[PATCH 5/6] fs/ntfs3: walk $FILE_NAME under ni_lock in ni_write_parents()

[Kamil Chojnowski via B4 Relay]

From: Kamil Chojnowski <contact@xxxxxxxxx>

ni_write_parents() (fsync) walks the inode's attributes without ni_lock.
When the record is rearranged at the same time - a name removed by
unlink, $DATA moved out of the record by a write through another file
descriptor - it can read a half-moved attribute.  It then either follows
a garbage parent reference:

  ntfs3(vda): MFT: r=0, expect seq=0 instead of 1!
  ntfs3(vda): ino=1e6, failed to open parent directory r=0 to write

or mi_enum_attr() sees an inconsistent record and marks a healthy inode
bad and the volume dirty, after which every operation on the file fails
with -EINVAL:

  ntfs3(vda): ino=100, "res2" mi_enum_attr

Collect the parent references under ni_lock and open and write the
parents after dropping it.

Reproducer: overwrite a small file through a shared mmap, grow it past
the MFT record with pwrite(2) and fsync(2) it, while another thread
calls fdatasync(2) on it in a loop.

Fixes: dcd9d6a47199 ("fs/ntfs3: fsync files by syncing parent inodes")
Cc: stable@xxxxxxxxxxxxxxx
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Kamil Chojnowski <contact@xxxxxxxxx>
---
 fs/ntfs3/frecord.c | 42 +++++++++++++++++++++++++++++++++++-------
 1 file changed, 35 insertions(+), 7 deletions(-)

diff --git a/fs/ntfs3/frecord.c b/fs/ntfs3/frecord.c
index 3d10b1519c86..a121ca4db0c0 100644
--- a/fs/ntfs3/frecord.c
+++ b/fs/ntfs3/frecord.c
@@ -2997,13 +2997,33 @@ int ni_write_parents(struct ntfs_inode *ni, int sync)
 	struct ATTR_LIST_ENTRY *le = NULL;
 	struct ntfs_sb_info *sbi = ni->mi.sbi;
 	struct super_block *sb = sbi->sb;
+	struct MFT_REF *homes;
+	u32 i, n = 0, cnt = 0;
 
 	if (!is_ni_base(ni))
 		return 0;
 
+	/*
+	 * Collect the parent references under ni_lock: the record may be
+	 * rearranged at the same time (a name removed, $DATA moved out of the
+	 * record), and mi_enum_attr() marks the inode bad if it walks a
+	 * half-moved attribute.  Open and write the parents without ni_lock.
+	 */
+	ni_lock(ni);
 	while ((attr = ni_find_attr(ni, attr, &le, ATTR_NAME, NULL, 0, NULL,
-				    NULL))) {
-		struct inode *dir;
+				    NULL)))
+		cnt++;
+
+	homes = kmalloc_objs(*homes, cnt, GFP_NOFS);
+	if (cnt && !homes) {
+		ni_unlock(ni);
+		return -ENOMEM;
+	}
+
+	attr = NULL;
+	le = NULL;
+	while (n < cnt && (attr = ni_find_attr(ni, attr, &le, ATTR_NAME, NULL,
+					       0, NULL, NULL))) {
 		struct ATTR_FILE_NAME *fname;
 
 		fname = resident_data_ex(attr, SIZEOF_ATTRIBUTE_FILENAME);
@@ -3012,19 +3032,26 @@ int ni_write_parents(struct ntfs_inode *ni, int sync)
 
 		/* Check simple case when parent inode equals current inode. */
 		if (ino_get(&fname->home) == ni->vfs_inode.i_ino) {
-			if (MFT_REC_ROOT != ni->vfs_inode.i_ino) {
-				ntfs_set_state(sbi, NTFS_DIRTY_ERROR);
+			if (ni->vfs_inode.i_ino != MFT_REC_ROOT)
 				err = -EINVAL;
-			}
 			continue;
 		}
 
-		dir = ntfs_iget5(sb, &fname->home, NULL);
+		homes[n++] = fname->home;
+	}
+	ni_unlock(ni);
+
+	if (err)
+		ntfs_set_state(sbi, NTFS_DIRTY_ERROR);
+
+	for (i = 0; i < n; i++) {
+		struct inode *dir = ntfs_iget5(sb, &homes[i], NULL);
+
 		if (IS_ERR(dir)) {
 			ntfs_inode_warn(
 				&ni->vfs_inode,
 				"failed to open parent directory r=%llx to write",
-				(u64)ino_get(&fname->home));
+				(u64)ino_get(&homes[i]));
 			continue;
 		}
 
@@ -3035,6 +3062,7 @@ int ni_write_parents(struct ntfs_inode *ni, int sync)
 		}
 		iput(dir);
 	}
+	kfree(homes);
 
 	return err;
 }

-- 
2.56.0