Make PQcancel use the PGconn's tcp_user_timeout and keepalives settings.

tglsfdc · tglsfdc · commit 5987feb70b5b · 2022-01-18T14:13:13.000-05:00
If connectivity to the server has been lost or become flaky, the user might well try to send a query cancel. It's highly annoying if PQcancel hangs up in such a case, but that's exactly what's likely to happen. To ameliorate this problem, apply the PGconn's tcp_user_timeout and keepalives settings to the TCP connection used to send the cancel. This should be safe on Unix machines, since POSIX specifies that setsockopt() is async-signal-safe. We are guessing that WSAIoctl(SIO_KEEPALIVE_VALS) is similarly safe on Windows. (Note that at least in psql and our other frontend programs, there's no safety issue involved anyway, since we run PQcancel in its own thread rather than in a signal handler.) Most of the value here comes from the expectation that tcp_user_timeout will be applied as a connection timeout. That appears to happen on Linux, even though its tcp(7) man page claims differently. The keepalive options probably won't help much, but as long as we can apply them for not much code, we might as well. Jelte Fennema, reviewed by Fujii Masao and myself Discussion: https://postgr.es/m/AM5PR83MB017870DE81FC84D5E21E9D1EF7AA9@AM5PR83MB0178.EURPRD83.prod.outlook.com
diff --git a/doc/src/sgml/libpq.sgml b/doc/src/sgml/libpq.sgml
@@ -5707,8 +5707,8 @@ int PQrequestCancel(PGconn *conn);
        <structname>PGconn</structname> object, and in case of failure stores the
        error message in the <structname>PGconn</structname> object (whence it can
        be retrieved by <xref linkend="libpq-PQerrorMessage"/>).  Although
-       the functionality is the same, this approach creates hazards for
-       multiple-thread programs and signal handlers, since it is possible
+       the functionality is the same, this approach is not safe within
+       multiple-thread programs or signal handlers, since it is possible
        that overwriting the <structname>PGconn</structname>'s error message will
        mess up the operation currently in progress on the connection.
       </para>
diff --git a/src/interfaces/libpq/fe-connect.c b/src/interfaces/libpq/fe-connect.c
@@ -1944,34 +1944,25 @@ setKeepalivesCount(PGconn *conn)
 /*
  * Enable keepalives and set the keepalive values on Win32,
  * where they are always set in one batch.
+ *
+ * CAUTION: This needs to be signal safe, since it's used by PQcancel.
  */
 static int
-setKeepalivesWin32(PGconn *conn)
+setKeepalivesWin32(pgsocket sock, int idle, int interval)
 {
 	struct tcp_keepalive ka;
 	DWORD		retsize;
-	int			idle = 0;
-	int			interval = 0;
 
-	if (conn->keepalives_idle &&
-		!parse_int_param(conn->keepalives_idle, &idle, conn,
-						 "keepalives_idle"))
-		return 0;
 	if (idle <= 0)
 		idle = 2 * 60 * 60;		/* 2 hours = default */
-
-	if (conn->keepalives_interval &&
-		!parse_int_param(conn->keepalives_interval, &interval, conn,
-						 "keepalives_interval"))
-		return 0;
 	if (interval <= 0)
 		interval = 1;			/* 1 second = default */
 
 	ka.onoff = 1;
 	ka.keepalivetime = idle * 1000;
 	ka.keepaliveinterval = interval * 1000;
 
-	if (WSAIoctl(conn->sock,
+	if (WSAIoctl(sock,
 				 SIO_KEEPALIVE_VALS,
 				 (LPVOID) &ka,
 				 sizeof(ka),
@@ -1981,6 +1972,26 @@ setKeepalivesWin32(PGconn *conn)
 				 NULL,
 				 NULL)
 		!= 0)
+		return 0;
+	return 1;
+}
+
+static int
+prepKeepalivesWin32(PGconn *conn)
+{
+	int			idle = -1;
+	int			interval = -1;
+
+	if (conn->keepalives_idle &&
+		!parse_int_param(conn->keepalives_idle, &idle, conn,
+						 "keepalives_idle"))
+		return 0;
+	if (conn->keepalives_interval &&
+		!parse_int_param(conn->keepalives_interval, &interval, conn,
+						 "keepalives_interval"))
+		return 0;
+
+	if (!setKeepalivesWin32(conn->sock, idle, interval))
 	{
 		appendPQExpBuffer(&conn->errorMessage,
 						  libpq_gettext("%s(%s) failed: error code %d\n"),
@@ -2644,7 +2655,7 @@ PQconnectPoll(PGconn *conn)
 							err = 1;
 #else							/* WIN32 */
 #ifdef SIO_KEEPALIVE_VALS
-						else if (!setKeepalivesWin32(conn))
+						else if (!prepKeepalivesWin32(conn))
 							err = 1;
 #endif							/* SIO_KEEPALIVE_VALS */
 #endif							/* WIN32 */
@@ -4380,8 +4391,53 @@ PQgetCancel(PGconn *conn)
 	memcpy(&cancel->raddr, &conn->raddr, sizeof(SockAddr));
 	cancel->be_pid = conn->be_pid;
 	cancel->be_key = conn->be_key;
+	/* We use -1 to indicate an unset connection option */
+	cancel->pgtcp_user_timeout = -1;
+	cancel->keepalives = -1;
+	cancel->keepalives_idle = -1;
+	cancel->keepalives_interval = -1;
+	cancel->keepalives_count = -1;
+	if (conn->pgtcp_user_timeout != NULL)
+	{
+		if (!parse_int_param(conn->pgtcp_user_timeout,
+							 &cancel->pgtcp_user_timeout,
+							 conn, "tcp_user_timeout"))
+			goto fail;
+	}
+	if (conn->keepalives != NULL)
+	{
+		if (!parse_int_param(conn->keepalives,
+							 &cancel->keepalives,
+							 conn, "keepalives"))
+			goto fail;
+	}
+	if (conn->keepalives_idle != NULL)
+	{
+		if (!parse_int_param(conn->keepalives_idle,
+							 &cancel->keepalives_idle,
+							 conn, "keepalives_idle"))
+			goto fail;
+	}
+	if (conn->keepalives_interval != NULL)
+	{
+		if (!parse_int_param(conn->keepalives_interval,
+							 &cancel->keepalives_interval,
+							 conn, "keepalives_interval"))
+			goto fail;
+	}
+	if (conn->keepalives_count != NULL)
+	{
+		if (!parse_int_param(conn->keepalives_count,
+							 &cancel->keepalives_count,
+							 conn, "keepalives_count"))
+			goto fail;
+	}
 
 	return cancel;
+
+fail:
+	free(cancel);
+	return NULL;
 }
 
 /* PQfreeCancel: free a cancel structure */
@@ -4394,28 +4450,46 @@ PQfreeCancel(PGcancel *cancel)
 
 
 /*
- * PQcancel and PQrequestCancel: attempt to request cancellation of the
- * current operation.
+ * Sets an integer socket option on a TCP socket, if the provided value is
+ * not negative.  Returns false if setsockopt fails for some reason.
+ *
+ * CAUTION: This needs to be signal safe, since it's used by PQcancel.
+ */
+#if defined(TCP_USER_TIMEOUT) || !defined(WIN32)
+static bool
+optional_setsockopt(int fd, int protoid, int optid, int value)
+{
+	if (value < 0)
+		return true;
+	if (setsockopt(fd, protoid, optid, (char *) &value, sizeof(value)) < 0)
+		return false;
+	return true;
+}
+#endif
+
+
+/*
+ * PQcancel: request query cancel
  *
  * The return value is true if the cancel request was successfully
  * dispatched, false if not (in which case an error message is available).
  * Note: successful dispatch is no guarantee that there will be any effect at
  * the backend.  The application must read the operation result as usual.
  *
+ * On failure, an error message is stored in *errbuf, which must be of size
+ * errbufsize (recommended size is 256 bytes).  *errbuf is not changed on
+ * success return.
+ *
  * CAUTION: we want this routine to be safely callable from a signal handler
  * (for example, an application might want to call it in a SIGINT handler).
  * This means we cannot use any C library routine that might be non-reentrant.
  * malloc/free are often non-reentrant, and anything that might call them is
  * just as dangerous.  We avoid sprintf here for that reason.  Building up
  * error messages with strcpy/strcat is tedious but should be quite safe.
  * We also save/restore errno in case the signal handler support doesn't.
- *
- * internal_cancel() is an internal helper function to make code-sharing
- * between the two versions of the cancel function possible.
  */
-static int
-internal_cancel(SockAddr *raddr, int be_pid, int be_key,
-				char *errbuf, int errbufsize)
+int
+PQcancel(PGcancel *cancel, char *errbuf, int errbufsize)
 {
 	int			save_errno = SOCK_ERRNO;
 	pgsocket	tmpsock = PGINVALID_SOCKET;
@@ -4426,18 +4500,98 @@ internal_cancel(SockAddr *raddr, int be_pid, int be_key,
 		CancelRequestPacket cp;
 	}			crp;
 
+	if (!cancel)
+	{
+		strlcpy(errbuf, "PQcancel() -- no cancel object supplied", errbufsize);
+		/* strlcpy probably doesn't change errno, but be paranoid */
+		SOCK_ERRNO_SET(save_errno);
+		return false;
+	}
+
 	/*
 	 * We need to open a temporary connection to the postmaster. Do this with
 	 * only kernel calls.
 	 */
-	if ((tmpsock = socket(raddr->addr.ss_family, SOCK_STREAM, 0)) == PGINVALID_SOCKET)
+	if ((tmpsock = socket(cancel->raddr.addr.ss_family, SOCK_STREAM, 0)) == PGINVALID_SOCKET)
 	{
 		strlcpy(errbuf, "PQcancel() -- socket() failed: ", errbufsize);
 		goto cancel_errReturn;
 	}
+
+	/*
+	 * Since this connection will only be used to send a single packet of
+	 * data, we don't need NODELAY.  We also don't set the socket to
+	 * nonblocking mode, because the API definition of PQcancel requires the
+	 * cancel to be sent in a blocking way.
+	 *
+	 * We do set socket options related to keepalives and other TCP timeouts.
+	 * This ensures that this function does not block indefinitely when
+	 * reasonable keepalive and timeout settings have been provided.
+	 */
+	if (!IS_AF_UNIX(cancel->raddr.addr.ss_family) &&
+		cancel->keepalives != 0)
+	{
+#ifndef WIN32
+		if (!optional_setsockopt(tmpsock, SOL_SOCKET, SO_KEEPALIVE, 1))
+		{
+			strlcpy(errbuf, "PQcancel() -- setsockopt(SO_KEEPALIVE) failed: ", errbufsize);
+			goto cancel_errReturn;
+		}
+
+#ifdef PG_TCP_KEEPALIVE_IDLE
+		if (!optional_setsockopt(tmpsock, IPPROTO_TCP, PG_TCP_KEEPALIVE_IDLE,
+								 cancel->keepalives_idle))
+		{
+			strlcpy(errbuf, "PQcancel() -- setsockopt(" PG_TCP_KEEPALIVE_IDLE_STR ") failed: ", errbufsize);
+			goto cancel_errReturn;
+		}
+#endif
+
+#ifdef TCP_KEEPINTVL
+		if (!optional_setsockopt(tmpsock, IPPROTO_TCP, TCP_KEEPINTVL,
+								 cancel->keepalives_interval))
+		{
+			strlcpy(errbuf, "PQcancel() -- setsockopt(TCP_KEEPINTVL) failed: ", errbufsize);
+			goto cancel_errReturn;
+		}
+#endif
+
+#ifdef TCP_KEEPCNT
+		if (!optional_setsockopt(tmpsock, IPPROTO_TCP, TCP_KEEPCNT,
+								 cancel->keepalives_count))
+		{
+			strlcpy(errbuf, "PQcancel() -- setsockopt(TCP_KEEPCNT) failed: ", errbufsize);
+			goto cancel_errReturn;
+		}
+#endif
+
+#else							/* WIN32 */
+
+#ifdef SIO_KEEPALIVE_VALS
+		if (!setKeepalivesWin32(tmpsock,
+								cancel->keepalives_idle,
+								cancel->keepalives_interval))
+		{
+			strlcpy(errbuf, "PQcancel() -- WSAIoctl(SIO_KEEPALIVE_VALS) failed: ", errbufsize);
+			goto cancel_errReturn;
+		}
+#endif							/* SIO_KEEPALIVE_VALS */
+#endif							/* WIN32 */
+
+		/* TCP_USER_TIMEOUT works the same way on Unix and Windows */
+#ifdef TCP_USER_TIMEOUT
+		if (!optional_setsockopt(tmpsock, IPPROTO_TCP, TCP_USER_TIMEOUT,
+								 cancel->pgtcp_user_timeout))
+		{
+			strlcpy(errbuf, "PQcancel() -- setsockopt(TCP_USER_TIMEOUT) failed: ", errbufsize);
+			goto cancel_errReturn;
+		}
+#endif
+	}
+
 retry3:
-	if (connect(tmpsock, (struct sockaddr *) &raddr->addr,
-				raddr->salen) < 0)
+	if (connect(tmpsock, (struct sockaddr *) &cancel->raddr.addr,
+				cancel->raddr.salen) < 0)
 	{
 		if (SOCK_ERRNO == EINTR)
 			/* Interrupted system call - we'll just try again */
@@ -4446,16 +4600,12 @@ internal_cancel(SockAddr *raddr, int be_pid, int be_key,
 		goto cancel_errReturn;
 	}
 
-	/*
-	 * We needn't set nonblocking I/O or NODELAY options here.
-	 */
-
 	/* Create and send the cancel request packet. */
 
 	crp.packetlen = pg_hton32((uint32) sizeof(crp));
 	crp.cp.cancelRequestCode = (MsgType) pg_hton32(CANCEL_REQUEST_CODE);
-	crp.cp.backendPID = pg_hton32(be_pid);
-	crp.cp.cancelAuthCode = pg_hton32(be_key);
+	crp.cp.backendPID = pg_hton32(cancel->be_pid);
+	crp.cp.cancelAuthCode = pg_hton32(cancel->be_key);
 
 retry4:
 	if (send(tmpsock, (char *) &crp, sizeof(crp), 0) != (int) sizeof(crp))
@@ -4524,27 +4674,6 @@ internal_cancel(SockAddr *raddr, int be_pid, int be_key,
 	return false;
 }
 
-/*
- * PQcancel: request query cancel
- *
- * Returns true if able to send the cancel request, false if not.
- *
- * On failure, an error message is stored in *errbuf, which must be of size
- * errbufsize (recommended size is 256 bytes).  *errbuf is not changed on
- * success return.
- */
-int
-PQcancel(PGcancel *cancel, char *errbuf, int errbufsize)
-{
-	if (!cancel)
-	{
-		strlcpy(errbuf, "PQcancel() -- no cancel object supplied", errbufsize);
-		return false;
-	}
-
-	return internal_cancel(&cancel->raddr, cancel->be_pid, cancel->be_key,
-						   errbuf, errbufsize);
-}
 
 /*
  * PQrequestCancel: old, not thread-safe function for requesting query cancel
@@ -4562,6 +4691,7 @@ int
 PQrequestCancel(PGconn *conn)
 {
 	int			r;
+	PGcancel   *cancel;
 
 	/* Check we have an open connection */
 	if (!conn)
@@ -4577,8 +4707,19 @@ PQrequestCancel(PGconn *conn)
 		return false;
 	}
 
-	r = internal_cancel(&conn->raddr, conn->be_pid, conn->be_key,
-						conn->errorMessage.data, conn->errorMessage.maxlen);
+	cancel = PQgetCancel(conn);
+	if (cancel)
+	{
+		r = PQcancel(cancel, conn->errorMessage.data,
+					 conn->errorMessage.maxlen);
+		PQfreeCancel(cancel);
+	}
+	else
+	{
+		strlcpy(conn->errorMessage.data, "out of memory",
+				conn->errorMessage.maxlen);
+		r = false;
+	}
 
 	if (!r)
 		conn->errorMessage.len = strlen(conn->errorMessage.data);
diff --git a/src/interfaces/libpq/libpq-int.h b/src/interfaces/libpq/libpq-int.h
@@ -583,6 +583,13 @@ struct pg_cancel
 	SockAddr	raddr;			/* Remote address */
 	int			be_pid;			/* PID of backend --- needed for cancels */
 	int			be_key;			/* key of backend --- needed for cancels */
+	int			pgtcp_user_timeout; /* tcp user timeout */
+	int			keepalives;		/* use TCP keepalives? */
+	int			keepalives_idle;	/* time between TCP keepalives */
+	int			keepalives_interval;	/* time between TCP keepalive
+										 * retransmits */
+	int			keepalives_count;	/* maximum number of TCP keepalive
+									 * retransmits */
 };
 
 
