net/mlx5: refactor vectorized Rx
[dpdk.git] / drivers / net / mlx5 / mlx5_rxtx_vec.h
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2017 6WIND S.A.
3  * Copyright 2017 Mellanox Technologies, Ltd
4  */
5
6 #ifndef RTE_PMD_MLX5_RXTX_VEC_H_
7 #define RTE_PMD_MLX5_RXTX_VEC_H_
8
9 #include <rte_common.h>
10 #include <rte_mbuf.h>
11
12 #include <mlx5_prm.h>
13
14 #include "mlx5_autoconf.h"
15
16 #include "mlx5_mr.h"
17
18 /* HW checksum offload capabilities of vectorized Tx. */
19 #define MLX5_VEC_TX_CKSUM_OFFLOAD_CAP \
20         (DEV_TX_OFFLOAD_IPV4_CKSUM | \
21          DEV_TX_OFFLOAD_UDP_CKSUM | \
22          DEV_TX_OFFLOAD_TCP_CKSUM | \
23          DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
24
25 /*
26  * Compile time sanity check for vectorized functions.
27  */
28
29 #define S_ASSERT_RTE_MBUF(s) \
30         static_assert(s, "A field of struct rte_mbuf is changed")
31 #define S_ASSERT_MLX5_CQE(s) \
32         static_assert(s, "A field of struct mlx5_cqe is changed")
33
34 /* rxq_cq_decompress_v() */
35 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, pkt_len) ==
36                   offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
37 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, data_len) ==
38                   offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
39 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, hash) ==
40                   offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
41
42 /* rxq_cq_to_ptype_oflags_v() */
43 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, ol_flags) ==
44                   offsetof(struct rte_mbuf, rearm_data) + 8);
45 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, rearm_data) ==
46                   RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
47
48 /* rxq_burst_v() */
49 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, pkt_len) ==
50                   offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
51 S_ASSERT_RTE_MBUF(offsetof(struct rte_mbuf, data_len) ==
52                   offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
53 #if (RTE_CACHE_LINE_SIZE == 128)
54 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, pkt_info) == 64);
55 #else
56 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, pkt_info) == 0);
57 #endif
58 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, rx_hash_res) ==
59                   offsetof(struct mlx5_cqe, pkt_info) + 12);
60 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, rsvd1) + 11 ==
61                   offsetof(struct mlx5_cqe, hdr_type_etc));
62 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, vlan_info) ==
63                   offsetof(struct mlx5_cqe, hdr_type_etc) + 2);
64 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, lro_num_seg) + 12 ==
65                   offsetof(struct mlx5_cqe, byte_cnt));
66 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, sop_drop_qpn) ==
67                   RTE_ALIGN(offsetof(struct mlx5_cqe, sop_drop_qpn), 8));
68 S_ASSERT_MLX5_CQE(offsetof(struct mlx5_cqe, op_own) ==
69                   offsetof(struct mlx5_cqe, sop_drop_qpn) + 7);
70
71 /**
72  * Replenish buffers for RX in bulk.
73  *
74  * @param rxq
75  *   Pointer to RX queue structure.
76  */
77 static inline void
78 mlx5_rx_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq)
79 {
80         const uint16_t q_n = 1 << rxq->elts_n;
81         const uint16_t q_mask = q_n - 1;
82         uint16_t n = q_n - (rxq->rq_ci - rxq->rq_pi);
83         uint16_t elts_idx = rxq->rq_ci & q_mask;
84         struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
85         volatile struct mlx5_wqe_data_seg *wq =
86                 &((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[elts_idx];
87         unsigned int i;
88
89         if (n >= rxq->rq_repl_thresh) {
90                 MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n));
91                 MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n) >
92                             MLX5_VPMD_DESCS_PER_LOOP);
93                 /* Not to cross queue end. */
94                 n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, q_n - elts_idx);
95                 if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
96                         rxq->stats.rx_nombuf += n;
97                         return;
98                 }
99                 for (i = 0; i < n; ++i) {
100                         void *buf_addr;
101
102                         /*
103                          * In order to support the mbufs with external attached
104                          * data buffer we should use the buf_addr pointer
105                          * instead of rte_mbuf_buf_addr(). It touches the mbuf
106                          * itself and may impact the performance.
107                          */
108                         buf_addr = elts[i]->buf_addr;
109                         wq[i].addr = rte_cpu_to_be_64((uintptr_t)buf_addr +
110                                                       RTE_PKTMBUF_HEADROOM);
111                         /* If there's a single MR, no need to replace LKey. */
112                         if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh)
113                                      > 1))
114                                 wq[i].lkey = mlx5_rx_mb2mr(rxq, elts[i]);
115                 }
116                 rxq->rq_ci += n;
117                 /* Prevent overflowing into consumed mbufs. */
118                 elts_idx = rxq->rq_ci & q_mask;
119                 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
120                         (*rxq->elts)[elts_idx + i] = &rxq->fake_mbuf;
121                 rte_io_wmb();
122                 *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
123         }
124 }
125
126 #endif /* RTE_PMD_MLX5_RXTX_VEC_H_ */