[PATCH] ceph: keep dentry in cache when inode still holds caps

From: Xiubo Li <[email protected]>

ceph_d_delete() drops a dentry once its lease expires, which
releases the inode and destroys its page cache at the last close.
A buffered write that outlasts the dentry lease duration (30s by
default on the MDS side) is flushed and closed, yet the reopen
re-reads everything from the OSDs instead of the page cache, even
though the caps and the pages are still valid.

Keep the dentry when the inode still holds caps: the caps guarantee
the inode object is still valid, and the name linkage is revalidated
against the MDS by ceph_d_revalidate() on the next lookup anyway.

Keeping the dentry also keeps the inode and its page cache around
longer than before.  This is bounded: caps no longer needed by any
open file or by dirty data are released once the delayed release
window expires (caps_wanted_delay_max, 60s by default) via
ceph_check_delayed_caps(), after which the dentry is just an
ordinary cached dentry.  It is also fully reclaimable: ->d_delete()
is only consulted on the last dput (in retain_dentry()), while the
dcache shrinker (shrink_dentry_list() -> __dentry_kill()) does not
call it, so memory pressure frees these dentries like any others.

The check is deliberately racy: i_ceph_lock cannot be taken under
dentry->d_lock, but a false result only means keeping or dropping a
dentry that could have gone the other way, which is safe.

Signed-off-by: Xiubo Li <[email protected]>
---
 fs/ceph/dir.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
index f4e0bf244fd2..a7d33ad9d2c7 100644
--- a/fs/ceph/dir.c
+++ b/fs/ceph/dir.c
@@ -2090,6 +2090,20 @@ static int ceph_d_delete(const struct dentry *dentry)
 		if (__dir_lease_try_check(dentry))
 			return 0;
 	}
+	/*
+	 * The lease has expired, but if the inode still holds caps, keep
+	 * the dentry: dropping it would release the inode and destroy its
+	 * page cache (e.g. on the last close after a long write).  The
+	 * caps guarantee the inode itself is still valid, and the name
+	 * linkage is revalidated against the MDS on the next lookup
+	 * anyway.
+	 *
+	 * This is deliberately racy: we can't take i_ceph_lock under
+	 * dentry->d_lock, but a false result only means we keep or drop a
+	 * dentry we could have done the opposite with, which is safe.
+	 */
+	if (__ceph_is_any_real_caps(ceph_inode(d_inode(dentry))))
+		return 0;
 	return 1;
 }
 

---
base-commit: 6a8322d32e2e2d3e364d31fc86c784e0e0105c08
change-id: 20260821-b4-b4-ceph-dentry-caps-ad44734dea85

Best regards,
--  
Xiubo Li <[email protected]>