net/mlx5: support inline send
[dpdk.git] / drivers / net / mlx5 / mlx5_rxtx.c
index 1e9972b..2ca133e 100644 (file)
@@ -375,6 +375,139 @@ mlx5_wqe_write_vlan(struct txq *txq, volatile union mlx5_wqe *wqe,
        ++txq->wqe_ci;
 }
 
+/**
+ * Write a inline WQE.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param wqe
+ *   Pointer to the WQE to fill.
+ * @param addr
+ *   Buffer data address.
+ * @param length
+ *   Packet length.
+ * @param lkey
+ *   Memory region lkey.
+ */
+static inline void
+mlx5_wqe_write_inline(struct txq *txq, volatile union mlx5_wqe *wqe,
+                     uintptr_t addr, uint32_t length)
+{
+       uint32_t size;
+       uint16_t wqe_cnt = txq->wqe_n - 1;
+       uint16_t wqe_ci = txq->wqe_ci + 1;
+
+       /* Copy the first 16 bytes into inline header. */
+       rte_memcpy((void *)(uintptr_t)wqe->inl.eseg.inline_hdr_start,
+                  (void *)(uintptr_t)addr,
+                  MLX5_ETH_INLINE_HEADER_SIZE);
+       addr += MLX5_ETH_INLINE_HEADER_SIZE;
+       length -= MLX5_ETH_INLINE_HEADER_SIZE;
+       size = 3 + ((4 + length + 15) / 16);
+       wqe->inl.byte_cnt = htonl(length | MLX5_INLINE_SEG);
+       rte_memcpy((void *)(uintptr_t)&wqe->inl.data[0],
+                  (void *)addr, MLX5_WQE64_INL_DATA);
+       addr += MLX5_WQE64_INL_DATA;
+       length -= MLX5_WQE64_INL_DATA;
+       while (length) {
+               volatile union mlx5_wqe *wqe_next =
+                       &(*txq->wqes)[wqe_ci & wqe_cnt];
+               uint32_t copy_bytes = (length > sizeof(*wqe)) ?
+                                     sizeof(*wqe) :
+                                     length;
+
+               rte_mov64((uint8_t *)(uintptr_t)&wqe_next->data[0],
+                         (uint8_t *)addr);
+               addr += copy_bytes;
+               length -= copy_bytes;
+               ++wqe_ci;
+       }
+       assert(size < 64);
+       wqe->inl.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
+       wqe->inl.ctrl.data[1] = htonl(txq->qp_num_8s | size);
+       wqe->inl.ctrl.data[3] = 0;
+       wqe->inl.eseg.rsvd0 = 0;
+       wqe->inl.eseg.rsvd1 = 0;
+       wqe->inl.eseg.mss = 0;
+       wqe->inl.eseg.rsvd2 = 0;
+       wqe->inl.eseg.inline_hdr_sz = htons(MLX5_ETH_INLINE_HEADER_SIZE);
+       /* Increment consumer index. */
+       txq->wqe_ci = wqe_ci;
+}
+
+/**
+ * Write a inline WQE with VLAN.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param wqe
+ *   Pointer to the WQE to fill.
+ * @param addr
+ *   Buffer data address.
+ * @param length
+ *   Packet length.
+ * @param lkey
+ *   Memory region lkey.
+ * @param vlan_tci
+ *   VLAN field to insert in packet.
+ */
+static inline void
+mlx5_wqe_write_inline_vlan(struct txq *txq, volatile union mlx5_wqe *wqe,
+                          uintptr_t addr, uint32_t length, uint16_t vlan_tci)
+{
+       uint32_t size;
+       uint32_t wqe_cnt = txq->wqe_n - 1;
+       uint16_t wqe_ci = txq->wqe_ci + 1;
+       uint32_t vlan = htonl(0x81000000 | vlan_tci);
+
+       /*
+        * Copy 12 bytes of source & destination MAC address.
+        * Copy 4 bytes of VLAN.
+        * Copy 2 bytes of Ether type.
+        */
+       rte_memcpy((uint8_t *)(uintptr_t)wqe->inl.eseg.inline_hdr_start,
+                  (uint8_t *)addr, 12);
+       rte_memcpy((uint8_t *)(uintptr_t)wqe->inl.eseg.inline_hdr_start + 12,
+                  &vlan, sizeof(vlan));
+       rte_memcpy((uint8_t *)(uintptr_t)wqe->inl.eseg.inline_hdr_start + 16,
+                  ((uint8_t *)addr + 12), 2);
+       addr += MLX5_ETH_VLAN_INLINE_HEADER_SIZE - sizeof(vlan);
+       length -= MLX5_ETH_VLAN_INLINE_HEADER_SIZE - sizeof(vlan);
+       size = (sizeof(wqe->inl.ctrl.ctrl) +
+               sizeof(wqe->inl.eseg) +
+               sizeof(wqe->inl.byte_cnt) +
+               length + 15) / 16;
+       wqe->inl.byte_cnt = htonl(length | MLX5_INLINE_SEG);
+       rte_memcpy((void *)(uintptr_t)&wqe->inl.data[0],
+                  (void *)addr, MLX5_WQE64_INL_DATA);
+       addr += MLX5_WQE64_INL_DATA;
+       length -= MLX5_WQE64_INL_DATA;
+       while (length) {
+               volatile union mlx5_wqe *wqe_next =
+                       &(*txq->wqes)[wqe_ci & wqe_cnt];
+               uint32_t copy_bytes = (length > sizeof(*wqe)) ?
+                                     sizeof(*wqe) :
+                                     length;
+
+               rte_mov64((uint8_t *)(uintptr_t)&wqe_next->data[0],
+                         (uint8_t *)addr);
+               addr += copy_bytes;
+               length -= copy_bytes;
+               ++wqe_ci;
+       }
+       assert(size < 64);
+       wqe->inl.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
+       wqe->inl.ctrl.data[1] = htonl(txq->qp_num_8s | size);
+       wqe->inl.ctrl.data[3] = 0;
+       wqe->inl.eseg.rsvd0 = 0;
+       wqe->inl.eseg.rsvd1 = 0;
+       wqe->inl.eseg.mss = 0;
+       wqe->inl.eseg.rsvd2 = 0;
+       wqe->inl.eseg.inline_hdr_sz = htons(MLX5_ETH_VLAN_INLINE_HEADER_SIZE);
+       /* Increment consumer index. */
+       txq->wqe_ci = wqe_ci;
+}
+
 /**
  * Ring TX queue doorbell.
  *
@@ -416,6 +549,23 @@ tx_prefetch_cqe(struct txq *txq, uint16_t ci)
        rte_prefetch0(cqe);
 }
 
+/**
+ * Prefetch a WQE.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param  wqe_ci
+ *   WQE consumer index.
+ */
+static inline void
+tx_prefetch_wqe(struct txq *txq, uint16_t ci)
+{
+       volatile union mlx5_wqe *wqe;
+
+       wqe = &(*txq->wqes)[ci & (txq->wqe_n - 1)];
+       rte_prefetch0(wqe);
+}
+
 /**
  * DPDK callback for TX.
  *
@@ -528,6 +678,129 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        return i;
 }
 
+/**
+ * DPDK callback for TX with inline support.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       struct txq *txq = (struct txq *)dpdk_txq;
+       uint16_t elts_head = txq->elts_head;
+       const unsigned int elts_n = txq->elts_n;
+       unsigned int i;
+       unsigned int max;
+       unsigned int comp;
+       volatile union mlx5_wqe *wqe;
+       struct rte_mbuf *buf;
+       unsigned int max_inline = txq->max_inline;
+
+       if (unlikely(!pkts_n))
+               return 0;
+       buf = pkts[0];
+       /* Prefetch first packet cacheline. */
+       tx_prefetch_cqe(txq, txq->cq_ci);
+       tx_prefetch_cqe(txq, txq->cq_ci + 1);
+       rte_prefetch0(buf);
+       /* Start processing. */
+       txq_complete(txq);
+       max = (elts_n - (elts_head - txq->elts_tail));
+       if (max > elts_n)
+               max -= elts_n;
+       assert(max >= 1);
+       assert(max <= elts_n);
+       /* Always leave one free entry in the ring. */
+       --max;
+       if (max == 0)
+               return 0;
+       if (max > pkts_n)
+               max = pkts_n;
+       for (i = 0; (i != max); ++i) {
+               unsigned int elts_head_next = (elts_head + 1) & (elts_n - 1);
+               uintptr_t addr;
+               uint32_t length;
+               uint32_t lkey;
+
+               wqe = &(*txq->wqes)[txq->wqe_ci & (txq->wqe_n - 1)];
+               tx_prefetch_wqe(txq, txq->wqe_ci);
+               tx_prefetch_wqe(txq, txq->wqe_ci + 1);
+               if (i + 1 < max)
+                       rte_prefetch0(pkts[i + 1]);
+               /* Should we enable HW CKSUM offload */
+               if (buf->ol_flags &
+                   (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+                       wqe->inl.eseg.cs_flags =
+                               MLX5_ETH_WQE_L3_CSUM |
+                               MLX5_ETH_WQE_L4_CSUM;
+               } else {
+                       wqe->inl.eseg.cs_flags = 0;
+               }
+               /* Retrieve buffer information. */
+               addr = rte_pktmbuf_mtod(buf, uintptr_t);
+               length = DATA_LEN(buf);
+               /* Update element. */
+               (*txq->elts)[elts_head] = buf;
+               /* Prefetch next buffer data. */
+               if (i + 1 < max)
+                       rte_prefetch0(rte_pktmbuf_mtod(pkts[i + 1],
+                                                      volatile void *));
+               if (length <= max_inline) {
+                       if (buf->ol_flags & PKT_TX_VLAN_PKT)
+                               mlx5_wqe_write_inline_vlan(txq, wqe,
+                                                          addr, length,
+                                                          buf->vlan_tci);
+                       else
+                               mlx5_wqe_write_inline(txq, wqe, addr, length);
+               } else {
+                       /* Retrieve Memory Region key for this memory pool. */
+                       lkey = txq_mp2mr(txq, txq_mb2mp(buf));
+                       if (buf->ol_flags & PKT_TX_VLAN_PKT)
+                               mlx5_wqe_write_vlan(txq, wqe, addr, length,
+                                                   lkey, buf->vlan_tci);
+                       else
+                               mlx5_wqe_write(txq, wqe, addr, length, lkey);
+               }
+               wqe->inl.ctrl.data[2] = 0;
+               elts_head = elts_head_next;
+               buf = pkts[i + 1];
+#ifdef MLX5_PMD_SOFT_COUNTERS
+               /* Increment sent bytes counter. */
+               txq->stats.obytes += length;
+#endif
+       }
+       /* Take a shortcut if nothing must be sent. */
+       if (unlikely(i == 0))
+               return 0;
+       /* Check whether completion threshold has been reached. */
+       comp = txq->elts_comp + i;
+       if (comp >= MLX5_TX_COMP_THRESH) {
+               /* Request completion on last WQE. */
+               wqe->inl.ctrl.data[2] = htonl(8);
+               /* Save elts_head in unused "immediate" field of WQE. */
+               wqe->inl.ctrl.data[3] = elts_head;
+               txq->elts_comp = 0;
+       } else {
+               txq->elts_comp = comp;
+       }
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       /* Increment sent packets counter. */
+       txq->stats.opackets += i;
+#endif
+       /* Ring QP doorbell. */
+       mlx5_tx_dbrec(txq);
+       txq->elts_head = elts_head;
+       return i;
+}
+
 /**
  * Translate RX completion flags to packet type.
  *