Max mtu

docs/grout.8.md

@@ -24,6 +24,7 @@ Grout is a software router based on DPDK __rte_graph__.
 [**-s** _PATH_]
 [**-t**]
 [**-T** _REGEXP_]
+[**-u** _MTU_]
 [**-v**]
 [**-V**]
 [**-x**]
@@ -130,6 +131,12 @@ Global trace configuration for ALL the components:
 
 Can be specified multiple times up to 32 times.
 
+#### **-u**, **--max-mtu** _MTU_
+
+Maximum Transmission Unit.
+
+Default: _1800_.
+
 #### **-v**, **--verbose**
 
 Increase verbosity. Can be specified multiple times.

main/dpdk.c

@@ -143,7 +143,10 @@ int dpdk_init(void) {
 		gr_vec_add(eal_args, "--no-shconf");
 		gr_vec_add(eal_args, "--no-huge");
 		gr_vec_add(eal_args, "-m");
-		gr_vec_add(eal_args, "2048");
+		if (gr_config.max_mtu > 2048)
+			gr_vec_add(eal_args, "4096");
+		else
+			gr_vec_add(eal_args, "2048");
 	} else {
 		gr_vec_add(eal_args, "--in-memory");
 	}

main/gr_config.h

@@ -15,6 +15,7 @@ struct gr_config {
 	gid_t api_sock_gid;
 	mode_t api_sock_mode;
 	unsigned log_level;
+	unsigned max_mtu;
 	bool test_mode;
 	bool poll_mode;
 	bool log_syslog;

main/main.c

@@ -31,8 +31,8 @@
 // Please keep options/flags in alphabetical order.
 
 static void usage(const char *prog) {
-	printf("Usage: %s [-h] [-L <type>:<lvl>] [-p] [-s <path>] [-t] [-T <regexp>]\n", prog);
-	printf("       %*s [-B <size>] [-D <path>] [-M <mode>] [-x] [-v] [-V]\n",
+	printf("Usage: %s [-h] [-L <type>:<lvl>] [-p] [-s <path>] [-t] [-u <mtu>]\n", prog);
+	printf("       %*s [-T <regexp>] [-B <size>] [-D <path>] [-M <mode>] [-x] [-v] [-V]\n",
 	       (int)strlen(prog),
 	       "");
 	puts("");
@@ -50,6 +50,7 @@ static void usage(const char *prog) {
 	puts("                                 Default: GROUT_SOCK_PATH from env or");
 	printf("                                 %s).\n", GR_DEFAULT_SOCK_PATH);
 	puts("  -t, --test-mode                Run in test mode (no hugepages).");
+	puts("  -u, --max-mtu <mtu>            Maximum Transmission Unit (default 1800).");
 	puts("  -T, --trace <regexp>           Enable trace matching the regular expression.");
 	puts("  -B, --trace-bufsz <size>       Maximum size of allocated memory for trace output.");
 	puts("  -D, --trace-dir <path>         Change path for trace output.");
@@ -127,13 +128,37 @@ static int parse_sock_mode(const char *perm_str) {
 	return 0;
 }
 
+static int parse_max_mtu(const char *mtu) {
+	unsigned long val;
+	char *endptr;
+
+	errno = 0;
+	val = strtoul(mtu, &endptr, 10);
+
+	if (errno != 0 || *endptr != '\0' || val < 512 || val > 16384) {
+		if (errno == 0) {
+			if (*endptr != '\0')
+				errno = EINVAL;
+			else
+				errno = ERANGE;
+		}
+		fprintf(stderr, "error: invalid max-mtu '%s': %s\n", mtu, strerror(errno));
+		return -1;
+	}
+
+	gr_config.max_mtu = (unsigned)val;
+
+	return 0;
+}
+
 static int parse_args(int argc, char **argv) {
 	int c;
 
-#define FLAGS ":B:D:L:M:T:Vhm:o:pSs:tvx"
+#define FLAGS ":B:D:L:M:T:Vhm:o:pSs:tu:vx"
 	static struct option long_options[] = {
 		{"help", no_argument, NULL, 'h'},
 		{"log-level", required_argument, NULL, 'L'},
+		{"max-mtu", required_argument, NULL, 'u'},
 		{"poll-mode", no_argument, NULL, 'p'},
 		{"syslog", no_argument, NULL, 'S'},
 		{"socket", required_argument, NULL, 's'},
@@ -158,6 +183,7 @@ static int parse_args(int argc, char **argv) {
 	gr_config.api_sock_uid = getuid();
 	gr_config.api_sock_gid = getgid();
 	gr_config.api_sock_mode = 0660;
+	gr_config.max_mtu = 1800;
 	gr_config.log_level = RTE_LOG_NOTICE;
 	gr_config.eal_extra_args = NULL;
 
@@ -205,6 +231,10 @@ static int parse_args(int argc, char **argv) {
 		case 'x':
 			gr_config.log_packets = true;
 			break;
+		case 'u':
+			if (parse_max_mtu(optarg) < 0)
+				return errno_set(EINVAL);
+			break;
 		case 'v':
 			gr_config.log_level++;
 			break;

modules/infra/control/mempool.c

@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: BSD-3-Clause
 // Copyright (c) 2024 Christophe Fontaine
 
+#include <gr_config.h>
 #include <gr_log.h>
 #include <gr_mbuf.h>
 #include <gr_mempool.h>
@@ -14,6 +15,8 @@ struct mempool_tracker {
 
 #define MAX_MEMPOOL_PER_NUMA 32
 #define MEMPOOL_DEFAULT_SIZE (1 << 16) - 1
+#define ETHER_HDR_SIZE 14
+#define VLAN_HDR_SIZE 4
 
 static int mt_sort(const void *p1, const void *p2) {
 	const struct mempool_tracker *mt1 = p1;
@@ -37,10 +40,15 @@ struct rte_mempool *gr_pktmbuf_pool_get(int8_t socket_id, uint32_t count) {
 	char mp_name[RTE_MEMPOOL_NAMESIZE];
 	struct rte_mempool *mp = NULL;
 	uint32_t alloc_size;
+	uint32_t mbuf_size;
 
 	if (socket_id < SOCKET_ID_ANY || socket_id >= RTE_MAX_NUMA_NODES)
 		return errno_set_null(EINVAL);
 
+	mbuf_size = rte_align32pow2(
+		RTE_PKTMBUF_HEADROOM + ETHER_HDR_SIZE + VLAN_HDR_SIZE + gr_config.max_mtu
+	);
+
 	for (int i = 0; i < MAX_MEMPOOL_PER_NUMA; i++) {
 		unsigned mt_index = socket_id == SOCKET_ID_ANY ? 0 : socket_id + 1;
 		struct mempool_tracker *mt = &trackers[mt_index][i];
@@ -54,16 +62,17 @@ struct rte_mempool *gr_pktmbuf_pool_get(int8_t socket_id, uint32_t count) {
 			}
 			sprintf(mp_name, "mbuf_%d:%d", socket_id, i);
 			LOG(DEBUG,
-			    "allocate mempool %s reserved %u (size %u)",
+			    "allocate mempool %s reserved %u (size %u, mbuf_size %u)",
 			    mp_name,
 			    count,
-			    alloc_size);
+			    alloc_size,
+			    mbuf_size);
 			mt->mp = rte_pktmbuf_pool_create(
 				mp_name,
 				alloc_size,
 				RTE_MEMPOOL_CACHE_MAX_SIZE,
 				GR_MBUF_PRIV_MAX_SIZE,
-				RTE_MBUF_DEFAULT_BUF_SIZE,
+				mbuf_size,
 				socket_id
 			);
 			if (mt->mp == NULL)
@@ -73,11 +82,12 @@ struct rte_mempool *gr_pktmbuf_pool_get(int8_t socket_id, uint32_t count) {
 			break;
 		} else if ((count + mt->reserved) <= mt->mp->size) {
 			LOG(DEBUG,
-			    "reuse mempool %s reserved %u -> %u (size %u)",
+			    "reuse mempool %s reserved %u -> %u (size %u, mbuf_size %u)",
 			    mt->mp->name,
 			    mt->reserved,
 			    mt->reserved + count,
-			    mt->mp->size);
+			    mt->mp->size,
+			    mbuf_size);
 			mt->reserved += count;
 			mp = mt->mp;
 			break;

modules/infra/control/port.c

@@ -14,6 +14,7 @@
 #include <gr_port.h>
 #include <gr_queue.h>
 #include <gr_vec.h>
+#include <gr_vlan.h>
 #include <gr_worker.h>
 
 #include <numa.h>
@@ -72,12 +73,7 @@ static struct rte_eth_conf default_port_config = {
 		},
 	},
 	.rxmode = {
-		.offloads = RTE_ETH_RX_OFFLOAD_CHECKSUM
-			| RTE_ETH_RX_OFFLOAD_VLAN
-			| RTE_ETH_RX_OFFLOAD_SCATTER,
-	},
-	.txmode = {
-		.offloads = RTE_ETH_TX_OFFLOAD_MULTI_SEGS,
+		.offloads = RTE_ETH_RX_OFFLOAD_CHECKSUM | RTE_ETH_RX_OFFLOAD_VLAN,
 	},
 };
 
@@ -128,7 +124,6 @@ int port_configure(struct iface_info_port *p, uint16_t n_txq_min) {
 	else
 		conf.rxmode.mq_mode = RTE_ETH_MQ_RX_RSS;
 	conf.rxmode.offloads &= info.rx_offload_capa;
-	conf.txmode.offloads &= info.tx_offload_capa;
 	if (info.dev_flags != NULL && *info.dev_flags & RTE_ETH_DEV_INTR_LSC) {
 		conf.intr_conf.lsc = 1;
 	}
@@ -162,6 +157,9 @@ static int iface_port_reconfig(
 	bool needs_configure = false;
 	int ret;
 
+	if ((set_attrs & GR_IFACE_SET_MTU) && conf->mtu > gr_config.max_mtu)
+		return errno_set(ERANGE);
+
 	if ((ret = port_unplug(p->port_id)) < 0)
 		return ret;
 
@@ -254,6 +252,13 @@ static int iface_port_reconfig(
 			return errno_log(-ret, "rte_eth_dev_get_mtu");
 	}
 
+	struct iface *v = NULL;
+	while ((v = iface_next(GR_IFACE_TYPE_VLAN, v)) != NULL) {
+		struct iface_info_vlan *vlan = (struct iface_info_vlan *)v->info;
+		if (vlan->parent_id == iface->id)
+			v->mtu = iface->mtu;
+	}
+
 	if (set_attrs & GR_IFACE_SET_MODE)
 		iface->mode = conf->mode;
 

modules/infra/control/vlan.c

@@ -124,10 +124,9 @@ static int iface_vlan_reconfig(
 
 	if (set_attrs & GR_IFACE_SET_FLAGS)
 		iface->flags = conf->flags;
-	if (set_attrs & GR_IFACE_SET_MTU)
-		iface->mtu = conf->mtu ? conf->mtu : iface_from_id(cur->parent_id)->mtu;
 	if (set_attrs & GR_IFACE_SET_VRF)
 		iface->vrf_id = conf->vrf_id;
+	iface->mtu = iface_from_id(cur->parent_id)->mtu;
 
 	gr_event_push(GR_EVENT_IFACE_POST_RECONFIG, iface);
 

modules/ip/datapath/ip_output.c

@@ -25,6 +25,7 @@ enum {
 	HOLD,
 	NO_ROUTE,
 	ERROR,
+	TOO_BIG,
 	EDGE_COUNT,
 };
 
@@ -83,6 +84,12 @@ ip_output_process(struct rte_graph *graph, struct rte_node *node, void **objs, u
 			edge = ERROR;
 			goto next;
 		}
+
+		if (rte_pktmbuf_pkt_len(mbuf) > iface->mtu) {
+			edge = TOO_BIG;
+			goto next;
+		}
+
 		// Determine what is the next node based on the output interface type
 		// By default, it will be eth_output unless another output node was registered.
 		edge = iface_type_edges[iface->type];
@@ -133,6 +140,7 @@ static struct rte_node_register output_node = {
 		[HOLD] = "ip_hold",
 		[NO_ROUTE] = "ip_error_dest_unreach",
 		[ERROR] = "ip_output_error",
+		[TOO_BIG] = "ip_output_too_big",
 	},
 };
 
@@ -144,3 +152,4 @@ static struct gr_node_info info = {
 GR_NODE_REGISTER(info);
 
 GR_DROP_REGISTER(ip_output_error);
+GR_DROP_REGISTER(ip_output_too_big);

modules/ip6/datapath/ip6_output.c

@@ -21,6 +21,7 @@ enum {
 	HOLD,
 	DEST_UNREACH,
 	ERROR,
+	TOO_BIG,
 	EDGE_COUNT,
 };
 
@@ -83,6 +84,12 @@ ip6_output_process(struct rte_graph *graph, struct rte_node *node, void **objs,
 			edge = ERROR;
 			goto next;
 		}
+
+		if (rte_pktmbuf_pkt_len(mbuf) > iface->mtu) {
+			edge = TOO_BIG;
+			goto next;
+		}
+
 		// Determine what is the next node based on the output interface type
 		// By default, it will be eth_output unless another output node was registered.
 		edge = iface_type_edges[iface->type];
@@ -128,6 +135,7 @@ static struct rte_node_register output_node = {
 		[HOLD] = "ip6_hold",
 		[ERROR] = "ip6_output_error",
 		[DEST_UNREACH] = "ip6_error_dest_unreach",
+		[TOO_BIG] = "ip6_output_too_big",
 	},
 };
 
@@ -139,3 +147,4 @@ static struct gr_node_info info = {
 GR_NODE_REGISTER(info);
 
 GR_DROP_REGISTER(ip6_output_error);
+GR_DROP_REGISTER(ip6_output_too_big);

modules/ipip/control.c

@@ -102,7 +102,7 @@ static int iface_ipip_fini(struct iface *iface) {
 static int iface_ipip_init(struct iface *iface, const void *api_info) {
 	const struct gr_iface conf = {
 		.flags = iface->flags,
-		.mtu = iface->mtu,
+		.mtu = iface->mtu ?: 1480,
 		.mode = iface->mode,
 		.vrf_id = iface->vrf_id
 	};

smoke/ip_forward_test.sh

@@ -7,8 +7,8 @@
 p0=${run_id}0
 p1=${run_id}1
 
-grcli add interface port $p0 devargs net_tap0,iface=$p0 mac f0:0d:ac:dc:00:00 mtu 9000
-grcli add interface port $p1 devargs net_tap1,iface=$p1 mac f0:0d:ac:dc:00:01 mtu 9000
+grcli add interface port $p0 devargs net_tap0,iface=$p0 mac f0:0d:ac:dc:00:00
+grcli add interface port $p1 devargs net_tap1,iface=$p1 mac f0:0d:ac:dc:00:01
 grcli add ip address 172.16.0.1/24 iface $p0
 grcli add ip address 172.16.1.1/24 iface $p1
 grcli add ip route 16.0.0.0/16 via 172.16.0.2
@@ -19,7 +19,7 @@ for n in 0 1; do
 	p=$run_id$n
 	netns_add $p
 	ip link set $p netns $p
-	ip -n $p link set $p address ba:d0:ca:ca:00:0$n mtu 9000
+	ip -n $p link set $p address ba:d0:ca:ca:00:0$n
 	ip -n $p link set $p up
 	ip -n $p link set lo up
 	ip -n $p addr add 172.16.$n.2/24 dev $p
@@ -28,8 +28,8 @@ for n in 0 1; do
 	ip -n $p addr show
 done
 
-ip netns exec $p0 ping -i0.01 -c3 -n 16.1.0.1 -s 5000
-ip netns exec $p1 ping -i0.01 -c3 -n 16.0.0.1 -s 5000
+ip netns exec $p0 ping -i0.01 -c3 -n 16.1.0.1
+ip netns exec $p1 ping -i0.01 -c3 -n 16.0.0.1
 ip netns exec $p0 ping -i0.01 -c3 -n 172.16.1.2
 ip netns exec $p1 ping -i0.01 -c3 -n 172.16.0.2
 ip netns exec $p0 ping -i0.01 -c3 -n 172.16.0.1