net/mlx5: re-add Tx gather support
authorAdrien Mazarguil <adrien.mazarguil@6wind.com>
Fri, 24 Jun 2016 13:18:00 +0000 (15:18 +0200)
committerBruce Richardson <bruce.richardson@intel.com>
Mon, 27 Jun 2016 14:17:52 +0000 (16:17 +0200)
Compared to its previous incarnation, the software limit on the number of
mbuf segments is no more (previously MLX5_PMD_SGE_WR_N, set to 4 by
default) hence no need for linearization code and related buffers that
permanently consumed a non negligible amount of memory to handle oversized
mbufs.

The resulting code is both lighter and faster.

With the addition of this code, older GCC versions (such
as 4.8.5) may complain about 'wqe' variable being uninitialized, so
initialize it preemptively, even though it is not necessary to do so.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
drivers/net/mlx5/mlx5_rxtx.c
drivers/net/mlx5/mlx5_txq.c

index f270ef9..be63d84 100644 (file)
@@ -303,6 +303,7 @@ mlx5_wqe_write(struct txq *txq, volatile union mlx5_wqe *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;
@@ -348,6 +349,7 @@ mlx5_wqe_write_vlan(struct txq *txq, volatile union mlx5_wqe *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;
@@ -425,6 +427,7 @@ mlx5_wqe_write_inline(struct txq *txq, volatile union mlx5_wqe *wqe,
        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;
@@ -498,6 +501,7 @@ mlx5_wqe_write_inline_vlan(struct txq *txq, volatile union mlx5_wqe *wqe,
        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;
@@ -586,9 +590,10 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        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;
+       volatile union mlx5_wqe *wqe = NULL;
 
        if (unlikely(!pkts_n))
                return 0;
@@ -602,23 +607,27 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        if (max > elts_n)
                max -= elts_n;
        do {
-               struct rte_mbuf *buf;
+               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.
                 */
-               if (max < 1 + 1)
+               assert(segs_n);
+               if (max < segs_n + 1)
                        break;
-               --max;
+               max -= segs_n;
                --pkts_n;
-               buf = *(pkts++);
                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);
                if (pkts_n)
                        rte_prefetch0(*pkts);
@@ -638,7 +647,6 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                                            buf->vlan_tci);
                else
                        mlx5_wqe_write(txq, wqe, addr, length, lkey);
-               wqe->wqe.ctrl.data[2] = 0;
                /* Should we enable HW CKSUM offload */
                if (buf->ol_flags &
                    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
@@ -648,6 +656,37 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                } else {
                        wqe->wqe.eseg.cs_flags = 0;
                }
+               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->wqe.ctrl.data[1] &= htonl(0xffffffc0);
+               wqe->wqe.ctrl.data[1] |= htonl(ds & 0x3f);
+               elts_head = elts_head_next;
 #ifdef MLX5_PMD_SOFT_COUNTERS
                /* Increment sent bytes counter. */
                txq->stats.obytes += length;
@@ -659,7 +698,7 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        if (unlikely(i == 0))
                return 0;
        /* Check whether completion threshold has been reached. */
-       comp = txq->elts_comp + i;
+       comp = txq->elts_comp + i + j;
        if (comp >= MLX5_TX_COMP_THRESH) {
                /* Request completion on last WQE. */
                wqe->wqe.ctrl.data[2] = htonl(8);
@@ -699,9 +738,10 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        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;
+       volatile union mlx5_wqe *wqe = NULL;
        unsigned int max_inline = txq->max_inline;
 
        if (unlikely(!pkts_n))
@@ -716,23 +756,27 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        if (max > elts_n)
                max -= elts_n;
        do {
-               struct rte_mbuf *buf;
+               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.
                 */
-               if (max < 1 + 1)
+               assert(segs_n);
+               if (max < segs_n + 1)
                        break;
-               --max;
+               max -= segs_n;
                --pkts_n;
-               buf = *(pkts++);
                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)
@@ -755,13 +799,14 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                if (pkts_n)
                        rte_prefetch0(rte_pktmbuf_mtod(*pkts,
                                                       volatile void *));
-               if (length <= max_inline) {
+               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));
@@ -771,7 +816,37 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        else
                                mlx5_wqe_write(txq, wqe, addr, length, lkey);
                }
-               wqe->inl.ctrl.data[2] = 0;
+               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. */
@@ -783,7 +858,7 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        if (unlikely(i == 0))
                return 0;
        /* Check whether completion threshold has been reached. */
-       comp = txq->elts_comp + i;
+       comp = txq->elts_comp + i + j;
        if (comp >= MLX5_TX_COMP_THRESH) {
                /* Request completion on last WQE. */
                wqe->inl.ctrl.data[2] = htonl(8);
@@ -890,6 +965,7 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        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;
        struct mlx5_mpw mpw = {
@@ -908,48 +984,69 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        if (max > elts_n)
                max -= elts_n;
        do {
-               struct rte_mbuf *buf;
-               volatile struct mlx5_wqe_data_seg *dseg;
+               struct rte_mbuf *buf = *(pkts++);
                unsigned int elts_head_next;
-               uintptr_t addr;
                uint32_t length;
+               unsigned int segs_n = buf->nb_segs;
                uint32_t cs_flags = 0;
 
                /*
                 * Make sure there is enough room to store this packet and
                 * that one ring entry remains unused.
                 */
-               if (max < 1 + 1)
+               assert(segs_n);
+               if (max < segs_n + 1)
+                       break;
+               /* Do not bother with large packets MPW cannot handle. */
+               if (segs_n > MLX5_MPW_DSEG_MAX)
                        break;
-               --max;
+               max -= segs_n;
                --pkts_n;
-               buf = *(pkts++);
-               elts_head_next = (elts_head + 1) & (elts_n - 1);
                /* Should we enable HW CKSUM offload */
                if (buf->ol_flags &
                    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
                        cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
-               /* Retrieve buffer information. */
-               addr = rte_pktmbuf_mtod(buf, uintptr_t);
-               length = DATA_LEN(buf);
-               /* Update element. */
-               (*txq->elts)[elts_head] = buf;
+               /* Retrieve packet information. */
+               length = PKT_LEN(buf);
+               assert(length);
                /* Start new session if packet differs. */
                if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
                    ((mpw.len != length) ||
+                    (segs_n != 1) ||
                     (mpw.wqe->mpw.eseg.cs_flags != cs_flags)))
                        mlx5_mpw_close(txq, &mpw);
                if (mpw.state == MLX5_MPW_STATE_CLOSED) {
                        mlx5_mpw_new(txq, &mpw, length);
                        mpw.wqe->mpw.eseg.cs_flags = cs_flags;
                }
-               dseg = mpw.data.dseg[mpw.pkts_n];
-               *dseg = (struct mlx5_wqe_data_seg){
-                       .byte_count = htonl(length),
-                       .lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
-                       .addr = htonll(addr),
-               };
-               ++mpw.pkts_n;
+               /* Multi-segment packets must be alone in their MPW. */
+               assert((segs_n == 1) || (mpw.pkts_n == 0));
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+               length = 0;
+#endif
+               do {
+                       volatile struct mlx5_wqe_data_seg *dseg;
+                       uintptr_t addr;
+
+                       elts_head_next = (elts_head + 1) & (elts_n - 1);
+                       assert(buf);
+                       (*txq->elts)[elts_head] = buf;
+                       dseg = mpw.data.dseg[mpw.pkts_n];
+                       addr = rte_pktmbuf_mtod(buf, uintptr_t);
+                       *dseg = (struct mlx5_wqe_data_seg){
+                               .byte_count = htonl(DATA_LEN(buf)),
+                               .lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
+                               .addr = htonll(addr),
+                       };
+                       elts_head = elts_head_next;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+                       length += DATA_LEN(buf);
+#endif
+                       buf = buf->next;
+                       ++mpw.pkts_n;
+                       ++j;
+               } while (--segs_n);
+               assert(length == mpw.len);
                if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
                        mlx5_mpw_close(txq, &mpw);
                elts_head = elts_head_next;
@@ -963,7 +1060,8 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        if (unlikely(i == 0))
                return 0;
        /* Check whether completion threshold has been reached. */
-       comp = txq->elts_comp + i;
+       /* "j" includes both packets and segments. */
+       comp = txq->elts_comp + j;
        if (comp >= MLX5_TX_COMP_THRESH) {
                volatile union mlx5_wqe *wqe = mpw.wqe;
 
@@ -1067,6 +1165,7 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
        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;
        unsigned int inline_room = txq->max_inline;
@@ -1086,38 +1185,40 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
        if (max > elts_n)
                max -= elts_n;
        do {
-               struct rte_mbuf *buf;
+               struct rte_mbuf *buf = *(pkts++);
                unsigned int elts_head_next;
                uintptr_t addr;
                uint32_t length;
+               unsigned int segs_n = buf->nb_segs;
                uint32_t cs_flags = 0;
 
                /*
                 * Make sure there is enough room to store this packet and
                 * that one ring entry remains unused.
                 */
-               if (max < 1 + 1)
+               assert(segs_n);
+               if (max < segs_n + 1)
+                       break;
+               /* Do not bother with large packets MPW cannot handle. */
+               if (segs_n > MLX5_MPW_DSEG_MAX)
                        break;
-               --max;
+               max -= segs_n;
                --pkts_n;
-               buf = *(pkts++);
-               elts_head_next = (elts_head + 1) & (elts_n - 1);
                /* Should we enable HW CKSUM offload */
                if (buf->ol_flags &
                    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
                        cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
-               /* Retrieve buffer information. */
-               addr = rte_pktmbuf_mtod(buf, uintptr_t);
-               length = DATA_LEN(buf);
-               /* Update element. */
-               (*txq->elts)[elts_head] = buf;
+               /* Retrieve packet information. */
+               length = PKT_LEN(buf);
                /* Start new session if packet differs. */
                if (mpw.state == MLX5_MPW_STATE_OPENED) {
                        if ((mpw.len != length) ||
+                           (segs_n != 1) ||
                            (mpw.wqe->mpw.eseg.cs_flags != cs_flags))
                                mlx5_mpw_close(txq, &mpw);
                } else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
                        if ((mpw.len != length) ||
+                           (segs_n != 1) ||
                            (length > inline_room) ||
                            (mpw.wqe->mpw_inl.eseg.cs_flags != cs_flags)) {
                                mlx5_mpw_inline_close(txq, &mpw);
@@ -1125,7 +1226,8 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
                        }
                }
                if (mpw.state == MLX5_MPW_STATE_CLOSED) {
-                       if (length > inline_room) {
+                       if ((segs_n != 1) ||
+                           (length > inline_room)) {
                                mlx5_mpw_new(txq, &mpw, length);
                                mpw.wqe->mpw.eseg.cs_flags = cs_flags;
                        } else {
@@ -1133,17 +1235,36 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
                                mpw.wqe->mpw_inl.eseg.cs_flags = cs_flags;
                        }
                }
+               /* Multi-segment packets must be alone in their MPW. */
+               assert((segs_n == 1) || (mpw.pkts_n == 0));
                if (mpw.state == MLX5_MPW_STATE_OPENED) {
-                       volatile struct mlx5_wqe_data_seg *dseg;
-
                        assert(inline_room == txq->max_inline);
-                       dseg = mpw.data.dseg[mpw.pkts_n];
-                       *dseg = (struct mlx5_wqe_data_seg){
-                               .byte_count = htonl(length),
-                               .lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
-                               .addr = htonll(addr),
-                       };
-                       ++mpw.pkts_n;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+                       length = 0;
+#endif
+                       do {
+                               volatile struct mlx5_wqe_data_seg *dseg;
+
+                               elts_head_next =
+                                       (elts_head + 1) & (elts_n - 1);
+                               assert(buf);
+                               (*txq->elts)[elts_head] = buf;
+                               dseg = mpw.data.dseg[mpw.pkts_n];
+                               addr = rte_pktmbuf_mtod(buf, uintptr_t);
+                               *dseg = (struct mlx5_wqe_data_seg){
+                                       .byte_count = htonl(DATA_LEN(buf)),
+                                       .lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
+                                       .addr = htonll(addr),
+                               };
+                               elts_head = elts_head_next;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+                               length += DATA_LEN(buf);
+#endif
+                               buf = buf->next;
+                               ++mpw.pkts_n;
+                               ++j;
+                       } while (--segs_n);
+                       assert(length == mpw.len);
                        if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
                                mlx5_mpw_close(txq, &mpw);
                } else {
@@ -1151,6 +1272,10 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 
                        assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
                        assert(length <= inline_room);
+                       assert(length == DATA_LEN(buf));
+                       elts_head_next = (elts_head + 1) & (elts_n - 1);
+                       addr = rte_pktmbuf_mtod(buf, uintptr_t);
+                       (*txq->elts)[elts_head] = buf;
                        /* Maximum number of bytes before wrapping. */
                        max = ((uintptr_t)&(*txq->wqes)[txq->wqe_n] -
                               (uintptr_t)mpw.data.raw);
@@ -1175,6 +1300,7 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
                                mpw.data.raw =
                                        (volatile void *)&(*txq->wqes)[0];
                        ++mpw.pkts_n;
+                       ++j;
                        if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
                                mlx5_mpw_inline_close(txq, &mpw);
                                inline_room = txq->max_inline;
@@ -1194,7 +1320,8 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
        if (unlikely(i == 0))
                return 0;
        /* Check whether completion threshold has been reached. */
-       comp = txq->elts_comp + i;
+       /* "j" includes both packets and segments. */
+       comp = txq->elts_comp + j;
        if (comp >= MLX5_TX_COMP_THRESH) {
                volatile union mlx5_wqe *wqe = mpw.wqe;
 
index bae9f3d..6fe61c4 100644 (file)
@@ -320,7 +320,13 @@ txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
                        .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
                                        priv->device_attr.max_qp_wr :
                                        desc),
-                       /* Max number of scatter/gather elements in a WR. */
+                       /*
+                        * Max number of scatter/gather elements in a WR,
+                        * must be 1 to prevent libmlx5 from trying to affect
+                        * too much memory. TX gather is not impacted by the
+                        * priv->device_attr.max_sge limit and will still work
+                        * properly.
+                        */
                        .max_send_sge = 1,
                },
                .qp_type = IBV_QPT_RAW_PACKET,