Fixes following clang-9 compiler warning:
ubusd.c:36:19: error: comparison of integers of different signs: 'uint32_t' (aka 'unsigned int') and 'int' [-Werror,-Wsign-compare]
if (ub->refcount == ~0) {
~~~~~~~~~~~~ ^ ~~
Signed-off-by: Petr Štetiar <ynezz@true.cz>
This makes it clear that `ubus_msg_send()` is only
about sending and queue-ing messages, and has nothing
to do with free-ing.
It can be a bit misleading/confusing when trying to go
through the code and make assumptions about whether a
buffer is free'd in ubus_send_msg(), or is free'd outside.
In `ubusd_proto_receive_message()` the `ubus_msg_free()`
is now called before the `if (ret == -1)` check.
That way, all callbacks will have their messages free'd,
which is what's desired, but confusing, because:
* ubusd_handle_invoke() called ubus_msg_free() before returning -1
* ubusd_handle_notify() called ubus_msg_free() before returning -1
* ubusd_handle_response() called ubus_msg_send(,,free=true) before returning -1
* ubus_msg_send() would call ubus_msg_send(,,free=false)
* all other callback callers would `ubus_msg_send(,,free=true)`
(free the buffers in ubus_msg_send() )
In all other places, where `ubus_msg_send(,,free=true)`
an explicit `ubus_msg_free()` was added.
Signed-off-by: Alexandru Ardelean <ardeleanalex@gmail.com>
For a shared ubus_msg_buf, the ubus_msg_ref function will
create a copy for queue-ing.
Problem is, that during the dequeue (especially) in client_cb,
the header is 0-ed (because it's was a newly alloc-ed buffer).
And during ubus_msg_writev(), the header info will be ignored
by the client.
Signed-off-by: Alexandru Ardelean <ardeleanalex@gmail.com>
An -1 returned by ubus_msg_writev() will be interpreted as
UINT_MAX during a check to see how much data had could be
written on the socket.
Because sizeof() will return size_t it will promote the
comparsion to unsigned
Signed-off-by: Mihai Richard <mihairichard@live.com>
62cdfc3 added systemd units including a ubus.socket unit, but didn't
actually add socket activation support to ubusd. This would cause the first
connection that activated ubusd to hang (as ubusd ignored it), and stopping
ubusd would break it completely (as ubusd removed the socket file).
The ENABLE_SYSTEMD default is changed to OFF as the socket activation uses
libsystemd, so setting ENABLE_SYSTEMD to ON will now require libsystemd.
Signed-off-by: Matthias Schiffer <mschiffer@universe-factory.net>
Changing the ubus message header fields from 'host' order to 'network' order
allows passing ubus messages between hosts with different endianity.
Example use (creating a ubus proxy):
on host A (e.g. big endian router already running ubusd), run:
$ socat TCP-LISTEN:5699,fork UNIX:/var/run/ubus.sock &
On host B (e.g. little endian development PC) run:
$ socat UNIX-LISTEN:/var/run/ubus.sock,fork TCP:<host A IP>:5699 &
Now ubus applications can be run on host B and seamlessly interact with ubus
applications on host A.
Signed-off-by: Eyal Birger <eyal.birger@gmail.com>
The realloc is problematic mostly with large packets, as the pointer changes
so what eventually gets free'd is invalid.
Note that ub ptr param in the call will be passed on to a ubus_msg_free(),
right after ubus_msg_ref() finishes.
This bug stayed hidden the same way as the bug in libubus writev_retry().
Since the write/sendmsg function can send about ~200k the ubus_msg_enqueue()
call does not get triggered.
