Simon McVittie [Wed, 22 Jun 2011 10:59:32 +0000 (11:59 +0100)]
_dbus_connect_tcp_socket_with_nonce: don't create an extra fd (which is then leaked)
This block should have been deleted in 2007, when IPv6 support was added:
previously, the fd allocated at the beginning of the function was used
for connect(), but for IPv6 support, the socket() call has to be inside
the loop over getaddrinfo() results so its address family can be changed.
Do not allow eavedropping unless rule owner explicitely declare it
Adds "eavesdrop=true" as a match rule, meaning that the owner
intend to eavedrop.
Otherwise the owner will receive only broadcasted messages and the ones
meant to be delivered to it.
[plus a typo fix in an error message -smcv]
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=37890
Bug-NB: NB#269748 Reviewed-by: Simon McVittie <simon.mcvittie@collabora.co.uk>
Introduce static generic _dbus_connection_register_object_path()
function to remove duplicate code for registration object paths.
Having *four* almost the same functions is error-prone and hard
to follow as well.
Signed-off-by: Jiří Klimeš <jklimes@redhat.com> Reviewed-by: Simon McVittie <simon.mcvittie@collabora.co.uk>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=38874
Jiří Klimeš [Fri, 1 Jul 2011 10:58:11 +0000 (12:58 +0200)]
DBusConnection: use DBUS_ERROR_OBJECT_PATH_IN_USE instead of DBUS_ERROR_ADDRESS_IN_USE
While registering an object path the possible error is
DBUS_ERROR_OBJECT_PATH_IN_USE, not DBUS_ERROR_ADDRESS_IN_USE (the error
is set by _dbus_object_tree_register()).
Signed-off-by: Jiří Klimeš <jklimes@redhat.com> Reviewed-by: Simon McVittie <simon.mcvittie@collabora.co.uk>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=38874
Simon McVittie [Wed, 22 Jun 2011 13:57:56 +0000 (14:57 +0100)]
Remove maximum length field from DBusString
The source code says it's "a historical artifact from a feature that
turned out to be dumb". Respond accordingly!
This reduces sizeof (DBusString) by 20% on ILP32 architectures, which
can't hurt. (No reduction on LP64 architectures that align pointers
naturally, unfortunately.)
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=38570 Reviewed-by: Will Thompson <will.thompson@collabora.co.uk>
Simon McVittie [Thu, 7 Apr 2011 16:36:54 +0000 (17:36 +0100)]
DBusLoop: store watches in a hash table keyed by fd
This means we don't need to build up the watches_for_fds array, because
it's quick to find the watches based on the fd.
This also fixes fd.o #23194, because we only put each fd in the
array of DBusPollFD once, with the union of all its watches' flags; we
can no longer get more entries in the array than our number of file
descriptors, which is capped at our RLIMIT_NOFILE, so we can't get
EINVAL from poll(2).
Simon McVittie [Mon, 24 Jan 2011 14:38:13 +0000 (14:38 +0000)]
DBusLoop: inline add_callback, remove_callback into their callers
The watch and timeout code paths will diverge completely when we change
WatchCallback * to just be a DBusWatch *, removing the benefit of having
common code.
Simon McVittie [Tue, 25 Jan 2011 12:37:01 +0000 (12:37 +0000)]
DBusLoop: move OOM flag in watches inside the DBusWatch
This will eventually let us maintain a DBusPollFD[] of just the active
watches, between several iterations. The more immediate benefit is that
WatchCallback can go away, because it only contains a refcount, a
now-useless type, and a watch that already has its own refcount.
This doesn't take any more memory for DBusWatch when not using DBusLoop
(e.g. in client/service code or bindings), because we're just using more
bits in an existing bitfield.
Simon McVittie [Fri, 21 Jan 2011 18:54:33 +0000 (18:54 +0000)]
DBusLoop: remove second layer of watch callbacks where possible
Similar to the previous commit, almost every use of DBusWatch can just
have the main loop call dbus_watch_handle.
The one exception is the bus activation code; it's had a comment
explaining why it's wrong since 2003. We should fix that one day, but for
now, just migrate it to a new _dbus_loop_add_watch_full which preserves
the second-layer callback.
Simon McVittie [Fri, 21 Jan 2011 18:54:09 +0000 (18:54 +0000)]
DBusLoop: remove a layer of pointless abstraction around timeouts
Instead of supplying 8 tiny wrapper functions around dbus_timeout_handle,
each with a user_data parameter that's a potentially unsafe borrowed
pointer but isn't actually used, we can call dbus_timeout_handle directly
and save a lot of trouble.
One of the wrappers previously called dbus_timeout_handle repeatedly
if it returned FALSE to indicate OOM, but that timeout's handler never
actually returned FALSE, so there was no practical effect. The rest just
ignore the return, which is documented as OK to do.
Simon McVittie [Thu, 20 Jan 2011 15:48:07 +0000 (15:48 +0000)]
DBusConnection: ref the connection in the timeout handler
client_timeout_callback in bus/test.c refs the connection across the
timeout invocation, which looks suspiciously like a workaround. If we
make the timeout handler itself ref the connection, we won't need that,
and can simplify timeout handling drastically.
Simon McVittie [Mon, 13 Jun 2011 14:59:14 +0000 (15:59 +0100)]
bus signal_handler: don't use _dbus_warn, and don't pretend to be portable
_dbus_warn isn't async-signal-safe, so that's out. We can use write()
instead; it's POSIX but not ISO C, but then again, so are signals.
Accordingly, guard it with DBUS_UNIX.
dbus-sysdeps-util-win doesn't actually implement _dbus_set_signal_handler
anyway, so not compiling this code on non-Unix seems more honest.
Simon McVittie [Fri, 10 Jun 2011 16:38:14 +0000 (17:38 +0100)]
Add and use DBUS_TIMEOUT_INFINITE and DBUS_TIMEOUT_USE_DEFAULT
The documentation claimed that INT_MAX (whatever that means) meant the
default, but the value that has actually always been checked for is
0x7fffffff (aka INT32_MAX on the competent platforms we sadly don't
restrict our portability to), so we should use that. (In practice D-Bus
probably never worked on platforms where int wasn't 32 bits, though.)
Reviewed-by: Will Thompson <will.thompson@collabora.co.uk>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=34570
Simon McVittie [Tue, 22 Feb 2011 13:43:18 +0000 (13:43 +0000)]
Test nonce-tcp transport
Regression test for https://bugs.freedesktop.org/show_bug.cgi?id=34569
Reviewed-by: David Zeuthen <davidz@redhat.com> Reviewed-by: Will Thompson <will.thompson@collabora.co.uk>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=34570
Simon McVittie [Tue, 22 Feb 2011 11:43:11 +0000 (11:43 +0000)]
Add support for building "modular" tests, which require GLib and dbus-glib
For the moment, the CMake build system only knows about the existing
"embedded tests"; make it define both symbols, though.
We use GLib because it has GTester (and life's too short to write yet another
JUnit clone), and dbus-glib for the main-loop integration only (see
fd.o #31515 for thoughts on incorporating just those two functions in a
separate library in the dbus tarball).
I'm not using DBusLoop for the main loop because I specifically don't
want to use non-public API or ABI of libdbus in the modular tests. If we make
sure they work against a shared libdbus, we can use them to test the
installed system, with "make installcheck".
Reviewed-by: Will Thompson <will.thompson@collabora.co.uk>
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=34570
Simon McVittie [Fri, 10 Jun 2011 17:08:11 +0000 (18:08 +0100)]
Consistently use atomic operations for all access to DBusConnection refcount
Trying to mix atomic operations with locked non-atomic operations is
broken: the atomic ops aren't necessarily atomic with respect to the
locked non-atomic ops, and the non-atomic ops aren't protected by the
lock because the atomic ops can change the refcount behind their back.
In theory we could use the connection lock if atomic ops aren't supported
(making a per-connection lock cheaper than the global lock used to
implement atomic ops) *and* our mutexes are recursive (making it safe
against deadlocks)... but life's too short.
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=38005 Tested-by: Will Manley <freedesktop williammanley net> Reviewed-by: Will Thompson <will.thompson@collabora.co.uk>
Simon McVittie [Tue, 7 Jun 2011 13:01:39 +0000 (14:01 +0100)]
bus: use ln -fs to enable dbus in systemd, not $(LN_S)
Using $(LN_S) is inappropriate because it could in theory mean either
ln -s, ln or cp -p depending on autoconf checks.
Not using -f breaks reinstallation directly from source (DESTDIR unset),
because the symlinks will already exist.
Because systemd isn't currently portable to non-Linux, let alone
non-SUS-compliant systems, it seems safe to assume that ln -fs behaves
as specified by SUS if systemd was found.
Bug: https://bugs.freedesktop.org/show_bug.cgi?id=37870 Reviewed-by: Colin Walters <walters@verbum.org>
projects / thirdparty / dbus.git / log
