mlx5: support checksum offload
authorAdrien Mazarguil <adrien.mazarguil@6wind.com>
Fri, 30 Oct 2015 18:52:41 +0000 (19:52 +0100)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Fri, 30 Oct 2015 21:41:01 +0000 (22:41 +0100)
This is the same implementation as mlx4.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
drivers/net/mlx5/mlx5_rxq.c
drivers/net/mlx5/mlx5_rxtx.c
drivers/net/mlx5/mlx5_rxtx.h
drivers/net/mlx5/mlx5_utils.h

index 1cd28c2..5a55886 100644 (file)
@@ -565,6 +565,15 @@ rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
        /* Number of descriptors and mbufs currently allocated. */
        desc_n = (tmpl.elts_n * (tmpl.sp ? MLX5_PMD_SGE_WR_N : 1));
        mbuf_n = desc_n;
+       /* Toggle RX checksum offload if hardware supports it. */
+       if (priv->hw_csum) {
+               tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
+               rxq->csum = tmpl.csum;
+       }
+       if (priv->hw_csum_l2tun) {
+               tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
+               rxq->csum_l2tun = tmpl.csum_l2tun;
+       }
        /* Enable scattered packets support for this queue if necessary. */
        if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
            (dev->data->dev_conf.rxmode.max_rx_pkt_len >
@@ -785,6 +794,11 @@ rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
                rte_pktmbuf_tailroom(buf)) == tmpl.mb_len);
        assert(rte_pktmbuf_headroom(buf) == RTE_PKTMBUF_HEADROOM);
        rte_pktmbuf_free(buf);
+       /* Toggle RX checksum offload if hardware supports it. */
+       if (priv->hw_csum)
+               tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
+       if (priv->hw_csum_l2tun)
+               tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
        /* Enable scattered packets support for this queue if necessary. */
        if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
            (dev->data->dev_conf.rxmode.max_rx_pkt_len >
index eccbbb9..623219d 100644 (file)
@@ -390,6 +390,17 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        ++elts_comp;
                        send_flags |= IBV_EXP_QP_BURST_SIGNALED;
                }
+               /* Should we enable HW CKSUM offload */
+               if (buf->ol_flags &
+                   (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+                       send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
+                       /* HW does not support checksum offloads at arbitrary
+                        * offsets but automatically recognizes the packet
+                        * type. For inner L3/L4 checksums, only VXLAN (UDP)
+                        * tunnels are currently supported. */
+                       if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
+                               send_flags |= IBV_EXP_QP_BURST_TUNNEL;
+               }
                if (likely(segs == 1)) {
                        uintptr_t addr;
                        uint32_t length;
@@ -490,6 +501,85 @@ stop:
        return i;
 }
 
+/**
+ * Translate RX completion flags to packet type.
+ *
+ * @param flags
+ *   RX completion flags returned by poll_length_flags().
+ *
+ * @return
+ *   Packet type for struct rte_mbuf.
+ */
+static inline uint32_t
+rxq_cq_to_pkt_type(uint32_t flags)
+{
+       uint32_t pkt_type;
+
+       if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
+               pkt_type =
+                       TRANSPOSE(flags,
+                                 IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
+                                 RTE_PTYPE_L3_IPV4) |
+                       TRANSPOSE(flags,
+                                 IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
+                                 RTE_PTYPE_L3_IPV6) |
+                       TRANSPOSE(flags,
+                                 IBV_EXP_CQ_RX_IPV4_PACKET,
+                                 RTE_PTYPE_INNER_L3_IPV4) |
+                       TRANSPOSE(flags,
+                                 IBV_EXP_CQ_RX_IPV6_PACKET,
+                                 RTE_PTYPE_INNER_L3_IPV6);
+       else
+               pkt_type =
+                       TRANSPOSE(flags,
+                                 IBV_EXP_CQ_RX_IPV4_PACKET,
+                                 RTE_PTYPE_L3_IPV4) |
+                       TRANSPOSE(flags,
+                                 IBV_EXP_CQ_RX_IPV6_PACKET,
+                                 RTE_PTYPE_L3_IPV6);
+       return pkt_type;
+}
+
+/**
+ * Translate RX completion flags to offload flags.
+ *
+ * @param[in] rxq
+ *   Pointer to RX queue structure.
+ * @param flags
+ *   RX completion flags returned by poll_length_flags().
+ *
+ * @return
+ *   Offload flags (ol_flags) for struct rte_mbuf.
+ */
+static inline uint32_t
+rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
+{
+       uint32_t ol_flags = 0;
+
+       if (rxq->csum)
+               ol_flags |=
+                       TRANSPOSE(~flags,
+                                 IBV_EXP_CQ_RX_IP_CSUM_OK,
+                                 PKT_RX_IP_CKSUM_BAD) |
+                       TRANSPOSE(~flags,
+                                 IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
+                                 PKT_RX_L4_CKSUM_BAD);
+       /*
+        * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
+        * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
+        * (its value is 0).
+        */
+       if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
+               ol_flags |=
+                       TRANSPOSE(~flags,
+                                 IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
+                                 PKT_RX_IP_CKSUM_BAD) |
+                       TRANSPOSE(~flags,
+                                 IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
+                                 PKT_RX_L4_CKSUM_BAD);
+       return ol_flags;
+}
+
 /**
  * DPDK callback for RX with scattered packets support.
  *
@@ -669,6 +759,8 @@ mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
                NB_SEGS(pkt_buf) = j;
                PORT(pkt_buf) = rxq->port_id;
                PKT_LEN(pkt_buf) = pkt_buf_len;
+               pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
+               pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 
                /* Return packet. */
                *(pkts++) = pkt_buf;
@@ -829,6 +921,8 @@ mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
                NEXT(seg) = NULL;
                PKT_LEN(seg) = len;
                DATA_LEN(seg) = len;
+               seg->packet_type = rxq_cq_to_pkt_type(flags);
+               seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 
                /* Return packet. */
                *(pkts++) = seg;
index 521aee0..d86d623 100644 (file)
@@ -116,6 +116,8 @@ struct rxq {
                struct rxq_elt (*no_sp)[]; /* RX elements. */
        } elts;
        unsigned int sp:1; /* Use scattered RX elements. */
+       unsigned int csum:1; /* Enable checksum offloading. */
+       unsigned int csum_l2tun:1; /* Same for L2 tunnels. */
        uint32_t mb_len; /* Length of a mp-issued mbuf. */
        struct mlx5_rxq_stats stats; /* RX queue counters. */
        unsigned int socket; /* CPU socket ID for allocations. */
index e48e6b6..8ff075b 100644 (file)
@@ -149,6 +149,12 @@ pmd_drv_log_basename(const char *s)
 #define NB_SEGS(m) ((m)->nb_segs)
 #define PORT(m) ((m)->port)
 
+/* Transpose flags. Useful to convert IBV to DPDK flags. */
+#define TRANSPOSE(val, from, to) \
+       (((from) >= (to)) ? \
+        (((val) & (from)) / ((from) / (to))) : \
+        (((val) & (from)) * ((to) / (from))))
+
 /* Allocate a buffer on the stack and fill it with a printf format string. */
 #define MKSTR(name, ...) \
        char name[snprintf(NULL, 0, __VA_ARGS__) + 1]; \