On Tue, Mar 18, 2025 at 08:04:59AM +0100, Ard Biesheuvel wrote: > the latter is only needed when it is mounted to begin with, and as a > VFS non-expert, I struggle to understand why it is a) ok and b) > preferred to create a new mount to pass to kernel_file_open(). Could > we add a paragraph to the commit log that explains this? I'm not at all convinced that iterate_dir() is the right thing to use there, but *IF* we go that way, yes, we need a reference to struct mount. We are not going to introduce a very special kind of struct file, along with the arseloads of checking for that crap all over the place - not for the sake of one weird case in one weird filesystem. file->f_path is a valid struct path, which means that ->mnt->mnt_sb == ->dentry->d_sb and refcount of ->mnt is positive as long as struct file exists. Keeping a persistent internal struct mount is, of course, possible, but it will make the damn thing impossible to rmmod, etc. - it will remain in place until the reboot. It might be possible to put together something like "grab a reference to superblock and allocate a temporary struct mount refering to it" (which is what that vfs_kern_mount() boils down to). But I would very much prefer to have it go over the list of children of ->s_root manually, instead of playing silly buggers with iterate_dir(). And yes, it would require exclusion with dissolving dentry tree on umount, for obvious reasons. Which might be done with ->s_active or simply by unregistering that notifier chain as the very first step in ->kill_sb() there.
