If we don't trap EEXIST and its Windows equivalent, we are unable to
detect the situation where we create an ostensibly unique
subdirectory in a shared /tmp, but an attacker has already created it.
This affects dbus-nonce (the nonce-tcp transport) and the activation
reload test.

Add a new _dbus_ensure_directory() for the one case where we want it to
succeed even on EEXIST: the DBUS_COOKIE_SHA1 keyring, which we know
we are creating in our own trusted "official" $HOME. In the new
transient service support on Bug #99825, ensure_owned_directory()
would need the same treatment.

We are not treating this as a serious security problem, because the
nonce-tcp transport is rarely enabled on Unix and there are multiple
mitigations.

The nonce-tcp transport creates a new unique file with O_EXCL and 0600
(private to user) permissions, then overwrites the requested filename
via atomic-overwrite, so the worst that could happen there is that an
attacker could place a symbolic link matching the name of a directory
we are going to create, causing a dbus-daemon configured for nonce-tcp
to traverse the symlink and atomically overwrite a file named "nonce"
in a directory of the attacker's choice, with new random contents that
are not known to the attacker. This seems unlikely to be exploitable
for anything worse than denial of service in practice. In mainline
Linux since 3.6, this attack is also defeated by the
fs.protected_symlinks sysctl, which many distributions enable by default.

The activation reload test suffers from a classic symlink attack
due to time-of-check/time-of-use errors in its implementation, but as
part of the developer-only "embedded tests" that are only intended
to be run on a trusted machine, it is not treated as security-sensitive.
That code path will be fixed in a subsequent commit.

Bug: https://bugs.freedesktop.org/show_bug.cgi?id=99828


Signed-off-by: Simon McVittie <simon.mcvittie@collabora.co.uk>
Reviewed-by: Philip Withnall <withnall@endlessm.com>