The crappy epoll API stroke again with reloads and transferred FDs.
Indeed, when listening sockets are retrieved by a new worker from a
previous one, and the old one finally stops listening on them, it
closes the FDs. But in this case, since the sockets themselves were
not closed, epoll will not unregister them and will continue to
report new activity for these in the old process, which can only
observe, count an fd_poll_drop event and not unregister them since
they're not reachable anymore.
The unfortunate effect is that long-lasting old processes are woken
up at the same rate as the new process when accepting new connections,
and can waste a lot of CPU. Accept rates divided by 8 were observed on
a small test involving a slow transfer on 10 connections facing a reload
every second so that 10 processes were busy dealing with them while
another process was hammering the service with new connections.
Fortunately, years ago we implemented a flag FD_CLONED exactly for
similar purposes. Let's simply mark transferred FDs with FD_CLONED
so that the process knows that these ones require special treatment
and have to be manually unregistered before being closed. This does
the job fine, now old processes correctly unregister the FD before
closing it and no longer receive accept events for the new process.
This needs to be backported to all stable versions. It only affects
epoll, as usual, and this time in combination with transferred FDs
(typically reloads in master-worker mode). Thanks to Damien Claisse
for providing all detailed measurements and statistics allowing to
understand and reproduce the problem.
if (!(fdtab[cur_fd].state & FD_EXPORTED))
continue;
+ /* this FD is now shared between processes */
+ HA_ATOMIC_OR(&fdtab[cur_fd].state, FD_CLONED);
+
ns_name = if_name = "";
ns_nlen = if_nlen = 0;