nfs: don't skip revalidate on directory delegation when attrs flagged stale

On a local directory mutation (rename/create/unlink) the client marks
CHANGE / MTIME / CTIME as invalid in NFS_I(dir)->cache_validity.  When
a subsequent stat(2) enters __nfs_revalidate_inode() and finds a
directory delegation held, the function currently early-exits and
returns the cached (now stale) mtime to userspace without sending a
GETATTR RPC.

Keep the early-exit for the fast path, but take the RPC when CHANGE,
MTIME, or CTIME are already marked invalid.  The delegation alone is
not a guarantee of cached-attr freshness once the code itself has
flagged the cache as stale.

Assisted-by: Claude:claude-opus-4-7 [bpftrace] [tshark]
Signed-off-by: Tom Haynes <loghyr@gmail.com>
[Anna: Use NFS_INO_INVALID_ATTR insteado of individual NFS_INO_INVALID_* flags]
Signed-off-by: Anna Schumaker <anna@kernel.org>

diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
index 0d9451a2ad8ed9a17419747b2ca92b65a5947ec2..c6c939a76722181295a3fe1a747b16c731106f3f 100644 (file)
--- a/fs/nfs/inode.c
+++ b/fs/nfs/inode.c
@@ -1357,7 +1357,8 @@ __nfs_revalidate_inode(struct nfs_server *server, struct inode *inode)
                status = pnfs_sync_inode(inode, false);
                if (status)
                        goto out;
-       } else if (nfs_have_directory_delegation(inode)) {
+       } else if (nfs_have_directory_delegation(inode) &&
+                  !(NFS_I(inode)->cache_validity & NFS_INO_INVALID_ATTR)) {
                status = 0;
                goto out;
        }