send_fds(): Make sure no other thread interrupts us

Two threads trying to send fds at the same time could interfere. To guarantee a
correct ordering, we have to use correct locking. The code in send_fds() missed
one case: If there was another thread already writing requests, we slept on the
"done with writing" condition variable (c->out.cond). This would allow other
threads to re-acquire the iolock before us and could cause fds to be sent out of
order.

To fix this, at the beginning of send_fds() we now make sure that no other
thread is already writing requests. This is what prepare_socket_request() does.
Additionally, it gets the socket back in case xcb_take_socket() was called,
which is a good thing, too, since fds are only sent with corresponding requests.
This commit is contained in:
Uli Schlachter 2015-05-18 21:40:34 +02:00
parent 25f9e7e45a
commit cc04cfb41b

View File

@ -186,6 +186,15 @@ static void close_fds(int *fds, unsigned int num_fds)
static void send_fds(xcb_connection_t *c, int *fds, unsigned int num_fds)
{
#if HAVE_SENDMSG
/* Calling _xcb_out_flush_to() can drop the iolock and wait on a condition
* variable if another thread is currently writing (c->out.writing > 0).
* This call waits for writers to be done and thus _xcb_out_flush_to() will
* do the work itself (in which case we are a writer and
* prepare_socket_request() will wait for us to be done if another threads
* tries to send fds, too). Thanks to this, we can atomically write out FDs.
*/
prepare_socket_request(c);
while (num_fds > 0) {
/* FIXME: This will busy-loop when XCB_MAX_PASS_FD fds are sent at once */
while (c->out.out_fd.nfd == XCB_MAX_PASS_FD && !c->has_error) {