Re: [RFC][PATCH] fanotify: allow setting FAN_CREATE in mount mark mask

On Tue, Mar 30, 2021 at 3:12 PM Christian Brauner
[off-list ref] wrote:

On Sun, Mar 28, 2021 at 06:56:24PM +0300, Amir Goldstein wrote:

quoted

Add a high level hook fsnotify_path_create() which is called from
syscall context where mount context is available, so that FAN_CREATE
event can be added to a mount mark mask.

This high level hook is called in addition to fsnotify_create(),
fsnotify_mkdir() and fsnotify_link() hooks in vfs helpers where the mount
context is not available.

In the context where fsnotify_path_create() will be called, a dentry flag
flag is set on the new dentry the suppress the FS_CREATE event in the vfs
level hooks.

Ok, just to make sure this scheme would also work for overlay-style
filesystems like ecryptfs where you possible generate two notify events:
- in the ecryptfs layer
- in the lower fs layer
at least when you set a regular inode watch.

If you set a mount watch you ideally would generate two events in both
layers too, right? But afaict that wouldn't work.

Say, someone creates a new link in ecryptfs the DENTRY_PATH_CREATE
flag will be set on the new ecryptfs dentry and so no notify event will
be generated for the ecryptfs layer again. Then ecryptfs calls
vfs_link() to create a new dentry in the lower layer. The new dentry in
the lower layer won't have DCACHE_PATH_CREATE set. Ok, that makes sense.

But since vfs_link() doesn't have access to the mnt context itself you
can't generate a notify event for the mount associated with the lower
fs. This would cause people who a FAN_MARK_MOUNT watch on that lower fs
mount to not get notified about creation events going through the
ecryptfs layer. Is that right?  Seems like this could be a problem.

Not sure I follow what the problem might be.

FAN_MARK_MOUNT subscribes to get only events that were
generated via that vfsmount - that has been that way forever.

A listener may subscribe to (say) FAN_CREATE on a certain
mount AND also also on a specific parent directory.

If the listener is watching the entire ecryptfs mount and the
specific lower directory where said vfs_link() happens, both
events will be reported. One from fsnotify_create_path() and
the lower from fsnotify_create().

If one listener is watching the ecryptfs mount and another
listener is watching the specific ecryptfs directory, both
listeners will get a single event each. They will both get
the event that is emitted from fsnotify_path_create().

Besides I am not sure about ecryptfs, but overlayfs uses
private mount clone for accessing lower layer, so by definition
users cannot watch the underlying overlayfs operations using
a mount mark. Also, overlayfs suppresses fsnotify events on
underlying files intentionally with FMODE_NONOTIFY.

Hope that answers your question.

Thanks,
Amir.