http: implement idle client expiry under FD pressure

In case there are some wacky setups which open way too many
connections, we will nuke old HTTP connections (older than
5 seconds) under high FD pressure.  We do not drop connections
under normal loads.

This costs nothing unless we hit ENFILE/EMFILE/ENOMEM when
accepting or opening files.

This only affects Linux, but we expect most deployments to be
Linux anyways.

1 files changed, 1 insertions, 0 deletions

diff --git a/accept_loop.c b/accept_loop.c
index 92ca342..498d05f 100644
--- a/accept_loop.c
+++ b/accept_loop.c
@@ -39,6 +39,7 @@ MOG_NOINLINE static void accept_error_check(struct mog_accept *ac)
         case ENOBUFS:
         case ENOMEM:
                 syslog(LOG_ERR, MOG_ACCEPT_FN" failed with: %m");
+                mog_fdmap_expire(ac->svc->idle_timeout);
                 return;
         case ENOSYS:
                 syslog(LOG_CRIT, ENOSYS_msg);