From: Marko Kreen <markokr@gmail.com>
Date: Fri, 16 Jan 2009 12:18:24 +0000 (+0000)
Subject: sbuf_close: dont retry event_del() if it fails, just proceed with cleanup.
X-Git-Tag: pgbouncer_1_3_rc1~11
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=cdf2f665e6f3516718823b7f06dc87e423850e9e;p=pgbouncer

sbuf_close: dont retry event_del() if it fails, just proceed with cleanup.

The retry could have been bug in pgbouncer - although then the
question is that how to survive ENOMEM from event handlers that
may need allocation on event_del().

But this does not explain why it fails first time.  Seems there
is something funny going on in libevent.
---

diff --git a/src/sbuf.c b/src/sbuf.c
index f8dcb61..e46b8d1 100644
--- a/src/sbuf.c
+++ b/src/sbuf.c
@@ -236,14 +236,20 @@ bool sbuf_continue_with_callback(SBuf *sbuf, sbuf_libevent_cb user_cb)
 	return true;
 }
 
-/* socket cleanup & close */
+/* socket cleanup & close: keeps .handler and .arg values */
 bool sbuf_close(SBuf *sbuf)
 {
-	/* keep handler & arg values */
 	if (sbuf->sock > 0) {
+		/* event_del() acts funny occasionally, debug it */
+		errno = 0;
 		if (event_del(&sbuf->ev) < 0) {
-			log_warning("event_del: %s", strerror(errno));
-			return false;
+			if (errno)
+				log_warning("event_del: %s", strerror(errno));
+			else
+				log_warning("event_del: libevent error");
+
+			/* we can retry whole sbuf_close() if needed */
+			/* if (errno == ENOMEM) return false; */
 		}
 		safe_close(sbuf->sock);
 	}