Issue 104 cannot exec setuids #106
Conversation
Previously, misinterpreted that os.access(file, X_OK) always returns True on Solaris. Yes, but only for the uid of 0. Python issue #13706 closed "not a bug" reads to "just use os.stat()", so we went to great lengths to do so quite exhaustively. But this is wrong -- *only* when root, should we check the file modes -- os.access of X_OK works perfectly fine for non-root users. And, we should only check if any of the executable bits are set. Alas, it is true, you may execute that which you may not read -- because as root, you can always read it anyway. Verified similar solution in NetBSD test.c (/bin/test), OpenBSD ksh for its built-in test, and what FreeBSD/Darwin for their implementation of which.c.
'isroot' check.
|
From what I've read recently, it's possible to have an executable but genuinely not readable binary. |
| ``````````` | ||
| * Fix regression that prevented executable, but unreadable files from | ||
| being found when not specified by absolute path -- such as | ||
| /usr/bin/sudo (:ghissue:`104`). |
There was a problem hiding this comment.
Unfortunately, this conflicts with PR #109, since I merged that. We may want to keep personal scratch files of the changelog, and merge them before release. For IPython, we keep a directory that PRs add snippet files into, and they're merged into release notes by a script, but that may be overkill for pexpect.
There was a problem hiding this comment.
Sorry for long delay, I should be able to address this.
I was aware of generating such conflict but not sure how to address -- as though we needed a header prior to commits, or some base 3.4-prepare branch to base both changes from.
|
+1 for thorough tests. Note that this needs a rebase because of the conflict in the change log. |
…ec-setuids Conflicts: doc/history.rst
|
oi! lost track of this old thing, merging and cleaning up now. |
…ec-setuids Conflicts: pexpect/__init__.py
|
This is ready for review & merge @takluyver, thanks. |
|
Thanks, looks good to me. |
Issue 104 cannot exec setuids
Previously, misinterpreted that os.access(file, X_OK)
always returns True on Solaris. Yes, but only for
the uid of 0. Python issue #13706 closed "not a bug"
reads to "just use os.stat()", so we went to great
lengths to do so quite exhaustively.
But this is wrong -- only when root, should we check
the file modes -- os.access of X_OK works perfectly
fine for non-root users.
And, we should only check if any of the executable bits
are set. Alas, it is true, you may execute that which
you may not read -- because as root, you can always read
it anyway.
Verified similar solution in NetBSD test.c (/bin/test),
OpenBSD ksh for its built-in test, and what FreeBSD/Darwin
for their implementation of which.c.
Tested on OSX and Solaris, travis on Linux, of course.
Manually tested pexpect.which('sudo') on OSX and Solaris,
especially on Solaris as root and user nobody.