On Wed, Sep 21, 2022 at 10:26:12AM +0200, Christian Brauner wrote: > - /* From now path refers to the tmpfile */ > + > + /* prepare tmp path */ > + path.mnt = cache->mnt; > path.dentry = file->f_path.dentry; Do we even want that struct path from that point on? Look: d_backing_inode(path.dentry) is a weird way to spell file_inode(file). cachefiles_mark_inode_in_use() is an overkill here - it *can't* fail here, so all we want is inode_lock(inode); inode->i_flags |= S_KERNEL_FILE; trace_cachefiles_mark_active(object, inode); inode_unlock(inode); where inode is, again, file_inode(file). cachefiles_do_unmark_inode_in_use() uses only inode. vfs_truncate() could use &file->f_path, but there's a potentially nastier problem - theoretically, there are filesystems where we might want struct file available for operation, especially for opened-and-unlinked equivalents. In any case, &file->f_path would do just as well as its copy.
