Merge pull request #5150 from qlyoung/bgp-vector-io-4

BGP vector I/O - Redux

bgpd/bgp_io.c

@@ -22,6 +22,7 @@
 /* clang-format off */
 #include <zebra.h>
 #include <pthread.h>		// for pthread_mutex_unlock, pthread_mutex_lock
+#include <sys/uio.h>		// for writev
 
 #include "frr_pthread.h"
 #include "linklist.h"		// for list_delete, list_delete_all_node, lis...
@@ -275,21 +276,45 @@ static uint16_t bgp_write(struct peer *peer)
 {
 	uint8_t type;
 	struct stream *s;
-	int num;
 	int update_last_write = 0;
-	unsigned int count = 0;
+	unsigned int count;
 	uint32_t uo = 0;
 	uint16_t status = 0;
 	uint32_t wpkt_quanta_old;
 
+	int writenum = 0;
+	int num;
+	unsigned int iovsz;
+	unsigned int strmsz;
+	unsigned int total_written;
+
 	wpkt_quanta_old = atomic_load_explicit(&peer->bgp->wpkt_quanta,
 					       memory_order_relaxed);
+	struct stream *ostreams[wpkt_quanta_old];
+	struct stream **streams = ostreams;
+	struct iovec iov[wpkt_quanta_old];
+
+	s = stream_fifo_head(peer->obuf);
+
+	if (!s)
+		goto done;
+
+	count = iovsz = 0;
+	while (count < wpkt_quanta_old && iovsz < array_size(iov) && s) {
+		ostreams[iovsz] = s;
+		iov[iovsz].iov_base = stream_pnt(s);
+		iov[iovsz].iov_len = STREAM_READABLE(s);
+		writenum += STREAM_READABLE(s);
+		s = s->next;
+		++iovsz;
+		++count;
+	}
+
+	strmsz = iovsz;
+	total_written = 0;
 
-	while (count < wpkt_quanta_old && (s = stream_fifo_head(peer->obuf))) {
-		int writenum;
 	do {
-			writenum = stream_get_endp(s) - stream_get_getp(s);
+		num = writev(peer->fd, iov, iovsz);
-			num = write(peer->fd, stream_pnt(s), writenum);
 
 		if (num < 0) {
 			if (!ERRNO_IO_RETRY(errno)) {
@@ -299,12 +324,49 @@ static uint16_t bgp_write(struct peer *peer)
 				SET_FLAG(status, BGP_IO_TRANS_ERR);
 			}
 
-				goto done;
+			break;
-			} else if (num != writenum)
+		} else if (num != writenum) {
-				stream_forward_getp(s, num);
+			unsigned int msg_written = 0;
+			unsigned int ic = iovsz;
+
+			for (unsigned int i = 0; i < ic; i++) {
+				size_t ss = iov[i].iov_len;
+
+				if (ss > (unsigned int) num)
+					break;
+
+				msg_written++;
+				iovsz--;
+				writenum -= ss;
+				num -= ss;
+			}
+
+			total_written += msg_written;
+
+			assert(total_written < count);
+
+			memmove(&iov, &iov[msg_written],
+				sizeof(iov[0]) * iovsz);
+			streams = &streams[msg_written];
+			stream_forward_getp(streams[0], num);
+			iov[0].iov_base = stream_pnt(streams[0]);
+			iov[0].iov_len = STREAM_READABLE(streams[0]);
+
+			writenum -= num;
+			num = 0;
+			assert(writenum > 0);
+		} else {
+			total_written = strmsz;
+		}
 
 	} while (num != writenum);
 
+	/* Handle statistics */
+	for (unsigned int i = 0; i < total_written; i++) {
+		s = stream_fifo_pop(peer->obuf);
+
+		assert(s == ostreams[i]);
+
 		/* Retrieve BGP packet type. */
 		stream_set_getp(s, BGP_MARKER_SIZE + 2);
 		type = stream_getc(s);
@@ -351,9 +413,8 @@ static uint16_t bgp_write(struct peer *peer)
 			break;
 		}
 
-		count++;
+		stream_free(s);
-
+		ostreams[i] = NULL;
-		stream_free(stream_fifo_pop(peer->obuf));
 		update_last_write = 1;
 	}
 

bgpd/bgp_io.h

@@ -22,7 +22,7 @@
 #ifndef _FRR_BGP_IO_H
 #define _FRR_BGP_IO_H
 
-#define BGP_WRITE_PACKET_MAX 10U
+#define BGP_WRITE_PACKET_MAX 64U
 #define BGP_READ_PACKET_MAX  10U
 
 #include "bgpd/bgpd.h"

bgpd/bgp_vty.c

@@ -1586,36 +1586,24 @@ DEFUN (no_bgp_update_delay,
 }
 
 
-static int bgp_wpkt_quanta_config_vty(struct vty *vty, const char *num,
+static int bgp_wpkt_quanta_config_vty(struct vty *vty, uint32_t quanta,
-				      char set)
+				      bool set)
 {
 	VTY_DECLVAR_CONTEXT(bgp, bgp);
 
-	if (set) {
+	quanta = set ? quanta : BGP_WRITE_PACKET_MAX;
-		uint32_t quanta = strtoul(num, NULL, 10);
+	atomic_store_explicit(&bgp->wpkt_quanta, quanta, memory_order_relaxed);
-		atomic_store_explicit(&bgp->wpkt_quanta, quanta,
-				      memory_order_relaxed);
-	} else {
-		atomic_store_explicit(&bgp->wpkt_quanta, BGP_WRITE_PACKET_MAX,
-				      memory_order_relaxed);
-	}
 
 	return CMD_SUCCESS;
 }
 
-static int bgp_rpkt_quanta_config_vty(struct vty *vty, const char *num,
+static int bgp_rpkt_quanta_config_vty(struct vty *vty, uint32_t quanta,
-				      char set)
+				      bool set)
 {
 	VTY_DECLVAR_CONTEXT(bgp, bgp);
 
-	if (set) {
+	quanta = set ? quanta : BGP_READ_PACKET_MAX;
-		uint32_t quanta = strtoul(num, NULL, 10);
+	atomic_store_explicit(&bgp->rpkt_quanta, quanta, memory_order_relaxed);
-		atomic_store_explicit(&bgp->rpkt_quanta, quanta,
-				      memory_order_relaxed);
-	} else {
-		atomic_store_explicit(&bgp->rpkt_quanta, BGP_READ_PACKET_MAX,
-				      memory_order_relaxed);
-	}
 
 	return CMD_SUCCESS;
 }
@@ -1636,47 +1624,32 @@ void bgp_config_write_rpkt_quanta(struct vty *vty, struct bgp *bgp)
 		vty_out(vty, " read-quanta %d\n", quanta);
 }
 
-/* Packet quanta configuration */
+/* Packet quanta configuration
-DEFUN (bgp_wpkt_quanta,
+ *
+ * XXX: The value set here controls the size of a stack buffer in the IO
+ * thread. When changing these limits be careful to prevent stack overflow.
+ *
+ * Furthermore, the maximums used here should correspond to
+ * BGP_WRITE_PACKET_MAX and BGP_READ_PACKET_MAX.
+ */
+DEFPY (bgp_wpkt_quanta,
        bgp_wpkt_quanta_cmd,
-       "write-quanta (1-10)",
+       "[no] write-quanta (1-64)$quanta",
+       NO_STR
        "How many packets to write to peer socket per run\n"
        "Number of packets\n")
 {
-	int idx_number = 1;
+	return bgp_wpkt_quanta_config_vty(vty, quanta, !no);
-	return bgp_wpkt_quanta_config_vty(vty, argv[idx_number]->arg, 1);
 }
 
-DEFUN (no_bgp_wpkt_quanta,
+DEFPY (bgp_rpkt_quanta,
-       no_bgp_wpkt_quanta_cmd,
-       "no write-quanta (1-10)",
-       NO_STR
-       "How many packets to write to peer socket per I/O cycle\n"
-       "Number of packets\n")
-{
-	int idx_number = 2;
-	return bgp_wpkt_quanta_config_vty(vty, argv[idx_number]->arg, 0);
-}
-
-DEFUN (bgp_rpkt_quanta,
        bgp_rpkt_quanta_cmd,
-       "read-quanta (1-10)",
+       "[no] read-quanta (1-10)$quanta",
-       "How many packets to read from peer socket per I/O cycle\n"
-       "Number of packets\n")
-{
-	int idx_number = 1;
-	return bgp_rpkt_quanta_config_vty(vty, argv[idx_number]->arg, 1);
-}
-
-DEFUN (no_bgp_rpkt_quanta,
-       no_bgp_rpkt_quanta_cmd,
-       "no read-quanta (1-10)",
        NO_STR
        "How many packets to read from peer socket per I/O cycle\n"
        "Number of packets\n")
 {
-	int idx_number = 2;
+	return bgp_rpkt_quanta_config_vty(vty, quanta, !no);
-	return bgp_rpkt_quanta_config_vty(vty, argv[idx_number]->arg, 0);
 }
 
 void bgp_config_write_coalesce_time(struct vty *vty, struct bgp *bgp)
@@ -13072,9 +13045,7 @@ void bgp_vty_init(void)
 	install_element(BGP_NODE, &bgp_update_delay_establish_wait_cmd);
 
 	install_element(BGP_NODE, &bgp_wpkt_quanta_cmd);
-	install_element(BGP_NODE, &no_bgp_wpkt_quanta_cmd);
 	install_element(BGP_NODE, &bgp_rpkt_quanta_cmd);
-	install_element(BGP_NODE, &no_bgp_rpkt_quanta_cmd);
 
 	install_element(BGP_NODE, &bgp_coalesce_time_cmd);
 	install_element(BGP_NODE, &no_bgp_coalesce_time_cmd);

doc/user/bgp.rst

@@ -2167,6 +2167,8 @@ Dumping Messages and Routing Tables
 Other BGP Commands
 ------------------
 
+The following are available in the top level *enable* mode:
+
 .. index:: clear bgp \*
 .. clicmd:: clear bgp \*
 
@@ -2202,6 +2204,24 @@ Other BGP Commands
 
    Clear peer using soft reconfiguration in this address-family and sub-address-family.
 
+The following are available in the ``router bgp`` mode:
+
+.. index:: write-quanta (1-64)
+.. clicmd:: write-quanta (1-64)
+
+   BGP message Tx I/O is vectored. This means that multiple packets are written
+   to the peer socket at the same time each I/O cycle, in order to minimize
+   system call overhead. This value controls how many are written at a time.
+   Under certain load conditions, reducing this value could make peer traffic
+   less 'bursty'. In practice, leave this settings on the default (64) unless
+   you truly know what you are doing.
+
+.. index:: read-quanta (1-10)
+.. index:: read-quanta (1-10)
+
+   Unlike Tx, BGP Rx traffic is not vectored. Packets are read off the wire one
+   at a time in a loop. This setting controls how many iterations the loop runs
+   for. As with write-quanta, it is best to leave this setting on the default.
 
 .. _bgp-displaying-bgp-information: