In particular, if libbdus is available, we test interoperability with
a libdbus client: see GNOME/glib#1831. Because that issue describes a
race condition, we do each test repeatedly to try to hit the failing
case.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Conceptually, a D-Bus server is really trying to determine the credentials
of (the process that initiated) a connection, not the credentials that
the process had when it sent a particular message. Ideally, it does
this with a getsockopt()-style API that queries the credentials of the
connection's initiator without requiring any particular cooperation from
that process, avoiding a class of possible failures.
The leading '\0' in the D-Bus protocol is primarily a workaround
for platforms where the message-based credentials-passing API is
strictly better than the getsockopt()-style API (for example, on
FreeBSD, SCM_CREDS includes a process ID but getpeereid() does not),
or where the getsockopt()-style API does not exist at all. As a result
libdbus, the reference implementation of D-Bus, does not implement
Linux SCM_CREDENTIALS at all - it has no reason to do so, because the
SO_PEERCRED socket option is equally informative.
This change makes GDBusServer on Linux more closely match the behaviour
of libdbus.
In particular, GNOME/glib#1831 indicates that when a libdbus client
connects to a GDBus server, recvmsg() sometimes yields a SCM_CREDENTIALS
message with cmsg_data={pid=0, uid=65534, gid=65534}. I think this is
most likely a race condition in the early steps to connect:
client server
connect
accept
send '\0' <- race -> set SO_PASSCRED = 1
receive '\0'
If the server wins the race:
client server
connect
accept
set SO_PASSCRED = 1
send '\0'
receive '\0'
then everything is fine. However, if the client wins the race:
client server
connect
accept
send '\0'
set SO_PASSCRED = 1
receive '\0'
then the kernel does not record credentials for the message containing
'\0' (because SO_PASSCRED was 0 at the time). However, by the time the
server receives the message, the kernel knows that credentials are
desired. I would have expected the kernel to omit the credentials header
in this case, but it seems that instead, it synthesizes a credentials
structure with a dummy process ID 0, a dummy uid derived from
/proc/sys/kernel/overflowuid and a dummy gid derived from
/proc/sys/kernel/overflowgid.
In an unconfigured GDBusServer, hitting this race condition results in
falling back to DBUS_COOKIE_SHA1 authentication, which in practice usually
succeeds in authenticating the peer's uid. However, we encourage AF_UNIX
servers on Unix platforms to allow only EXTERNAL authentication as a
security-hardening measure, because DBUS_COOKIE_SHA1 relies on a series
of assumptions including a cryptographically strong PRNG and a shared
home directory with no write access by others, which are not necessarily
true for all operating systems and users. EXTERNAL authentication will
fail if the server cannot determine the client's credentials.
In particular, this caused a regression when CVE-2019-14822 was fixed
in ibus, which appears to be resolved by this commit. Qt clients
(which use libdbus) intermittently fail to connect to an ibus server
(which uses GDBusServer), because ibus no longer allows DBUS_COOKIE_SHA1
authentication or non-matching uids.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Closes: https://gitlab.gnome.org/GNOME/glib/issues/1831
On Linux, if getsockopt SO_PEERCRED is used on a TCP socket, one
might expect it to fail with an appropriate error like ENOTSUP or
EPROTONOSUPPORT. However, it appears that in fact it succeeds, but
yields a credentials structure with pid 0, uid -1 and gid -1. These
are not real process, user and group IDs that can be allocated to a
real process (pid 0 needs to be reserved to give kill(0) its documented
special semantics, and similarly uid and gid -1 need to be reserved for
setresuid() and setresgid()) so it is not meaningful to signal them to
high-level API users.
An API user with Linux-specific knowledge can still inspect these fields
via g_credentials_get_native() if desired.
Similarly, if SO_PASSCRED is used to receive a SCM_CREDENTIALS message
on a receiving Unix socket, but the sending socket had not enabled
SO_PASSCRED at the time that the message was sent, it is possible
for it to succeed but yield a credentials structure with pid 0, uid
/proc/sys/kernel/overflowuid and gid /proc/sys/kernel/overflowgid. Even
if we were to read those pseudo-files, we cannot distinguish between
the overflow IDs and a real process that legitimately has the same IDs
(typically they are set to 'nobody' and 'nogroup', which can be used
by a real process), so we detect this situation by noticing that
pid == 0, and to save syscalls we do not read the overflow IDs from
/proc at all.
This results in a small API change: g_credentials_is_same_user() now
returns FALSE if we compare two credentials structures that are both
invalid. This seems like reasonable, conservative behaviour: if we cannot
prove that they are the same user, we should assume they are not.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Previously we were keeping a pointer to the `GFileMonitor` in a
`GFileMonitorSource` instance, but since we weren’t keeping a strong
reference, that `GFileMonitor` instance could be finalised from another
thread at any point while the source was referring to it. Not good.
Use a weak reference, and upgrade it to a strong reference whenever the
`GFileMonitorSource` is referring to the file monitor.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Helps: #1903
It’s not enough to unref the monitor, since the GLib worker thread might
still hold a reference to it.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Helps: #1903
`DesktopFileDir` pointers are passed around between threads: they are
initially created on the main thread, but a pointer to them is passed to
the GLib worker thread in the file monitor callback
(`desktop_file_dir_changed()`).
Accordingly, the `DesktopFileDir` objects either have to be
(1) immutable;
(2) reference counted; or
(3) synchronised between the two threads
to avoid one of them being used by one thread after being freed on
another. Option (1) changed with commit 99bc33b6 and is no longer an
option. Option (3) would mean blocking the main thread on the worker
thread, which would be hard to achieve and is against the point of
having a worker thread. So that leaves option (2), which is implemented
here.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Fixes: #1903
When the _g_dbus_worker_flush_sync() schedules the 'data' and releases
the worker->write_lock, it is possible for the GDBus worker thread thread
to finish the D-Bus call and acquire the worker->write_lock before
the _g_dbus_worker_flush_sync() re-acquires it in the if (data != NULL) body.
When that happens, the ostream_flush_cb() increases the worker->write_num_messages_flushed
and then releases the worker->write_lock. The write lock is reacquired by
the _g_dbus_worker_flush_sync(), which sees that the while condition is satisfied,
thus it doesn't enter the loop body and immediately clears the data members and
frees the data structure itself. The ostream_flush_cb() is still ongoing, possibly
inside flush_data_list_complete(), where it accesses the FlushData, which can be
in any stage of being freed.
Instead, add an explicit boolean flag indicating when the flush is truly finished.
Closes#1896
`-1` isn’t a valid member of the enum, so cast to `int` first. This
fixes a compiler warning on Android.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Mounts are currently completed only if the prefix looks like scheme,
however, this doesn't work well if the mounts have also path component.
Let's always include them to fix this issue. The mounts are cached by the
volume monitors, so it should not significantly affect the performance.
Currently, "gio mount google-drive<tab>" isn't completed even though
that volume exists for google-drive://oholy@redhat.com/. Let's use
"gio mount -li" output to complete also activation roots of volumes.
Currently, "gio list file:///h<tab>" doesn't complete "file:///home"
because the result of "dirname file:///h" is not "file:///" but "file:/",
which breaks the consequent logic. Let's subtract basename from the
path in order to workaround this issue.
Fixes build failure:
../gio/gunixmounts.c: In function ‘_g_get_unix_mounts’:
../gio/gunixmounts.c:742:53: error: ‘struct mnttab’ has no member named ‘mnt_opts’; did you mean ‘mnt_mntopts’?
742 | mntent.mnt_opts,
| ^~~~~~~~
| mnt_mntopts
Signed-off-by: Alan Coopersmith <alan.coopersmith@oracle.com>
g_date_time_add_seconds() and g_date_time_add_full() use floating-point
seconds, which can result in the value varying slightly from what's
actually on disk. This causes intermittent test failures in
gio/tests/g-file-info.c on Debian i386, where we set a file's mtime
to be 50µs later, then read it back and sometimes find that it is only
49µs later than the previous value.
I've only seen this happen on i386, which means it might be to do with
different floating-point rounding when a value is stored in the 80-bit
legacy floating point registers rather than in double precision.
g_date_time_add() takes a GTimeSpan, which is in microseconds;
conveniently, that's exactly what we get from the GFileInfo.
Bug-Debian: https://bugs.debian.org/941547
Signed-off-by: Simon McVittie <smcv@collabora.com>
If we're cross-compiling, the installed-tests are useful even if we
can't run them on the build machine: we can copy them to the host
machine (possibly via a distro package like Debian's libglib2.0-tests)
and run them there.
While I'm changing the build-tests condition anyway, deduplicate it.
Based on a patch by Helmut Grohne.
Bug-Debian: https://bugs.debian.org/941509
Signed-off-by: Simon McVittie <smcv@collabora.com>
Skip it on systems which don’t support it, rather than compiling it out.
That gives us more information from test runs about which tests are
being run on which architectures.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
As with the previous commit, `st_mode` contains both the file type
(regular file, directory, symlink, special, etc.) and the file mode. For
`G_FILE_ATTRIBUTE_ID_UNIX_MODE`, we only want the file mode — so mask
`st_mode` with `~S_IFMT`.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
chmod() technically only accepts file modes, not the file type and mode
as returned by stat(). Filter by `S_IFMT` to avoid sending the file
type (regular file, directory, symbolic link, etc.).
In practice, chmod() ignores anything except the file mode, but we might
as well comply with the specification.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
`GFile` always checks whether these vfuncs are `NULL` before calling
them, so document that it’s safe for implementations of `GFile` to not
implement them.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
The caller assumes that an unimplemented vfunc means that copying is
unsupported (and falls back to its internal copy implementation), so
there’s no point in implementing the vfunc just to unconditionally
return `G_IO_ERROR_NOT_SUPPORTED`.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Rather than defining a vfunc which only ever returns
`G_IO_ERROR_NOT_SUPPORTED`, just don’t define the vfunc at all. The
caller in `GFile` interprets this as symlinks not being supported — so
we get the same behaviour, but without spending a vfunc call on it.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
The string is already translated in `GLocalFile`, so this doesn’t
introduce a new translatable string.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
This sets the `G_FILE_COPY_DEFAULT_PERMS` flag on the operation,
creating the copied file with default permissions rather than the same
permissions as the source file.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Fixes: #174