Make all mbufs single segment

cps/kni.c

@@ -44,10 +44,15 @@ static int
 setup_dpdk_interface(struct cps_kni *kni, const struct gatekeeper_if *iface,
 	struct rte_mempool *mp, uint16_t queue_size)
 {
+	/*
+	 * DO NOT add RTE_ETH_RX_OFFLOAD_SCATTER to port_conf.rxmode.offloads
+	 * to guarantee that all mbufs have a single segment.
+	 *
+	 * If more dataroom in mbufs is needed, review create_pktmbuf_pool().
+	 */
 	struct rte_eth_conf port_conf = {
 		.rxmode = {
 			.mtu = iface->mtu,
-			.offloads = RTE_ETH_RX_OFFLOAD_SCATTER,
 		},
 	};
 

cps/main.c

@@ -808,7 +808,7 @@ assign_cps_queue_ids(struct cps_config *cps_conf)
 	num_mbuf = calculate_mempool_config_para("cps",
 		cps_conf->net, total_pkt_burst);
 	cps_conf->mp = create_pktmbuf_pool("cps",
-		cps_conf->lcore_id, num_mbuf);
+		cps_conf->lcore_id, num_mbuf, cps_conf->net);
 	if (cps_conf->mp == NULL) {
 		ret = -1;
 		goto fail;

ggu/main.c

@@ -633,7 +633,7 @@ ggu_stage1(void *arg)
 			ggu_conf->net, ggu_conf->total_pkt_burst);
 
 		ggu_conf->mp = create_pktmbuf_pool("ggu",
-			ggu_conf->lcore_id, num_mbuf);
+			ggu_conf->lcore_id, num_mbuf, ggu_conf->net);
 		if (ggu_conf->mp == NULL)
 			return -1;
 

gk/main.c

@@ -2550,7 +2550,8 @@ gk_stage1(void *arg)
 		unsigned int lcore = gk_conf->lcores[i];
 		struct gk_instance *inst_ptr = &gk_conf->instances[i];
 
-		inst_ptr->mp = create_pktmbuf_pool("gk", lcore, num_mbuf);
+		inst_ptr->mp = create_pktmbuf_pool("gk", lcore, num_mbuf,
+			gk_conf->net);
 		if (inst_ptr->mp == NULL)
 			goto cleanup;
 

gt/main.c

@@ -2075,7 +2075,8 @@ init_gt_instances(struct gt_config *gt_conf)
 		unsigned int lcore = gt_conf->lcores[i];
 		inst_ptr = &gt_conf->instances[i];
 
-		inst_ptr->mp = create_pktmbuf_pool("gt", lcore, num_mbuf);
+		inst_ptr->mp = create_pktmbuf_pool("gt", lcore, num_mbuf,
+			gt_conf->net);
 		if (inst_ptr->mp == NULL) {
 			ret = -1;
 			goto free_gt_instance;

include/gatekeeper_net.h

@@ -241,6 +241,16 @@ struct gatekeeper_if {
 	/* Whether IPv4 (L3) checksums should be enabled in hardware. */
 	bool		ipv4_hw_cksum;
 
+	/*
+	 * XXX #730 Due to a bug in some Poll Mode Drivers (PMD),
+	 * the MTU parameter passed during the initialization must include
+	 * length of the link layer header(s).
+	 *
+	 * The ICE drivers is known to require this workaround.
+	 * See check_if_mtu() for mare information.
+	 */
+	bool		pmd_mtu_workaround;
+
 	/*
 	 * This field decides if the flag GRND_RANDOM is passed to getradom(2)
 	 * while initializing field @rss_key.
@@ -609,7 +619,8 @@ bool ipv6_configured(struct net_config *net_conf);
 unsigned int calculate_mempool_config_para(const char *block_name,
 	struct net_config *net_conf, unsigned int total_pkt_burst);
 struct rte_mempool *create_pktmbuf_pool(const char *block_name,
-	unsigned int lcore, unsigned int num_mbuf);
+	unsigned int lcore, unsigned int num_mbuf,
+	const struct net_config *net);
 
 /*
  * No cleanup for this step, since DPDK

lib/net.c

@@ -26,6 +26,7 @@
 #include <pwd.h>
 #include <grp.h>
 #include <net/if.h>
+#include <linux/virtio_net.h> /* For sizeof(struct virtio_net_hdr_mrg_rxbuf) */
 
 #include <rte_mbuf.h>
 #include <rte_thash.h>
@@ -1362,12 +1363,21 @@ check_if_mtu(struct gatekeeper_if *iface,
 	/*
 	 * Set up device MTU.
 	 *
-	 * If greater than the size of the mbufs, then add the
-	 * multi-segment buffer flag.
+	 * DO NOT add RTE_ETH_TX_OFFLOAD_MULTI_SEGS to
+	 * port_conf->txmode.offloads because all mbufs in Gatekeeper are
+	 * single segment and RTE_ETH_TX_OFFLOAD_MULTI_SEGS may disable
+	 * vectorial transmission; see ice_set_tx_function() in the ICE driver
+	 * for an example.
+	 *
+	 * If more dataroom in mbufs is needed, review create_pktmbuf_pool().
 	 */
 	port_conf->rxmode.mtu = iface->mtu;
-	if (iface->mtu > RTE_MBUF_DEFAULT_BUF_SIZE)
-		port_conf->txmode.offloads |= RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
+
+	if (iface->pmd_mtu_workaround) {
+		port_conf->rxmode.mtu += RTE_ETHER_HDR_LEN; /* Ethernet */
+		if (iface->vlan_insert)
+			port_conf->rxmode.mtu += RTE_VLAN_HLEN; /* VLAN */
+	}
 
 	if (unlikely(dev_info->min_mtu > port_conf->rxmode.mtu)) {
 		G_LOG(ERR, "%s(%s): the minimum MTU %u is larger than the configured MTU %"PRIu32"\n",
@@ -1383,15 +1393,6 @@ check_if_mtu(struct gatekeeper_if *iface,
 		return -EINVAL;
 	}
 
-	if (unlikely((port_conf->txmode.offloads &
-				RTE_ETH_TX_OFFLOAD_MULTI_SEGS) &&
-			!(dev_info->tx_offload_capa &
-			RTE_ETH_TX_OFFLOAD_MULTI_SEGS))) {
-		G_LOG(NOTICE, "%s(%s): interface does not support offloading multi-segment TX buffers\n",
-			__func__, iface->name);
-		port_conf->txmode.offloads &= ~RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
-	}
-
 	return 0;
 }
 
@@ -2378,17 +2379,57 @@ calculate_mempool_config_para(const char *block_name,
 	return num_mbuf;
 }
 
+#define LINK_LAYER_HEADROOM (32)
+
+/*
+ * Computes the dataroom size needed for struct mbuf's to be used at front,
+ * back, and KNI interfaces.
+ */
+static uint16_t
+dataroom_from_net(const struct net_config *net)
+{
+	uint16_t max_mtu = net->front.mtu;
+	if (net->back_iface_enabled && net->back.mtu > max_mtu)
+		max_mtu = net->back.mtu;
+
+	/*
+	 * Even Ethernet NICs do not agree on the exact headroom for
+	 * the link layer, so the test below favors Virtio (i.e., KNI
+	 * interfaces) because it is the most demaing part.
+	 * See virtio_ethdev.c:virtio_mtu_set() for details.
+	 */
+	RTE_BUILD_BUG_ON(LINK_LAYER_HEADROOM < (
+			RTE_ETHER_HDR_LEN	+ /* Ethernet */
+			RTE_VLAN_HLEN		+ /* VLAN */
+			sizeof(struct virtio_net_hdr_mrg_rxbuf) /* Virtio */
+		)
+	);
+
+	/*
+	 * Having RTE_MBUF_DEFAULT_BUF_SIZE in the RTE_MAX() below
+	 * guarantees the minimum size that some NICs require to avoid
+	 * splitting frames into multiple segments.
+	 *
+	 * In the second expression:
+	 *   @RTE_PKTMBUF_HEADROOM is required by DPDK libraries.
+	 *   @LINK_LAYER_HEADROOM accounts for link layers headers.
+	 *   @max_mtu makes the expression valid for front and back interfaces.
+	 */
+	return RTE_MAX(RTE_MBUF_DEFAULT_BUF_SIZE,
+		RTE_PKTMBUF_HEADROOM + LINK_LAYER_HEADROOM + max_mtu);
+}
+
 struct rte_mempool *
 create_pktmbuf_pool(const char *block_name, unsigned int lcore,
-	unsigned int num_mbuf)
+	unsigned int num_mbuf, const struct net_config *net)
 {
 	struct rte_mempool *mp;
 	char pool_name[64];
 	int ret = snprintf(pool_name, sizeof(pool_name), "pktmbuf_pool_%s_%u",
 		block_name, lcore);
 	RTE_VERIFY(ret > 0 && ret < (int)sizeof(pool_name));
 	mp = rte_pktmbuf_pool_create_by_ops(pool_name, num_mbuf, 0,
-		sizeof(struct sol_mbuf_priv), RTE_MBUF_DEFAULT_BUF_SIZE,
+		sizeof(struct sol_mbuf_priv), dataroom_from_net(net),
 		rte_lcore_to_socket_id(lcore), "ring_mp_sc");
 	if (mp == NULL) {
 		G_LOG(ERR,

lls/main.c

@@ -836,7 +836,7 @@ assign_lls_queue_ids(struct lls_config *lls_conf)
 	num_mbuf = calculate_mempool_config_para("lls",
 		lls_conf->net, total_pkt_burst);
 	lls_conf->mp = create_pktmbuf_pool("lls",
-		lls_conf->lcore_id, num_mbuf);
+		lls_conf->lcore_id, num_mbuf, lls_conf->net);
 	if (lls_conf->mp == NULL) {
 		ret = -1;
 		goto fail;

lua/gatekeeper/staticlib.lua

@@ -283,6 +283,7 @@ struct gatekeeper_if {
 	bool     ipv4_hw_udp_cksum;
 	bool     ipv6_hw_udp_cksum;
 	bool     ipv4_hw_cksum;
+	bool	 pmd_mtu_workaround;
 	bool     guarantee_random_entropy;
 	bool     alternative_rss_hash;
 	/* This struct has hidden fields. */

lua/net.lua

@@ -46,6 +46,8 @@ return function (gatekeeper_server)
 	local back_ipv6_hw_udp_cksum = true
 	local front_ipv4_hw_cksum = true
 	local back_ipv4_hw_cksum = true
+	local front_pmd_mtu_workaround = false
+	local back_pmd_mtu_workaround = false
 	local front_alternative_rss_hash = false
 	local back_alternative_rss_hash = false
 
@@ -75,6 +77,7 @@ return function (gatekeeper_server)
 	front_iface.ipv4_hw_udp_cksum = front_ipv4_hw_udp_cksum
 	front_iface.ipv6_hw_udp_cksum = front_ipv6_hw_udp_cksum
 	front_iface.ipv4_hw_cksum = front_ipv4_hw_cksum
+	front_iface.pmd_mtu_workaround = front_pmd_mtu_workaround
 	front_iface.guarantee_random_entropy = guarantee_random_entropy
 	front_iface.alternative_rss_hash = front_alternative_rss_hash
 	local ret = staticlib.init_iface(front_iface, "front",
@@ -99,6 +102,7 @@ return function (gatekeeper_server)
 		back_iface.ipv4_hw_udp_cksum = back_ipv4_hw_udp_cksum
 		back_iface.ipv6_hw_udp_cksum = back_ipv6_hw_udp_cksum
 		back_iface.ipv4_hw_cksum = back_ipv4_hw_cksum
+		back_iface.pmd_mtu_workaround = back_pmd_mtu_workaround
 		back_iface.guarantee_random_entropy = guarantee_random_entropy
 		back_iface.alternative_rss_hash = back_alternative_rss_hash
 		ret = staticlib.init_iface(back_iface, "back",