net/mlx5: zero LRO mbuf headroom

LRO packet may consume all the stride memory, hence the PMD cannot
guaranty head-room for the LRO mbuf.

The issue is lack in HW support to write the packet in offset from the
stride start.

A new striding RQ feature may be added in CX6 DX to allow head-room and
tail-room for the LRO strides.

Signed-off-by: Matan Azrad <matan@mellanox.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>

diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst
index ecd25dc..92f1b97 100644 (file)
--- a/doc/guides/nics/mlx5.rst
+++ b/doc/guides/nics/mlx5.rst
@@ -165,6 +165,7 @@ Limitations
 
 - LRO:
 
+  - No mbuf headroom space is created for RX packets when LRO is configured.
   - ``scatter_fcs`` is disabled when LRO is configured.
 
 Statistics

diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 0538caf..edfcdd1 100644 (file)
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -1565,6 +1565,12 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
        unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
        unsigned int mprq_stride_size;
        struct mlx5_dev_config *config = &priv->config;
+       /*
+        * LRO packet may consume all the stride memory, hence we cannot
+        * guaranty head-room. A new striding RQ feature may be added in CX6 DX
+        * to allow head-room and tail-room for the LRO packets.
+        */
+       unsigned int strd_headroom_en = mlx5_lro_on(dev) ? 0 : 1;
        /*
         * Always allocate extra slots, even if eventually
         * the vector Rx will not be used.
@@ -1600,9 +1606,9 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
         *    stride.
         *  Otherwise, enable Rx scatter if necessary.
         */
-       assert(mb_len >= RTE_PKTMBUF_HEADROOM);
+       assert(mb_len >= RTE_PKTMBUF_HEADROOM * strd_headroom_en);
        mprq_stride_size = dev->data->dev_conf.rxmode.max_rx_pkt_len +
-                               RTE_PKTMBUF_HEADROOM;
+                               RTE_PKTMBUF_HEADROOM * strd_headroom_en;
        if (mprq_en &&
            desc > (1U << config->mprq.stride_num_n) &&
            mprq_stride_size <= (1U << config->mprq.max_stride_size_n)) {
@@ -1614,9 +1620,9 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                tmpl->rxq.strd_sz_n = RTE_MAX(log2above(mprq_stride_size),
                                              config->mprq.min_stride_size_n);
                tmpl->rxq.strd_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT;
-               tmpl->rxq.mprq_max_memcpy_len =
-                       RTE_MIN(mb_len - RTE_PKTMBUF_HEADROOM,
-                               config->mprq.max_memcpy_len);
+               tmpl->rxq.strd_headroom_en = strd_headroom_en;
+               tmpl->rxq.mprq_max_memcpy_len = RTE_MIN(mb_len -
+                           RTE_PKTMBUF_HEADROOM, config->mprq.max_memcpy_len);
                DRV_LOG(DEBUG,
                        "port %u Rx queue %u: Multi-Packet RQ is enabled"
                        " strd_num_n = %u, strd_sz_n = %u",

diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index fc9090e..007df8f 100644 (file)
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -1540,6 +1540,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
        unsigned int i = 0;
        uint32_t rq_ci = rxq->rq_ci;
        uint16_t consumed_strd = rxq->consumed_strd;
+       uint16_t headroom_sz = rxq->strd_headroom_en * RTE_PKTMBUF_HEADROOM;
        struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
 
        while (i < pkts_n) {
@@ -1650,7 +1651,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        rte_atomic16_add_return(&buf->refcnt, 1);
                        assert((uint16_t)rte_atomic16_read(&buf->refcnt) <=
                               strd_n + 1);
-                       buf_addr = RTE_PTR_SUB(addr, RTE_PKTMBUF_HEADROOM);
+                       buf_addr = RTE_PTR_SUB(addr, headroom_sz);
                        /*
                         * MLX5 device doesn't use iova but it is necessary in a
                         * case where the Rx packet is transmitted via a
@@ -1668,7 +1669,8 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
                         */
                        rte_pktmbuf_attach_extbuf(pkt, buf_addr, buf_iova,
                                                  buf_len, shinfo);
-                       rte_pktmbuf_reset_headroom(pkt);
+                       /* Set mbuf head-room. */
+                       pkt->data_off = headroom_sz;
                        assert(pkt->ol_flags == EXT_ATTACHED_MBUF);
                        /*
                         * Prevent potential overflow due to MTU change through

diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index bbd9b31..4252832 100644 (file)
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -114,7 +114,8 @@ struct mlx5_rxq_data {
        unsigned int strd_sz_n:4; /* Log 2 of stride size. */
        unsigned int strd_shift_en:1; /* Enable 2bytes shift on a stride. */
        unsigned int err_state:2; /* enum mlx5_rxq_err_state. */
-       unsigned int :4; /* Remaining bits. */
+       unsigned int strd_headroom_en:1; /* Enable mbuf headroom in MPRQ. */
+       unsigned int :3; /* Remaining bits. */
        volatile uint32_t *rq_db;
        volatile uint32_t *cq_db;
        uint16_t port_id;