net/mlx5: fix inline logic
authorNélio Laranjeiro <nelio.laranjeiro@6wind.com>
Wed, 14 Sep 2016 11:53:55 +0000 (13:53 +0200)
committerBruce Richardson <bruce.richardson@intel.com>
Fri, 30 Sep 2016 10:27:18 +0000 (12:27 +0200)
To improve performance the NIC expects for large packets to have a pointer
to a cache aligned address, old inline code could break this assumption
which hurts performance.

Fixes: 2a66cf378954 ("net/mlx5: support inline send")

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Vasily Philipov <vasilyf@mellanox.com>
drivers/net/mlx5/mlx5_ethdev.c
drivers/net/mlx5/mlx5_rxtx.c
drivers/net/mlx5/mlx5_rxtx.h
drivers/net/mlx5/mlx5_txq.c

index 47f323e..1ae80e5 100644 (file)
@@ -1398,10 +1398,6 @@ priv_select_tx_function(struct priv *priv)
        } else if ((priv->sriov == 0) && priv->mps) {
                priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
                DEBUG("selected MPW TX function");
-       } else if (priv->txq_inline && (priv->txqs_n >= priv->txqs_inline)) {
-               priv->dev->tx_pkt_burst = mlx5_tx_burst_inline;
-               DEBUG("selected inline TX function (%u >= %u queues)",
-                     priv->txqs_n, priv->txqs_inline);
        }
 }
 
index c7e538f..ecc76ad 100644 (file)
@@ -297,179 +297,99 @@ txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
  *   Buffer.
  * @param length
  *   Packet length.
- * @param lkey
- *   Memory region lkey.
+ *
+ * @return ds
+ *   Number of DS elements consumed.
  */
-static inline void
+static inline unsigned int
 mlx5_wqe_write(struct txq *txq, volatile union mlx5_wqe *wqe,
-              struct rte_mbuf *buf, uint32_t length, uint32_t lkey)
+              struct rte_mbuf *buf, uint32_t length)
 {
+       uintptr_t raw = (uintptr_t)&wqe->wqe.eseg.inline_hdr_start;
+       uint16_t ds;
+       uint16_t pkt_inline_sz = 16;
        uintptr_t addr = rte_pktmbuf_mtod(buf, uintptr_t);
+       struct mlx5_wqe_data_seg *dseg = NULL;
 
-       rte_mov16((uint8_t *)&wqe->wqe.eseg.inline_hdr_start,
-                 (uint8_t *)addr);
-       addr += 16;
+       assert(length >= 16);
+       /* Start the know and common part of the WQE structure. */
+       wqe->wqe.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
+       wqe->wqe.ctrl.data[2] = 0;
+       wqe->wqe.ctrl.data[3] = 0;
+       wqe->wqe.eseg.rsvd0 = 0;
+       wqe->wqe.eseg.rsvd1 = 0;
+       wqe->wqe.eseg.mss = 0;
+       wqe->wqe.eseg.rsvd2 = 0;
+       /* Start by copying the Ethernet Header. */
+       rte_mov16((uint8_t *)raw, (uint8_t *)addr);
        length -= 16;
-       /* Need to insert VLAN ? */
+       addr += 16;
+       /* Replace the Ethernet type by the VLAN if necessary. */
        if (buf->ol_flags & PKT_TX_VLAN_PKT) {
                uint32_t vlan = htonl(0x81000000 | buf->vlan_tci);
 
-               memcpy((uint8_t *)&wqe->wqe.eseg.inline_hdr_start + 12,
+               memcpy((uint8_t *)(raw + 16 - sizeof(vlan)),
                       &vlan, sizeof(vlan));
                addr -= sizeof(vlan);
                length += sizeof(vlan);
        }
-       /* Write the WQE. */
-       wqe->wqe.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
-       wqe->wqe.ctrl.data[1] = htonl((txq->qp_num_8s) | 4);
-       wqe->wqe.ctrl.data[2] = 0;
-       wqe->wqe.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->wqe.eseg.inline_hdr_sz = htons(16);
-       /* Store remaining data in data segment. */
-       wqe->wqe.dseg.byte_count = htonl(length);
-       wqe->wqe.dseg.lkey = lkey;
-       wqe->wqe.dseg.addr = htonll(addr);
-       /* Increment consumer index. */
-       ++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[2] = 0;
-       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;
+       /* Inline if enough room. */
+       if (txq->max_inline != 0) {
+               uintptr_t end = (uintptr_t)&(*txq->wqes)[txq->wqe_n];
+               uint16_t max_inline = txq->max_inline * RTE_CACHE_LINE_SIZE;
+               uint16_t room;
+
+               raw += 16;
+               room = end - (uintptr_t)raw;
+               if (room > max_inline) {
+                       uintptr_t addr_end = (addr + max_inline) &
+                               ~(RTE_CACHE_LINE_SIZE - 1);
+                       uint16_t copy_b = ((addr_end - addr) > length) ?
+                                         length :
+                                         (addr_end - addr);
+
+                       rte_memcpy((void *)raw, (void *)addr, copy_b);
+                       addr += copy_b;
+                       length -= copy_b;
+                       pkt_inline_sz += copy_b;
+                       /* Sanity check. */
+                       assert(addr <= addr_end);
+               }
+               /* Store the inlined packet size in the WQE. */
+               wqe->wqe.eseg.inline_hdr_sz = htons(pkt_inline_sz);
+               /*
+                * 2 DWORDs consumed by the WQE header + 1 DSEG +
+                * the size of the inline part of the packet.
+                */
+               ds = 2 + ((pkt_inline_sz - 2 + 15) / 16);
+               if (length > 0) {
+                       dseg = (struct mlx5_wqe_data_seg *)
+                               ((uintptr_t)wqe + (ds * 16));
+                       if ((uintptr_t)dseg >= end)
+                               dseg = (struct mlx5_wqe_data_seg *)
+                                       ((uintptr_t)&(*txq->wqes)[0]);
+                       goto use_dseg;
+               }
+       } else {
+               /* Add the remaining packet as a simple ds. */
+               ds = 3;
+               /*
+                * No inline has been done in the packet, only the Ethernet
+                * Header as been stored.
+                */
+               wqe->wqe.eseg.inline_hdr_sz = htons(16);
+               dseg = (struct mlx5_wqe_data_seg *)
+                       ((uintptr_t)wqe + (ds * 16));
+use_dseg:
+               *dseg = (struct mlx5_wqe_data_seg) {
+                       .addr = htonll(addr),
+                       .byte_count = htonl(length),
+                       .lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
+               };
+               ++ds;
        }
-       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[2] = 0;
-       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;
+       wqe->wqe.ctrl.data[1] = htonl(txq->qp_num_8s | ds);
+       return ds;
 }
 
 /**
@@ -570,7 +490,6 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                struct rte_mbuf *buf = *(pkts++);
                unsigned int elts_head_next;
                uint32_t length;
-               uint32_t lkey;
                unsigned int segs_n = buf->nb_segs;
                volatile struct mlx5_wqe_data_seg *dseg;
                unsigned int ds = sizeof(*wqe) / 16;
@@ -586,8 +505,8 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                --pkts_n;
                elts_head_next = (elts_head + 1) & (elts_n - 1);
                wqe = &(*txq->wqes)[txq->wqe_ci & (txq->wqe_n - 1)];
-               dseg = &wqe->wqe.dseg;
-               rte_prefetch0(wqe);
+               tx_prefetch_wqe(txq, txq->wqe_ci);
+               tx_prefetch_wqe(txq, txq->wqe_ci + 1);
                if (pkts_n)
                        rte_prefetch0(*pkts);
                length = DATA_LEN(buf);
@@ -597,9 +516,6 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                if (pkts_n)
                        rte_prefetch0(rte_pktmbuf_mtod(*pkts,
                                                       volatile void *));
-               /* Retrieve Memory Region key for this memory pool. */
-               lkey = txq_mp2mr(txq, txq_mb2mp(buf));
-               mlx5_wqe_write(txq, wqe, buf, length, lkey);
                /* Should we enable HW CKSUM offload */
                if (buf->ol_flags &
                    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
@@ -609,6 +525,11 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                } else {
                        wqe->wqe.eseg.cs_flags = 0;
                }
+               ds = mlx5_wqe_write(txq, wqe, buf, length);
+               if (segs_n == 1)
+                       goto skip_segs;
+               dseg = (volatile struct mlx5_wqe_data_seg *)
+                       (((uintptr_t)wqe) + ds * 16);
                while (--segs_n) {
                        /*
                         * Spill on next WQE when the current one does not have
@@ -639,11 +560,13 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                /* Update DS field in WQE. */
                wqe->wqe.ctrl.data[1] &= htonl(0xffffffc0);
                wqe->wqe.ctrl.data[1] |= htonl(ds & 0x3f);
-               elts_head = elts_head_next;
+skip_segs:
 #ifdef MLX5_PMD_SOFT_COUNTERS
                /* Increment sent bytes counter. */
                txq->stats.obytes += length;
 #endif
+               /* Increment consumer index. */
+               txq->wqe_ci += (ds + 3) / 4;
                elts_head = elts_head_next;
                ++i;
        } while (pkts_n);
@@ -671,162 +594,6 @@ 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 = 0;
-       unsigned int j = 0;
-       unsigned int max;
-       unsigned int comp;
-       volatile union mlx5_wqe *wqe = NULL;
-       unsigned int max_inline = txq->max_inline;
-
-       if (unlikely(!pkts_n))
-               return 0;
-       /* Prefetch first packet cacheline. */
-       tx_prefetch_cqe(txq, txq->cq_ci);
-       tx_prefetch_cqe(txq, txq->cq_ci + 1);
-       rte_prefetch0(*pkts);
-       /* Start processing. */
-       txq_complete(txq);
-       max = (elts_n - (elts_head - txq->elts_tail));
-       if (max > elts_n)
-               max -= elts_n;
-       do {
-               struct rte_mbuf *buf = *(pkts++);
-               unsigned int elts_head_next;
-               uintptr_t addr;
-               uint32_t length;
-               uint32_t lkey;
-               unsigned int segs_n = buf->nb_segs;
-               volatile struct mlx5_wqe_data_seg *dseg;
-               unsigned int ds = sizeof(*wqe) / 16;
-
-               /*
-                * Make sure there is enough room to store this packet and
-                * that one ring entry remains unused.
-                */
-               assert(segs_n);
-               if (max < segs_n + 1)
-                       break;
-               max -= segs_n;
-               --pkts_n;
-               elts_head_next = (elts_head + 1) & (elts_n - 1);
-               wqe = &(*txq->wqes)[txq->wqe_ci & (txq->wqe_n - 1)];
-               dseg = &wqe->wqe.dseg;
-               tx_prefetch_wqe(txq, txq->wqe_ci);
-               tx_prefetch_wqe(txq, txq->wqe_ci + 1);
-               if (pkts_n)
-                       rte_prefetch0(*pkts);
-               /* 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 (pkts_n)
-                       rte_prefetch0(rte_pktmbuf_mtod(*pkts,
-                                                      volatile void *));
-               if ((length <= max_inline) && (segs_n == 1)) {
-                       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);
-                       goto skip_segs;
-               } else {
-                       /* Retrieve Memory Region key for this memory pool. */
-                       lkey = txq_mp2mr(txq, txq_mb2mp(buf));
-                       mlx5_wqe_write(txq, wqe, buf, length, lkey);
-               }
-               while (--segs_n) {
-                       /*
-                        * Spill on next WQE when the current one does not have
-                        * enough room left. Size of WQE must a be a multiple
-                        * of data segment size.
-                        */
-                       assert(!(sizeof(*wqe) % sizeof(*dseg)));
-                       if (!(ds % (sizeof(*wqe) / 16)))
-                               dseg = (volatile void *)
-                                       &(*txq->wqes)[txq->wqe_ci++ &
-                                                     (txq->wqe_n - 1)];
-                       else
-                               ++dseg;
-                       ++ds;
-                       buf = buf->next;
-                       assert(buf);
-                       /* Store segment information. */
-                       dseg->byte_count = htonl(DATA_LEN(buf));
-                       dseg->lkey = txq_mp2mr(txq, txq_mb2mp(buf));
-                       dseg->addr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
-                       (*txq->elts)[elts_head_next] = buf;
-                       elts_head_next = (elts_head_next + 1) & (elts_n - 1);
-#ifdef MLX5_PMD_SOFT_COUNTERS
-                       length += DATA_LEN(buf);
-#endif
-                       ++j;
-               }
-               /* Update DS field in WQE. */
-               wqe->inl.ctrl.data[1] &= htonl(0xffffffc0);
-               wqe->inl.ctrl.data[1] |= htonl(ds & 0x3f);
-skip_segs:
-               elts_head = elts_head_next;
-#ifdef MLX5_PMD_SOFT_COUNTERS
-               /* Increment sent bytes counter. */
-               txq->stats.obytes += length;
-#endif
-               ++i;
-       } while (pkts_n);
-       /* 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 + j;
-       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;
-}
-
 /**
  * Open a MPW session.
  *
@@ -1117,7 +884,7 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
        unsigned int j = 0;
        unsigned int max;
        unsigned int comp;
-       unsigned int inline_room = txq->max_inline;
+       unsigned int inline_room = txq->max_inline * RTE_CACHE_LINE_SIZE;
        struct mlx5_mpw mpw = {
                .state = MLX5_MPW_STATE_CLOSED,
        };
@@ -1171,7 +938,8 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
                            (length > inline_room) ||
                            (mpw.wqe->mpw_inl.eseg.cs_flags != cs_flags)) {
                                mlx5_mpw_inline_close(txq, &mpw);
-                               inline_room = txq->max_inline;
+                               inline_room =
+                                       txq->max_inline * RTE_CACHE_LINE_SIZE;
                        }
                }
                if (mpw.state == MLX5_MPW_STATE_CLOSED) {
@@ -1187,7 +955,8 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
                /* Multi-segment packets must be alone in their MPW. */
                assert((segs_n == 1) || (mpw.pkts_n == 0));
                if (mpw.state == MLX5_MPW_STATE_OPENED) {
-                       assert(inline_room == txq->max_inline);
+                       assert(inline_room ==
+                              txq->max_inline * RTE_CACHE_LINE_SIZE);
 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
                        length = 0;
 #endif
@@ -1252,7 +1021,8 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
                        ++j;
                        if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
                                mlx5_mpw_inline_close(txq, &mpw);
-                               inline_room = txq->max_inline;
+                               inline_room =
+                                       txq->max_inline * RTE_CACHE_LINE_SIZE;
                        } else {
                                inline_room -= length;
                        }
index c8a93c0..8c568ad 100644 (file)
@@ -249,7 +249,7 @@ struct txq {
        uint16_t wqe_n; /* Number of WQ elements. */
        uint16_t bf_offset; /* Blueflame offset. */
        uint16_t bf_buf_size; /* Blueflame size. */
-       uint16_t max_inline; /* Maximum size to inline in a WQE. */
+       uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
        uint32_t qp_num_8s; /* QP number shifted by 8. */
        volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
        volatile union mlx5_wqe (*wqes)[]; /* Work queue. */
@@ -314,7 +314,6 @@ uint16_t mlx5_tx_burst_secondary_setup(void *, struct rte_mbuf **, uint16_t);
 /* mlx5_rxtx.c */
 
 uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
-uint16_t mlx5_tx_burst_inline(void *, struct rte_mbuf **, uint16_t);
 uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
 uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
 uint16_t mlx5_rx_burst(void *, struct rte_mbuf **, uint16_t);
index 6fe61c4..5ddd2fb 100644 (file)
@@ -338,9 +338,12 @@ txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
                .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
                              IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
        };
-       if (priv->txq_inline && priv->txqs_n >= priv->txqs_inline) {
-               tmpl.txq.max_inline = priv->txq_inline;
-               attr.init.cap.max_inline_data = tmpl.txq.max_inline;
+       if (priv->txq_inline && (priv->txqs_n >= priv->txqs_inline)) {
+               tmpl.txq.max_inline =
+                       ((priv->txq_inline + (RTE_CACHE_LINE_SIZE - 1)) /
+                        RTE_CACHE_LINE_SIZE);
+               attr.init.cap.max_inline_data =
+                       tmpl.txq.max_inline * RTE_CACHE_LINE_SIZE;
        }
        tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
        if (tmpl.qp == NULL) {