v3-0002-Use-WL_SOCKET_CLOSED-for-client_connection_check_.patch

From 2d191c0d3bbc240e125633d51f9f1e003630b866 Mon Sep 17 00:00:00 2001
From: Thomas Munro <thomas.munro@gmail.com>
Date: Fri, 30 Apr 2021 10:48:32 +1200
Subject: [PATCH v3 2/2] Use WL_SOCKET_CLOSED for
 client_connection_check_interval.

Previously we used poll() directly to check for a POLLRDHUP event.
Instead, use WaitEventSet to poll the socket for WL_SOCKET_CLOSED, which
knows how to detect equivalent events on many more operating systems.

Reviewed-by: Zhihong Yu <zyu@yugabyte.com>
Reviewed-by: Maksim Milyutin <milyutinma@gmail.com>
Discussion: https://postgr.es/m/77def86b27e41f0efcba411460e929ae%40postgrespro.ru
---
 doc/src/sgml/config.sgml     |  6 ++---
 src/backend/libpq/pqcomm.c   | 44 ++++++++++++++++++++++++------------
 src/backend/utils/misc/guc.c |  7 ++----
 3 files changed, 34 insertions(+), 23 deletions(-)

diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 4ac617615c..ab52d4bd2c 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1012,9 +1012,9 @@ include_dir 'conf.d'
         the kernel reports that the connection is closed.
        </para>
        <para>
-        This option is currently available only on systems that support the
-        non-standard <symbol>POLLRDHUP</symbol> extension to the
-        <symbol>poll</symbol> system call, including Linux.
+        This option relies on kernel events exposed by Linux, macOS, illumos
+        and the BSD family of operating systems, and is not currently available
+        on other systems.
        </para>
        <para>
         If the value is specified without units, it is taken as milliseconds.
diff --git a/src/backend/libpq/pqcomm.c b/src/backend/libpq/pqcomm.c
index 4c37df09cf..c235a682c2 100644
--- a/src/backend/libpq/pqcomm.c
+++ b/src/backend/libpq/pqcomm.c
@@ -1959,22 +1959,36 @@ pq_settcpusertimeout(int timeout, Port *port)
 bool
 pq_check_connection(void)
 {
-#if defined(POLLRDHUP)
-	/*
-	 * POLLRDHUP is a Linux extension to poll(2) to detect sockets closed by
-	 * the other end.  We don't have a portable way to do that without
-	 * actually trying to read or write data on other systems.  We don't want
-	 * to read because that would be confused by pipelined queries and COPY
-	 * data. Perhaps in future we'll try to write a heartbeat message instead.
-	 */
-	struct pollfd pollfd;
+	WaitEvent	event;
 	int			rc;
 
-	pollfd.fd = MyProcPort->sock;
-	pollfd.events = POLLOUT | POLLIN | POLLRDHUP;
-	pollfd.revents = 0;
+	/*
+	 * Temporarily ignore the latch, so that we can poll for just the one
+	 * event we care about.
+	 */
+	ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetLatchPos, WL_LATCH_SET, NULL);
 
-	rc = poll(&pollfd, 1, 0);
+	PG_TRY();
+	{
+		/*
+		 * It's OK to clobber the socket event to report only the event we're
+		 * interested in without restoring the previous state afterwards,
+		 * because every FeBeWaitSet wait site does the same.
+		 */
+		ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetSocketPos, WL_SOCKET_CLOSED,
+						NULL);
+		rc = WaitEventSetWait(FeBeWaitSet, 0, &event, 1, 0);
+	}
+	PG_FINALLY();
+	{
+		/*
+		 * Restore the latch, so we can't leave FeBeWaitSet in a broken state
+		 * that ignores latches.
+		 */
+		ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetLatchPos, WL_LATCH_SET,
+						MyLatch);
+	}
+	PG_END_TRY();
 
 	if (rc < 0)
 	{
@@ -1983,9 +1997,9 @@ pq_check_connection(void)
 				 errmsg("could not poll socket: %m")));
 		return false;
 	}
-	else if (rc == 1 && (pollfd.revents & (POLLHUP | POLLRDHUP)))
+
+	if (rc == 1 && (event.events & WL_SOCKET_CLOSED))
 		return false;
-#endif
 
 	return true;
 }
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index ee6a838b3a..ca45b36754 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -12169,14 +12169,11 @@ check_huge_page_size(int *newval, void **extra, GucSource source)
 static bool
 check_client_connection_check_interval(int *newval, void **extra, GucSource source)
 {
-#ifndef POLLRDHUP
-	/* Linux only, for now.  See pq_check_connection(). */
-	if (*newval != 0)
+	if (!WaitEventSetCanReportClosed() && *newval != 0)
 	{
-		GUC_check_errdetail("client_connection_check_interval must be set to 0 on platforms that lack POLLRDHUP.");
+		GUC_check_errdetail("client_connection_check_interval must be set to 0 on this platform");
 		return false;
 	}
-#endif
 	return true;
 }
 
-- 
2.33.1