dpdk: lro support
Change-Id: Ifa09dcdfd567ac3c2afab4bd3bd468260f766e1d
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
Type: improvement
diff --git a/src/plugins/dpdk/device/dpdk.h b/src/plugins/dpdk/device/dpdk.h
index 9df7f39..dd43137 100644
--- a/src/plugins/dpdk/device/dpdk.h
+++ b/src/plugins/dpdk/device/dpdk.h
@@ -251,11 +251,12 @@
#define DPDK_LINK_POLL_INTERVAL (3.0)
#define DPDK_MIN_LINK_POLL_INTERVAL (0.001) /* 1msec */
-#define foreach_dpdk_device_config_item \
- _ (num_rx_queues) \
- _ (num_tx_queues) \
- _ (num_rx_desc) \
- _ (num_tx_desc) \
+#define foreach_dpdk_device_config_item \
+ _ (num_rx_queues) \
+ _ (num_tx_queues) \
+ _ (num_rx_desc) \
+ _ (num_tx_desc) \
+ _ (max_lro_pkt_size) \
_ (rss_fn)
typedef enum
@@ -301,6 +302,7 @@
u8 *eal_init_args_str;
u8 *uio_driver_name;
u8 no_multi_seg;
+ u8 enable_lro;
u8 enable_tcp_udp_checksum;
u8 no_tx_checksum_offload;
u8 enable_telemetry;
@@ -347,7 +349,7 @@
u32 buffers[DPDK_RX_BURST_SZ];
u16 next[DPDK_RX_BURST_SZ];
u16 etype[DPDK_RX_BURST_SZ];
- u16 flags[DPDK_RX_BURST_SZ];
+ u32 flags[DPDK_RX_BURST_SZ];
vlib_buffer_t buffer_template;
} dpdk_per_thread_data_t;
diff --git a/src/plugins/dpdk/device/dpdk_priv.h b/src/plugins/dpdk/device/dpdk_priv.h
index a5a8a2a..7f595ae 100644
--- a/src/plugins/dpdk/device/dpdk_priv.h
+++ b/src/plugins/dpdk/device/dpdk_priv.h
@@ -15,6 +15,7 @@
#define DPDK_NB_RX_DESC_DEFAULT 1024
#define DPDK_NB_TX_DESC_DEFAULT 1024
+#define DPDK_MAX_LRO_SIZE_DEFAULT 65536
#define DPDK_NB_RX_DESC_VIRTIO 256
#define DPDK_NB_TX_DESC_VIRTIO 256
diff --git a/src/plugins/dpdk/device/format.c b/src/plugins/dpdk/device/format.c
index 71f0d89..eb98730 100644
--- a/src/plugins/dpdk/device/format.c
+++ b/src/plugins/dpdk/device/format.c
@@ -60,6 +60,7 @@
_ (PKT_RX_L4_CKSUM_GOOD, "L4 cksum of RX pkt. is valid") \
_ (PKT_RX_IEEE1588_PTP, "RX IEEE1588 L2 Ethernet PT Packet") \
_ (PKT_RX_IEEE1588_TMST, "RX IEEE1588 L2/L4 timestamped packet") \
+ _ (PKT_RX_LRO, "LRO packet") \
_ (PKT_RX_QINQ_STRIPPED, "RX packet QinQ tags stripped")
#define foreach_dpdk_pkt_type \
diff --git a/src/plugins/dpdk/device/init.c b/src/plugins/dpdk/device/init.c
index f923da6..8c49ab8 100644
--- a/src/plugins/dpdk/device/init.c
+++ b/src/plugins/dpdk/device/init.c
@@ -372,6 +372,19 @@
xd->port_conf.rxmode.offloads |= DEV_RX_OFFLOAD_TCP_CKSUM;
}
+ if (dm->conf->enable_lro)
+ {
+ if (dev_info.rx_offload_capa & DEV_RX_OFFLOAD_TCP_LRO)
+ {
+ xd->port_conf.rxmode.offloads |= DEV_RX_OFFLOAD_TCP_LRO;
+ if (devconf->max_lro_pkt_size)
+ xd->port_conf.rxmode.max_lro_pkt_size =
+ devconf->max_lro_pkt_size;
+ else
+ xd->port_conf.rxmode.max_lro_pkt_size =
+ DPDK_MAX_LRO_SIZE_DEFAULT;
+ }
+ }
if (dm->conf->no_multi_seg)
{
xd->port_conf.txmode.offloads &= ~DEV_TX_OFFLOAD_MULTI_SEGS;
@@ -474,9 +487,16 @@
xd->port_conf.intr_conf.rxq = 1;
break;
+ case VNET_DPDK_PMD_MLX5:
+ if (dm->conf->no_tx_checksum_offload == 0)
+ {
+ xd->port_conf.txmode.offloads |= DEV_TX_OFFLOAD_TCP_CKSUM;
+ xd->port_conf.txmode.offloads |= DEV_TX_OFFLOAD_UDP_CKSUM;
+ xd->flags |= DPDK_DEVICE_FLAG_TX_OFFLOAD |
+ DPDK_DEVICE_FLAG_INTEL_PHDR_CKSUM;
+ }
case VNET_DPDK_PMD_CXGBE:
case VNET_DPDK_PMD_MLX4:
- case VNET_DPDK_PMD_MLX5:
case VNET_DPDK_PMD_QEDE:
case VNET_DPDK_PMD_BNXT:
xd->port_type = port_type_from_speed_capa (&dev_info);
@@ -1270,6 +1290,9 @@
else if (unformat (input, "rss-queues %U",
unformat_bitmap_list, &devconf->rss_queues))
;
+ else if (unformat (input, "max-lro-pkt-size %u",
+ &devconf->max_lro_pkt_size))
+ ;
else
{
error = clib_error_return (0, "unknown input `%U'",
@@ -1376,6 +1399,8 @@
else if (unformat (input, "no-multi-seg"))
conf->no_multi_seg = 1;
+ else if (unformat (input, "enable-lro"))
+ conf->enable_lro = 1;
else if (unformat (input, "max-simd-bitwidth %U",
unformat_max_simd_bitwidth, &conf->max_simd_bitwidth))
;
diff --git a/src/plugins/dpdk/device/node.c b/src/plugins/dpdk/device/node.c
index 2522285..c725681 100644
--- a/src/plugins/dpdk/device/node.c
+++ b/src/plugins/dpdk/device/node.c
@@ -36,9 +36,10 @@
#undef _
};
-/* make sure all flags we need are stored in lower 8 bits */
-STATIC_ASSERT ((PKT_RX_IP_CKSUM_BAD | PKT_RX_FDIR) <
- 256, "dpdk flags not un lower byte, fix needed");
+/* make sure all flags we need are stored in lower 32 bits */
+STATIC_ASSERT ((u64) (PKT_RX_IP_CKSUM_BAD | PKT_RX_FDIR | PKT_RX_LRO) <
+ (1ULL << 32),
+ "dpdk flags not in lower word, fix needed");
static_always_inline uword
dpdk_process_subseq_segs (vlib_main_t * vm, vlib_buffer_t * b,
@@ -145,31 +146,30 @@
<code>xd->per_interface_next_index</code>
*/
-static_always_inline u16
-dpdk_ol_flags_extract (struct rte_mbuf **mb, u16 * flags, int count)
+static_always_inline u32
+dpdk_ol_flags_extract (struct rte_mbuf **mb, u32 *flags, int count)
{
- u16 rv = 0;
+ u32 rv = 0;
int i;
for (i = 0; i < count; i++)
{
/* all flags we are interested in are in lower 8 bits but
that might change */
- flags[i] = (u16) mb[i]->ol_flags;
+ flags[i] = (u32) mb[i]->ol_flags;
rv |= flags[i];
}
return rv;
}
static_always_inline uword
-dpdk_process_rx_burst (vlib_main_t * vm, dpdk_per_thread_data_t * ptd,
- uword n_rx_packets, int maybe_multiseg,
- u16 * or_flagsp)
+dpdk_process_rx_burst (vlib_main_t *vm, dpdk_per_thread_data_t *ptd,
+ uword n_rx_packets, int maybe_multiseg, u32 *or_flagsp)
{
u32 n_left = n_rx_packets;
vlib_buffer_t *b[4];
struct rte_mbuf **mb = ptd->mbufs;
uword n_bytes = 0;
- u16 *flags, or_flags = 0;
+ u32 *flags, or_flags = 0;
vlib_buffer_t bt;
mb = ptd->mbufs;
@@ -277,6 +277,65 @@
}
}
+static_always_inline u16
+dpdk_lro_find_l4_hdr_sz (vlib_buffer_t *b)
+{
+ u16 l4_hdr_sz = 0;
+ u16 current_offset = 0;
+ ethernet_header_t *e;
+ tcp_header_t *tcp;
+ u8 *data = vlib_buffer_get_current (b);
+ u16 ethertype;
+ e = (void *) data;
+ current_offset += sizeof (e[0]);
+ ethertype = clib_net_to_host_u16 (e->type);
+ if (ethernet_frame_is_tagged (ethertype))
+ {
+ ethernet_vlan_header_t *vlan = (ethernet_vlan_header_t *) (e + 1);
+ ethertype = clib_net_to_host_u16 (vlan->type);
+ current_offset += sizeof (*vlan);
+ if (ethertype == ETHERNET_TYPE_VLAN)
+ {
+ vlan++;
+ current_offset += sizeof (*vlan);
+ ethertype = clib_net_to_host_u16 (vlan->type);
+ }
+ }
+ data += current_offset;
+ if (ethertype == ETHERNET_TYPE_IP4)
+ {
+ data += sizeof (ip4_header_t);
+ tcp = (void *) data;
+ l4_hdr_sz = tcp_header_bytes (tcp);
+ }
+ else
+ {
+ /* FIXME: extension headers...*/
+ data += sizeof (ip6_header_t);
+ tcp = (void *) data;
+ l4_hdr_sz = tcp_header_bytes (tcp);
+ }
+ return l4_hdr_sz;
+}
+
+static_always_inline void
+dpdk_process_lro_offload (dpdk_device_t *xd, dpdk_per_thread_data_t *ptd,
+ uword n_rx_packets)
+{
+ uword n;
+ vlib_buffer_t *b0;
+ for (n = 0; n < n_rx_packets; n++)
+ {
+ b0 = vlib_buffer_from_rte_mbuf (ptd->mbufs[n]);
+ if (ptd->flags[n] & PKT_RX_LRO)
+ {
+ b0->flags |= VNET_BUFFER_F_GSO;
+ vnet_buffer2 (b0)->gso_size = ptd->mbufs[n]->tso_segsz;
+ vnet_buffer2 (b0)->gso_l4_hdr_sz = dpdk_lro_find_l4_hdr_sz (b0);
+ }
+ }
+}
+
static_always_inline u32
dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
vlib_node_runtime_t * node, u32 thread_index, u16 queue_id)
@@ -289,7 +348,7 @@
struct rte_mbuf **mb;
vlib_buffer_t *b0;
u16 *next;
- u16 or_flags;
+ u32 or_flags;
u32 n;
int single_next = 0;
@@ -339,6 +398,9 @@
else
n_rx_bytes = dpdk_process_rx_burst (vm, ptd, n_rx_packets, 0, &or_flags);
+ if (PREDICT_FALSE ((or_flags & PKT_RX_LRO)))
+ dpdk_process_lro_offload (xd, ptd, n_rx_packets);
+
if (PREDICT_FALSE (or_flags & PKT_RX_FDIR))
{
/* some packets will need to go to different next nodes */