Michael Orlitzky [Tue, 11 Dec 2018 21:13:31 +0000 (16:13 -0500)]
src/libadacl.c: fix a clang-tidy warning by adding a redundant check.
The clang-tidy program is complaining about a potential null pointer
dereference where none is possible. To convince it of that fact, I've
added a redundant equality check: we have a case where (a == b) and (a
== c), and I've added (b == c) explicitly. This fixes the warning, and
should have very little performance impact, so everyone is happy.
Michael Orlitzky [Tue, 11 Dec 2018 20:23:28 +0000 (15:23 -0500)]
src/libadacl.c: rename variables in acl_update_entry() for clarity.
The updated copy of the entry in the acl_update_entry() function was
called simply "entry", which makes it hard to mentally separate from,
say, the "existing_entry" that is used in the same place. The same
problem exists with names like "entry_tag" with respect to
"existing_tag", and so on.
This commit renames "entry" to "updated_entry" and all of the other
variables like "entry_tag" to "updated_tag". That should make it a
little clearer which entry is the new one.
Michael Orlitzky [Tue, 11 Dec 2018 05:40:17 +0000 (00:40 -0500)]
src/libadacl.c: only update entries with matching qualifiers.
This fixes the bug reported by Michał Bartoszkiewicz wherein multiple
named user/group entries clobber one another. Now, when updating an
existing ACL entry with a given one, we check that their qualifiers
match. This can mean one of three things:
1. Both entries are named user entries, and their UIDs match.
2. Both entries are named group entries, and their GIDs match.
3. Both entries are neither named user or group entries, and
their qualifiers "match vacuously." That is to say: they
don't have any qualifiers that should match in the first
place, so we want to update the entry regardless.
This passes the regression test case that I recently added, and (I
hope) fixes the problem entirely.
Michael Orlitzky [Tue, 11 Dec 2018 05:16:21 +0000 (00:16 -0500)]
stc/libadacl.c: use a "cleanup" routine in acl_update_entry().
This is in preparation for comparing the qualifiers of the given and
existing ACL entries. Since acl_get_qualifier() can allocate memory,
we need to be sure that memory gets freed, even if an error occurs. A
"cleanup" routine and liberal use of "goto" is the standard pattern
throughout the rest of the library to deal with that problem.
Michael Orlitzky [Tue, 11 Dec 2018 03:12:23 +0000 (22:12 -0500)]
run-tests.sh: add regression test for bug with multiple named entities.
Michał Bartoszkiewicz just reported a very bad bug in the latest
release. When multiple named-user and named-group entries exist, the
later entries clobber earlier ones in the list. So if there are two
default ACL entries on a parent directory for group:bar and group:foo,
then apply-default-acl will create two entries on a child but both
wind up with the permissions of the group:foo entry. The full test
case he provided is as follows:
This commit adds a new regression test that creates multiple default
named-user and named-group entries at once (with different
permissions!) and checks that they get applied correctly.
Michael Orlitzky [Mon, 18 Jun 2018 01:48:40 +0000 (21:48 -0400)]
src/libadacl.c: fix handling of "./" and "../" as paths.
The recent fixes for the paths "." and ".." ignored the other obvious
cases, where those paths have a trailing slash appended. The trailing
slash is now handled by comparing the basename of the path against "."
and ".." rather than the path itself. This allows the test suite,
which now contains tests for "./" and "../", to pass.
Michael Orlitzky [Mon, 18 Jun 2018 01:28:32 +0000 (21:28 -0400)]
src/libadacl.c: simplify the "." and ".." path handling.
The initial fix for the path arguments "." and ".." was a little
hacky, but necessary to get the test suite passing. Now the logic is a
little cleaner, and both paths are handled in one special-case rather
than two separate "if" statements.
Michael Orlitzky [Mon, 18 Jun 2018 00:57:24 +0000 (20:57 -0400)]
src/libadacl.c: add a special case for the path ".." as an argument.
In a similar vein, the path ".." needs special-case handling when
opening its parent and child file descriptots. With the special-case
added, the test suite once again passes.
Michael Orlitzky [Sun, 17 Jun 2018 23:43:21 +0000 (19:43 -0400)]
src/libadacl.c: add a special case for the path "." as an argument.
There's a bug (exposed by the most recent test case) in the way the
path "." is handled. Specifically, the dirname() function miscomputes
its parent path as ".", which is clearly not correct.
In this commit, a special case is added for the path ".", and the test
suite passes once more. The implementation is a bit of a hack, however,
and will be improved once the same issue with ".." has been dealt with.
Michael Orlitzky [Thu, 29 Mar 2018 01:59:51 +0000 (21:59 -0400)]
src/libadacl.c: fix a memory leak found by clang-tidy.
There was an error path in apply_default_acl() that was returning
directly rather than jumping to the cleanup function where memory is
freed. Thanks, clang-tidy.
Michael Orlitzky [Thu, 29 Mar 2018 01:41:31 +0000 (21:41 -0400)]
src/libadacl.c: cast two fgetxattr() and fsetxattr() params to size_t.
The two functions fgetxattr() and fsetxattr() take an unsigned "size"
parameter as arguments. We are passing them signed integers that we
happen to know are nonnegative, since we have ruled out the one
possible negative value -- but the compiler doesn't know that. To
avoid a warning from clang, we now cast the parameters to the
(unsigned) size_t type.
Michael Orlitzky [Wed, 28 Mar 2018 03:50:41 +0000 (23:50 -0400)]
Drop the "--no-exec-mask" flag and function parameters.
Nobody needs the "--no-exec-mask" flag, and it's uglying up the
library's API. Sayonara:
* Update the man page:
* Remove all mentions of the flag.
* Update the algorithm description.
* Reword the general description.
* Remove all --no-exec-mask tests.
* Bump the program version in configure.ac.
* Make apply_default_acl() work as if no_exec_mask == false.
* Remove all no_exec_mask function parameters.
* Bump the soname major version in src/Makefile.am.
Michael Orlitzky [Wed, 28 Mar 2018 01:03:01 +0000 (21:03 -0400)]
src/libadacl.c: use asprintf() instead of snprintf() for paths.
When constructing a path, there is an ancient problem: how do you
ensure that your path-name buffer is large enough, and what do you do
if it isn't? The existing solution was to use the PATH_MAX constant
from limits.h, which is often a big number, but need not actually be
defined. If a path exceeded PATH_MAX bytes, we would fail.
Now the GNU/BSD extension asprintf() is used instead. The path-name
buffer is constructed on-the-fly to be as large as necessary, and if
allocation fails, an error is returned. This solution is a little
cleaner, and is not too much less portable considering that we only
work on Linux anyway.
Michael Orlitzky [Wed, 28 Mar 2018 00:23:16 +0000 (20:23 -0400)]
doc: document the apply-default-acl algorithm.
It's nice to have a high-level overview of what the ACL application
actually does, so I have added one to the man page in a new section
titled "ALGORITHM". The manual now also explains how apply-default-acl
differs from the kernel when, for example, you "touch" a file in a
directory with a default ACL.
autotools: replace my busted header checks with something that works.
My existing AC_CHECK_HEADERS checks were failing silently. Oops. I've
now defined my own macro in m4/ac_header_required.m4 that successfully
fails when a required header is missing.
Replace nftw() with manual recursion in apply_default_acl().
The nftw() tree walk worked well for a while; in particular, before we
handled symlinks safely, it was empirically faster than a hand-written
recursive descent. But recently, the very slow safe_open() function
was being called on the path that was passed to apply_default_acl(),
and nftw() fed that function a whole bunch of paths.
The apply_default_acl() function now takes a third "recusive"
parameter, and implements the recursion on its own. This lets us pass
down the old child file descriptor as the new parent file descriptor,
and avoid calling safe_open() more than once when we're operating
recursively. The result is a big speed improvement with --recursive,
tested for example on the Linux kernel source tree.
The hand-written recursion also allows us to fix a lingering exit code
bug. Now --recursive acts as if all of the targets were passed (in the
right order) on the command-line.
The new parameter affects the public API, so in the next release the
library will get a new version. The upside to this is that now it's
easy for other programs to operate recursively, simply by passing
"true" as the third parameter to apply_default_acl().
Update docs and tests for the --recursive exit code.
With the nftw() implementation, there was some bugginess in our exit
code that was both documented and tested. Well now I plan on fixing
that, so the documentation has been updated to state what the exit
code _should_ be, and the tests now check for the correct behavior
(meaning that they fail, for the moment).
The apply_default_acl_ex() function was an "extended" version of the
apply_default_acl() function that, in addition, took a stat structure
pointer to the target path. The extended function was used by nftw(),
which usually has such a stat structure handy. However, the provenance
of that stat structure is not clear: does nftw() obtain it in a safe
way?
Since I don't know the answer to that question, I would rather stat()
the descriptor that I know was obtained safely. Thus there's no reason
to take the pointer as an argument, and then no reason to keep the
extended function around at all.
src/libadacl.c: remove two NULL checks around acl_free() calls.
The acl_free() function will return ACL_ERROR and set errno to EINVAL
if we pass it a null pointer; but aside from that, nothing bad
happens, and removing the checks makes the code marginally cleaner.
src/libadacl.c: rewrite acl_set_entry() as acl_update_entry().
It turns out we only need to update existing entries in our ACLs, to
mask the execute permissions. Since we never need to create new
entries, the name "acl_set_entry" was not quite right. The
new-entry-creation code has been removed from the bottom half of the
function, and it has been renamed to "acl_update_entry".
Michael Orlitzky [Wed, 28 Feb 2018 22:33:17 +0000 (17:33 -0500)]
Eliminate the last bit of pathname usage.
A lot of work has been done recently to make apply-default-acl safe
from symlink and hardlink attacks. A big part of that work was the
recent switch to using file descriptors instead of pathnames; but,
pathnames still lingered in a few places due to a shortcoming in
libacl. Through the use of a new function, acl_copy_xattr(), I've
finally eliminated those last few bits.
The apply_default_acl_ex() function now uses path names only as
arguments to safe_open(), which hopefully is safe. Afterwards, the
file descriptors obtained from safe_open() are used. Thus the hard and
symlink attacks should finally be fixed, modulo a tiny race condition
between safe_open() and fstat() that has no known solution.
These changes rely on the Linux xattr implementation and kill our
portability, but I don't think we ever had any to begin with.
Michael Orlitzky [Mon, 26 Feb 2018 18:27:18 +0000 (13:27 -0500)]
Rename apply_default_acl() to apply_default_acl_ex() and add a wrapper.
The old apply_default_acl() function has a weird second argument that
will usually be NULL for other users of the library. Instead of making
them deal with that design choice, the old apply_default_acl()
function was renamed t apply_default_acl_ex(), and a new
apply_default_acl() was added with no second argument to wrap the
former.