How do the UNIX commands mv and rm work with open files?

Question

If I am reading a file stored on an NTFS filesystem, and I try to move/rename that file while it is still being read, I am prevented from doing so. If I try this on a UNIX filesystem such as EXT3, it succeeds, and the process doing the reading is unaffected. I can even rm the file and reading processes are unaffected. How does this work? Could somebody explain to me why this behaviour is supported under UNIX filesystems but not NTFS? I have a vague feeling it has to do with hard links and inodes, but I would appreciate a good explanation.

Answer 1

Unix filesystems use reference counting and a two-layer architecture for finding files.

The filename refers to something called an inode, for information node or index node. The inode stores (a pointer to) the file contents as well as some metadata, such as the file's type (ordinary, directory, device, etc.) and who owns it.

Multiple filenames can refer to the same inode; they are then called hard links. In addition, a file descriptor (fd) refers to an inode. An fd is the type of object a process gets when it opens a file.

A file in a Unix filesystem only disappears when the last reference to it is gone, so when there are no more names (hard links) or fd's referencing it. So, rm does not actually remove a file; it removes a reference to a file.

This filesystem setup may seem confusing and it sometimes poses problems (esp. with NFS), but it has the benefit that locking is not necessary for a lot of applications. Many Unix programs also use the situation to their advantage by opening a temporary file and deleting it immediately after. As soon as they terminate, even if they crash, the temporary file is gone.

Answer 2

On unix, a filename is simply a link to the actual file(inode). Opening a file also creates a (temporary) link to the actual file. When all links to a file have disappeared (rm and close()) then the file is removed.

On NTFS, logically the filename is the file. There's no indirection layer from the filename to the file metainfo, they're the same object. If you open it, it's in use and can't be removed, just as the actual file(inode) on unix can't be removed while it's in use.

Unix:   Filename ➜ FileInfo ➜ File Data

NTFS: FileName + FileInfo ➜ File Data