On Sat, Nov 14, 2026 at 9:06 PM Chris Down <[email protected]> wrote: > > Hi Christian, > > Christian Brauner writes: > >+ /* Look through stacking filesystems at the inode with the data. */ > >+ if (d_real_inode(file->f_path.dentry)->i_sb->s_iflags & SB_I_PSEUDO) > >+ return -EINVAL; > > This looks good in general, but what the bots said about O_PATH files > here is real. With FMODE_PATH this is reached for a file that was never > opened, so on overlayfs nothing has been through ovl_open() to look up > the lowerdata of a metacopy file yet. ovl_d_real() then does the lookup > itself, and when that fails it hits the WARN(1) at its end and returns > the overlayfs dentry, so the check passes as well. > > Only fuse passes FMODE_PATH, so it takes CAP_SYS_ADMIN to get there, but > it's easy to trigger with that. From a quick test with a metacopy file > whose redirect into a data-only layer points at nothing, registering an > O_PATH fd of it with FUSE_DEV_IOC_BACKING_OPEN gives > > ovl_d_real(/f, 0): real dentry not found > WARNING: fs/overlayfs/super.c:79 at ovl_d_real+0xc9/0x210 > [...] > file_check_backing+0x308/0x5e0 > fuse_backing_new+0x219/0x4a0 > fuse_backing_open+0xe8/0x390 > fuse_dev_ioctl+0x53a/0x7d0 > > and the ioctl still succeeds. A normal open of that file fails with > ENOENT, so nothing could have used it as a backing file anyway, but with > panic_on_warn this is now a panic. > > Maybe ovl_d_real() should just stop warning when the lookup fails, now > that it has a caller that isn't trace_uprobe?
Yeh I guess that this needs to be a Fixes for lazy data lookup. The correct fix would be to teach all the callers to deal with an error. > The check would still pass > for an O_PATH file whose lowerdata couldn't be looked up at that point, > though. One observation relevant to this patch set is that a datadir layer has no business being a pseudo fs and that can and should be blocked explicitly at overlayfs mount time. The other observation is that fuse_passthrough_open() should probably check the backing_file that was opened for io in addition to checking in fuse_backing_open(), which could have different FMODE and specifically could be FMODE_PATH. Thanks, Amir.
