I could understand what you are trying to do from the description, but then I completely confused by the patch. FreeBSD st_dev is mnt_stat.f_fsid, then linuxolator st_dev is lower 16bits of it. Now read the comment before vfs_getnewfsid().

In other words, I think it worked for you just because you only had less than 255 mounts total during the system life.

Hm, you're right. How about adding a reserved fsid (just pick a value and make vfs_getnewfsid() never use it), and using that for all S_ISCHR vnodes instead?

In D25559#565342, @trasz wrote:

Hm, you're right. How about adding a reserved fsid (just pick a value and make vfs_getnewfsid() never use it), and using that for all S_ISCHR vnodes instead?

Better find somehow the canonical /dev mount and use its fsid (if found).

Relying on ISCHR/ISBLK is too broad. For instance, you can have character device on UFS of ext2fs or nfs. They are not interpreted as special vnode by FreeBSD kernel, but might have some meaning for linux userspace.

Instead of this, I suggest you to check the type of the backing filesystem in linux stat callback (you still have the vnode referenced there), and if it is from devfs mount, replace st_dev as you do.

You are still translate based on the vnode type instead of type of the owning filesystem.

True. This also makes Linux stat(2) report different st_dev on devfs subdirectories and devfs device nodes.

What's the suggested way to determine if vnode belongs to devfs? Compare vp->v_mount->mnt_stat.f_fstypename with "devfs", and if matches cache vp->v_mount->mnt_stat.f_type to avoid string comparisons later on?

You can do something like vp->v_mount->mnt_vfc == rootdevmp->v_mount->mnt_vfc. In the case when you do not own the vnode lock, it is important to fetch vp->v_mount into a local variable and test it for NULL before dereferencing.

(Tinderboxed.)