647
 x23: 0000000000000002 x22: ffff07ff9236d800 x21: 0000000000001203
 x20: ffff07ff9237b000 x19: ffff088b8296afc0 x18: 00000000f3c93365
 x17: 0000000000070000 x16: ffff08faffcbdfe8 x15: ffff08faffcbdfec
 x14: 0000000000000000 x13: 45445f65645f3037 x12: 34385f6369706f74
 x11: 0000a2653104bb20 x10: ffffd85f26d73290 x9 : ffffd85f25664f94
 x8 : 00000000000000c0 x7 : 0000000000000000 x6 : 0000000000000002
 x5 : 0000000000000081 x4 : 0000000000000481 x3 : 0000000000000000
 x2 : 0000000000000000 x1 : 0000000000000000 x0 : ffff08727d3f91e8
 Call trace:
  drop_nlink+0x50/0x68 (P)
  vfs_unlink+0xb0/0x2e8
  do_unlinkat+0x204/0x288
  __arm64_sys_unlinkat+0x3c/0x80
  invoke_syscall.constprop.0+0x54/0xe8
  do_el0_svc+0xa4/0xc8
  el0_svc+0x18/0x58
  el0t_64_sync_handler+0x104/0x130
  el0t_64_sync+0x154/0x158

In ceph_unlink(), a call to ceph_mdsc_submit_request() submits the
CEPH_MDS_OP_UNLINK to the MDS, but does not wait for completion.

Meanwhile, between this call and the following drop_nlink() call, a
worker thread may process a CEPH_CAP_OP_IMPORT, CEPH_CAP_OP_GRANT or
just a CEPH_MSG_CLIENT_REPLY (the latter of which could be our own
completion).  These will lead to a set_nlink() call, updating the
`i_nlink` counter to the value received from the MDS.  If that new
`i_nlink` value happens to be zero, it is illegal to decrement it
further.  But that is exactly what ceph_unlink() will do then.

The WARNING can be reproduced this way:

1. Force async unlink; only the async code path is affected.  Having
   no real clue about Ceph internals, I was unable to find out why the
   MDS wouldn't give me the "Fxr" capabilities, so I patched
   get_caps_for_async_unlink() to always succeed.

   (Note that the WARNING dump above was found on an unpatched kernel,
   without this kludge - this is not a theoretical bug.)

2. Add a sleep call after ceph_mdsc_submit_request() so the unlink
   completion gets handled by a worker thread before drop_nlink() is
   called.  This guarantees that the `i_nlink` is already zero before
   drop_nlink() runs.

The solution is to skip the counter decrement when it is already zero,
but doing so without a lock is still racy (TOCTOU).  Since
ceph_fill_inode() and handle_cap_grant() both hold the
`ceph_inode_info.i_ceph_lock` spinlock while set_nlink() runs, this
seems like the proper lock to protect the `i_nlink` updates.

I found prior art in NFS and SMB (using `inode.i_lock`) and AFS (using
`afs_vnode.cb_lock`).  All three have the zero check as well.

Fixes: 2ccb45462aea ("ceph: perform asynchronous unlink if we have sufficient caps")
Cc: stable@vger.kernel.org
Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
---
 fs/ceph/dir.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
index 8478e7e75df6..67f04e23f78a 100644
--- a/fs/ceph/dir.c
+++ b/fs/ceph/dir.c
@@ -1341,6 +1341,7 @@ static int ceph_unlink(struct inode *dir, struct dentry *dentry)
 	struct ceph_client *cl = fsc->client;
 	struct ceph_mds_client *mdsc = fsc->mdsc;
 	struct inode *inode = d_inode(dentry);
+	struct ceph_inode_info *ci = ceph_inode(inode);
 	struct ceph_mds_request *req;
 	bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 	struct dentry *dn;
@@ -1427,7 +1428,19 @@ static int ceph_unlink(struct inode *dir, struct dentry *dentry)
 			 * We have enough caps, so we assume that the unlink
 			 * will succeed. Fix up the target inode and dcache.
 			 */
-			drop_nlink(inode);
+
+			/*
+			 * Protect the i_nlink update with i_ceph_lock
+			 * to precent racing against ceph_fill_inode()
+			 * handling our completion on a worker thread
+			 * and don't decrement if i_nlink has already
+			 * been updated to zero by this completion.
+			 */
+			spin_lock(&ci->i_ceph_lock);
+			if (inode->i_nlink > 0)
+				drop_nlink(inode);
+			spin_unlock(&ci->i_ceph_lock);
+
 			d_delete(dentry);
 		} else {
 			spin_lock(&fsc->async_unlink_conflict_lock);
-- 
2.47.2