dbus-daemon fails to launch on macOS 10.5 and above because of a breaking
change in setrlimit, in which RLIM_INFINITY is no longer supported
for RLIMIT_NOFILE. Instead we must use OPEN_MAX.
Resolves: #309
(cherry picked from commit 691946dabc)
Otherwise, dbus doesn't compile on FreeBSD if the GLib-based tests
are enabled (which suggests that no FreeBSD user has run those tests
successfully).
We already include <netinet/in.h> in other places with no conditions
or checks other than "is Unix", so apparently it's portable enough that
specifically testing for its presence is not necessary. POSIX requires it
to exist.
Signed-off-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit f0e526bca8)
Previously, the hash table indexed by uid (or gid) took ownership of the
single reference to the heap-allocated struct, and the hash table
indexed by username (or group name) had a borrowed pointer to the same
struct that exists in the other hash table.
However, this can break down if you have two or more distinct usernames
that share a numeric identifier. This is generally a bad idea, because
the user-space model in such situations does not match the kernel-space
reality, and in particular there is no effective kernel-level security
boundary between such users, but it is sometimes done anyway.
In this case, when the second username is looked up in the userdb, it
overwrites (replaces) the entry in the hash table that is indexed by
uid, freeing the DBusUserInfo. This results in both the key and the
value in the hash table that is indexed by username becoming dangling
pointers (use-after-free), leading to undefined behaviour, which is
certainly not what we want to see when doing access control.
An equivalent situation can occur with groups, in the rare case where
a numeric group ID has two names (although I have not heard of this
being done in practice).
Solve this by reference-counting the data structure. There are up to
three references in practice: one held temporarily while the lookup
function is populating and storing it, one held by the hash table that
is indexed by uid, and one held by the hash table that is indexed by
name.
Closes: dbus#305
Signed-off-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit 2b7948ef90)
This makes it more obvious that the returned pointer points to a
struct owned by the userdb, which must not be freed or have its
contents modified, and is only valid to dereference until the next
modification to the userdb's underlying hash tables (which in practice
means until the lock is released, because after that we have no
guarantees about what might be going on in another thread).
Signed-off-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit 6ee66ff7bc)
This version is for the dbus-1.12 branch, and doesn't rely on dbus!153
or dbus!120.
Reproduces: dbus#294
Reproduces: CVE-2020-12049
Reproduces: GHSL-2020-057
Signed-off-by: Simon McVittie <smcv@collabora.com>
MSG_CTRUNC indicates that we have received fewer fds that we should
have done because the buffer was too small, but we were treating it
as though it indicated that we received *no* fds. If we received any,
we still have to make sure we close them, otherwise they will be leaked.
On the system bus, if an attacker can induce us to leak fds in this
way, that's a local denial of service via resource exhaustion.
Reported-by: Kevin Backhouse, GitHub Security Lab
Fixes: dbus#294
Fixes: CVE-2020-12049
Fixes: GHSL-2020-057
In an unprivileged container, uid 0 doesn't have CAP_SYS_RESOURCE, so
we can't expect the dbus-daemon to be able to escalate its fd limit.
This can be reproduced using bubblewrap:
sudo bwrap \
--cap-drop CAP_SYS_RESOURCE \
--ro-bind / / \
--dev /dev \
env \
DBUS_TEST_DAEMON=.../bus/dbus-daemon \
DBUS_TEST_DATA=.../test/data \
.../test/test-dbus-daemon \
-p /fd-limit \
--verbose
Bug-Debian: https://bugs.debian.org/908092
Commit 724adb2f6 mangled the dbus_message_append_args() code example.
This fixes it by breaking the lines and aligning at the right places.
Signed-off-by: Felipe Franciosi <felipe@nutanix.com>
Each connection that is an active monitor holds a pointer to its own
link in this list, via BusConnectionData.link_in_monitors. We can't
validly free the list while these pointers exist: that would be a
use-after-free, when each connection gets disconnected and tries to
remove itself from the list.
Instead, let each connection remove itself from the list, then assert
that the list has become empty.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Resolves: https://gitlab.freedesktop.org/dbus/dbus/issues/291
(cherry picked from commit b034b83b59)
Spotted by Mubin. This documentation relates to the code in
add_bus_environment() in bus/activation.c.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Fixes: #275
(cherry picked from commit bf71a58e4a)
We don't actually complete successful authentication, because that
would require us to generate a cookie and compute the correct SHA1,
which is difficult to do in a deterministic authentication script.
However, we do assert that dbus#269 (CVE-2019-12749) has been fixed.
Signed-off-by: Simon McVittie <smcv@collabora.com>
The DBUS_COOKIE_SHA1 authentication mechanism aims to prove ownership
of a shared home directory by having the server write a secret "cookie"
into a .dbus-keyrings subdirectory of the desired identity's home
directory with 0700 permissions, and having the client prove that it can
read the cookie. This never actually worked for non-malicious clients in
the case where server uid != client uid (unless the server and client
both have privileges, such as Linux CAP_DAC_OVERRIDE or traditional
Unix uid 0) because an unprivileged server would fail to write out the
cookie, and an unprivileged client would be unable to read the resulting
file owned by the server.
Additionally, since dbus 1.7.10 we have checked that ~/.dbus-keyrings
is owned by the uid of the server (a side-effect of a check added to
harden our use of XDG_RUNTIME_DIR), further ruling out successful use
by a non-malicious client with a uid differing from the server's.
Joe Vennix of Apple Information Security discovered that the
implementation of DBUS_COOKIE_SHA1 was susceptible to a symbolic link
attack: a malicious client with write access to its own home directory
could manipulate a ~/.dbus-keyrings symlink to cause the DBusServer to
read and write in unintended locations. In the worst case this could
result in the DBusServer reusing a cookie that is known to the
malicious client, and treating that cookie as evidence that a subsequent
client connection came from an attacker-chosen uid, allowing
authentication bypass.
This is mitigated by the fact that by default, the well-known system
dbus-daemon (since 2003) and the well-known session dbus-daemon (in
stable releases since dbus 1.10.0 in 2015) only accept the EXTERNAL
authentication mechanism, and as a result will reject DBUS_COOKIE_SHA1
at an early stage, before manipulating cookies. As a result, this
vulnerability only applies to:
* system or session dbus-daemons with non-standard configuration
* third-party dbus-daemon invocations such as at-spi2-core (although
in practice at-spi2-core also only accepts EXTERNAL by default)
* third-party uses of DBusServer such as the one in Upstart
Avoiding symlink attacks in a portable way is difficult, because APIs
like openat() and Linux /proc/self/fd are not universally available.
However, because DBUS_COOKIE_SHA1 already doesn't work in practice for
a non-matching uid, we can solve this vulnerability in an easier way
without regressions, by rejecting it early (before looking at
~/.dbus-keyrings) whenever the requested identity doesn't match the
identity of the process hosting the DBusServer.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Closes: https://gitlab.freedesktop.org/dbus/dbus/issues/269
Closes: CVE-2019-12749
Linux systems have traditionally set the soft limit to 1024 and the hard
limit to 4096. Recent versions of systemd keep the soft fd limit at
1024 to avoid breaking programs that still use select(), but raise the
hard limit to 512*1024, while in recent Debian versions a complicated
interaction between components gives a soft limit of 1024 and a hard
limit of 1024*1024. If we can, we might as well elevate our soft limit
to match the hard limit, minimizing the chance that we will run out of
file descriptor slots.
Unlike the previous code to raise the hard and soft limits to at least
65536, we do this even if we don't have privileges: privileges are
unnecessary to raise the soft limit up to the hard limit.
If we *do* have privileges, we also continue to raise the hard and soft
limits to at least 65536 if they weren't already that high, making
it harder to carry out a denial of service attack on the system bus on
systems that use the traditional limit (CVE-2014-7824).
As was previously the case on the system bus, we'll drop the limits back
to our initial limits before we execute a subprocess for traditional
(non-systemd) activation, if enabled.
systemd activation doesn't involve us starting subprocesses at all,
so in both cases activated services will still inherit the same limits
they did previously.
This change also fixes a bug when the hard limit is very large but
the soft limit is not, for example seen as a regression when upgrading
to systemd >= 240 (Debian #928877). In such environments, dbus-daemon
would previously have changed its fd limit to 64K soft/64K hard. Because
this hard limit is less than its original hard limit, it was unable to
restore its original hard limit as intended when carrying out traditional
activation, leaving activated subprocesses with unintended limits (while
logging a warning).
Reviewed-by: Lennart Poettering <lennart@poettering.net>
[smcv: Correct a comment based on Lennart's review, reword commit message]
Signed-off-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit 7eacbfece7)
[smcv: Mention that this also fixes Debian #928877]
The CMake config file installed by DBus will run in the context of other
projects. Consequently, changing the value of the PKG_CONFIG_DIR,
PKG_CONFIG_PATH or PKG_CONFIG_LIBDIR environment variables will affect
any further calls to pkg-config made by such projects, which can cause
problems.
A common case of this happening are pkg-config files installed in
usr/share/pkgconfig for .pc files that are architecture-independent, as
for example systemd does.
Avoid clobbering the environment variables by saving and restoring their
values. Note that for some of the variables, setting them to an empty
string is different from not setting them at all.
Signed-off-by: Clemens Lang <clemens.lang@bmw-carit.de>
(cherry picked from commit 3525cc045d)
Closes: dbus#267
MSYS2 has enough of a Unixish environment to run Autotools, but
apparently not enough of a Unixish environment to have functional
permissions.
Closes: dbus#216
(cherry picked from commit 14f46d14a0)
EAGAIN and EWOULDBLOCK are documented to possibly be numerically equal,
for instance in errno(3), and a simple logical OR check will trigger the
-Wlogical-op warning of GCC. The GCC developers consider the warning to
work as-designed in this case:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69602
Avoid such a warning by explicitly checking if the values are identical.
Fixes: https://gitlab.freedesktop.org/dbus/dbus/issues/225
Signed-off-by: David King <dking@redhat.com>
Reviewed-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit a653191342)
Switch the order of the argument checks to avoid the
-Wduplicated-branches warning.
Signed-off-by: David King <dking@redhat.com>
Reviewed-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit c0bf0d185d)
Group names in desktop files may contain all ASCII characters, except
control characters and '[' and ']'. Rather than accepting all values,
thanks to a logical operator confusion found by GCC warning
-Wlogical-op, instead explicitly reject the invalid values.
Signed-off-by: David King <dking@redhat.com>
Fixes: https://gitlab.freedesktop.org/dbus/dbus/issues/208
(cherry picked from commit 3ef9e789c1)
AX_CODE_COVERAGE recently changed the way it embedded its Makefile rules
in the output file: instead of using @CODE_COVERAGE_RULES@, users
are now meant to include aminclude_static.am.
The new AX_CODE_COVERAGE is only in the latest autoconf-archive release,
version 2019.01.06, which is inconveniently new, so bundle everything
we need for the moment.
This requires us to stop using the deprecated CODE_COVERAGE_LDFLAGS
(which we still used to support older versions of autoconf-archive)
and replace them with CODE_COVERAGE_LIBS.
Signed-off-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit 2938c2125e)
Closes: dbus#265
We want to make autoconf fail early and with a user-comprehensible
message if autoconf-archive isn't installed, rather than generating
a configure script with syntax errors, or a configure script that runs
successfully but doesn't do what we intended.
However, autoconf-archive doesn't actually guarantee not to use
AX_-prefixed shell variable names without m4_pattern_allow'ing them
(unlike Autoconf, Automake, Libtool and pkg-config, which explicitly use
m4_pattern_allow for variables with AC_, AM_, LT_ and PKG_ prefixes), so
it isn't safe to assume that they won't be used. In particular, recent
versions of AX_CHECK_GNU_MAKE appear to be using
$AX_CHECK_GNU_MAKE_HEADLINE as a shell variable.
Instead, specifically forbid the names of the finite list of macros
that we actually use.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Resolves: dbus#249
(cherry picked from commit ee09cc0acd)
This gives us a way to build on a more recent host OS if we want to.
For Gitlab-CI it's disabled by default.
Signed-off-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit 28c27349e2)
The version of gcc in trusty is too old for AddressSanitizer, which we
want to be able to start using, and Travis-CI finally supports Ubuntu
16.04 'xenial' now. This lets us remove some workarounds, but we need
to update others.
Signed-off-by: Simon McVittie <smcv@collabora.com>
(cherry picked from commit 0c553afcd9)
