Eric Blake wrote:
> On 02/11/2012 04:23 AM, Jim Meyering wrote:
>> +++ b/NEWS
...
>> + "mv A B" could succeed, yet A would remain. This would happen only when
>> + both A and B were hard links to the same symlink, and with a kernel for
>> + which rename("A","B") would do nothing and return 0. Now, in this
>> + unusual case, mv does not call rename, and instead simply unlinks A.
>
> You make it sound like a kernel where rename("A","B") returns 0 is
> unusual;
Thank you for the review and suggestions.
Such kernels *should* be unusual. This rename-is-sometimes-a-no-op
exception makes it hard to use rename in an application that must
reliably produce results that make sense even to people who don't
care what inodes and invariants are.
> on the contrary, that is normal, since it is the POSIX-mandated
> behavior for kernels. What is unusual is having two hardlinks to the
> same symlink. Maybe we should reword this slightly, to attach the
> "unusual" modifier to the correct phrase, or even take "kernel" out of
> the description:
>
> "mv A B" could succeed, yet A would remain. This would only happen
> in the unusual case when both A and B were hard links to the same
> symlink, due to the standard behavior of rename. Now, mv recognizes
> the case and simply unlinks A.
This is the NEWS file, where we prefer to stick to the facts, but I
feel I have to make a small statement, so have adjusted it like this:
"mv A B" could succeed, yet A would remain. This would happen only when
both A and B were hard links to the same symlink, and with a kernel for
which rename("A","B") does nothing and returns 0 (POSIX mandates this
surprising rename no-op behavior). Now, mv handles this case by skipping
the usually-useless rename and simply unlinking A.
>> +++ b/tests/mv/symlink-onto-hardlink-to-self
>> @@ -0,0 +1,56 @@
>> +#!/bin/sh
>> +# Demonstrate that when moving a symlink onto a hardlink-to-that-symlink,
>> the
>> +# source symlink is removed. Depending on your kernel (e.g., with the linux
>> +# kernel), prior to coreutils-8.16, the mv would successfully perform a
>> no-op.
>
> Again, this is the POSIX-required behavior of ALL kernels, and not
> something special to Linux.
NetBSD 5.1 has the sensible kernel rename behavior, i.e.,
what one would expect in the absence of standardized legacy:
netbsd$ : > f; ln f g; perl -e 'rename qw(f g) or die "$!"'; ls f g
ls: cannot access f: No such file or directory
g
[Exit 2]
Programs like mv should not have to jump through hoops like copy.c's
same_file_ok function just to avoid the surprising (nonsensical, to most)
behavior of the standardized rename syscall.
I adjusted that comment, too, and pushed the result:
# Demonstrate that when moving a symlink onto a hardlink-to-that-symlink, the
# source symlink is removed. Depending on your kernel (e.g., Linux, Solaris,
# but not NetBSD), prior to coreutils-8.16, the mv would successfully perform
# a no-op. I.e., surprisingly, mv s1 s2 would succeed, yet fail to remove s1.
