When fanotify reports a permission event, copy_event_to_user() creates the event fd with dentry_open_nonotify(), so that the listener's use of that fd does not generate further events. If the event's path is on a stacked filesystem, the stacked filesystem's ->open() opens the real file with backing_file_open(). Since commit ebe559609d78 ("fs: get rid of __FMODE_NONOTIFY kludge"), FMODE_NONOTIFY lives in f_mode only and the backing file does not inherit it: it is allocated with the default fsnotify mode, and fsnotify_open_perm_and_set_mode() raises FS_OPEN_PERM on the real path if the backing filesystem has permission event watchers. Before that commit __FMODE_NONOTIFY travelled in f_flags, and overlayfs copies the user file's f_flags into the backing open (ovl_open_realfile()), so the backing file was exempt implicitly. Commit bc2473c90fca ("ovl: enable fsnotify events on underlying real files") relied on that when it dropped FMODE_NONOTIFY from OVL_OPEN_FLAGS. The nested event is queued to the same group, from inside the listener's read(2), and can only be answered by another reader thread of that group. A listener holding filesystem marks on both the stacked and the backing filesystem therefore deadlocks as soon as all of its reader threads are inside such a read(2) at the same time, and every other task that touches a marked filesystem queues up behind it. Observed with a commercial anti-malware agent (16 reader threads) on an ostree/composefs host: / is an overlayfs whose data layer is the ostree object store on the same btrfs that holds /etc and /home, and a systemd-sysext overlay sits on /usr. Every on-demand scan of /usr wedged the whole machine until the group was killed. Stack of a blocked reader thread on 7.2.7: fanotify_get_response fanotify_handle_event send_to_group fsnotify __fsnotify_parent do_dentry_open vfs_open backing_file_open ovl_open_realfile [overlay] ovl_open [overlay] do_dentry_open vfs_open backing_file_open ovl_open_realfile [overlay] ovl_open [overlay] do_dentry_open vfs_open dentry_open_nonotify copy_event_to_user fanotify_read vfs_read ksys_read Userspace cannot avoid this: the backing open happens while the kernel builds the fd, before the event is reported, and fid reporting is not available for permission event classes. Make a backing file inherit FMODE_NONOTIFY when the user file it is opened for must not generate events. This restores the pre-6.14 behaviour for the fds that fanotify creates (and for any other FMODE_NONOTIFY file), while ordinary opens through a stacked filesystem keep generating events on the real path, as intended by bc2473c90fca. Verified with a self-contained reproducer (private tmpfs lower layers, an overlayfs over them, FAN_OPEN_PERM filesystem marks on the overlay and on one lower layer, a child opening a file through the overlay) in a qemu VM on this tree built with defconfig plus OVERLAY_FS, FANOTIFY and FANOTIFY_ACCESS_PERMISSIONS: without this patch the listener's read(2) never returns and the child's open is released only by the listener's death; with it read(2) returns, the event is answered and the open completes. backing_tmpfile_open() is left alone: its only caller is ovl_create_tmpfile(), whose user file is a regular overlayfs file. Fixes: ebe559609d78 ("fs: get rid of __FMODE_NONOTIFY kludge") Cc: stable@vger.kernel.org Reviewed-by: Amir Goldstein Signed-off-by: Edvard Conrad --- v3: - Back to the v1 change in backing_file_open(). The init_file() variant posted as v2 was larger without being more general; Amir withdrew that suggestion and reviewed v1 instead. - Add Amir's Reviewed-by. Drop the Assisted-by and Suggested-by trailers and the blank line in the trailer block, as asked. - Keep the subject line from v2, which is the prefix used for fs/backing-file.c. - The reproducer posted inline with v1 is now an LTP test, under review as syscalls/fanotify26: https://lore.kernel.org/ltp/P50G-yE--F-9@tuta.io/ v2: https://lore.kernel.org/all/P4wV1J9--F-9@tuta.io/ v1: https://lore.kernel.org/all/P4sl5iz--F-9@tuta.io/ Tool disclosure (Documentation/process/generated-content.rst): the stack analysis, the git-history search for the regressing commit, the patch text, the commit message and the reproducer were produced with an AI coding assistant (Anthropic Claude Code) under my direction over interactive sessions; I reviewed, built and tested the result. No static analysis tools beyond checkpatch were used. fs/backing-file.c | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/fs/backing-file.c b/fs/backing-file.c index 5614cb7801e..24357bf9328 100644 --- a/fs/backing-file.c +++ b/fs/backing-file.c @@ -43,6 +43,17 @@ struct file *backing_file_open(const struct file *user_file, int flags, if (IS_ERR(f)) return f; + /* + * A user file that must not generate fsnotify events (e.g. the fd + * that fanotify creates for a permission event with + * dentry_open_nonotify()) must not generate them through its backing + * file either. Otherwise a listener watching the backing filesystem + * gets a nested permission event from its own read(2), which only + * another reader of the same group can answer. + */ + if (FMODE_FSNOTIFY_NONE(user_file->f_mode)) + file_set_fsnotify_mode(f, FMODE_NONOTIFY); + path_get(user_path); backing_file_set_user_path(f, user_path); error = vfs_open(real_path, f); -- 2.56.0