Skip to contentSkip to conversation

gfs2: mark the VFS superblock read-only before quota cleanup

PATCH In review Tao Yu 2 participants 2 messages Aug 4 – Aug 27 GFS2 file system

lore
Share & tools
lore

Tao Yu

author
#0 lore raw

PATCH 1/1 gfs2: mark the VFS superblock read-only before quota cleanup

During unmount, gfs2_put_super() calls gfs2_make_fs_ro(), which shuts down quota/statfs activity and then frees the quota bitmap via gfs2_quota_cleanup().

However, gfs2_make_fs_ro() did not mark the VFS superblock read-only before that cleanup. As a result, later inode eviction could still enter the deallocation path through gfs2_evict_inode() -> gfs2_dinode_dealloc() -> gfs2_quota_hold() -> slot_get(), even though sd_quota_bitmap had already been freed and cleared. That leads to a NULL pointer dereference in find_first_zero_bit().

Set SB_RDONLY after the final quota/statfs sync and log shutdown, but before gfs2_quota_cleanup(). This keeps subsequent inode eviction from re-entering write-side quota/deallocation paths once quota state has been torn down.

This fixes a crash reported by syzbot:

  Oops: general protection fault in _find_first_zero_bit
  KASAN: null-ptr-deref in slot_get()
---
 fs/gfs2/super.c | 8 ++++++++
 1 file changed, 8 insertions(+)
fs/gfs2/super.c +8 -0
index 4d854556b5299..6b5457f965673 100644
@@ -565,6 +565,14 @@ void gfs2_make_fs_ro(struct gfs2_sbd *sdp)
 				   HZ * 5);
 		gfs2_assert_warn(sdp, gfs2_log_is_empty(sdp));
 	}
+
+	/*
+	 * Once the final quota/statfs sync and log shutdown are complete, the VFS
+	 * must see the filesystem as read-only so later inode eviction cannot
+	 * re-enter deallocation paths that still expect live quota state.
+	 */
+	sdp->sd_vfs->s_flags |= SB_RDONLY;
+
 	gfs2_quota_cleanup(sdp);
 }
 
Signature · 1 line
2.34.1
#1 lore raw
Quoted 48 lines · Tao Yu

During unmount, gfs2_put_super() calls gfs2_make_fs_ro(), which shuts down quota/statfs activity and then frees the quota bitmap via gfs2_quota_cleanup().

However, gfs2_make_fs_ro() did not mark the VFS superblock read-only before that cleanup. As a result, later inode eviction could still enter the deallocation path through gfs2_evict_inode() -> gfs2_dinode_dealloc() -> gfs2_quota_hold() -> slot_get(), even though sd_quota_bitmap had already been freed and cleared. That leads to a NULL pointer dereference in find_first_zero_bit().

Set SB_RDONLY after the final quota/statfs sync and log shutdown, but before gfs2_quota_cleanup(). This keeps subsequent inode eviction from re-entering write-side quota/deallocation paths once quota state has been torn down.

This fixes a crash reported by syzbot:

  Oops: general protection fault in _find_first_zero_bit
  KASAN: null-ptr-deref in slot_get()
---
 fs/gfs2/super.c | 8 ++++++++
 1 file changed, 8 insertions(+)
fs/gfs2/super.c +8 -1
index 4d854556b5299..6b5457f965673 100644
@@ -565,6 +565,14 @@ void gfs2_make_fs_ro(struct gfs2_sbd *sdp)
                                   HZ * 5);
                gfs2_assert_warn(sdp, gfs2_log_is_empty(sdp));
        }
+
+       /*
+        * Once the final quota/statfs sync and log shutdown are complete, the VFS
+        * must see the filesystem as read-only so later inode eviction cannot
+        * re-enter deallocation paths that still expect live quota state.
+        */
+       sdp->sd_vfs->s_flags |= SB_RDONLY;
+
        gfs2_quota_cleanup(sdp);
 }
 
--
Quoted 2 lines

2.34.1

Thanks for the patch but no, that's not it. Instead, I've added this:

https://lore.kernel.org/gfs2/[email protected]

Andreas