summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorHolger Hans Peter Freyther <holger@moiji-mobile.com>2017-03-07 07:59:00 +0100
committerHolger Hans Peter Freyther <holger@moiji-mobile.com>2017-03-08 10:32:13 +0100
commit37c2254f5dd0f29bf1073ee3d983eb270f4f8429 (patch)
tree17392a3c128ccb60ddfca9b5a2021450784a59f9
parent49880ddf740a7af47a1247dbff9acb34087b6afc (diff)
evpoll: Add workaround for usage of IP_RECVERR in sofia-sip
We are using glib to benefit from the sofia-sip-glib eventloop integration and set a poll func (evpoll) to be called by glib to integrate with the rest of libosmocore. Sofia-sip will use IP_RECVERR to enable error reporting on the socket and then sets SU_WAIT_ERR (mapped to POLLERR if not using kqueue) in the internal events flag of the socket. This will be registered with a su_wait (mapped to struct pollfd) and then glib will be called with g_source_add_poll. At this point the the fd->events will still have the POLLERR bit set. Before glib is calling its internal poll routine or our one it will copy all registered fd into an array and mask the events flags: /* In direct contradiction to the Unix98 spec, IRIX runs into * difficulty if you pass in POLLERR, POLLHUP or POLLNVAL * flags in the events field of the pollfd while it should * just ignoring them. So we mask them out here. */ events = pollrec->fd->events & ~(G_IO_ERR|G_IO_HUP|G_IO_NVAL); This leads to the POLLERR flag never been set in the revents of the struct poll_fd and as such we never put them in the exceptionset and as such: static int tport_base_wakeup(tport_t *self, int events) { int error = 0; if (events & SU_WAIT_ERR) error = tport_error_event(self); tport_base_wakeup will never call tport_error_event. And the error will be stuck in the socket data forever and recvmsg will return a zero length packet. And this will repeat until the end of time. As a first hack I mapped SU_WAIT_ERR to POLLPRI but when using select the Linux kernel will not put the socket error into the except queue unless the sockopt SOCK_SELECT_ERR_QUEUE is used. One option is to use poll and then map the select requirements to poll. Right now I just signal POLLERR as well to trigger tport_error_event. This will result in extra syscalls for each received UDP message right now. Change-Id: I5bec4a7b70f421ab670e649e5bc1ea6faf59707c
-rw-r--r--src/evpoll.c2
1 files changed, 1 insertions, 1 deletions
diff --git a/src/evpoll.c b/src/evpoll.c
index 3273bd0..087955f 100644
--- a/src/evpoll.c
+++ b/src/evpoll.c
@@ -98,7 +98,7 @@ int evpoll(struct pollfd *fds, nfds_t nfds, int timeout)
continue;
if (FD_ISSET(fds[i].fd, &readset))
- fds[i].revents = POLLIN;
+ fds[i].revents = POLLIN | POLLERR;
if (FD_ISSET(fds[i].fd, &writeset))
fds[i].revents |= POLLOUT;
if (FD_ISSET(fds[i].fd, &exceptset))