Hello,
If "l" is a symlink to a directory, rm -r l/ recursively removes the contents
of the target directory and only then fails, because removing the entry "l/"
itself is rejected with ENOTDIR. With -f the failure is silent and the exit
status is 0. I first noticed this on MSYS2 (Windows) and then reproduced it
with GNU coreutils 9.4 on Linux:
```
$ rm --version | head -1
rm (GNU coreutils) 9.4
$ mkdir d
$ touch d/f
$ ln -s d l
$ rm l/
rm: cannot remove 'l/': Is a directory
$ ls d
f
$ rm -r l/
rm: cannot remove 'l/': Not a directory
$ ls -A d
$ # d is now empty; both d and l still exist
$ touch d/f
$ rm -rf l/; echo $?
0
$ ls -A d
$ # d/f is gone, with no diagnostic at all
```
I understand that a trailing slash makes "l/" resolve to the directory, and
that this was discussed in bug#12339. I am not asking to change how "l/"
resolves. My concern is narrower:
1. The contents are destroyed before any diagnostic is issued, and with -f
there is never one. The end state (contents gone, directory kept, link kept,
error reported) is neither "remove the link" nor "remove the directory".
Removing a directory's contents through a slash-terminated symlink is almost
never what the user intended, and it is increasingly easy to hit with scripts
and automatically generated commands that append "/" to directory paths.
2. The diagnostics contradict each other: rm l/ says "Is a directory" while rm
-r l/ says "Not a directory". This was already pointed out on this list in 2003
("Bug: Symlink removal").
Proposal: when recursive removal is requested and an operand ends in a slash
while its last component (after stripping the slashes) is a symbolic link,
refuse that operand before descending into it, with a diagnostic such as
```
rm: refusing to remove 'l/': it is a symbolic link to a directory;
use 'rm l' to remove the link
```
and a non-zero exit status, regardless of -f. This would be similar in spirit
to the --preserve-root safeguard. If an opt-in or opt-out option is preferred,
that would work for me as well.
Would such a change be acceptable? If so, I can try to prepare a patch with a
test.
Thanks,
Bayes