This gives us feature parity with the Autotools build system for this
particular area, and in particular means a system dbus-daemon built
with cmake can expand its fd limit.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=105165
Startup ordering was changed in #92832 to ensure that SELinux audit
messages could be sent. As a side effect, the raising of file descriptor
limits was moved to after the dropping of root privileges, resulting in
the limit change always failing.
Move the raise_file_descriptor_limit() call to ensure that it is called
before dropping root privileges.
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=105165
Bug-RedHat: https://bugzilla.redhat.com/show_bug.cgi?id=1529044
[smcv: Call raise_file_descriptor_limit() even if !context->user]
Reviewed-by: Simon McVittie <smcv@collabora.com>
In the bus daemon, don't pass through the container instance path:
if there's any value here at all, we want to be able to guarantee that
we sent it (in a later commit).
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=101899
We'll need this if we want to stamp optional header fields on the
message according to the preferences of the recipient(s).
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=101899
Based on code contributed by Manish Narang. This is not included in the
automated test suite, because it isn't reliable on heavily-loaded
automatic test infrastructure like Travis-CI.
Reviewed-by: Philip Withnall <withnall@endlessm.com>
[smcv: Add the test to the CMake build system too, as requested]
[smcv: Convert into a manual test]
Signed-off-by: Simon McVittie <smcv@collabora.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=102839
DBusLoop isn't thread-safe, so we can't use it to test multi-threaded
situations.
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=102839
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
If one thread is blocking on a pending call, and another thread is
dispatching the connection, then we need them to agree on the value
of the completed flag by protecting all accesses with a lock. Reads
for this member seem to have the connection lock already, so it's
sufficient to make sure that the only write also happens under the
connection lock.
We already set the completed flag before calling the callback, so it
seems OK to stretch it to meaning that some thread has merely *taken
responsibility for* calling the callback.
The completed flag shares a bitfield with timeout_added, but that
flag is protected by the connection lock already.
Based on suggestions from Simon McVittie on
<https://bugs.freedesktop.org/show_bug.cgi?id=102839>.
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=102839
[smcv: Revert indentation changes; add commit message]
Reviewed-by: Simon McVittie <smcv@collabora.com>
If a pending call is provided, _dbus_connection_do_iteration_unlocked
checks whether it has completed or has a reply ready as soon as it
acquires the I/O path. If that's the case, then the iteration
terminates without trying to carry out I/O, so that the pending call
can be dispatched immediately, without blocking until a timeout is
reached. This change is believed to be necessary, but not sufficient,
to resolve#102839.
Based on part of a patch from Michael Searle on
<https://bugs.freedesktop.org/show_bug.cgi?id=102839>.
Commit message added by Simon McVittie.
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=102839
Reviewed-by: Simon McVittie <smcv@collabora.com>
This file is the Unix counterpart of dbus-spawn-win.c, so it's less
confusing for it to have an indicative name.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104588
Otherwise the pre-commit hook won't let me rename it.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104588
dbus-spawn.c and dbus-userdb* don't have obviously-Unix-specific names,
but are Unix-specific anyway.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104588
Previously, USERID_HEX and USERNAME_HEX were both replaced by the hex
encoding of the numeric uid, something like 31303030 for "1000".
Now USERNAME_HEX is something like 736d6376 for "smcv". This is only
supported on Unix, but no authentication mechanisms use usernames on
Windows anyway.
This would require changing the tests that make use of USERNAME_HEX
if we had any, but we currently don't.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104588
In the well-known system dbus-daemon, it's desirable to avoid looking
up non-numeric authorization identities in the user database, because
that could deadlock with NSS modules that directly or indirectly
require the system bus. Add a flag for whether the username will be
looked up in the userdb, and don't set that flag for EXTERNAL auth
(which is what we use on the system bus, and on the session bus
if not configured otherwise).
DBUS_COOKIE_SHA1 authentication is documented in terms of the
username (although in fact libdbus sends a numeric uid there too,
and GDBus only accepts a numeric uid) so continue to use the userdb
for that mechanism. DBUS_COOKIE_SHA1 needs to use the userdb on Unix
anyway, otherwise it won't find the user's home directory.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104588
While I'm changing its signature anyway, I might as well fix a
long-standing FIXME.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104588
The very common case for this function is that during AUTH EXTERNAL,
it receives a Unix uid encoded as an ASCII decimal integer. There is
no need to look up such uids in the system's user database
(/etc/password or NSS) when the only information we are going to use
from the DBusUserInfo struct is the uid anyway. This avoids taking
the lock and performing a potentially time-consuming NSS lookup.
This changes behaviour in one corner case: if a privileged process has
used one of the set*uid family of functions to set its effective uid
to a numeric uid that does not exist in the system's user database,
we would previously fail. Now, we succeed anyway: it is true to say
in the DBusCredentials that the process has uid 12345, even if uid
12345 does not correspond to any named user.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104588
This provides the necessary information for services to make an
informed decision about how far they should trust the container type,
name and metadata fields.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104610
On the session bus, the container type and name might be
uncontroversial, but on the system bus, it's questionable how far
they can be trusted: they're supplied by the initiator of the
per-container server, so we only have their word for it. While we
think about what to do about this, remove them, leaving only the
instance (which can be used to look up the rest).
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104610
On the session bus, the container type and name might be
uncontroversial, but on the system bus, it's questionable how far
they can be trusted: they're supplied by the initiator of the
per-container server, so we only have their word for it.
Signed-off-by: Simon McVittie <smcv@collabora.com>
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=104610
Signed-off-by: Simon McVittie <smcv@collabora.com>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=100317
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Signed-off-by: Simon McVittie <smcv@collabora.com>
This is an interpretation of the existing text. There are two plausible
ways a relaying server could interpret "must ignore [new] fields":
it could pass them through as-is, or it could delete them before
relaying. Until now, the reference implementation has done the former.
However, this behaviour is difficult to defend. If a server relays
messages without filtering out header fields that it doesn't
understand, then a client can't know whether the header field was
supplied by the server, or whether it was supplied by a (possibly
malicious) fellow client.
We can't introduce useful round-trip-reducing header fields like
SENDER_UNIX_USER_ID or SENDER_LINUX_SECURITY_LABEL until the
message bus filters them out, *and* provides a way for clients to
know for sure that it has done so. This is a step towards that
feature.
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=100317
Reviewed-by: Philip Withnall <withnall@endlessm.com>
Signed-off-by: Simon McVittie <smcv@collabora.com>
