xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision 6c02043e9967a9d8f6e8c058256e257efe1d6d1a)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015-2019 Mellanox Technologies, Ltd
4  */
5 
6 #include <stdint.h>
7 #include <string.h>
8 #include <stdlib.h>
9 
10 #include <rte_mbuf.h>
11 #include <rte_mempool.h>
12 #include <rte_prefetch.h>
13 #include <rte_common.h>
14 #include <rte_branch_prediction.h>
15 #include <rte_ether.h>
16 #include <rte_cycles.h>
17 #include <rte_flow.h>
18 
19 #include <mlx5_prm.h>
20 #include <mlx5_common.h>
21 
22 #include "mlx5_autoconf.h"
23 #include "mlx5_defs.h"
24 #include "mlx5.h"
25 #include "mlx5_mr.h"
26 #include "mlx5_utils.h"
27 #include "mlx5_rxtx.h"
28 
29 /* TX burst subroutines return codes. */
30 enum mlx5_txcmp_code {
31 	MLX5_TXCMP_CODE_EXIT = 0,
32 	MLX5_TXCMP_CODE_ERROR,
33 	MLX5_TXCMP_CODE_SINGLE,
34 	MLX5_TXCMP_CODE_MULTI,
35 	MLX5_TXCMP_CODE_TSO,
36 	MLX5_TXCMP_CODE_EMPW,
37 };
38 
39 /*
40  * These defines are used to configure Tx burst routine option set
41  * supported at compile time. The not specified options are optimized out
42  * out due to if conditions can be explicitly calculated at compile time.
43  * The offloads with bigger runtime check (require more CPU cycles to
44  * skip) overhead should have the bigger index - this is needed to
45  * select the better matching routine function if no exact match and
46  * some offloads are not actually requested.
47  */
48 #define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/
49 #define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/
50 #define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/
51 #define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */
52 #define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */
53 #define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/
54 #define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */
55 #define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/
56 #define MLX5_TXOFF_CONFIG_MPW (1u << 9) /* Legacy MPW supported.*/
57 #define MLX5_TXOFF_CONFIG_TXPP (1u << 10) /* Scheduling on timestamp.*/
58 
59 /* The most common offloads groups. */
60 #define MLX5_TXOFF_CONFIG_NONE 0
61 #define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \
62 				MLX5_TXOFF_CONFIG_TSO | \
63 				MLX5_TXOFF_CONFIG_SWP | \
64 				MLX5_TXOFF_CONFIG_CSUM | \
65 				MLX5_TXOFF_CONFIG_INLINE | \
66 				MLX5_TXOFF_CONFIG_VLAN | \
67 				MLX5_TXOFF_CONFIG_METADATA)
68 
69 #define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask)
70 
71 #define MLX5_TXOFF_DECL(func, olx) \
72 static uint16_t mlx5_tx_burst_##func(void *txq, \
73 				     struct rte_mbuf **pkts, \
74 				    uint16_t pkts_n) \
75 { \
76 	return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \
77 		    pkts, pkts_n, (olx)); \
78 }
79 
80 #define MLX5_TXOFF_INFO(func, olx) {mlx5_tx_burst_##func, olx},
81 
82 /* static asserts */
83 static_assert(MLX5_CQE_STATUS_HW_OWN < 0, "Must be negative value");
84 static_assert(MLX5_CQE_STATUS_SW_OWN < 0, "Must be negative value");
85 static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
86 		(sizeof(uint16_t) +
87 		 sizeof(rte_v128u32_t)),
88 		"invalid Ethernet Segment data size");
89 static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
90 		(sizeof(uint16_t) +
91 		 sizeof(struct rte_vlan_hdr) +
92 		 2 * RTE_ETHER_ADDR_LEN),
93 		"invalid Ethernet Segment data size");
94 static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
95 		(sizeof(uint16_t) +
96 		 sizeof(rte_v128u32_t)),
97 		"invalid Ethernet Segment data size");
98 static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
99 		(sizeof(uint16_t) +
100 		 sizeof(struct rte_vlan_hdr) +
101 		 2 * RTE_ETHER_ADDR_LEN),
102 		"invalid Ethernet Segment data size");
103 static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
104 		(sizeof(uint16_t) +
105 		 sizeof(rte_v128u32_t)),
106 		"invalid Ethernet Segment data size");
107 static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
108 		(sizeof(uint16_t) +
109 		 sizeof(struct rte_vlan_hdr) +
110 		 2 * RTE_ETHER_ADDR_LEN),
111 		"invalid Ethernet Segment data size");
112 static_assert(MLX5_DSEG_MIN_INLINE_SIZE ==
113 		(2 * RTE_ETHER_ADDR_LEN),
114 		"invalid Data Segment data size");
115 static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
116 static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
117 static_assert((sizeof(struct rte_vlan_hdr) +
118 			sizeof(struct rte_ether_hdr)) ==
119 		MLX5_ESEG_MIN_INLINE_SIZE,
120 		"invalid min inline data size");
121 static_assert(MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE <=
122 		MLX5_DSEG_MAX, "invalid WQE max size");
123 static_assert(MLX5_WQE_CSEG_SIZE == MLX5_WSEG_SIZE,
124 		"invalid WQE Control Segment size");
125 static_assert(MLX5_WQE_ESEG_SIZE == MLX5_WSEG_SIZE,
126 		"invalid WQE Ethernet Segment size");
127 static_assert(MLX5_WQE_DSEG_SIZE == MLX5_WSEG_SIZE,
128 		"invalid WQE Data Segment size");
129 static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE,
130 		"invalid WQE size");
131 
132 static __rte_always_inline uint32_t
133 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
134 				   volatile struct mlx5_mini_cqe8 *mcqe);
135 
136 static __rte_always_inline int
137 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
138 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
139 
140 static __rte_always_inline uint32_t
141 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
142 
143 static __rte_always_inline void
144 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
145 	       volatile struct mlx5_cqe *cqe,
146 	       volatile struct mlx5_mini_cqe8 *mcqe);
147 
148 static int
149 mlx5_queue_state_modify(struct rte_eth_dev *dev,
150 			struct mlx5_mp_arg_queue_state_modify *sm);
151 
152 static inline void
153 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
154 			volatile struct mlx5_cqe *__rte_restrict cqe,
155 			uint32_t phcsum, uint8_t l4_type);
156 
157 static inline void
158 mlx5_lro_update_hdr(uint8_t *__rte_restrict padd,
159 		    volatile struct mlx5_cqe *__rte_restrict cqe,
160 			volatile struct mlx5_mini_cqe8 *mcqe,
161 		    struct mlx5_rxq_data *rxq, uint32_t len);
162 
163 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
164 	[0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
165 };
166 
167 uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
168 uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
169 
170 uint64_t rte_net_mlx5_dynf_inline_mask;
171 #define PKT_TX_DYNF_NOINLINE rte_net_mlx5_dynf_inline_mask
172 
173 /**
174  * Build a table to translate Rx completion flags to packet type.
175  *
176  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
177  */
178 void
179 mlx5_set_ptype_table(void)
180 {
181 	unsigned int i;
182 	uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
183 
184 	/* Last entry must not be overwritten, reserved for errored packet. */
185 	for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
186 		(*p)[i] = RTE_PTYPE_UNKNOWN;
187 	/*
188 	 * The index to the array should have:
189 	 * bit[1:0] = l3_hdr_type
190 	 * bit[4:2] = l4_hdr_type
191 	 * bit[5] = ip_frag
192 	 * bit[6] = tunneled
193 	 * bit[7] = outer_l3_type
194 	 */
195 	/* L2 */
196 	(*p)[0x00] = RTE_PTYPE_L2_ETHER;
197 	/* L3 */
198 	(*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
199 		     RTE_PTYPE_L4_NONFRAG;
200 	(*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
201 		     RTE_PTYPE_L4_NONFRAG;
202 	/* Fragmented */
203 	(*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
204 		     RTE_PTYPE_L4_FRAG;
205 	(*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
206 		     RTE_PTYPE_L4_FRAG;
207 	/* TCP */
208 	(*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
209 		     RTE_PTYPE_L4_TCP;
210 	(*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
211 		     RTE_PTYPE_L4_TCP;
212 	(*p)[0x0d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
213 		     RTE_PTYPE_L4_TCP;
214 	(*p)[0x0e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
215 		     RTE_PTYPE_L4_TCP;
216 	(*p)[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
217 		     RTE_PTYPE_L4_TCP;
218 	(*p)[0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
219 		     RTE_PTYPE_L4_TCP;
220 	/* UDP */
221 	(*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
222 		     RTE_PTYPE_L4_UDP;
223 	(*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
224 		     RTE_PTYPE_L4_UDP;
225 	/* Repeat with outer_l3_type being set. Just in case. */
226 	(*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
227 		     RTE_PTYPE_L4_NONFRAG;
228 	(*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
229 		     RTE_PTYPE_L4_NONFRAG;
230 	(*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
231 		     RTE_PTYPE_L4_FRAG;
232 	(*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
233 		     RTE_PTYPE_L4_FRAG;
234 	(*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
235 		     RTE_PTYPE_L4_TCP;
236 	(*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
237 		     RTE_PTYPE_L4_TCP;
238 	(*p)[0x8d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
239 		     RTE_PTYPE_L4_TCP;
240 	(*p)[0x8e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
241 		     RTE_PTYPE_L4_TCP;
242 	(*p)[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
243 		     RTE_PTYPE_L4_TCP;
244 	(*p)[0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
245 		     RTE_PTYPE_L4_TCP;
246 	(*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
247 		     RTE_PTYPE_L4_UDP;
248 	(*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
249 		     RTE_PTYPE_L4_UDP;
250 	/* Tunneled - L3 */
251 	(*p)[0x40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
252 	(*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
253 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
254 		     RTE_PTYPE_INNER_L4_NONFRAG;
255 	(*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
256 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
257 		     RTE_PTYPE_INNER_L4_NONFRAG;
258 	(*p)[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
259 	(*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
260 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
261 		     RTE_PTYPE_INNER_L4_NONFRAG;
262 	(*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
263 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
264 		     RTE_PTYPE_INNER_L4_NONFRAG;
265 	/* Tunneled - Fragmented */
266 	(*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
267 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
268 		     RTE_PTYPE_INNER_L4_FRAG;
269 	(*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
270 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
271 		     RTE_PTYPE_INNER_L4_FRAG;
272 	(*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
273 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
274 		     RTE_PTYPE_INNER_L4_FRAG;
275 	(*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
276 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
277 		     RTE_PTYPE_INNER_L4_FRAG;
278 	/* Tunneled - TCP */
279 	(*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
280 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
281 		     RTE_PTYPE_INNER_L4_TCP;
282 	(*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
283 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
284 		     RTE_PTYPE_INNER_L4_TCP;
285 	(*p)[0x4d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
286 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
287 		     RTE_PTYPE_INNER_L4_TCP;
288 	(*p)[0x4e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
289 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
290 		     RTE_PTYPE_INNER_L4_TCP;
291 	(*p)[0x51] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
292 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
293 		     RTE_PTYPE_INNER_L4_TCP;
294 	(*p)[0x52] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
295 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
296 		     RTE_PTYPE_INNER_L4_TCP;
297 	(*p)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
298 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
299 		     RTE_PTYPE_INNER_L4_TCP;
300 	(*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
301 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
302 		     RTE_PTYPE_INNER_L4_TCP;
303 	(*p)[0xcd] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
304 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
305 		     RTE_PTYPE_INNER_L4_TCP;
306 	(*p)[0xce] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
307 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
308 		     RTE_PTYPE_INNER_L4_TCP;
309 	(*p)[0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
310 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
311 		     RTE_PTYPE_INNER_L4_TCP;
312 	(*p)[0xd2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
313 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
314 		     RTE_PTYPE_INNER_L4_TCP;
315 	/* Tunneled - UDP */
316 	(*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
317 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
318 		     RTE_PTYPE_INNER_L4_UDP;
319 	(*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
320 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
321 		     RTE_PTYPE_INNER_L4_UDP;
322 	(*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
323 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
324 		     RTE_PTYPE_INNER_L4_UDP;
325 	(*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
326 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
327 		     RTE_PTYPE_INNER_L4_UDP;
328 }
329 
330 /**
331  * Build a table to translate packet to checksum type of Verbs.
332  */
333 void
334 mlx5_set_cksum_table(void)
335 {
336 	unsigned int i;
337 	uint8_t v;
338 
339 	/*
340 	 * The index should have:
341 	 * bit[0] = PKT_TX_TCP_SEG
342 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
343 	 * bit[4] = PKT_TX_IP_CKSUM
344 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
345 	 * bit[9] = tunnel
346 	 */
347 	for (i = 0; i < RTE_DIM(mlx5_cksum_table); ++i) {
348 		v = 0;
349 		if (i & (1 << 9)) {
350 			/* Tunneled packet. */
351 			if (i & (1 << 8)) /* Outer IP. */
352 				v |= MLX5_ETH_WQE_L3_CSUM;
353 			if (i & (1 << 4)) /* Inner IP. */
354 				v |= MLX5_ETH_WQE_L3_INNER_CSUM;
355 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
356 				v |= MLX5_ETH_WQE_L4_INNER_CSUM;
357 		} else {
358 			/* No tunnel. */
359 			if (i & (1 << 4)) /* IP. */
360 				v |= MLX5_ETH_WQE_L3_CSUM;
361 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
362 				v |= MLX5_ETH_WQE_L4_CSUM;
363 		}
364 		mlx5_cksum_table[i] = v;
365 	}
366 }
367 
368 /**
369  * Build a table to translate packet type of mbuf to SWP type of Verbs.
370  */
371 void
372 mlx5_set_swp_types_table(void)
373 {
374 	unsigned int i;
375 	uint8_t v;
376 
377 	/*
378 	 * The index should have:
379 	 * bit[0:1] = PKT_TX_L4_MASK
380 	 * bit[4] = PKT_TX_IPV6
381 	 * bit[8] = PKT_TX_OUTER_IPV6
382 	 * bit[9] = PKT_TX_OUTER_UDP
383 	 */
384 	for (i = 0; i < RTE_DIM(mlx5_swp_types_table); ++i) {
385 		v = 0;
386 		if (i & (1 << 8))
387 			v |= MLX5_ETH_WQE_L3_OUTER_IPV6;
388 		if (i & (1 << 9))
389 			v |= MLX5_ETH_WQE_L4_OUTER_UDP;
390 		if (i & (1 << 4))
391 			v |= MLX5_ETH_WQE_L3_INNER_IPV6;
392 		if ((i & 3) == (PKT_TX_UDP_CKSUM >> 52))
393 			v |= MLX5_ETH_WQE_L4_INNER_UDP;
394 		mlx5_swp_types_table[i] = v;
395 	}
396 }
397 
398 /**
399  * Set Software Parser flags and offsets in Ethernet Segment of WQE.
400  * Flags must be preliminary initialized to zero.
401  *
402  * @param loc
403  *   Pointer to burst routine local context.
404  * @param swp_flags
405  *   Pointer to store Software Parser flags
406  * @param olx
407  *   Configured Tx offloads mask. It is fully defined at
408  *   compile time and may be used for optimization.
409  *
410  * @return
411  *   Software Parser offsets packed in dword.
412  *   Software Parser flags are set by pointer.
413  */
414 static __rte_always_inline uint32_t
415 txq_mbuf_to_swp(struct mlx5_txq_local *__rte_restrict loc,
416 		uint8_t *swp_flags,
417 		unsigned int olx)
418 {
419 	uint64_t ol, tunnel;
420 	unsigned int idx, off;
421 	uint32_t set;
422 
423 	if (!MLX5_TXOFF_CONFIG(SWP))
424 		return 0;
425 	ol = loc->mbuf->ol_flags;
426 	tunnel = ol & PKT_TX_TUNNEL_MASK;
427 	/*
428 	 * Check whether Software Parser is required.
429 	 * Only customized tunnels may ask for.
430 	 */
431 	if (likely(tunnel != PKT_TX_TUNNEL_UDP && tunnel != PKT_TX_TUNNEL_IP))
432 		return 0;
433 	/*
434 	 * The index should have:
435 	 * bit[0:1] = PKT_TX_L4_MASK
436 	 * bit[4] = PKT_TX_IPV6
437 	 * bit[8] = PKT_TX_OUTER_IPV6
438 	 * bit[9] = PKT_TX_OUTER_UDP
439 	 */
440 	idx = (ol & (PKT_TX_L4_MASK | PKT_TX_IPV6 | PKT_TX_OUTER_IPV6)) >> 52;
441 	idx |= (tunnel == PKT_TX_TUNNEL_UDP) ? (1 << 9) : 0;
442 	*swp_flags = mlx5_swp_types_table[idx];
443 	/*
444 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
445 	 * complements HW parser. SW parser starts to engage only if HW parser
446 	 * can't reach a header. For the older devices, HW parser will not kick
447 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
448 	 * should be set regardless of HW offload.
449 	 */
450 	off = loc->mbuf->outer_l2_len;
451 	if (MLX5_TXOFF_CONFIG(VLAN) && ol & PKT_TX_VLAN_PKT)
452 		off += sizeof(struct rte_vlan_hdr);
453 	set = (off >> 1) << 8; /* Outer L3 offset. */
454 	off += loc->mbuf->outer_l3_len;
455 	if (tunnel == PKT_TX_TUNNEL_UDP)
456 		set |= off >> 1; /* Outer L4 offset. */
457 	if (ol & (PKT_TX_IPV4 | PKT_TX_IPV6)) { /* Inner IP. */
458 		const uint64_t csum = ol & PKT_TX_L4_MASK;
459 			off += loc->mbuf->l2_len;
460 		set |= (off >> 1) << 24; /* Inner L3 offset. */
461 		if (csum == PKT_TX_TCP_CKSUM ||
462 		    csum == PKT_TX_UDP_CKSUM ||
463 		    (MLX5_TXOFF_CONFIG(TSO) && ol & PKT_TX_TCP_SEG)) {
464 			off += loc->mbuf->l3_len;
465 			set |= (off >> 1) << 16; /* Inner L4 offset. */
466 		}
467 	}
468 	set = rte_cpu_to_le_32(set);
469 	return set;
470 }
471 
472 /**
473  * Convert the Checksum offloads to Verbs.
474  *
475  * @param buf
476  *   Pointer to the mbuf.
477  *
478  * @return
479  *   Converted checksum flags.
480  */
481 static __rte_always_inline uint8_t
482 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
483 {
484 	uint32_t idx;
485 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
486 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
487 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
488 
489 	/*
490 	 * The index should have:
491 	 * bit[0] = PKT_TX_TCP_SEG
492 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
493 	 * bit[4] = PKT_TX_IP_CKSUM
494 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
495 	 * bit[9] = tunnel
496 	 */
497 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
498 	return mlx5_cksum_table[idx];
499 }
500 
501 /**
502  * Internal function to compute the number of used descriptors in an RX queue
503  *
504  * @param rxq
505  *   The Rx queue.
506  *
507  * @return
508  *   The number of used rx descriptor.
509  */
510 static uint32_t
511 rx_queue_count(struct mlx5_rxq_data *rxq)
512 {
513 	struct rxq_zip *zip = &rxq->zip;
514 	volatile struct mlx5_cqe *cqe;
515 	const unsigned int cqe_n = (1 << rxq->cqe_n);
516 	const unsigned int sges_n = (1 << rxq->sges_n);
517 	const unsigned int elts_n = (1 << rxq->elts_n);
518 	const unsigned int strd_n = (1 << rxq->strd_num_n);
519 	const unsigned int cqe_cnt = cqe_n - 1;
520 	unsigned int cq_ci, used;
521 
522 	/* if we are processing a compressed cqe */
523 	if (zip->ai) {
524 		used = zip->cqe_cnt - zip->ai;
525 		cq_ci = zip->cq_ci;
526 	} else {
527 		used = 0;
528 		cq_ci = rxq->cq_ci;
529 	}
530 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
531 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
532 		int8_t op_own;
533 		unsigned int n;
534 
535 		op_own = cqe->op_own;
536 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
537 			n = rte_be_to_cpu_32(cqe->byte_cnt);
538 		else
539 			n = 1;
540 		cq_ci += n;
541 		used += n;
542 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
543 	}
544 	used = RTE_MIN(used * sges_n, elts_n * strd_n);
545 	return used;
546 }
547 
548 /**
549  * DPDK callback to check the status of a rx descriptor.
550  *
551  * @param rx_queue
552  *   The Rx queue.
553  * @param[in] offset
554  *   The index of the descriptor in the ring.
555  *
556  * @return
557  *   The status of the tx descriptor.
558  */
559 int
560 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
561 {
562 	struct mlx5_rxq_data *rxq = rx_queue;
563 	struct mlx5_rxq_ctrl *rxq_ctrl =
564 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
565 	struct rte_eth_dev *dev = ETH_DEV(rxq_ctrl->priv);
566 
567 	if (dev->rx_pkt_burst == NULL ||
568 	    dev->rx_pkt_burst == removed_rx_burst) {
569 		rte_errno = ENOTSUP;
570 		return -rte_errno;
571 	}
572 	if (offset >= (1 << rxq->cqe_n)) {
573 		rte_errno = EINVAL;
574 		return -rte_errno;
575 	}
576 	if (offset < rx_queue_count(rxq))
577 		return RTE_ETH_RX_DESC_DONE;
578 	return RTE_ETH_RX_DESC_AVAIL;
579 }
580 
581 /**
582  * DPDK callback to get the RX queue information
583  *
584  * @param dev
585  *   Pointer to the device structure.
586  *
587  * @param rx_queue_id
588  *   Rx queue identificator.
589  *
590  * @param qinfo
591  *   Pointer to the RX queue information structure.
592  *
593  * @return
594  *   None.
595  */
596 
597 void
598 mlx5_rxq_info_get(struct rte_eth_dev *dev, uint16_t rx_queue_id,
599 		  struct rte_eth_rxq_info *qinfo)
600 {
601 	struct mlx5_priv *priv = dev->data->dev_private;
602 	struct mlx5_rxq_data *rxq = (*priv->rxqs)[rx_queue_id];
603 	struct mlx5_rxq_ctrl *rxq_ctrl =
604 		container_of(rxq, struct mlx5_rxq_ctrl, rxq);
605 
606 	if (!rxq)
607 		return;
608 	qinfo->mp = mlx5_rxq_mprq_enabled(rxq) ?
609 					rxq->mprq_mp : rxq->mp;
610 	qinfo->conf.rx_thresh.pthresh = 0;
611 	qinfo->conf.rx_thresh.hthresh = 0;
612 	qinfo->conf.rx_thresh.wthresh = 0;
613 	qinfo->conf.rx_free_thresh = rxq->rq_repl_thresh;
614 	qinfo->conf.rx_drop_en = 1;
615 	qinfo->conf.rx_deferred_start = rxq_ctrl ? 0 : 1;
616 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
617 	qinfo->scattered_rx = dev->data->scattered_rx;
618 	qinfo->nb_desc = mlx5_rxq_mprq_enabled(rxq) ?
619 		(1 << rxq->elts_n) * (1 << rxq->strd_num_n) :
620 		(1 << rxq->elts_n);
621 }
622 
623 /**
624  * DPDK callback to get the RX packet burst mode information
625  *
626  * @param dev
627  *   Pointer to the device structure.
628  *
629  * @param rx_queue_id
630  *   Rx queue identificatior.
631  *
632  * @param mode
633  *   Pointer to the burts mode information.
634  *
635  * @return
636  *   0 as success, -EINVAL as failure.
637  */
638 
639 int
640 mlx5_rx_burst_mode_get(struct rte_eth_dev *dev,
641 		       uint16_t rx_queue_id __rte_unused,
642 		       struct rte_eth_burst_mode *mode)
643 {
644 	eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
645 	struct mlx5_priv *priv = dev->data->dev_private;
646 	struct mlx5_rxq_data *rxq;
647 
648 	rxq = (*priv->rxqs)[rx_queue_id];
649 	if (!rxq) {
650 		rte_errno = EINVAL;
651 		return -rte_errno;
652 	}
653 	if (pkt_burst == mlx5_rx_burst) {
654 		snprintf(mode->info, sizeof(mode->info), "%s", "Scalar");
655 	} else if (pkt_burst == mlx5_rx_burst_mprq) {
656 		snprintf(mode->info, sizeof(mode->info), "%s", "Multi-Packet RQ");
657 	} else if (pkt_burst == mlx5_rx_burst_vec) {
658 #if defined RTE_ARCH_X86_64
659 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector SSE");
660 #elif defined RTE_ARCH_ARM64
661 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector Neon");
662 #elif defined RTE_ARCH_PPC_64
663 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector AltiVec");
664 #else
665 		return -EINVAL;
666 #endif
667 	} else if (pkt_burst == mlx5_rx_burst_mprq_vec) {
668 #if defined RTE_ARCH_X86_64
669 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector SSE");
670 #elif defined RTE_ARCH_ARM64
671 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector Neon");
672 #elif defined RTE_ARCH_PPC_64
673 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector AltiVec");
674 #else
675 		return -EINVAL;
676 #endif
677 	} else {
678 		return -EINVAL;
679 	}
680 	return 0;
681 }
682 
683 /**
684  * DPDK callback to get the number of used descriptors in a RX queue
685  *
686  * @param dev
687  *   Pointer to the device structure.
688  *
689  * @param rx_queue_id
690  *   The Rx queue.
691  *
692  * @return
693  *   The number of used rx descriptor.
694  *   -EINVAL if the queue is invalid
695  */
696 uint32_t
697 mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
698 {
699 	struct mlx5_priv *priv = dev->data->dev_private;
700 	struct mlx5_rxq_data *rxq;
701 
702 	if (dev->rx_pkt_burst == NULL ||
703 	    dev->rx_pkt_burst == removed_rx_burst) {
704 		rte_errno = ENOTSUP;
705 		return -rte_errno;
706 	}
707 	rxq = (*priv->rxqs)[rx_queue_id];
708 	if (!rxq) {
709 		rte_errno = EINVAL;
710 		return -rte_errno;
711 	}
712 	return rx_queue_count(rxq);
713 }
714 
715 #define MLX5_SYSTEM_LOG_DIR "/var/log"
716 /**
717  * Dump debug information to log file.
718  *
719  * @param fname
720  *   The file name.
721  * @param hex_title
722  *   If not NULL this string is printed as a header to the output
723  *   and the output will be in hexadecimal view.
724  * @param buf
725  *   This is the buffer address to print out.
726  * @param len
727  *   The number of bytes to dump out.
728  */
729 void
730 mlx5_dump_debug_information(const char *fname, const char *hex_title,
731 			    const void *buf, unsigned int hex_len)
732 {
733 	FILE *fd;
734 
735 	MKSTR(path, "%s/%s", MLX5_SYSTEM_LOG_DIR, fname);
736 	fd = fopen(path, "a+");
737 	if (!fd) {
738 		DRV_LOG(WARNING, "cannot open %s for debug dump", path);
739 		MKSTR(path2, "./%s", fname);
740 		fd = fopen(path2, "a+");
741 		if (!fd) {
742 			DRV_LOG(ERR, "cannot open %s for debug dump", path2);
743 			return;
744 		}
745 		DRV_LOG(INFO, "New debug dump in file %s", path2);
746 	} else {
747 		DRV_LOG(INFO, "New debug dump in file %s", path);
748 	}
749 	if (hex_title)
750 		rte_hexdump(fd, hex_title, buf, hex_len);
751 	else
752 		fprintf(fd, "%s", (const char *)buf);
753 	fprintf(fd, "\n\n\n");
754 	fclose(fd);
755 }
756 
757 /**
758  * Move QP from error state to running state and initialize indexes.
759  *
760  * @param txq_ctrl
761  *   Pointer to TX queue control structure.
762  *
763  * @return
764  *   0 on success, else -1.
765  */
766 static int
767 tx_recover_qp(struct mlx5_txq_ctrl *txq_ctrl)
768 {
769 	struct mlx5_mp_arg_queue_state_modify sm = {
770 			.is_wq = 0,
771 			.queue_id = txq_ctrl->txq.idx,
772 	};
773 
774 	if (mlx5_queue_state_modify(ETH_DEV(txq_ctrl->priv), &sm))
775 		return -1;
776 	txq_ctrl->txq.wqe_ci = 0;
777 	txq_ctrl->txq.wqe_pi = 0;
778 	txq_ctrl->txq.elts_comp = 0;
779 	return 0;
780 }
781 
782 /* Return 1 if the error CQE is signed otherwise, sign it and return 0. */
783 static int
784 check_err_cqe_seen(volatile struct mlx5_err_cqe *err_cqe)
785 {
786 	static const uint8_t magic[] = "seen";
787 	int ret = 1;
788 	unsigned int i;
789 
790 	for (i = 0; i < sizeof(magic); ++i)
791 		if (!ret || err_cqe->rsvd1[i] != magic[i]) {
792 			ret = 0;
793 			err_cqe->rsvd1[i] = magic[i];
794 		}
795 	return ret;
796 }
797 
798 /**
799  * Handle error CQE.
800  *
801  * @param txq
802  *   Pointer to TX queue structure.
803  * @param error_cqe
804  *   Pointer to the error CQE.
805  *
806  * @return
807  *   Negative value if queue recovery failed, otherwise
808  *   the error completion entry is handled successfully.
809  */
810 static int
811 mlx5_tx_error_cqe_handle(struct mlx5_txq_data *__rte_restrict txq,
812 			 volatile struct mlx5_err_cqe *err_cqe)
813 {
814 	if (err_cqe->syndrome != MLX5_CQE_SYNDROME_WR_FLUSH_ERR) {
815 		const uint16_t wqe_m = ((1 << txq->wqe_n) - 1);
816 		struct mlx5_txq_ctrl *txq_ctrl =
817 				container_of(txq, struct mlx5_txq_ctrl, txq);
818 		uint16_t new_wqe_pi = rte_be_to_cpu_16(err_cqe->wqe_counter);
819 		int seen = check_err_cqe_seen(err_cqe);
820 
821 		if (!seen && txq_ctrl->dump_file_n <
822 		    txq_ctrl->priv->config.max_dump_files_num) {
823 			MKSTR(err_str, "Unexpected CQE error syndrome "
824 			      "0x%02x CQN = %u SQN = %u wqe_counter = %u "
825 			      "wq_ci = %u cq_ci = %u", err_cqe->syndrome,
826 			      txq->cqe_s, txq->qp_num_8s >> 8,
827 			      rte_be_to_cpu_16(err_cqe->wqe_counter),
828 			      txq->wqe_ci, txq->cq_ci);
829 			MKSTR(name, "dpdk_mlx5_port_%u_txq_%u_index_%u_%u",
830 			      PORT_ID(txq_ctrl->priv), txq->idx,
831 			      txq_ctrl->dump_file_n, (uint32_t)rte_rdtsc());
832 			mlx5_dump_debug_information(name, NULL, err_str, 0);
833 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
834 						    (const void *)((uintptr_t)
835 						    txq->cqes),
836 						    sizeof(*err_cqe) *
837 						    (1 << txq->cqe_n));
838 			mlx5_dump_debug_information(name, "MLX5 Error SQ:",
839 						    (const void *)((uintptr_t)
840 						    txq->wqes),
841 						    MLX5_WQE_SIZE *
842 						    (1 << txq->wqe_n));
843 			txq_ctrl->dump_file_n++;
844 		}
845 		if (!seen)
846 			/*
847 			 * Count errors in WQEs units.
848 			 * Later it can be improved to count error packets,
849 			 * for example, by SQ parsing to find how much packets
850 			 * should be counted for each WQE.
851 			 */
852 			txq->stats.oerrors += ((txq->wqe_ci & wqe_m) -
853 						new_wqe_pi) & wqe_m;
854 		if (tx_recover_qp(txq_ctrl)) {
855 			/* Recovering failed - retry later on the same WQE. */
856 			return -1;
857 		}
858 		/* Release all the remaining buffers. */
859 		txq_free_elts(txq_ctrl);
860 	}
861 	return 0;
862 }
863 
864 /**
865  * Translate RX completion flags to packet type.
866  *
867  * @param[in] rxq
868  *   Pointer to RX queue structure.
869  * @param[in] cqe
870  *   Pointer to CQE.
871  *
872  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
873  *
874  * @return
875  *   Packet type for struct rte_mbuf.
876  */
877 static inline uint32_t
878 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
879 				   volatile struct mlx5_mini_cqe8 *mcqe)
880 {
881 	uint8_t idx;
882 	uint8_t ptype;
883 	uint8_t pinfo = (cqe->pkt_info & 0x3) << 6;
884 
885 	/* Get l3/l4 header from mini-CQE in case L3/L4 format*/
886 	if (mcqe == NULL ||
887 	    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
888 		ptype = (cqe->hdr_type_etc & 0xfc00) >> 10;
889 	else
890 		ptype = mcqe->hdr_type >> 2;
891 	/*
892 	 * The index to the array should have:
893 	 * bit[1:0] = l3_hdr_type
894 	 * bit[4:2] = l4_hdr_type
895 	 * bit[5] = ip_frag
896 	 * bit[6] = tunneled
897 	 * bit[7] = outer_l3_type
898 	 */
899 	idx = pinfo | ptype;
900 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
901 }
902 
903 /**
904  * Initialize Rx WQ and indexes.
905  *
906  * @param[in] rxq
907  *   Pointer to RX queue structure.
908  */
909 void
910 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
911 {
912 	const unsigned int wqe_n = 1 << rxq->elts_n;
913 	unsigned int i;
914 
915 	for (i = 0; (i != wqe_n); ++i) {
916 		volatile struct mlx5_wqe_data_seg *scat;
917 		uintptr_t addr;
918 		uint32_t byte_count;
919 
920 		if (mlx5_rxq_mprq_enabled(rxq)) {
921 			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
922 
923 			scat = &((volatile struct mlx5_wqe_mprq *)
924 				rxq->wqes)[i].dseg;
925 			addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
926 							 1 << rxq->strd_num_n);
927 			byte_count = (1 << rxq->strd_sz_n) *
928 					(1 << rxq->strd_num_n);
929 		} else {
930 			struct rte_mbuf *buf = (*rxq->elts)[i];
931 
932 			scat = &((volatile struct mlx5_wqe_data_seg *)
933 					rxq->wqes)[i];
934 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
935 			byte_count = DATA_LEN(buf);
936 		}
937 		/* scat->addr must be able to store a pointer. */
938 		MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t));
939 		*scat = (struct mlx5_wqe_data_seg){
940 			.addr = rte_cpu_to_be_64(addr),
941 			.byte_count = rte_cpu_to_be_32(byte_count),
942 			.lkey = mlx5_rx_addr2mr(rxq, addr),
943 		};
944 	}
945 	rxq->consumed_strd = 0;
946 	rxq->decompressed = 0;
947 	rxq->rq_pi = 0;
948 	rxq->zip = (struct rxq_zip){
949 		.ai = 0,
950 	};
951 	rxq->elts_ci = mlx5_rxq_mprq_enabled(rxq) ?
952 		(wqe_n >> rxq->sges_n) * (1 << rxq->strd_num_n) : 0;
953 	/* Update doorbell counter. */
954 	rxq->rq_ci = wqe_n >> rxq->sges_n;
955 	rte_io_wmb();
956 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
957 }
958 
959 /**
960  * Modify a Verbs/DevX queue state.
961  * This must be called from the primary process.
962  *
963  * @param dev
964  *   Pointer to Ethernet device.
965  * @param sm
966  *   State modify request parameters.
967  *
968  * @return
969  *   0 in case of success else non-zero value and rte_errno is set.
970  */
971 int
972 mlx5_queue_state_modify_primary(struct rte_eth_dev *dev,
973 			const struct mlx5_mp_arg_queue_state_modify *sm)
974 {
975 	int ret;
976 	struct mlx5_priv *priv = dev->data->dev_private;
977 
978 	if (sm->is_wq) {
979 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[sm->queue_id];
980 		struct mlx5_rxq_ctrl *rxq_ctrl =
981 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
982 
983 		ret = priv->obj_ops.rxq_obj_modify(rxq_ctrl->obj, sm->state);
984 		if (ret) {
985 			DRV_LOG(ERR, "Cannot change Rx WQ state to %u  - %s",
986 					sm->state, strerror(errno));
987 			rte_errno = errno;
988 			return ret;
989 		}
990 	} else {
991 		struct mlx5_txq_data *txq = (*priv->txqs)[sm->queue_id];
992 		struct mlx5_txq_ctrl *txq_ctrl =
993 			container_of(txq, struct mlx5_txq_ctrl, txq);
994 
995 		ret = priv->obj_ops.txq_obj_modify(txq_ctrl->obj,
996 						   MLX5_TXQ_MOD_ERR2RDY,
997 						   (uint8_t)priv->dev_port);
998 		if (ret)
999 			return ret;
1000 	}
1001 	return 0;
1002 }
1003 
1004 /**
1005  * Modify a Verbs queue state.
1006  *
1007  * @param dev
1008  *   Pointer to Ethernet device.
1009  * @param sm
1010  *   State modify request parameters.
1011  *
1012  * @return
1013  *   0 in case of success else non-zero value.
1014  */
1015 static int
1016 mlx5_queue_state_modify(struct rte_eth_dev *dev,
1017 			struct mlx5_mp_arg_queue_state_modify *sm)
1018 {
1019 	struct mlx5_priv *priv = dev->data->dev_private;
1020 	int ret = 0;
1021 
1022 	switch (rte_eal_process_type()) {
1023 	case RTE_PROC_PRIMARY:
1024 		ret = mlx5_queue_state_modify_primary(dev, sm);
1025 		break;
1026 	case RTE_PROC_SECONDARY:
1027 		ret = mlx5_mp_req_queue_state_modify(&priv->mp_id, sm);
1028 		break;
1029 	default:
1030 		break;
1031 	}
1032 	return ret;
1033 }
1034 
1035 /**
1036  * Handle a Rx error.
1037  * The function inserts the RQ state to reset when the first error CQE is
1038  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
1039  * it moves the RQ state to ready and initializes the RQ.
1040  * Next CQE identification and error counting are in the caller responsibility.
1041  *
1042  * @param[in] rxq
1043  *   Pointer to RX queue structure.
1044  * @param[in] vec
1045  *   1 when called from vectorized Rx burst, need to prepare mbufs for the RQ.
1046  *   0 when called from non-vectorized Rx burst.
1047  *
1048  * @return
1049  *   -1 in case of recovery error, otherwise the CQE status.
1050  */
1051 int
1052 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
1053 {
1054 	const uint16_t cqe_n = 1 << rxq->cqe_n;
1055 	const uint16_t cqe_mask = cqe_n - 1;
1056 	const uint16_t wqe_n = 1 << rxq->elts_n;
1057 	const uint16_t strd_n = 1 << rxq->strd_num_n;
1058 	struct mlx5_rxq_ctrl *rxq_ctrl =
1059 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
1060 	union {
1061 		volatile struct mlx5_cqe *cqe;
1062 		volatile struct mlx5_err_cqe *err_cqe;
1063 	} u = {
1064 		.cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
1065 	};
1066 	struct mlx5_mp_arg_queue_state_modify sm;
1067 	int ret;
1068 
1069 	switch (rxq->err_state) {
1070 	case MLX5_RXQ_ERR_STATE_NO_ERROR:
1071 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
1072 		/* Fall-through */
1073 	case MLX5_RXQ_ERR_STATE_NEED_RESET:
1074 		sm.is_wq = 1;
1075 		sm.queue_id = rxq->idx;
1076 		sm.state = IBV_WQS_RESET;
1077 		if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv), &sm))
1078 			return -1;
1079 		if (rxq_ctrl->dump_file_n <
1080 		    rxq_ctrl->priv->config.max_dump_files_num) {
1081 			MKSTR(err_str, "Unexpected CQE error syndrome "
1082 			      "0x%02x CQN = %u RQN = %u wqe_counter = %u"
1083 			      " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
1084 			      rxq->cqn, rxq_ctrl->wqn,
1085 			      rte_be_to_cpu_16(u.err_cqe->wqe_counter),
1086 			      rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
1087 			MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
1088 			      rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
1089 			mlx5_dump_debug_information(name, NULL, err_str, 0);
1090 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
1091 						    (const void *)((uintptr_t)
1092 								    rxq->cqes),
1093 						    sizeof(*u.cqe) * cqe_n);
1094 			mlx5_dump_debug_information(name, "MLX5 Error RQ:",
1095 						    (const void *)((uintptr_t)
1096 								    rxq->wqes),
1097 						    16 * wqe_n);
1098 			rxq_ctrl->dump_file_n++;
1099 		}
1100 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
1101 		/* Fall-through */
1102 	case MLX5_RXQ_ERR_STATE_NEED_READY:
1103 		ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
1104 		if (ret == MLX5_CQE_STATUS_HW_OWN) {
1105 			rte_io_wmb();
1106 			*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1107 			rte_io_wmb();
1108 			/*
1109 			 * The RQ consumer index must be zeroed while moving
1110 			 * from RESET state to RDY state.
1111 			 */
1112 			*rxq->rq_db = rte_cpu_to_be_32(0);
1113 			rte_io_wmb();
1114 			sm.is_wq = 1;
1115 			sm.queue_id = rxq->idx;
1116 			sm.state = IBV_WQS_RDY;
1117 			if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv),
1118 						    &sm))
1119 				return -1;
1120 			if (vec) {
1121 				const uint32_t elts_n =
1122 					mlx5_rxq_mprq_enabled(rxq) ?
1123 					wqe_n * strd_n : wqe_n;
1124 				const uint32_t e_mask = elts_n - 1;
1125 				uint32_t elts_ci =
1126 					mlx5_rxq_mprq_enabled(rxq) ?
1127 					rxq->elts_ci : rxq->rq_ci;
1128 				uint32_t elt_idx;
1129 				struct rte_mbuf **elt;
1130 				int i;
1131 				unsigned int n = elts_n - (elts_ci -
1132 							  rxq->rq_pi);
1133 
1134 				for (i = 0; i < (int)n; ++i) {
1135 					elt_idx = (elts_ci + i) & e_mask;
1136 					elt = &(*rxq->elts)[elt_idx];
1137 					*elt = rte_mbuf_raw_alloc(rxq->mp);
1138 					if (!*elt) {
1139 						for (i--; i >= 0; --i) {
1140 							elt_idx = (elts_ci +
1141 								   i) & elts_n;
1142 							elt = &(*rxq->elts)
1143 								[elt_idx];
1144 							rte_pktmbuf_free_seg
1145 								(*elt);
1146 						}
1147 						return -1;
1148 					}
1149 				}
1150 				for (i = 0; i < (int)elts_n; ++i) {
1151 					elt = &(*rxq->elts)[i];
1152 					DATA_LEN(*elt) =
1153 						(uint16_t)((*elt)->buf_len -
1154 						rte_pktmbuf_headroom(*elt));
1155 				}
1156 				/* Padding with a fake mbuf for vec Rx. */
1157 				for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
1158 					(*rxq->elts)[elts_n + i] =
1159 								&rxq->fake_mbuf;
1160 			}
1161 			mlx5_rxq_initialize(rxq);
1162 			rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
1163 		}
1164 		return ret;
1165 	default:
1166 		return -1;
1167 	}
1168 }
1169 
1170 /**
1171  * Get size of the next packet for a given CQE. For compressed CQEs, the
1172  * consumer index is updated only once all packets of the current one have
1173  * been processed.
1174  *
1175  * @param rxq
1176  *   Pointer to RX queue.
1177  * @param cqe
1178  *   CQE to process.
1179  * @param[out] mcqe
1180  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
1181  *   written.
1182  *
1183  * @return
1184  *   0 in case of empty CQE, otherwise the packet size in bytes.
1185  */
1186 static inline int
1187 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1188 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
1189 {
1190 	struct rxq_zip *zip = &rxq->zip;
1191 	uint16_t cqe_n = cqe_cnt + 1;
1192 	int len;
1193 	uint16_t idx, end;
1194 
1195 	do {
1196 		len = 0;
1197 		/* Process compressed data in the CQE and mini arrays. */
1198 		if (zip->ai) {
1199 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1200 				(volatile struct mlx5_mini_cqe8 (*)[8])
1201 				(uintptr_t)(&(*rxq->cqes)[zip->ca &
1202 							  cqe_cnt].pkt_info);
1203 			len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt &
1204 					       rxq->byte_mask);
1205 			*mcqe = &(*mc)[zip->ai & 7];
1206 			if ((++zip->ai & 7) == 0) {
1207 				/* Invalidate consumed CQEs */
1208 				idx = zip->ca;
1209 				end = zip->na;
1210 				while (idx != end) {
1211 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1212 						MLX5_CQE_INVALIDATE;
1213 					++idx;
1214 				}
1215 				/*
1216 				 * Increment consumer index to skip the number
1217 				 * of CQEs consumed. Hardware leaves holes in
1218 				 * the CQ ring for software use.
1219 				 */
1220 				zip->ca = zip->na;
1221 				zip->na += 8;
1222 			}
1223 			if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1224 				/* Invalidate the rest */
1225 				idx = zip->ca;
1226 				end = zip->cq_ci;
1227 
1228 				while (idx != end) {
1229 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1230 						MLX5_CQE_INVALIDATE;
1231 					++idx;
1232 				}
1233 				rxq->cq_ci = zip->cq_ci;
1234 				zip->ai = 0;
1235 			}
1236 		/*
1237 		 * No compressed data, get next CQE and verify if it is
1238 		 * compressed.
1239 		 */
1240 		} else {
1241 			int ret;
1242 			int8_t op_own;
1243 			uint32_t cq_ci;
1244 
1245 			ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1246 			if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
1247 				if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
1248 					     rxq->err_state)) {
1249 					ret = mlx5_rx_err_handle(rxq, 0);
1250 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
1251 					    ret == -1)
1252 						return 0;
1253 				} else {
1254 					return 0;
1255 				}
1256 			}
1257 			/*
1258 			 * Introduce the local variable to have queue cq_ci
1259 			 * index in queue structure always consistent with
1260 			 * actual CQE boundary (not pointing to the middle
1261 			 * of compressed CQE session).
1262 			 */
1263 			cq_ci = rxq->cq_ci + 1;
1264 			op_own = cqe->op_own;
1265 			if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1266 				volatile struct mlx5_mini_cqe8 (*mc)[8] =
1267 					(volatile struct mlx5_mini_cqe8 (*)[8])
1268 					(uintptr_t)(&(*rxq->cqes)
1269 						[cq_ci & cqe_cnt].pkt_info);
1270 
1271 				/* Fix endianness. */
1272 				zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1273 				/*
1274 				 * Current mini array position is the one
1275 				 * returned by check_cqe64().
1276 				 *
1277 				 * If completion comprises several mini arrays,
1278 				 * as a special case the second one is located
1279 				 * 7 CQEs after the initial CQE instead of 8
1280 				 * for subsequent ones.
1281 				 */
1282 				zip->ca = cq_ci;
1283 				zip->na = zip->ca + 7;
1284 				/* Compute the next non compressed CQE. */
1285 				zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1286 				/* Get packet size to return. */
1287 				len = rte_be_to_cpu_32((*mc)[0].byte_cnt &
1288 						       rxq->byte_mask);
1289 				*mcqe = &(*mc)[0];
1290 				zip->ai = 1;
1291 				/* Prefetch all to be invalidated */
1292 				idx = zip->ca;
1293 				end = zip->cq_ci;
1294 				while (idx != end) {
1295 					rte_prefetch0(&(*rxq->cqes)[(idx) &
1296 								    cqe_cnt]);
1297 					++idx;
1298 				}
1299 			} else {
1300 				rxq->cq_ci = cq_ci;
1301 				len = rte_be_to_cpu_32(cqe->byte_cnt);
1302 			}
1303 		}
1304 		if (unlikely(rxq->err_state)) {
1305 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1306 			++rxq->stats.idropped;
1307 		} else {
1308 			return len;
1309 		}
1310 	} while (1);
1311 }
1312 
1313 /**
1314  * Translate RX completion flags to offload flags.
1315  *
1316  * @param[in] cqe
1317  *   Pointer to CQE.
1318  *
1319  * @return
1320  *   Offload flags (ol_flags) for struct rte_mbuf.
1321  */
1322 static inline uint32_t
1323 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
1324 {
1325 	uint32_t ol_flags = 0;
1326 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1327 
1328 	ol_flags =
1329 		TRANSPOSE(flags,
1330 			  MLX5_CQE_RX_L3_HDR_VALID,
1331 			  PKT_RX_IP_CKSUM_GOOD) |
1332 		TRANSPOSE(flags,
1333 			  MLX5_CQE_RX_L4_HDR_VALID,
1334 			  PKT_RX_L4_CKSUM_GOOD);
1335 	return ol_flags;
1336 }
1337 
1338 /**
1339  * Fill in mbuf fields from RX completion flags.
1340  * Note that pkt->ol_flags should be initialized outside of this function.
1341  *
1342  * @param rxq
1343  *   Pointer to RX queue.
1344  * @param pkt
1345  *   mbuf to fill.
1346  * @param cqe
1347  *   CQE to process.
1348  * @param rss_hash_res
1349  *   Packet RSS Hash result.
1350  */
1351 static inline void
1352 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
1353 	       volatile struct mlx5_cqe *cqe,
1354 	       volatile struct mlx5_mini_cqe8 *mcqe)
1355 {
1356 	/* Update packet information. */
1357 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe, mcqe);
1358 
1359 	if (rxq->rss_hash) {
1360 		uint32_t rss_hash_res = 0;
1361 
1362 		/* If compressed, take hash result from mini-CQE. */
1363 		if (mcqe == NULL ||
1364 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_HASH)
1365 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
1366 		else
1367 			rss_hash_res = rte_be_to_cpu_32(mcqe->rx_hash_result);
1368 		if (rss_hash_res) {
1369 			pkt->hash.rss = rss_hash_res;
1370 			pkt->ol_flags |= PKT_RX_RSS_HASH;
1371 		}
1372 	}
1373 	if (rxq->mark) {
1374 		uint32_t mark = 0;
1375 
1376 		/* If compressed, take flow tag from mini-CQE. */
1377 		if (mcqe == NULL ||
1378 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
1379 			mark = cqe->sop_drop_qpn;
1380 		else
1381 			mark = ((mcqe->byte_cnt_flow & 0xff) << 8) |
1382 				(mcqe->flow_tag_high << 16);
1383 		if (MLX5_FLOW_MARK_IS_VALID(mark)) {
1384 			pkt->ol_flags |= PKT_RX_FDIR;
1385 			if (mark != RTE_BE32(MLX5_FLOW_MARK_DEFAULT)) {
1386 				pkt->ol_flags |= PKT_RX_FDIR_ID;
1387 				pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
1388 			}
1389 		}
1390 	}
1391 	if (rxq->dynf_meta && cqe->flow_table_metadata) {
1392 		pkt->ol_flags |= rxq->flow_meta_mask;
1393 		*RTE_MBUF_DYNFIELD(pkt, rxq->flow_meta_offset, uint32_t *) =
1394 			cqe->flow_table_metadata;
1395 	}
1396 	if (rxq->csum)
1397 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
1398 	if (rxq->vlan_strip) {
1399 		bool vlan_strip;
1400 
1401 		if (mcqe == NULL ||
1402 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
1403 			vlan_strip = cqe->hdr_type_etc &
1404 				     RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
1405 		else
1406 			vlan_strip = mcqe->hdr_type &
1407 				     RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
1408 		if (vlan_strip) {
1409 			pkt->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1410 			pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
1411 		}
1412 	}
1413 	if (rxq->hw_timestamp) {
1414 		uint64_t ts = rte_be_to_cpu_64(cqe->timestamp);
1415 
1416 		if (rxq->rt_timestamp)
1417 			ts = mlx5_txpp_convert_rx_ts(rxq->sh, ts);
1418 		mlx5_timestamp_set(pkt, rxq->timestamp_offset, ts);
1419 		pkt->ol_flags |= rxq->timestamp_rx_flag;
1420 	}
1421 }
1422 
1423 /**
1424  * DPDK callback for RX.
1425  *
1426  * @param dpdk_rxq
1427  *   Generic pointer to RX queue structure.
1428  * @param[out] pkts
1429  *   Array to store received packets.
1430  * @param pkts_n
1431  *   Maximum number of packets in array.
1432  *
1433  * @return
1434  *   Number of packets successfully received (<= pkts_n).
1435  */
1436 uint16_t
1437 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1438 {
1439 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1440 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1441 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1442 	const unsigned int sges_n = rxq->sges_n;
1443 	struct rte_mbuf *pkt = NULL;
1444 	struct rte_mbuf *seg = NULL;
1445 	volatile struct mlx5_cqe *cqe =
1446 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1447 	unsigned int i = 0;
1448 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1449 	int len = 0; /* keep its value across iterations. */
1450 
1451 	while (pkts_n) {
1452 		unsigned int idx = rq_ci & wqe_cnt;
1453 		volatile struct mlx5_wqe_data_seg *wqe =
1454 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
1455 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1456 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1457 
1458 		if (pkt)
1459 			NEXT(seg) = rep;
1460 		seg = rep;
1461 		rte_prefetch0(seg);
1462 		rte_prefetch0(cqe);
1463 		rte_prefetch0(wqe);
1464 		/* Allocate the buf from the same pool. */
1465 		rep = rte_mbuf_raw_alloc(seg->pool);
1466 		if (unlikely(rep == NULL)) {
1467 			++rxq->stats.rx_nombuf;
1468 			if (!pkt) {
1469 				/*
1470 				 * no buffers before we even started,
1471 				 * bail out silently.
1472 				 */
1473 				break;
1474 			}
1475 			while (pkt != seg) {
1476 				MLX5_ASSERT(pkt != (*rxq->elts)[idx]);
1477 				rep = NEXT(pkt);
1478 				NEXT(pkt) = NULL;
1479 				NB_SEGS(pkt) = 1;
1480 				rte_mbuf_raw_free(pkt);
1481 				pkt = rep;
1482 			}
1483 			break;
1484 		}
1485 		if (!pkt) {
1486 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1487 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
1488 			if (!len) {
1489 				rte_mbuf_raw_free(rep);
1490 				break;
1491 			}
1492 			pkt = seg;
1493 			MLX5_ASSERT(len >= (rxq->crc_present << 2));
1494 			pkt->ol_flags &= EXT_ATTACHED_MBUF;
1495 			rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
1496 			if (rxq->crc_present)
1497 				len -= RTE_ETHER_CRC_LEN;
1498 			PKT_LEN(pkt) = len;
1499 			if (cqe->lro_num_seg > 1) {
1500 				mlx5_lro_update_hdr
1501 					(rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
1502 					 mcqe, rxq, len);
1503 				pkt->ol_flags |= PKT_RX_LRO;
1504 				pkt->tso_segsz = len / cqe->lro_num_seg;
1505 			}
1506 		}
1507 		DATA_LEN(rep) = DATA_LEN(seg);
1508 		PKT_LEN(rep) = PKT_LEN(seg);
1509 		SET_DATA_OFF(rep, DATA_OFF(seg));
1510 		PORT(rep) = PORT(seg);
1511 		(*rxq->elts)[idx] = rep;
1512 		/*
1513 		 * Fill NIC descriptor with the new buffer.  The lkey and size
1514 		 * of the buffers are already known, only the buffer address
1515 		 * changes.
1516 		 */
1517 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1518 		/* If there's only one MR, no need to replace LKey in WQE. */
1519 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1520 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
1521 		if (len > DATA_LEN(seg)) {
1522 			len -= DATA_LEN(seg);
1523 			++NB_SEGS(pkt);
1524 			++rq_ci;
1525 			continue;
1526 		}
1527 		DATA_LEN(seg) = len;
1528 #ifdef MLX5_PMD_SOFT_COUNTERS
1529 		/* Increment bytes counter. */
1530 		rxq->stats.ibytes += PKT_LEN(pkt);
1531 #endif
1532 		/* Return packet. */
1533 		*(pkts++) = pkt;
1534 		pkt = NULL;
1535 		--pkts_n;
1536 		++i;
1537 		/* Align consumer index to the next stride. */
1538 		rq_ci >>= sges_n;
1539 		++rq_ci;
1540 		rq_ci <<= sges_n;
1541 	}
1542 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1543 		return 0;
1544 	/* Update the consumer index. */
1545 	rxq->rq_ci = rq_ci >> sges_n;
1546 	rte_io_wmb();
1547 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1548 	rte_io_wmb();
1549 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1550 #ifdef MLX5_PMD_SOFT_COUNTERS
1551 	/* Increment packets counter. */
1552 	rxq->stats.ipackets += i;
1553 #endif
1554 	return i;
1555 }
1556 
1557 /**
1558  * Update LRO packet TCP header.
1559  * The HW LRO feature doesn't update the TCP header after coalescing the
1560  * TCP segments but supplies information in CQE to fill it by SW.
1561  *
1562  * @param tcp
1563  *   Pointer to the TCP header.
1564  * @param cqe
1565  *   Pointer to the completion entry..
1566  * @param phcsum
1567  *   The L3 pseudo-header checksum.
1568  */
1569 static inline void
1570 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
1571 			volatile struct mlx5_cqe *__rte_restrict cqe,
1572 			uint32_t phcsum, uint8_t l4_type)
1573 {
1574 	/*
1575 	 * The HW calculates only the TCP payload checksum, need to complete
1576 	 * the TCP header checksum and the L3 pseudo-header checksum.
1577 	 */
1578 	uint32_t csum = phcsum + cqe->csum;
1579 
1580 	if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK ||
1581 	    l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) {
1582 		tcp->tcp_flags |= RTE_TCP_ACK_FLAG;
1583 		tcp->recv_ack = cqe->lro_ack_seq_num;
1584 		tcp->rx_win = cqe->lro_tcp_win;
1585 	}
1586 	if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK)
1587 		tcp->tcp_flags |= RTE_TCP_PSH_FLAG;
1588 	tcp->cksum = 0;
1589 	csum += rte_raw_cksum(tcp, (tcp->data_off >> 4) * 4);
1590 	csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff);
1591 	csum = (~csum) & 0xffff;
1592 	if (csum == 0)
1593 		csum = 0xffff;
1594 	tcp->cksum = csum;
1595 }
1596 
1597 /**
1598  * Update LRO packet headers.
1599  * The HW LRO feature doesn't update the L3/TCP headers after coalescing the
1600  * TCP segments but supply information in CQE to fill it by SW.
1601  *
1602  * @param padd
1603  *   The packet address.
1604  * @param cqe
1605  *   Pointer to the completion entry..
1606  * @param len
1607  *   The packet length.
1608  */
1609 static inline void
1610 mlx5_lro_update_hdr(uint8_t *__rte_restrict padd,
1611 		    volatile struct mlx5_cqe *__rte_restrict cqe,
1612 		    volatile struct mlx5_mini_cqe8 *mcqe,
1613 		    struct mlx5_rxq_data *rxq, uint32_t len)
1614 {
1615 	union {
1616 		struct rte_ether_hdr *eth;
1617 		struct rte_vlan_hdr *vlan;
1618 		struct rte_ipv4_hdr *ipv4;
1619 		struct rte_ipv6_hdr *ipv6;
1620 		struct rte_tcp_hdr *tcp;
1621 		uint8_t *hdr;
1622 	} h = {
1623 			.hdr = padd,
1624 	};
1625 	uint16_t proto = h.eth->ether_type;
1626 	uint32_t phcsum;
1627 	uint8_t l4_type;
1628 
1629 	h.eth++;
1630 	while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
1631 	       proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
1632 		proto = h.vlan->eth_proto;
1633 		h.vlan++;
1634 	}
1635 	if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
1636 		h.ipv4->time_to_live = cqe->lro_min_ttl;
1637 		h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1638 		h.ipv4->hdr_checksum = 0;
1639 		h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1640 		phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1641 		h.ipv4++;
1642 	} else {
1643 		h.ipv6->hop_limits = cqe->lro_min_ttl;
1644 		h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1645 						       sizeof(*h.ipv6));
1646 		phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1647 		h.ipv6++;
1648 	}
1649 	if (mcqe == NULL ||
1650 	    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
1651 		l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1652 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1653 	else
1654 		l4_type = (rte_be_to_cpu_16(mcqe->hdr_type) &
1655 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1656 	mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum, l4_type);
1657 }
1658 
1659 void
1660 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
1661 {
1662 	struct mlx5_mprq_buf *buf = opaque;
1663 
1664 	if (__atomic_load_n(&buf->refcnt, __ATOMIC_RELAXED) == 1) {
1665 		rte_mempool_put(buf->mp, buf);
1666 	} else if (unlikely(__atomic_sub_fetch(&buf->refcnt, 1,
1667 					       __ATOMIC_RELAXED) == 0)) {
1668 		__atomic_store_n(&buf->refcnt, 1, __ATOMIC_RELAXED);
1669 		rte_mempool_put(buf->mp, buf);
1670 	}
1671 }
1672 
1673 void
1674 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1675 {
1676 	mlx5_mprq_buf_free_cb(NULL, buf);
1677 }
1678 
1679 /**
1680  * DPDK callback for RX with Multi-Packet RQ support.
1681  *
1682  * @param dpdk_rxq
1683  *   Generic pointer to RX queue structure.
1684  * @param[out] pkts
1685  *   Array to store received packets.
1686  * @param pkts_n
1687  *   Maximum number of packets in array.
1688  *
1689  * @return
1690  *   Number of packets successfully received (<= pkts_n).
1691  */
1692 uint16_t
1693 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1694 {
1695 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1696 	const uint32_t strd_n = 1 << rxq->strd_num_n;
1697 	const uint32_t strd_sz = 1 << rxq->strd_sz_n;
1698 	const uint32_t cq_mask = (1 << rxq->cqe_n) - 1;
1699 	const uint32_t wq_mask = (1 << rxq->elts_n) - 1;
1700 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1701 	unsigned int i = 0;
1702 	uint32_t rq_ci = rxq->rq_ci;
1703 	uint16_t consumed_strd = rxq->consumed_strd;
1704 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1705 
1706 	while (i < pkts_n) {
1707 		struct rte_mbuf *pkt;
1708 		int ret;
1709 		uint32_t len;
1710 		uint16_t strd_cnt;
1711 		uint16_t strd_idx;
1712 		uint32_t byte_cnt;
1713 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1714 		enum mlx5_rqx_code rxq_code;
1715 
1716 		if (consumed_strd == strd_n) {
1717 			/* Replace WQE if the buffer is still in use. */
1718 			mprq_buf_replace(rxq, rq_ci & wq_mask);
1719 			/* Advance to the next WQE. */
1720 			consumed_strd = 0;
1721 			++rq_ci;
1722 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1723 		}
1724 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1725 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1726 		if (!ret)
1727 			break;
1728 		byte_cnt = ret;
1729 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1730 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1731 		if (rxq->crc_present)
1732 			len -= RTE_ETHER_CRC_LEN;
1733 		if (mcqe &&
1734 		    rxq->mcqe_format == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
1735 			strd_cnt = (len / strd_sz) + !!(len % strd_sz);
1736 		else
1737 			strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1738 				   MLX5_MPRQ_STRIDE_NUM_SHIFT;
1739 		MLX5_ASSERT(strd_cnt);
1740 		consumed_strd += strd_cnt;
1741 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1742 			continue;
1743 		strd_idx = rte_be_to_cpu_16(mcqe == NULL ?
1744 					cqe->wqe_counter :
1745 					mcqe->stride_idx);
1746 		MLX5_ASSERT(strd_idx < strd_n);
1747 		MLX5_ASSERT(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) &
1748 			    wq_mask));
1749 		pkt = rte_pktmbuf_alloc(rxq->mp);
1750 		if (unlikely(pkt == NULL)) {
1751 			++rxq->stats.rx_nombuf;
1752 			break;
1753 		}
1754 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1755 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1756 		if (rxq->crc_present)
1757 			len -= RTE_ETHER_CRC_LEN;
1758 		rxq_code = mprq_buf_to_pkt(rxq, pkt, len, buf,
1759 					   strd_idx, strd_cnt);
1760 		if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
1761 			rte_pktmbuf_free_seg(pkt);
1762 			if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
1763 				++rxq->stats.idropped;
1764 				continue;
1765 			}
1766 			if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
1767 				++rxq->stats.rx_nombuf;
1768 				break;
1769 			}
1770 		}
1771 		rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
1772 		if (cqe->lro_num_seg > 1) {
1773 			mlx5_lro_update_hdr(rte_pktmbuf_mtod(pkt, uint8_t *),
1774 					    cqe, mcqe, rxq, len);
1775 			pkt->ol_flags |= PKT_RX_LRO;
1776 			pkt->tso_segsz = len / cqe->lro_num_seg;
1777 		}
1778 		PKT_LEN(pkt) = len;
1779 		PORT(pkt) = rxq->port_id;
1780 #ifdef MLX5_PMD_SOFT_COUNTERS
1781 		/* Increment bytes counter. */
1782 		rxq->stats.ibytes += PKT_LEN(pkt);
1783 #endif
1784 		/* Return packet. */
1785 		*(pkts++) = pkt;
1786 		++i;
1787 	}
1788 	/* Update the consumer indexes. */
1789 	rxq->consumed_strd = consumed_strd;
1790 	rte_io_wmb();
1791 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1792 	if (rq_ci != rxq->rq_ci) {
1793 		rxq->rq_ci = rq_ci;
1794 		rte_io_wmb();
1795 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1796 	}
1797 #ifdef MLX5_PMD_SOFT_COUNTERS
1798 	/* Increment packets counter. */
1799 	rxq->stats.ipackets += i;
1800 #endif
1801 	return i;
1802 }
1803 
1804 /**
1805  * Dummy DPDK callback for TX.
1806  *
1807  * This function is used to temporarily replace the real callback during
1808  * unsafe control operations on the queue, or in case of error.
1809  *
1810  * @param dpdk_txq
1811  *   Generic pointer to TX queue structure.
1812  * @param[in] pkts
1813  *   Packets to transmit.
1814  * @param pkts_n
1815  *   Number of packets in array.
1816  *
1817  * @return
1818  *   Number of packets successfully transmitted (<= pkts_n).
1819  */
1820 uint16_t
1821 removed_tx_burst(void *dpdk_txq __rte_unused,
1822 		 struct rte_mbuf **pkts __rte_unused,
1823 		 uint16_t pkts_n __rte_unused)
1824 {
1825 	rte_mb();
1826 	return 0;
1827 }
1828 
1829 /**
1830  * Dummy DPDK callback for RX.
1831  *
1832  * This function is used to temporarily replace the real callback during
1833  * unsafe control operations on the queue, or in case of error.
1834  *
1835  * @param dpdk_rxq
1836  *   Generic pointer to RX queue structure.
1837  * @param[out] pkts
1838  *   Array to store received packets.
1839  * @param pkts_n
1840  *   Maximum number of packets in array.
1841  *
1842  * @return
1843  *   Number of packets successfully received (<= pkts_n).
1844  */
1845 uint16_t
1846 removed_rx_burst(void *dpdk_txq __rte_unused,
1847 		 struct rte_mbuf **pkts __rte_unused,
1848 		 uint16_t pkts_n __rte_unused)
1849 {
1850 	rte_mb();
1851 	return 0;
1852 }
1853 
1854 /*
1855  * Vectorized Rx/Tx routines are not compiled in when required vector
1856  * instructions are not supported on a target architecture. The following null
1857  * stubs are needed for linkage when those are not included outside of this file
1858  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1859  */
1860 
1861 __rte_weak uint16_t
1862 mlx5_rx_burst_vec(void *dpdk_txq __rte_unused,
1863 		  struct rte_mbuf **pkts __rte_unused,
1864 		  uint16_t pkts_n __rte_unused)
1865 {
1866 	return 0;
1867 }
1868 
1869 __rte_weak uint16_t
1870 mlx5_rx_burst_mprq_vec(void *dpdk_txq __rte_unused,
1871 		       struct rte_mbuf **pkts __rte_unused,
1872 		       uint16_t pkts_n __rte_unused)
1873 {
1874 	return 0;
1875 }
1876 
1877 __rte_weak int
1878 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1879 {
1880 	return -ENOTSUP;
1881 }
1882 
1883 __rte_weak int
1884 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1885 {
1886 	return -ENOTSUP;
1887 }
1888 
1889 /**
1890  * Free the mbufs from the linear array of pointers.
1891  *
1892  * @param txq
1893  *   Pointer to Tx queue structure.
1894  * @param pkts
1895  *   Pointer to array of packets to be free.
1896  * @param pkts_n
1897  *   Number of packets to be freed.
1898  * @param olx
1899  *   Configured Tx offloads mask. It is fully defined at
1900  *   compile time and may be used for optimization.
1901  */
1902 static __rte_always_inline void
1903 mlx5_tx_free_mbuf(struct mlx5_txq_data *__rte_restrict txq,
1904 		  struct rte_mbuf **__rte_restrict pkts,
1905 		  unsigned int pkts_n,
1906 		  unsigned int olx __rte_unused)
1907 {
1908 	struct rte_mempool *pool = NULL;
1909 	struct rte_mbuf **p_free = NULL;
1910 	struct rte_mbuf *mbuf;
1911 	unsigned int n_free = 0;
1912 
1913 	/*
1914 	 * The implemented algorithm eliminates
1915 	 * copying pointers to temporary array
1916 	 * for rte_mempool_put_bulk() calls.
1917 	 */
1918 	MLX5_ASSERT(pkts);
1919 	MLX5_ASSERT(pkts_n);
1920 	/*
1921 	 * Free mbufs directly to the pool in bulk
1922 	 * if fast free offload is engaged
1923 	 */
1924 	if (!MLX5_TXOFF_CONFIG(MULTI) && txq->fast_free) {
1925 		mbuf = *pkts;
1926 		pool = mbuf->pool;
1927 		rte_mempool_put_bulk(pool, (void *)pkts, pkts_n);
1928 		return;
1929 	}
1930 	for (;;) {
1931 		for (;;) {
1932 			/*
1933 			 * Decrement mbuf reference counter, detach
1934 			 * indirect and external buffers if needed.
1935 			 */
1936 			mbuf = rte_pktmbuf_prefree_seg(*pkts);
1937 			if (likely(mbuf != NULL)) {
1938 				MLX5_ASSERT(mbuf == *pkts);
1939 				if (likely(n_free != 0)) {
1940 					if (unlikely(pool != mbuf->pool))
1941 						/* From different pool. */
1942 						break;
1943 				} else {
1944 					/* Start new scan array. */
1945 					pool = mbuf->pool;
1946 					p_free = pkts;
1947 				}
1948 				++n_free;
1949 				++pkts;
1950 				--pkts_n;
1951 				if (unlikely(pkts_n == 0)) {
1952 					mbuf = NULL;
1953 					break;
1954 				}
1955 			} else {
1956 				/*
1957 				 * This happens if mbuf is still referenced.
1958 				 * We can't put it back to the pool, skip.
1959 				 */
1960 				++pkts;
1961 				--pkts_n;
1962 				if (unlikely(n_free != 0))
1963 					/* There is some array to free.*/
1964 					break;
1965 				if (unlikely(pkts_n == 0))
1966 					/* Last mbuf, nothing to free. */
1967 					return;
1968 			}
1969 		}
1970 		for (;;) {
1971 			/*
1972 			 * This loop is implemented to avoid multiple
1973 			 * inlining of rte_mempool_put_bulk().
1974 			 */
1975 			MLX5_ASSERT(pool);
1976 			MLX5_ASSERT(p_free);
1977 			MLX5_ASSERT(n_free);
1978 			/*
1979 			 * Free the array of pre-freed mbufs
1980 			 * belonging to the same memory pool.
1981 			 */
1982 			rte_mempool_put_bulk(pool, (void *)p_free, n_free);
1983 			if (unlikely(mbuf != NULL)) {
1984 				/* There is the request to start new scan. */
1985 				pool = mbuf->pool;
1986 				p_free = pkts++;
1987 				n_free = 1;
1988 				--pkts_n;
1989 				if (likely(pkts_n != 0))
1990 					break;
1991 				/*
1992 				 * This is the last mbuf to be freed.
1993 				 * Do one more loop iteration to complete.
1994 				 * This is rare case of the last unique mbuf.
1995 				 */
1996 				mbuf = NULL;
1997 				continue;
1998 			}
1999 			if (likely(pkts_n == 0))
2000 				return;
2001 			n_free = 0;
2002 			break;
2003 		}
2004 	}
2005 }
2006 /*
2007  * No inline version to free buffers for optimal call
2008  * on the tx_burst completion.
2009  */
2010 static __rte_noinline void
2011 __mlx5_tx_free_mbuf(struct mlx5_txq_data *__rte_restrict txq,
2012 		    struct rte_mbuf **__rte_restrict pkts,
2013 		    unsigned int pkts_n,
2014 		    unsigned int olx __rte_unused)
2015 {
2016 	mlx5_tx_free_mbuf(txq, pkts, pkts_n, olx);
2017 }
2018 
2019 /**
2020  * Free the mbuf from the elts ring buffer till new tail.
2021  *
2022  * @param txq
2023  *   Pointer to Tx queue structure.
2024  * @param tail
2025  *   Index in elts to free up to, becomes new elts tail.
2026  * @param olx
2027  *   Configured Tx offloads mask. It is fully defined at
2028  *   compile time and may be used for optimization.
2029  */
2030 static __rte_always_inline void
2031 mlx5_tx_free_elts(struct mlx5_txq_data *__rte_restrict txq,
2032 		  uint16_t tail,
2033 		  unsigned int olx __rte_unused)
2034 {
2035 	uint16_t n_elts = tail - txq->elts_tail;
2036 
2037 	MLX5_ASSERT(n_elts);
2038 	MLX5_ASSERT(n_elts <= txq->elts_s);
2039 	/*
2040 	 * Implement a loop to support ring buffer wraparound
2041 	 * with single inlining of mlx5_tx_free_mbuf().
2042 	 */
2043 	do {
2044 		unsigned int part;
2045 
2046 		part = txq->elts_s - (txq->elts_tail & txq->elts_m);
2047 		part = RTE_MIN(part, n_elts);
2048 		MLX5_ASSERT(part);
2049 		MLX5_ASSERT(part <= txq->elts_s);
2050 		mlx5_tx_free_mbuf(txq,
2051 				  &txq->elts[txq->elts_tail & txq->elts_m],
2052 				  part, olx);
2053 		txq->elts_tail += part;
2054 		n_elts -= part;
2055 	} while (n_elts);
2056 }
2057 
2058 /**
2059  * Store the mbuf being sent into elts ring buffer.
2060  * On Tx completion these mbufs will be freed.
2061  *
2062  * @param txq
2063  *   Pointer to Tx queue structure.
2064  * @param pkts
2065  *   Pointer to array of packets to be stored.
2066  * @param pkts_n
2067  *   Number of packets to be stored.
2068  * @param olx
2069  *   Configured Tx offloads mask. It is fully defined at
2070  *   compile time and may be used for optimization.
2071  */
2072 static __rte_always_inline void
2073 mlx5_tx_copy_elts(struct mlx5_txq_data *__rte_restrict txq,
2074 		  struct rte_mbuf **__rte_restrict pkts,
2075 		  unsigned int pkts_n,
2076 		  unsigned int olx __rte_unused)
2077 {
2078 	unsigned int part;
2079 	struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts;
2080 
2081 	MLX5_ASSERT(pkts);
2082 	MLX5_ASSERT(pkts_n);
2083 	part = txq->elts_s - (txq->elts_head & txq->elts_m);
2084 	MLX5_ASSERT(part);
2085 	MLX5_ASSERT(part <= txq->elts_s);
2086 	/* This code is a good candidate for vectorizing with SIMD. */
2087 	rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)),
2088 		   (void *)pkts,
2089 		   RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *));
2090 	txq->elts_head += pkts_n;
2091 	if (unlikely(part < pkts_n))
2092 		/* The copy is wrapping around the elts array. */
2093 		rte_memcpy((void *)elts, (void *)(pkts + part),
2094 			   (pkts_n - part) * sizeof(struct rte_mbuf *));
2095 }
2096 
2097 /**
2098  * Update completion queue consuming index via doorbell
2099  * and flush the completed data buffers.
2100  *
2101  * @param txq
2102  *   Pointer to TX queue structure.
2103  * @param valid CQE pointer
2104  *   if not NULL update txq->wqe_pi and flush the buffers
2105  * @param olx
2106  *   Configured Tx offloads mask. It is fully defined at
2107  *   compile time and may be used for optimization.
2108  */
2109 static __rte_always_inline void
2110 mlx5_tx_comp_flush(struct mlx5_txq_data *__rte_restrict txq,
2111 		   volatile struct mlx5_cqe *last_cqe,
2112 		   unsigned int olx __rte_unused)
2113 {
2114 	if (likely(last_cqe != NULL)) {
2115 		uint16_t tail;
2116 
2117 		txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter);
2118 		tail = txq->fcqs[(txq->cq_ci - 1) & txq->cqe_m];
2119 		if (likely(tail != txq->elts_tail)) {
2120 			mlx5_tx_free_elts(txq, tail, olx);
2121 			MLX5_ASSERT(tail == txq->elts_tail);
2122 		}
2123 	}
2124 }
2125 
2126 /**
2127  * Manage TX completions. This routine checks the CQ for
2128  * arrived CQEs, deduces the last accomplished WQE in SQ,
2129  * updates SQ producing index and frees all completed mbufs.
2130  *
2131  * @param txq
2132  *   Pointer to TX queue structure.
2133  * @param olx
2134  *   Configured Tx offloads mask. It is fully defined at
2135  *   compile time and may be used for optimization.
2136  *
2137  * NOTE: not inlined intentionally, it makes tx_burst
2138  * routine smaller, simple and faster - from experiments.
2139  */
2140 static void
2141 mlx5_tx_handle_completion(struct mlx5_txq_data *__rte_restrict txq,
2142 			  unsigned int olx __rte_unused)
2143 {
2144 	unsigned int count = MLX5_TX_COMP_MAX_CQE;
2145 	volatile struct mlx5_cqe *last_cqe = NULL;
2146 	bool ring_doorbell = false;
2147 	int ret;
2148 
2149 	do {
2150 		volatile struct mlx5_cqe *cqe;
2151 
2152 		cqe = &txq->cqes[txq->cq_ci & txq->cqe_m];
2153 		ret = check_cqe(cqe, txq->cqe_s, txq->cq_ci);
2154 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
2155 			if (likely(ret != MLX5_CQE_STATUS_ERR)) {
2156 				/* No new CQEs in completion queue. */
2157 				MLX5_ASSERT(ret == MLX5_CQE_STATUS_HW_OWN);
2158 				break;
2159 			}
2160 			/*
2161 			 * Some error occurred, try to restart.
2162 			 * We have no barrier after WQE related Doorbell
2163 			 * written, make sure all writes are completed
2164 			 * here, before we might perform SQ reset.
2165 			 */
2166 			rte_wmb();
2167 			ret = mlx5_tx_error_cqe_handle
2168 				(txq, (volatile struct mlx5_err_cqe *)cqe);
2169 			if (unlikely(ret < 0)) {
2170 				/*
2171 				 * Some error occurred on queue error
2172 				 * handling, we do not advance the index
2173 				 * here, allowing to retry on next call.
2174 				 */
2175 				return;
2176 			}
2177 			/*
2178 			 * We are going to fetch all entries with
2179 			 * MLX5_CQE_SYNDROME_WR_FLUSH_ERR status.
2180 			 * The send queue is supposed to be empty.
2181 			 */
2182 			ring_doorbell = true;
2183 			++txq->cq_ci;
2184 			txq->cq_pi = txq->cq_ci;
2185 			last_cqe = NULL;
2186 			continue;
2187 		}
2188 		/* Normal transmit completion. */
2189 		MLX5_ASSERT(txq->cq_ci != txq->cq_pi);
2190 #ifdef RTE_LIBRTE_MLX5_DEBUG
2191 		MLX5_ASSERT((txq->fcqs[txq->cq_ci & txq->cqe_m] >> 16) ==
2192 			    cqe->wqe_counter);
2193 #endif
2194 		ring_doorbell = true;
2195 		++txq->cq_ci;
2196 		last_cqe = cqe;
2197 		/*
2198 		 * We have to restrict the amount of processed CQEs
2199 		 * in one tx_burst routine call. The CQ may be large
2200 		 * and many CQEs may be updated by the NIC in one
2201 		 * transaction. Buffers freeing is time consuming,
2202 		 * multiple iterations may introduce significant
2203 		 * latency.
2204 		 */
2205 		if (likely(--count == 0))
2206 			break;
2207 	} while (true);
2208 	if (likely(ring_doorbell)) {
2209 		/* Ring doorbell to notify hardware. */
2210 		rte_compiler_barrier();
2211 		*txq->cq_db = rte_cpu_to_be_32(txq->cq_ci);
2212 		mlx5_tx_comp_flush(txq, last_cqe, olx);
2213 	}
2214 }
2215 
2216 /**
2217  * Check if the completion request flag should be set in the last WQE.
2218  * Both pushed mbufs and WQEs are monitored and the completion request
2219  * flag is set if any of thresholds is reached.
2220  *
2221  * @param txq
2222  *   Pointer to TX queue structure.
2223  * @param loc
2224  *   Pointer to burst routine local context.
2225  * @param olx
2226  *   Configured Tx offloads mask. It is fully defined at
2227  *   compile time and may be used for optimization.
2228  */
2229 static __rte_always_inline void
2230 mlx5_tx_request_completion(struct mlx5_txq_data *__rte_restrict txq,
2231 			   struct mlx5_txq_local *__rte_restrict loc,
2232 			   unsigned int olx)
2233 {
2234 	uint16_t head = txq->elts_head;
2235 	unsigned int part;
2236 
2237 	part = MLX5_TXOFF_CONFIG(INLINE) ?
2238 	       0 : loc->pkts_sent - loc->pkts_copy;
2239 	head += part;
2240 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
2241 	     (MLX5_TXOFF_CONFIG(INLINE) &&
2242 	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
2243 		volatile struct mlx5_wqe *last = loc->wqe_last;
2244 
2245 		MLX5_ASSERT(last);
2246 		txq->elts_comp = head;
2247 		if (MLX5_TXOFF_CONFIG(INLINE))
2248 			txq->wqe_comp = txq->wqe_ci;
2249 		/* Request unconditional completion on last WQE. */
2250 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
2251 					    MLX5_COMP_MODE_OFFSET);
2252 		/* Save elts_head in dedicated free on completion queue. */
2253 #ifdef RTE_LIBRTE_MLX5_DEBUG
2254 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head |
2255 			  (last->cseg.opcode >> 8) << 16;
2256 #else
2257 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head;
2258 #endif
2259 		/* A CQE slot must always be available. */
2260 		MLX5_ASSERT((txq->cq_pi - txq->cq_ci) <= txq->cqe_s);
2261 	}
2262 }
2263 
2264 /**
2265  * DPDK callback to check the status of a tx descriptor.
2266  *
2267  * @param tx_queue
2268  *   The tx queue.
2269  * @param[in] offset
2270  *   The index of the descriptor in the ring.
2271  *
2272  * @return
2273  *   The status of the tx descriptor.
2274  */
2275 int
2276 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
2277 {
2278 	struct mlx5_txq_data *__rte_restrict txq = tx_queue;
2279 	uint16_t used;
2280 
2281 	mlx5_tx_handle_completion(txq, 0);
2282 	used = txq->elts_head - txq->elts_tail;
2283 	if (offset < used)
2284 		return RTE_ETH_TX_DESC_FULL;
2285 	return RTE_ETH_TX_DESC_DONE;
2286 }
2287 
2288 /**
2289  * Build the Control Segment with specified opcode:
2290  * - MLX5_OPCODE_SEND
2291  * - MLX5_OPCODE_ENHANCED_MPSW
2292  * - MLX5_OPCODE_TSO
2293  *
2294  * @param txq
2295  *   Pointer to TX queue structure.
2296  * @param loc
2297  *   Pointer to burst routine local context.
2298  * @param wqe
2299  *   Pointer to WQE to fill with built Control Segment.
2300  * @param ds
2301  *   Supposed length of WQE in segments.
2302  * @param opcode
2303  *   SQ WQE opcode to put into Control Segment.
2304  * @param olx
2305  *   Configured Tx offloads mask. It is fully defined at
2306  *   compile time and may be used for optimization.
2307  */
2308 static __rte_always_inline void
2309 mlx5_tx_cseg_init(struct mlx5_txq_data *__rte_restrict txq,
2310 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
2311 		  struct mlx5_wqe *__rte_restrict wqe,
2312 		  unsigned int ds,
2313 		  unsigned int opcode,
2314 		  unsigned int olx __rte_unused)
2315 {
2316 	struct mlx5_wqe_cseg *__rte_restrict cs = &wqe->cseg;
2317 
2318 	/* For legacy MPW replace the EMPW by TSO with modifier. */
2319 	if (MLX5_TXOFF_CONFIG(MPW) && opcode == MLX5_OPCODE_ENHANCED_MPSW)
2320 		opcode = MLX5_OPCODE_TSO | MLX5_OPC_MOD_MPW << 24;
2321 	cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode);
2322 	cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2323 	cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
2324 			     MLX5_COMP_MODE_OFFSET);
2325 	cs->misc = RTE_BE32(0);
2326 }
2327 
2328 /**
2329  * Build the Synchronize Queue Segment with specified completion index.
2330  *
2331  * @param txq
2332  *   Pointer to TX queue structure.
2333  * @param loc
2334  *   Pointer to burst routine local context.
2335  * @param wqe
2336  *   Pointer to WQE to fill with built Control Segment.
2337  * @param wci
2338  *   Completion index in Clock Queue to wait.
2339  * @param olx
2340  *   Configured Tx offloads mask. It is fully defined at
2341  *   compile time and may be used for optimization.
2342  */
2343 static __rte_always_inline void
2344 mlx5_tx_wseg_init(struct mlx5_txq_data *restrict txq,
2345 		  struct mlx5_txq_local *restrict loc __rte_unused,
2346 		  struct mlx5_wqe *restrict wqe,
2347 		  unsigned int wci,
2348 		  unsigned int olx __rte_unused)
2349 {
2350 	struct mlx5_wqe_qseg *qs;
2351 
2352 	qs = RTE_PTR_ADD(wqe, MLX5_WSEG_SIZE);
2353 	qs->max_index = rte_cpu_to_be_32(wci);
2354 	qs->qpn_cqn = rte_cpu_to_be_32(txq->sh->txpp.clock_queue.cq_obj.cq->id);
2355 	qs->reserved0 = RTE_BE32(0);
2356 	qs->reserved1 = RTE_BE32(0);
2357 }
2358 
2359 /**
2360  * Build the Ethernet Segment without inlined data.
2361  * Supports Software Parser, Checksums and VLAN
2362  * insertion Tx offload features.
2363  *
2364  * @param txq
2365  *   Pointer to TX queue structure.
2366  * @param loc
2367  *   Pointer to burst routine local context.
2368  * @param wqe
2369  *   Pointer to WQE to fill with built Ethernet Segment.
2370  * @param olx
2371  *   Configured Tx offloads mask. It is fully defined at
2372  *   compile time and may be used for optimization.
2373  */
2374 static __rte_always_inline void
2375 mlx5_tx_eseg_none(struct mlx5_txq_data *__rte_restrict txq __rte_unused,
2376 		  struct mlx5_txq_local *__rte_restrict loc,
2377 		  struct mlx5_wqe *__rte_restrict wqe,
2378 		  unsigned int olx)
2379 {
2380 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2381 	uint32_t csum;
2382 
2383 	/*
2384 	 * Calculate and set check sum flags first, dword field
2385 	 * in segment may be shared with Software Parser flags.
2386 	 */
2387 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2388 	es->flags = rte_cpu_to_le_32(csum);
2389 	/*
2390 	 * Calculate and set Software Parser offsets and flags.
2391 	 * These flags a set for custom UDP and IP tunnel packets.
2392 	 */
2393 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2394 	/* Fill metadata field if needed. */
2395 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2396 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2397 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2398 	/* Engage VLAN tag insertion feature if requested. */
2399 	if (MLX5_TXOFF_CONFIG(VLAN) &&
2400 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
2401 		/*
2402 		 * We should get here only if device support
2403 		 * this feature correctly.
2404 		 */
2405 		MLX5_ASSERT(txq->vlan_en);
2406 		es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT |
2407 						  loc->mbuf->vlan_tci);
2408 	} else {
2409 		es->inline_hdr = RTE_BE32(0);
2410 	}
2411 }
2412 
2413 /**
2414  * Build the Ethernet Segment with minimal inlined data
2415  * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is
2416  * used to fill the gap in single WQEBB WQEs.
2417  * Supports Software Parser, Checksums and VLAN
2418  * insertion Tx offload features.
2419  *
2420  * @param txq
2421  *   Pointer to TX queue structure.
2422  * @param loc
2423  *   Pointer to burst routine local context.
2424  * @param wqe
2425  *   Pointer to WQE to fill with built Ethernet Segment.
2426  * @param vlan
2427  *   Length of VLAN tag insertion if any.
2428  * @param olx
2429  *   Configured Tx offloads mask. It is fully defined at
2430  *   compile time and may be used for optimization.
2431  */
2432 static __rte_always_inline void
2433 mlx5_tx_eseg_dmin(struct mlx5_txq_data *__rte_restrict txq __rte_unused,
2434 		  struct mlx5_txq_local *__rte_restrict loc,
2435 		  struct mlx5_wqe *__rte_restrict wqe,
2436 		  unsigned int vlan,
2437 		  unsigned int olx)
2438 {
2439 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2440 	uint32_t csum;
2441 	uint8_t *psrc, *pdst;
2442 
2443 	/*
2444 	 * Calculate and set check sum flags first, dword field
2445 	 * in segment may be shared with Software Parser flags.
2446 	 */
2447 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2448 	es->flags = rte_cpu_to_le_32(csum);
2449 	/*
2450 	 * Calculate and set Software Parser offsets and flags.
2451 	 * These flags a set for custom UDP and IP tunnel packets.
2452 	 */
2453 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2454 	/* Fill metadata field if needed. */
2455 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2456 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2457 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2458 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2459 	es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE);
2460 	es->inline_data = *(unaligned_uint16_t *)psrc;
2461 	psrc +=	sizeof(uint16_t);
2462 	pdst = (uint8_t *)(es + 1);
2463 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2464 		/* Implement VLAN tag insertion as part inline data. */
2465 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2466 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2467 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2468 		/* Insert VLAN ethertype + VLAN tag. */
2469 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2470 						((RTE_ETHER_TYPE_VLAN << 16) |
2471 						 loc->mbuf->vlan_tci);
2472 		pdst += sizeof(struct rte_vlan_hdr);
2473 		/* Copy the rest two bytes from packet data. */
2474 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2475 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2476 	} else {
2477 		/* Fill the gap in the title WQEBB with inline data. */
2478 		rte_mov16(pdst, psrc);
2479 	}
2480 }
2481 
2482 /**
2483  * Build the Ethernet Segment with entire packet
2484  * data inlining. Checks the boundary of WQEBB and
2485  * ring buffer wrapping, supports Software Parser,
2486  * Checksums and VLAN insertion Tx offload features.
2487  *
2488  * @param txq
2489  *   Pointer to TX queue structure.
2490  * @param loc
2491  *   Pointer to burst routine local context.
2492  * @param wqe
2493  *   Pointer to WQE to fill with built Ethernet Segment.
2494  * @param vlan
2495  *   Length of VLAN tag insertion if any.
2496  * @param inlen
2497  *   Length of data to inline (VLAN included, if any).
2498  * @param tso
2499  *   TSO flag, set mss field from the packet.
2500  * @param olx
2501  *   Configured Tx offloads mask. It is fully defined at
2502  *   compile time and may be used for optimization.
2503  *
2504  * @return
2505  *   Pointer to the next Data Segment (aligned and wrapped around).
2506  */
2507 static __rte_always_inline struct mlx5_wqe_dseg *
2508 mlx5_tx_eseg_data(struct mlx5_txq_data *__rte_restrict txq,
2509 		  struct mlx5_txq_local *__rte_restrict loc,
2510 		  struct mlx5_wqe *__rte_restrict wqe,
2511 		  unsigned int vlan,
2512 		  unsigned int inlen,
2513 		  unsigned int tso,
2514 		  unsigned int olx)
2515 {
2516 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2517 	uint32_t csum;
2518 	uint8_t *psrc, *pdst;
2519 	unsigned int part;
2520 
2521 	/*
2522 	 * Calculate and set check sum flags first, dword field
2523 	 * in segment may be shared with Software Parser flags.
2524 	 */
2525 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2526 	if (tso) {
2527 		csum <<= 24;
2528 		csum |= loc->mbuf->tso_segsz;
2529 		es->flags = rte_cpu_to_be_32(csum);
2530 	} else {
2531 		es->flags = rte_cpu_to_le_32(csum);
2532 	}
2533 	/*
2534 	 * Calculate and set Software Parser offsets and flags.
2535 	 * These flags a set for custom UDP and IP tunnel packets.
2536 	 */
2537 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2538 	/* Fill metadata field if needed. */
2539 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2540 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2541 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2542 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2543 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
2544 	es->inline_data = *(unaligned_uint16_t *)psrc;
2545 	psrc +=	sizeof(uint16_t);
2546 	pdst = (uint8_t *)(es + 1);
2547 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2548 		/* Implement VLAN tag insertion as part inline data. */
2549 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2550 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2551 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2552 		/* Insert VLAN ethertype + VLAN tag. */
2553 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2554 						((RTE_ETHER_TYPE_VLAN << 16) |
2555 						 loc->mbuf->vlan_tci);
2556 		pdst += sizeof(struct rte_vlan_hdr);
2557 		/* Copy the rest two bytes from packet data. */
2558 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2559 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2560 		psrc += sizeof(uint16_t);
2561 	} else {
2562 		/* Fill the gap in the title WQEBB with inline data. */
2563 		rte_mov16(pdst, psrc);
2564 		psrc += sizeof(rte_v128u32_t);
2565 	}
2566 	pdst = (uint8_t *)(es + 2);
2567 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2568 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2569 	inlen -= MLX5_ESEG_MIN_INLINE_SIZE;
2570 	if (!inlen) {
2571 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2572 		return (struct mlx5_wqe_dseg *)pdst;
2573 	}
2574 	/*
2575 	 * The WQEBB space availability is checked by caller.
2576 	 * Here we should be aware of WQE ring buffer wraparound only.
2577 	 */
2578 	part = (uint8_t *)txq->wqes_end - pdst;
2579 	part = RTE_MIN(part, inlen);
2580 	do {
2581 		rte_memcpy(pdst, psrc, part);
2582 		inlen -= part;
2583 		if (likely(!inlen)) {
2584 			/*
2585 			 * If return value is not used by the caller
2586 			 * the code below will be optimized out.
2587 			 */
2588 			pdst += part;
2589 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2590 			if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2591 				pdst = (uint8_t *)txq->wqes;
2592 			return (struct mlx5_wqe_dseg *)pdst;
2593 		}
2594 		pdst = (uint8_t *)txq->wqes;
2595 		psrc += part;
2596 		part = inlen;
2597 	} while (true);
2598 }
2599 
2600 /**
2601  * Copy data from chain of mbuf to the specified linear buffer.
2602  * Checksums and VLAN insertion Tx offload features. If data
2603  * from some mbuf copied completely this mbuf is freed. Local
2604  * structure is used to keep the byte stream state.
2605  *
2606  * @param pdst
2607  *   Pointer to the destination linear buffer.
2608  * @param loc
2609  *   Pointer to burst routine local context.
2610  * @param len
2611  *   Length of data to be copied.
2612  * @param must
2613  *   Length of data to be copied ignoring no inline hint.
2614  * @param olx
2615  *   Configured Tx offloads mask. It is fully defined at
2616  *   compile time and may be used for optimization.
2617  *
2618  * @return
2619  *   Number of actual copied data bytes. This is always greater than or
2620  *   equal to must parameter and might be lesser than len in no inline
2621  *   hint flag is encountered.
2622  */
2623 static __rte_always_inline unsigned int
2624 mlx5_tx_mseg_memcpy(uint8_t *pdst,
2625 		    struct mlx5_txq_local *__rte_restrict loc,
2626 		    unsigned int len,
2627 		    unsigned int must,
2628 		    unsigned int olx __rte_unused)
2629 {
2630 	struct rte_mbuf *mbuf;
2631 	unsigned int part, dlen, copy = 0;
2632 	uint8_t *psrc;
2633 
2634 	MLX5_ASSERT(len);
2635 	MLX5_ASSERT(must <= len);
2636 	do {
2637 		/* Allow zero length packets, must check first. */
2638 		dlen = rte_pktmbuf_data_len(loc->mbuf);
2639 		if (dlen <= loc->mbuf_off) {
2640 			/* Exhausted packet, just free. */
2641 			mbuf = loc->mbuf;
2642 			loc->mbuf = mbuf->next;
2643 			rte_pktmbuf_free_seg(mbuf);
2644 			loc->mbuf_off = 0;
2645 			MLX5_ASSERT(loc->mbuf_nseg > 1);
2646 			MLX5_ASSERT(loc->mbuf);
2647 			--loc->mbuf_nseg;
2648 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
2649 				unsigned int diff;
2650 
2651 				if (copy >= must) {
2652 					/*
2653 					 * We already copied the minimal
2654 					 * requested amount of data.
2655 					 */
2656 					return copy;
2657 				}
2658 				diff = must - copy;
2659 				if (diff <= rte_pktmbuf_data_len(loc->mbuf)) {
2660 					/*
2661 					 * Copy only the minimal required
2662 					 * part of the data buffer.
2663 					 */
2664 					len = diff;
2665 				}
2666 			}
2667 			continue;
2668 		}
2669 		dlen -= loc->mbuf_off;
2670 		psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2671 					       loc->mbuf_off);
2672 		part = RTE_MIN(len, dlen);
2673 		rte_memcpy(pdst, psrc, part);
2674 		copy += part;
2675 		loc->mbuf_off += part;
2676 		len -= part;
2677 		if (!len) {
2678 			if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) {
2679 				loc->mbuf_off = 0;
2680 				/* Exhausted packet, just free. */
2681 				mbuf = loc->mbuf;
2682 				loc->mbuf = mbuf->next;
2683 				rte_pktmbuf_free_seg(mbuf);
2684 				loc->mbuf_off = 0;
2685 				MLX5_ASSERT(loc->mbuf_nseg >= 1);
2686 				--loc->mbuf_nseg;
2687 			}
2688 			return copy;
2689 		}
2690 		pdst += part;
2691 	} while (true);
2692 }
2693 
2694 /**
2695  * Build the Ethernet Segment with inlined data from
2696  * multi-segment packet. Checks the boundary of WQEBB
2697  * and ring buffer wrapping, supports Software Parser,
2698  * Checksums and VLAN insertion Tx offload features.
2699  *
2700  * @param txq
2701  *   Pointer to TX queue structure.
2702  * @param loc
2703  *   Pointer to burst routine local context.
2704  * @param wqe
2705  *   Pointer to WQE to fill with built Ethernet Segment.
2706  * @param vlan
2707  *   Length of VLAN tag insertion if any.
2708  * @param inlen
2709  *   Length of data to inline (VLAN included, if any).
2710  * @param tso
2711  *   TSO flag, set mss field from the packet.
2712  * @param olx
2713  *   Configured Tx offloads mask. It is fully defined at
2714  *   compile time and may be used for optimization.
2715  *
2716  * @return
2717  *   Pointer to the next Data Segment (aligned and
2718  *   possible NOT wrapped around - caller should do
2719  *   wrapping check on its own).
2720  */
2721 static __rte_always_inline struct mlx5_wqe_dseg *
2722 mlx5_tx_eseg_mdat(struct mlx5_txq_data *__rte_restrict txq,
2723 		  struct mlx5_txq_local *__rte_restrict loc,
2724 		  struct mlx5_wqe *__rte_restrict wqe,
2725 		  unsigned int vlan,
2726 		  unsigned int inlen,
2727 		  unsigned int tso,
2728 		  unsigned int olx)
2729 {
2730 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2731 	uint32_t csum;
2732 	uint8_t *pdst;
2733 	unsigned int part, tlen = 0;
2734 
2735 	/*
2736 	 * Calculate and set check sum flags first, uint32_t field
2737 	 * in segment may be shared with Software Parser flags.
2738 	 */
2739 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2740 	if (tso) {
2741 		csum <<= 24;
2742 		csum |= loc->mbuf->tso_segsz;
2743 		es->flags = rte_cpu_to_be_32(csum);
2744 	} else {
2745 		es->flags = rte_cpu_to_le_32(csum);
2746 	}
2747 	/*
2748 	 * Calculate and set Software Parser offsets and flags.
2749 	 * These flags a set for custom UDP and IP tunnel packets.
2750 	 */
2751 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2752 	/* Fill metadata field if needed. */
2753 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2754 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2755 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2756 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2757 	pdst = (uint8_t *)&es->inline_data;
2758 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2759 		/* Implement VLAN tag insertion as part inline data. */
2760 		mlx5_tx_mseg_memcpy(pdst, loc,
2761 				    2 * RTE_ETHER_ADDR_LEN,
2762 				    2 * RTE_ETHER_ADDR_LEN, olx);
2763 		pdst += 2 * RTE_ETHER_ADDR_LEN;
2764 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2765 						((RTE_ETHER_TYPE_VLAN << 16) |
2766 						 loc->mbuf->vlan_tci);
2767 		pdst += sizeof(struct rte_vlan_hdr);
2768 		tlen += 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr);
2769 	}
2770 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2771 	/*
2772 	 * The WQEBB space availability is checked by caller.
2773 	 * Here we should be aware of WQE ring buffer wraparound only.
2774 	 */
2775 	part = (uint8_t *)txq->wqes_end - pdst;
2776 	part = RTE_MIN(part, inlen - tlen);
2777 	MLX5_ASSERT(part);
2778 	do {
2779 		unsigned int copy;
2780 
2781 		/*
2782 		 * Copying may be interrupted inside the routine
2783 		 * if run into no inline hint flag.
2784 		 */
2785 		copy = tlen >= txq->inlen_mode ? 0 : (txq->inlen_mode - tlen);
2786 		copy = mlx5_tx_mseg_memcpy(pdst, loc, part, copy, olx);
2787 		tlen += copy;
2788 		if (likely(inlen <= tlen) || copy < part) {
2789 			es->inline_hdr_sz = rte_cpu_to_be_16(tlen);
2790 			pdst += copy;
2791 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2792 			return (struct mlx5_wqe_dseg *)pdst;
2793 		}
2794 		pdst = (uint8_t *)txq->wqes;
2795 		part = inlen - tlen;
2796 	} while (true);
2797 }
2798 
2799 /**
2800  * Build the Data Segment of pointer type.
2801  *
2802  * @param txq
2803  *   Pointer to TX queue structure.
2804  * @param loc
2805  *   Pointer to burst routine local context.
2806  * @param dseg
2807  *   Pointer to WQE to fill with built Data Segment.
2808  * @param buf
2809  *   Data buffer to point.
2810  * @param len
2811  *   Data buffer length.
2812  * @param olx
2813  *   Configured Tx offloads mask. It is fully defined at
2814  *   compile time and may be used for optimization.
2815  */
2816 static __rte_always_inline void
2817 mlx5_tx_dseg_ptr(struct mlx5_txq_data *__rte_restrict txq,
2818 		 struct mlx5_txq_local *__rte_restrict loc,
2819 		 struct mlx5_wqe_dseg *__rte_restrict dseg,
2820 		 uint8_t *buf,
2821 		 unsigned int len,
2822 		 unsigned int olx __rte_unused)
2823 
2824 {
2825 	MLX5_ASSERT(len);
2826 	dseg->bcount = rte_cpu_to_be_32(len);
2827 	dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2828 	dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2829 }
2830 
2831 /**
2832  * Build the Data Segment of pointer type or inline
2833  * if data length is less than buffer in minimal
2834  * Data Segment size.
2835  *
2836  * @param txq
2837  *   Pointer to TX queue structure.
2838  * @param loc
2839  *   Pointer to burst routine local context.
2840  * @param dseg
2841  *   Pointer to WQE to fill with built Data Segment.
2842  * @param buf
2843  *   Data buffer to point.
2844  * @param len
2845  *   Data buffer length.
2846  * @param olx
2847  *   Configured Tx offloads mask. It is fully defined at
2848  *   compile time and may be used for optimization.
2849  */
2850 static __rte_always_inline void
2851 mlx5_tx_dseg_iptr(struct mlx5_txq_data *__rte_restrict txq,
2852 		  struct mlx5_txq_local *__rte_restrict loc,
2853 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
2854 		  uint8_t *buf,
2855 		  unsigned int len,
2856 		  unsigned int olx __rte_unused)
2857 
2858 {
2859 	uintptr_t dst, src;
2860 
2861 	MLX5_ASSERT(len);
2862 	if (len > MLX5_DSEG_MIN_INLINE_SIZE) {
2863 		dseg->bcount = rte_cpu_to_be_32(len);
2864 		dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2865 		dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2866 
2867 		return;
2868 	}
2869 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2870 	/* Unrolled implementation of generic rte_memcpy. */
2871 	dst = (uintptr_t)&dseg->inline_data[0];
2872 	src = (uintptr_t)buf;
2873 	if (len & 0x08) {
2874 #ifdef RTE_ARCH_STRICT_ALIGN
2875 		MLX5_ASSERT(dst == RTE_PTR_ALIGN(dst, sizeof(uint32_t)));
2876 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2877 		dst += sizeof(uint32_t);
2878 		src += sizeof(uint32_t);
2879 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2880 		dst += sizeof(uint32_t);
2881 		src += sizeof(uint32_t);
2882 #else
2883 		*(uint64_t *)dst = *(unaligned_uint64_t *)src;
2884 		dst += sizeof(uint64_t);
2885 		src += sizeof(uint64_t);
2886 #endif
2887 	}
2888 	if (len & 0x04) {
2889 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2890 		dst += sizeof(uint32_t);
2891 		src += sizeof(uint32_t);
2892 	}
2893 	if (len & 0x02) {
2894 		*(uint16_t *)dst = *(unaligned_uint16_t *)src;
2895 		dst += sizeof(uint16_t);
2896 		src += sizeof(uint16_t);
2897 	}
2898 	if (len & 0x01)
2899 		*(uint8_t *)dst = *(uint8_t *)src;
2900 }
2901 
2902 /**
2903  * Build the Data Segment of inlined data from single
2904  * segment packet, no VLAN insertion.
2905  *
2906  * @param txq
2907  *   Pointer to TX queue structure.
2908  * @param loc
2909  *   Pointer to burst routine local context.
2910  * @param dseg
2911  *   Pointer to WQE to fill with built Data Segment.
2912  * @param buf
2913  *   Data buffer to point.
2914  * @param len
2915  *   Data buffer length.
2916  * @param olx
2917  *   Configured Tx offloads mask. It is fully defined at
2918  *   compile time and may be used for optimization.
2919  *
2920  * @return
2921  *   Pointer to the next Data Segment after inlined data.
2922  *   Ring buffer wraparound check is needed. We do not
2923  *   do it here because it may not be needed for the
2924  *   last packet in the eMPW session.
2925  */
2926 static __rte_always_inline struct mlx5_wqe_dseg *
2927 mlx5_tx_dseg_empw(struct mlx5_txq_data *__rte_restrict txq,
2928 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
2929 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
2930 		  uint8_t *buf,
2931 		  unsigned int len,
2932 		  unsigned int olx __rte_unused)
2933 {
2934 	unsigned int part;
2935 	uint8_t *pdst;
2936 
2937 	if (!MLX5_TXOFF_CONFIG(MPW)) {
2938 		/* Store the descriptor byte counter for eMPW sessions. */
2939 		dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2940 		pdst = &dseg->inline_data[0];
2941 	} else {
2942 		/* The entire legacy MPW session counter is stored on close. */
2943 		pdst = (uint8_t *)dseg;
2944 	}
2945 	/*
2946 	 * The WQEBB space availability is checked by caller.
2947 	 * Here we should be aware of WQE ring buffer wraparound only.
2948 	 */
2949 	part = (uint8_t *)txq->wqes_end - pdst;
2950 	part = RTE_MIN(part, len);
2951 	do {
2952 		rte_memcpy(pdst, buf, part);
2953 		len -= part;
2954 		if (likely(!len)) {
2955 			pdst += part;
2956 			if (!MLX5_TXOFF_CONFIG(MPW))
2957 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2958 			/* Note: no final wraparound check here. */
2959 			return (struct mlx5_wqe_dseg *)pdst;
2960 		}
2961 		pdst = (uint8_t *)txq->wqes;
2962 		buf += part;
2963 		part = len;
2964 	} while (true);
2965 }
2966 
2967 /**
2968  * Build the Data Segment of inlined data from single
2969  * segment packet with VLAN insertion.
2970  *
2971  * @param txq
2972  *   Pointer to TX queue structure.
2973  * @param loc
2974  *   Pointer to burst routine local context.
2975  * @param dseg
2976  *   Pointer to the dseg fill with built Data Segment.
2977  * @param buf
2978  *   Data buffer to point.
2979  * @param len
2980  *   Data buffer length.
2981  * @param olx
2982  *   Configured Tx offloads mask. It is fully defined at
2983  *   compile time and may be used for optimization.
2984  *
2985  * @return
2986  *   Pointer to the next Data Segment after inlined data.
2987  *   Ring buffer wraparound check is needed.
2988  */
2989 static __rte_always_inline struct mlx5_wqe_dseg *
2990 mlx5_tx_dseg_vlan(struct mlx5_txq_data *__rte_restrict txq,
2991 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
2992 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
2993 		  uint8_t *buf,
2994 		  unsigned int len,
2995 		  unsigned int olx __rte_unused)
2996 
2997 {
2998 	unsigned int part;
2999 	uint8_t *pdst;
3000 
3001 	MLX5_ASSERT(len > MLX5_ESEG_MIN_INLINE_SIZE);
3002 	if (!MLX5_TXOFF_CONFIG(MPW)) {
3003 		/* Store the descriptor byte counter for eMPW sessions. */
3004 		dseg->bcount = rte_cpu_to_be_32
3005 				((len + sizeof(struct rte_vlan_hdr)) |
3006 				 MLX5_ETH_WQE_DATA_INLINE);
3007 		pdst = &dseg->inline_data[0];
3008 	} else {
3009 		/* The entire legacy MPW session counter is stored on close. */
3010 		pdst = (uint8_t *)dseg;
3011 	}
3012 	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
3013 	buf += MLX5_DSEG_MIN_INLINE_SIZE;
3014 	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
3015 	len -= MLX5_DSEG_MIN_INLINE_SIZE;
3016 	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
3017 	MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
3018 	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
3019 		pdst = (uint8_t *)txq->wqes;
3020 	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) |
3021 					      loc->mbuf->vlan_tci);
3022 	pdst += sizeof(struct rte_vlan_hdr);
3023 	/*
3024 	 * The WQEBB space availability is checked by caller.
3025 	 * Here we should be aware of WQE ring buffer wraparound only.
3026 	 */
3027 	part = (uint8_t *)txq->wqes_end - pdst;
3028 	part = RTE_MIN(part, len);
3029 	do {
3030 		rte_memcpy(pdst, buf, part);
3031 		len -= part;
3032 		if (likely(!len)) {
3033 			pdst += part;
3034 			if (!MLX5_TXOFF_CONFIG(MPW))
3035 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
3036 			/* Note: no final wraparound check here. */
3037 			return (struct mlx5_wqe_dseg *)pdst;
3038 		}
3039 		pdst = (uint8_t *)txq->wqes;
3040 		buf += part;
3041 		part = len;
3042 	} while (true);
3043 }
3044 
3045 /**
3046  * Build the Ethernet Segment with optionally inlined data with
3047  * VLAN insertion and following Data Segments (if any) from
3048  * multi-segment packet. Used by ordinary send and TSO.
3049  *
3050  * @param txq
3051  *   Pointer to TX queue structure.
3052  * @param loc
3053  *   Pointer to burst routine local context.
3054  * @param wqe
3055  *   Pointer to WQE to fill with built Ethernet/Data Segments.
3056  * @param vlan
3057  *   Length of VLAN header to insert, 0 means no VLAN insertion.
3058  * @param inlen
3059  *   Data length to inline. For TSO this parameter specifies
3060  *   exact value, for ordinary send routine can be aligned by
3061  *   caller to provide better WQE space saving and data buffer
3062  *   start address alignment. This length includes VLAN header
3063  *   being inserted.
3064  * @param tso
3065  *   Zero means ordinary send, inlined data can be extended,
3066  *   otherwise this is TSO, inlined data length is fixed.
3067  * @param olx
3068  *   Configured Tx offloads mask. It is fully defined at
3069  *   compile time and may be used for optimization.
3070  *
3071  * @return
3072  *   Actual size of built WQE in segments.
3073  */
3074 static __rte_always_inline unsigned int
3075 mlx5_tx_mseg_build(struct mlx5_txq_data *__rte_restrict txq,
3076 		   struct mlx5_txq_local *__rte_restrict loc,
3077 		   struct mlx5_wqe *__rte_restrict wqe,
3078 		   unsigned int vlan,
3079 		   unsigned int inlen,
3080 		   unsigned int tso,
3081 		   unsigned int olx __rte_unused)
3082 {
3083 	struct mlx5_wqe_dseg *__rte_restrict dseg;
3084 	unsigned int ds;
3085 
3086 	MLX5_ASSERT((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen);
3087 	loc->mbuf_nseg = NB_SEGS(loc->mbuf);
3088 	loc->mbuf_off = 0;
3089 
3090 	dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx);
3091 	if (!loc->mbuf_nseg)
3092 		goto dseg_done;
3093 	/*
3094 	 * There are still some mbuf remaining, not inlined.
3095 	 * The first mbuf may be partially inlined and we
3096 	 * must process the possible non-zero data offset.
3097 	 */
3098 	if (loc->mbuf_off) {
3099 		unsigned int dlen;
3100 		uint8_t *dptr;
3101 
3102 		/*
3103 		 * Exhausted packets must be dropped before.
3104 		 * Non-zero offset means there are some data
3105 		 * remained in the packet.
3106 		 */
3107 		MLX5_ASSERT(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf));
3108 		MLX5_ASSERT(rte_pktmbuf_data_len(loc->mbuf));
3109 		dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
3110 					       loc->mbuf_off);
3111 		dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off;
3112 		/*
3113 		 * Build the pointer/minimal data Data Segment.
3114 		 * Do ring buffer wrapping check in advance.
3115 		 */
3116 		if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3117 			dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3118 		mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx);
3119 		/* Store the mbuf to be freed on completion. */
3120 		MLX5_ASSERT(loc->elts_free);
3121 		txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3122 		--loc->elts_free;
3123 		++dseg;
3124 		if (--loc->mbuf_nseg == 0)
3125 			goto dseg_done;
3126 		loc->mbuf = loc->mbuf->next;
3127 		loc->mbuf_off = 0;
3128 	}
3129 	do {
3130 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3131 			struct rte_mbuf *mbuf;
3132 
3133 			/* Zero length segment found, just skip. */
3134 			mbuf = loc->mbuf;
3135 			loc->mbuf = loc->mbuf->next;
3136 			rte_pktmbuf_free_seg(mbuf);
3137 			if (--loc->mbuf_nseg == 0)
3138 				break;
3139 		} else {
3140 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3141 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3142 			mlx5_tx_dseg_iptr
3143 				(txq, loc, dseg,
3144 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3145 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3146 			MLX5_ASSERT(loc->elts_free);
3147 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3148 			--loc->elts_free;
3149 			++dseg;
3150 			if (--loc->mbuf_nseg == 0)
3151 				break;
3152 			loc->mbuf = loc->mbuf->next;
3153 		}
3154 	} while (true);
3155 
3156 dseg_done:
3157 	/* Calculate actual segments used from the dseg pointer. */
3158 	if ((uintptr_t)wqe < (uintptr_t)dseg)
3159 		ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE;
3160 	else
3161 		ds = (((uintptr_t)dseg - (uintptr_t)wqe) +
3162 		      txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE;
3163 	return ds;
3164 }
3165 
3166 /**
3167  * The routine checks timestamp flag in the current packet,
3168  * and push WAIT WQE into the queue if scheduling is required.
3169  *
3170  * @param txq
3171  *   Pointer to TX queue structure.
3172  * @param loc
3173  *   Pointer to burst routine local context.
3174  * @param olx
3175  *   Configured Tx offloads mask. It is fully defined at
3176  *   compile time and may be used for optimization.
3177  *
3178  * @return
3179  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3180  *   MLX5_TXCMP_CODE_SINGLE - continue processing with the packet.
3181  *   MLX5_TXCMP_CODE_MULTI - the WAIT inserted, continue processing.
3182  * Local context variables partially updated.
3183  */
3184 static __rte_always_inline enum mlx5_txcmp_code
3185 mlx5_tx_schedule_send(struct mlx5_txq_data *restrict txq,
3186 		      struct mlx5_txq_local *restrict loc,
3187 		      unsigned int olx)
3188 {
3189 	if (MLX5_TXOFF_CONFIG(TXPP) &&
3190 	    loc->mbuf->ol_flags & txq->ts_mask) {
3191 		struct mlx5_wqe *wqe;
3192 		uint64_t ts;
3193 		int32_t wci;
3194 
3195 		/*
3196 		 * Estimate the required space quickly and roughly.
3197 		 * We would like to ensure the packet can be pushed
3198 		 * to the queue and we won't get the orphan WAIT WQE.
3199 		 */
3200 		if (loc->wqe_free <= MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE ||
3201 		    loc->elts_free < NB_SEGS(loc->mbuf))
3202 			return MLX5_TXCMP_CODE_EXIT;
3203 		/* Convert the timestamp into completion to wait. */
3204 		ts = *RTE_MBUF_DYNFIELD(loc->mbuf, txq->ts_offset, uint64_t *);
3205 		wci = mlx5_txpp_convert_tx_ts(txq->sh, ts);
3206 		if (unlikely(wci < 0))
3207 			return MLX5_TXCMP_CODE_SINGLE;
3208 		/* Build the WAIT WQE with specified completion. */
3209 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3210 		mlx5_tx_cseg_init(txq, loc, wqe, 2, MLX5_OPCODE_WAIT, olx);
3211 		mlx5_tx_wseg_init(txq, loc, wqe, wci, olx);
3212 		++txq->wqe_ci;
3213 		--loc->wqe_free;
3214 		return MLX5_TXCMP_CODE_MULTI;
3215 	}
3216 	return MLX5_TXCMP_CODE_SINGLE;
3217 }
3218 
3219 /**
3220  * Tx one packet function for multi-segment TSO. Supports all
3221  * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs,
3222  * sends one packet per WQE.
3223  *
3224  * This routine is responsible for storing processed mbuf
3225  * into elts ring buffer and update elts_head.
3226  *
3227  * @param txq
3228  *   Pointer to TX queue structure.
3229  * @param loc
3230  *   Pointer to burst routine local context.
3231  * @param olx
3232  *   Configured Tx offloads mask. It is fully defined at
3233  *   compile time and may be used for optimization.
3234  *
3235  * @return
3236  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3237  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3238  * Local context variables partially updated.
3239  */
3240 static __rte_always_inline enum mlx5_txcmp_code
3241 mlx5_tx_packet_multi_tso(struct mlx5_txq_data *__rte_restrict txq,
3242 			struct mlx5_txq_local *__rte_restrict loc,
3243 			unsigned int olx)
3244 {
3245 	struct mlx5_wqe *__rte_restrict wqe;
3246 	unsigned int ds, dlen, inlen, ntcp, vlan = 0;
3247 
3248 	if (MLX5_TXOFF_CONFIG(TXPP)) {
3249 		enum mlx5_txcmp_code wret;
3250 
3251 		/* Generate WAIT for scheduling if requested. */
3252 		wret = mlx5_tx_schedule_send(txq, loc, olx);
3253 		if (wret == MLX5_TXCMP_CODE_EXIT)
3254 			return MLX5_TXCMP_CODE_EXIT;
3255 		if (wret == MLX5_TXCMP_CODE_ERROR)
3256 			return MLX5_TXCMP_CODE_ERROR;
3257 	}
3258 	/*
3259 	 * Calculate data length to be inlined to estimate
3260 	 * the required space in WQE ring buffer.
3261 	 */
3262 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3263 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3264 		vlan = sizeof(struct rte_vlan_hdr);
3265 	inlen = loc->mbuf->l2_len + vlan +
3266 		loc->mbuf->l3_len + loc->mbuf->l4_len;
3267 	if (unlikely((!inlen || !loc->mbuf->tso_segsz)))
3268 		return MLX5_TXCMP_CODE_ERROR;
3269 	if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3270 		inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len;
3271 	/* Packet must contain all TSO headers. */
3272 	if (unlikely(inlen > MLX5_MAX_TSO_HEADER ||
3273 		     inlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3274 		     inlen > (dlen + vlan)))
3275 		return MLX5_TXCMP_CODE_ERROR;
3276 	MLX5_ASSERT(inlen >= txq->inlen_mode);
3277 	/*
3278 	 * Check whether there are enough free WQEBBs:
3279 	 * - Control Segment
3280 	 * - Ethernet Segment
3281 	 * - First Segment of inlined Ethernet data
3282 	 * - ... data continued ...
3283 	 * - Data Segments of pointer/min inline type
3284 	 */
3285 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3286 				       MLX5_ESEG_MIN_INLINE_SIZE +
3287 				       MLX5_WSEG_SIZE +
3288 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3289 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3290 		return MLX5_TXCMP_CODE_EXIT;
3291 	/* Check for maximal WQE size. */
3292 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3293 		return MLX5_TXCMP_CODE_ERROR;
3294 #ifdef MLX5_PMD_SOFT_COUNTERS
3295 	/* Update sent data bytes/packets counters. */
3296 	ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) /
3297 		loc->mbuf->tso_segsz;
3298 	/*
3299 	 * One will be added for mbuf itself
3300 	 * at the end of the mlx5_tx_burst from
3301 	 * loc->pkts_sent field.
3302 	 */
3303 	--ntcp;
3304 	txq->stats.opackets += ntcp;
3305 	txq->stats.obytes += dlen + vlan + ntcp * inlen;
3306 #endif
3307 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3308 	loc->wqe_last = wqe;
3309 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx);
3310 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx);
3311 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3312 	txq->wqe_ci += (ds + 3) / 4;
3313 	loc->wqe_free -= (ds + 3) / 4;
3314 	return MLX5_TXCMP_CODE_MULTI;
3315 }
3316 
3317 /**
3318  * Tx one packet function for multi-segment SEND. Supports all
3319  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3320  * sends one packet per WQE, without any data inlining in
3321  * Ethernet Segment.
3322  *
3323  * This routine is responsible for storing processed mbuf
3324  * into elts ring buffer and update elts_head.
3325  *
3326  * @param txq
3327  *   Pointer to TX queue structure.
3328  * @param loc
3329  *   Pointer to burst routine local context.
3330  * @param olx
3331  *   Configured Tx offloads mask. It is fully defined at
3332  *   compile time and may be used for optimization.
3333  *
3334  * @return
3335  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3336  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3337  * Local context variables partially updated.
3338  */
3339 static __rte_always_inline enum mlx5_txcmp_code
3340 mlx5_tx_packet_multi_send(struct mlx5_txq_data *__rte_restrict txq,
3341 			  struct mlx5_txq_local *__rte_restrict loc,
3342 			  unsigned int olx)
3343 {
3344 	struct mlx5_wqe_dseg *__rte_restrict dseg;
3345 	struct mlx5_wqe *__rte_restrict wqe;
3346 	unsigned int ds, nseg;
3347 
3348 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3349 	if (MLX5_TXOFF_CONFIG(TXPP)) {
3350 		enum mlx5_txcmp_code wret;
3351 
3352 		/* Generate WAIT for scheduling if requested. */
3353 		wret = mlx5_tx_schedule_send(txq, loc, olx);
3354 		if (wret == MLX5_TXCMP_CODE_EXIT)
3355 			return MLX5_TXCMP_CODE_EXIT;
3356 		if (wret == MLX5_TXCMP_CODE_ERROR)
3357 			return MLX5_TXCMP_CODE_ERROR;
3358 	}
3359 	/*
3360 	 * No inline at all, it means the CPU cycles saving
3361 	 * is prioritized at configuration, we should not
3362 	 * copy any packet data to WQE.
3363 	 */
3364 	nseg = NB_SEGS(loc->mbuf);
3365 	ds = 2 + nseg;
3366 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3367 		return MLX5_TXCMP_CODE_EXIT;
3368 	/* Check for maximal WQE size. */
3369 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3370 		return MLX5_TXCMP_CODE_ERROR;
3371 	/*
3372 	 * Some Tx offloads may cause an error if
3373 	 * packet is not long enough, check against
3374 	 * assumed minimal length.
3375 	 */
3376 	if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE)
3377 		return MLX5_TXCMP_CODE_ERROR;
3378 #ifdef MLX5_PMD_SOFT_COUNTERS
3379 	/* Update sent data bytes counter. */
3380 	txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf);
3381 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3382 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3383 		txq->stats.obytes += sizeof(struct rte_vlan_hdr);
3384 #endif
3385 	/*
3386 	 * SEND WQE, one WQEBB:
3387 	 * - Control Segment, SEND opcode
3388 	 * - Ethernet Segment, optional VLAN, no inline
3389 	 * - Data Segments, pointer only type
3390 	 */
3391 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3392 	loc->wqe_last = wqe;
3393 	mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx);
3394 	mlx5_tx_eseg_none(txq, loc, wqe, olx);
3395 	dseg = &wqe->dseg[0];
3396 	do {
3397 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3398 			struct rte_mbuf *mbuf;
3399 
3400 			/*
3401 			 * Zero length segment found, have to
3402 			 * correct total size of WQE in segments.
3403 			 * It is supposed to be rare occasion, so
3404 			 * in normal case (no zero length segments)
3405 			 * we avoid extra writing to the Control
3406 			 * Segment.
3407 			 */
3408 			--ds;
3409 			wqe->cseg.sq_ds -= RTE_BE32(1);
3410 			mbuf = loc->mbuf;
3411 			loc->mbuf = mbuf->next;
3412 			rte_pktmbuf_free_seg(mbuf);
3413 			if (--nseg == 0)
3414 				break;
3415 		} else {
3416 			mlx5_tx_dseg_ptr
3417 				(txq, loc, dseg,
3418 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3419 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3420 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3421 			--loc->elts_free;
3422 			if (--nseg == 0)
3423 				break;
3424 			++dseg;
3425 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3426 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3427 			loc->mbuf = loc->mbuf->next;
3428 		}
3429 	} while (true);
3430 	txq->wqe_ci += (ds + 3) / 4;
3431 	loc->wqe_free -= (ds + 3) / 4;
3432 	return MLX5_TXCMP_CODE_MULTI;
3433 }
3434 
3435 /**
3436  * Tx one packet function for multi-segment SEND. Supports all
3437  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3438  * sends one packet per WQE, with data inlining in
3439  * Ethernet Segment and minimal Data Segments.
3440  *
3441  * This routine is responsible for storing processed mbuf
3442  * into elts ring buffer and update elts_head.
3443  *
3444  * @param txq
3445  *   Pointer to TX queue structure.
3446  * @param loc
3447  *   Pointer to burst routine local context.
3448  * @param olx
3449  *   Configured Tx offloads mask. It is fully defined at
3450  *   compile time and may be used for optimization.
3451  *
3452  * @return
3453  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3454  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3455  * Local context variables partially updated.
3456  */
3457 static __rte_always_inline enum mlx5_txcmp_code
3458 mlx5_tx_packet_multi_inline(struct mlx5_txq_data *__rte_restrict txq,
3459 			    struct mlx5_txq_local *__rte_restrict loc,
3460 			    unsigned int olx)
3461 {
3462 	struct mlx5_wqe *__rte_restrict wqe;
3463 	unsigned int ds, inlen, dlen, vlan = 0;
3464 
3465 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3466 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3467 	if (MLX5_TXOFF_CONFIG(TXPP)) {
3468 		enum mlx5_txcmp_code wret;
3469 
3470 		/* Generate WAIT for scheduling if requested. */
3471 		wret = mlx5_tx_schedule_send(txq, loc, olx);
3472 		if (wret == MLX5_TXCMP_CODE_EXIT)
3473 			return MLX5_TXCMP_CODE_EXIT;
3474 		if (wret == MLX5_TXCMP_CODE_ERROR)
3475 			return MLX5_TXCMP_CODE_ERROR;
3476 	}
3477 	/*
3478 	 * First calculate data length to be inlined
3479 	 * to estimate the required space for WQE.
3480 	 */
3481 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3482 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3483 		vlan = sizeof(struct rte_vlan_hdr);
3484 	inlen = dlen + vlan;
3485 	/* Check against minimal length. */
3486 	if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
3487 		return MLX5_TXCMP_CODE_ERROR;
3488 	MLX5_ASSERT(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
3489 	if (inlen > txq->inlen_send ||
3490 	    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
3491 		struct rte_mbuf *mbuf;
3492 		unsigned int nxlen;
3493 		uintptr_t start;
3494 
3495 		/*
3496 		 * Packet length exceeds the allowed inline
3497 		 * data length, check whether the minimal
3498 		 * inlining is required.
3499 		 */
3500 		if (txq->inlen_mode) {
3501 			MLX5_ASSERT(txq->inlen_mode >=
3502 				    MLX5_ESEG_MIN_INLINE_SIZE);
3503 			MLX5_ASSERT(txq->inlen_mode <= txq->inlen_send);
3504 			inlen = txq->inlen_mode;
3505 		} else {
3506 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE ||
3507 			    !vlan || txq->vlan_en) {
3508 				/*
3509 				 * VLAN insertion will be done inside by HW.
3510 				 * It is not utmost effective - VLAN flag is
3511 				 * checked twice, but we should proceed the
3512 				 * inlining length correctly and take into
3513 				 * account the VLAN header being inserted.
3514 				 */
3515 				return mlx5_tx_packet_multi_send
3516 							(txq, loc, olx);
3517 			}
3518 			inlen = MLX5_ESEG_MIN_INLINE_SIZE;
3519 		}
3520 		/*
3521 		 * Now we know the minimal amount of data is requested
3522 		 * to inline. Check whether we should inline the buffers
3523 		 * from the chain beginning to eliminate some mbufs.
3524 		 */
3525 		mbuf = loc->mbuf;
3526 		nxlen = rte_pktmbuf_data_len(mbuf);
3527 		if (unlikely(nxlen <= txq->inlen_send)) {
3528 			/* We can inline first mbuf at least. */
3529 			if (nxlen < inlen) {
3530 				unsigned int smlen;
3531 
3532 				/* Scan mbufs till inlen filled. */
3533 				do {
3534 					smlen = nxlen;
3535 					mbuf = NEXT(mbuf);
3536 					MLX5_ASSERT(mbuf);
3537 					nxlen = rte_pktmbuf_data_len(mbuf);
3538 					nxlen += smlen;
3539 				} while (unlikely(nxlen < inlen));
3540 				if (unlikely(nxlen > txq->inlen_send)) {
3541 					/* We cannot inline entire mbuf. */
3542 					smlen = inlen - smlen;
3543 					start = rte_pktmbuf_mtod_offset
3544 						    (mbuf, uintptr_t, smlen);
3545 					goto do_align;
3546 				}
3547 			}
3548 			do {
3549 				inlen = nxlen;
3550 				mbuf = NEXT(mbuf);
3551 				/* There should be not end of packet. */
3552 				MLX5_ASSERT(mbuf);
3553 				nxlen = inlen + rte_pktmbuf_data_len(mbuf);
3554 			} while (unlikely(nxlen < txq->inlen_send));
3555 		}
3556 		start = rte_pktmbuf_mtod(mbuf, uintptr_t);
3557 		/*
3558 		 * Check whether we can do inline to align start
3559 		 * address of data buffer to cacheline.
3560 		 */
3561 do_align:
3562 		start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1);
3563 		if (unlikely(start)) {
3564 			start += inlen;
3565 			if (start <= txq->inlen_send)
3566 				inlen = start;
3567 		}
3568 	}
3569 	/*
3570 	 * Check whether there are enough free WQEBBs:
3571 	 * - Control Segment
3572 	 * - Ethernet Segment
3573 	 * - First Segment of inlined Ethernet data
3574 	 * - ... data continued ...
3575 	 * - Data Segments of pointer/min inline type
3576 	 *
3577 	 * Estimate the number of Data Segments conservatively,
3578 	 * supposing no any mbufs is being freed during inlining.
3579 	 */
3580 	MLX5_ASSERT(inlen <= txq->inlen_send);
3581 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3582 				       MLX5_ESEG_MIN_INLINE_SIZE +
3583 				       MLX5_WSEG_SIZE +
3584 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3585 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3586 		return MLX5_TXCMP_CODE_EXIT;
3587 	/* Check for maximal WQE size. */
3588 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3589 		return MLX5_TXCMP_CODE_ERROR;
3590 #ifdef MLX5_PMD_SOFT_COUNTERS
3591 	/* Update sent data bytes/packets counters. */
3592 	txq->stats.obytes += dlen + vlan;
3593 #endif
3594 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3595 	loc->wqe_last = wqe;
3596 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx);
3597 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx);
3598 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3599 	txq->wqe_ci += (ds + 3) / 4;
3600 	loc->wqe_free -= (ds + 3) / 4;
3601 	return MLX5_TXCMP_CODE_MULTI;
3602 }
3603 
3604 /**
3605  * Tx burst function for multi-segment packets. Supports all
3606  * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs,
3607  * sends one packet per WQE. Function stops sending if it
3608  * encounters the single-segment packet.
3609  *
3610  * This routine is responsible for storing processed mbuf
3611  * into elts ring buffer and update elts_head.
3612  *
3613  * @param txq
3614  *   Pointer to TX queue structure.
3615  * @param[in] pkts
3616  *   Packets to transmit.
3617  * @param pkts_n
3618  *   Number of packets in array.
3619  * @param loc
3620  *   Pointer to burst routine local context.
3621  * @param olx
3622  *   Configured Tx offloads mask. It is fully defined at
3623  *   compile time and may be used for optimization.
3624  *
3625  * @return
3626  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3627  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3628  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3629  *   MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered.
3630  * Local context variables updated.
3631  */
3632 static __rte_always_inline enum mlx5_txcmp_code
3633 mlx5_tx_burst_mseg(struct mlx5_txq_data *__rte_restrict txq,
3634 		   struct rte_mbuf **__rte_restrict pkts,
3635 		   unsigned int pkts_n,
3636 		   struct mlx5_txq_local *__rte_restrict loc,
3637 		   unsigned int olx)
3638 {
3639 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3640 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3641 	pkts += loc->pkts_sent + 1;
3642 	pkts_n -= loc->pkts_sent;
3643 	for (;;) {
3644 		enum mlx5_txcmp_code ret;
3645 
3646 		MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3647 		/*
3648 		 * Estimate the number of free elts quickly but
3649 		 * conservatively. Some segment may be fully inlined
3650 		 * and freed, ignore this here - precise estimation
3651 		 * is costly.
3652 		 */
3653 		if (loc->elts_free < NB_SEGS(loc->mbuf))
3654 			return MLX5_TXCMP_CODE_EXIT;
3655 		if (MLX5_TXOFF_CONFIG(TSO) &&
3656 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) {
3657 			/* Proceed with multi-segment TSO. */
3658 			ret = mlx5_tx_packet_multi_tso(txq, loc, olx);
3659 		} else if (MLX5_TXOFF_CONFIG(INLINE)) {
3660 			/* Proceed with multi-segment SEND with inlining. */
3661 			ret = mlx5_tx_packet_multi_inline(txq, loc, olx);
3662 		} else {
3663 			/* Proceed with multi-segment SEND w/o inlining. */
3664 			ret = mlx5_tx_packet_multi_send(txq, loc, olx);
3665 		}
3666 		if (ret == MLX5_TXCMP_CODE_EXIT)
3667 			return MLX5_TXCMP_CODE_EXIT;
3668 		if (ret == MLX5_TXCMP_CODE_ERROR)
3669 			return MLX5_TXCMP_CODE_ERROR;
3670 		/* WQE is built, go to the next packet. */
3671 		++loc->pkts_sent;
3672 		--pkts_n;
3673 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3674 			return MLX5_TXCMP_CODE_EXIT;
3675 		loc->mbuf = *pkts++;
3676 		if (pkts_n > 1)
3677 			rte_prefetch0(*pkts);
3678 		if (likely(NB_SEGS(loc->mbuf) > 1))
3679 			continue;
3680 		/* Here ends the series of multi-segment packets. */
3681 		if (MLX5_TXOFF_CONFIG(TSO) &&
3682 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3683 			return MLX5_TXCMP_CODE_TSO;
3684 		return MLX5_TXCMP_CODE_SINGLE;
3685 	}
3686 	MLX5_ASSERT(false);
3687 }
3688 
3689 /**
3690  * Tx burst function for single-segment packets with TSO.
3691  * Supports all types of Tx offloads, except multi-packets.
3692  * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE.
3693  * Function stops sending if it encounters the multi-segment
3694  * packet or packet without TSO requested.
3695  *
3696  * The routine is responsible for storing processed mbuf
3697  * into elts ring buffer and update elts_head if inline
3698  * offloads is requested due to possible early freeing
3699  * of the inlined mbufs (can not store pkts array in elts
3700  * as a batch).
3701  *
3702  * @param txq
3703  *   Pointer to TX queue structure.
3704  * @param[in] pkts
3705  *   Packets to transmit.
3706  * @param pkts_n
3707  *   Number of packets in array.
3708  * @param loc
3709  *   Pointer to burst routine local context.
3710  * @param olx
3711  *   Configured Tx offloads mask. It is fully defined at
3712  *   compile time and may be used for optimization.
3713  *
3714  * @return
3715  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3716  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3717  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3718  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3719  * Local context variables updated.
3720  */
3721 static __rte_always_inline enum mlx5_txcmp_code
3722 mlx5_tx_burst_tso(struct mlx5_txq_data *__rte_restrict txq,
3723 		  struct rte_mbuf **__rte_restrict pkts,
3724 		  unsigned int pkts_n,
3725 		  struct mlx5_txq_local *__rte_restrict loc,
3726 		  unsigned int olx)
3727 {
3728 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3729 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3730 	pkts += loc->pkts_sent + 1;
3731 	pkts_n -= loc->pkts_sent;
3732 	for (;;) {
3733 		struct mlx5_wqe_dseg *__rte_restrict dseg;
3734 		struct mlx5_wqe *__rte_restrict wqe;
3735 		unsigned int ds, dlen, hlen, ntcp, vlan = 0;
3736 		uint8_t *dptr;
3737 
3738 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
3739 		if (MLX5_TXOFF_CONFIG(TXPP)) {
3740 			enum mlx5_txcmp_code wret;
3741 
3742 			/* Generate WAIT for scheduling if requested. */
3743 			wret = mlx5_tx_schedule_send(txq, loc, olx);
3744 			if (wret == MLX5_TXCMP_CODE_EXIT)
3745 				return MLX5_TXCMP_CODE_EXIT;
3746 			if (wret == MLX5_TXCMP_CODE_ERROR)
3747 				return MLX5_TXCMP_CODE_ERROR;
3748 		}
3749 		dlen = rte_pktmbuf_data_len(loc->mbuf);
3750 		if (MLX5_TXOFF_CONFIG(VLAN) &&
3751 		    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
3752 			vlan = sizeof(struct rte_vlan_hdr);
3753 		}
3754 		/*
3755 		 * First calculate the WQE size to check
3756 		 * whether we have enough space in ring buffer.
3757 		 */
3758 		hlen = loc->mbuf->l2_len + vlan +
3759 		       loc->mbuf->l3_len + loc->mbuf->l4_len;
3760 		if (unlikely((!hlen || !loc->mbuf->tso_segsz)))
3761 			return MLX5_TXCMP_CODE_ERROR;
3762 		if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3763 			hlen += loc->mbuf->outer_l2_len +
3764 				loc->mbuf->outer_l3_len;
3765 		/* Segment must contain all TSO headers. */
3766 		if (unlikely(hlen > MLX5_MAX_TSO_HEADER ||
3767 			     hlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3768 			     hlen > (dlen + vlan)))
3769 			return MLX5_TXCMP_CODE_ERROR;
3770 		/*
3771 		 * Check whether there are enough free WQEBBs:
3772 		 * - Control Segment
3773 		 * - Ethernet Segment
3774 		 * - First Segment of inlined Ethernet data
3775 		 * - ... data continued ...
3776 		 * - Finishing Data Segment of pointer type
3777 		 */
3778 		ds = 4 + (hlen - MLX5_ESEG_MIN_INLINE_SIZE +
3779 			  MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3780 		if (loc->wqe_free < ((ds + 3) / 4))
3781 			return MLX5_TXCMP_CODE_EXIT;
3782 #ifdef MLX5_PMD_SOFT_COUNTERS
3783 		/* Update sent data bytes/packets counters. */
3784 		ntcp = (dlen + vlan - hlen +
3785 			loc->mbuf->tso_segsz - 1) /
3786 			loc->mbuf->tso_segsz;
3787 		/*
3788 		 * One will be added for mbuf itself at the end
3789 		 * of the mlx5_tx_burst from loc->pkts_sent field.
3790 		 */
3791 		--ntcp;
3792 		txq->stats.opackets += ntcp;
3793 		txq->stats.obytes += dlen + vlan + ntcp * hlen;
3794 #endif
3795 		/*
3796 		 * Build the TSO WQE:
3797 		 * - Control Segment
3798 		 * - Ethernet Segment with hlen bytes inlined
3799 		 * - Data Segment of pointer type
3800 		 */
3801 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3802 		loc->wqe_last = wqe;
3803 		mlx5_tx_cseg_init(txq, loc, wqe, ds,
3804 				  MLX5_OPCODE_TSO, olx);
3805 		dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, hlen, 1, olx);
3806 		dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + hlen - vlan;
3807 		dlen -= hlen - vlan;
3808 		mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
3809 		/*
3810 		 * WQE is built, update the loop parameters
3811 		 * and go to the next packet.
3812 		 */
3813 		txq->wqe_ci += (ds + 3) / 4;
3814 		loc->wqe_free -= (ds + 3) / 4;
3815 		if (MLX5_TXOFF_CONFIG(INLINE))
3816 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3817 		--loc->elts_free;
3818 		++loc->pkts_sent;
3819 		--pkts_n;
3820 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3821 			return MLX5_TXCMP_CODE_EXIT;
3822 		loc->mbuf = *pkts++;
3823 		if (pkts_n > 1)
3824 			rte_prefetch0(*pkts);
3825 		if (MLX5_TXOFF_CONFIG(MULTI) &&
3826 		    unlikely(NB_SEGS(loc->mbuf) > 1))
3827 			return MLX5_TXCMP_CODE_MULTI;
3828 		if (likely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)))
3829 			return MLX5_TXCMP_CODE_SINGLE;
3830 		/* Continue with the next TSO packet. */
3831 	}
3832 	MLX5_ASSERT(false);
3833 }
3834 
3835 /**
3836  * Analyze the packet and select the best method to send.
3837  *
3838  * @param txq
3839  *   Pointer to TX queue structure.
3840  * @param loc
3841  *   Pointer to burst routine local context.
3842  * @param olx
3843  *   Configured Tx offloads mask. It is fully defined at
3844  *   compile time and may be used for optimization.
3845  * @param newp
3846  *   The predefined flag whether do complete check for
3847  *   multi-segment packets and TSO.
3848  *
3849  * @return
3850  *  MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3851  *  MLX5_TXCMP_CODE_TSO - TSO required, use TSO/LSO.
3852  *  MLX5_TXCMP_CODE_SINGLE - single-segment packet, use SEND.
3853  *  MLX5_TXCMP_CODE_EMPW - single-segment packet, use MPW.
3854  */
3855 static __rte_always_inline enum mlx5_txcmp_code
3856 mlx5_tx_able_to_empw(struct mlx5_txq_data *__rte_restrict txq,
3857 		     struct mlx5_txq_local *__rte_restrict loc,
3858 		     unsigned int olx,
3859 		     bool newp)
3860 {
3861 	/* Check for multi-segment packet. */
3862 	if (newp &&
3863 	    MLX5_TXOFF_CONFIG(MULTI) &&
3864 	    unlikely(NB_SEGS(loc->mbuf) > 1))
3865 		return MLX5_TXCMP_CODE_MULTI;
3866 	/* Check for TSO packet. */
3867 	if (newp &&
3868 	    MLX5_TXOFF_CONFIG(TSO) &&
3869 	    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3870 		return MLX5_TXCMP_CODE_TSO;
3871 	/* Check if eMPW is enabled at all. */
3872 	if (!MLX5_TXOFF_CONFIG(EMPW))
3873 		return MLX5_TXCMP_CODE_SINGLE;
3874 	/* Check if eMPW can be engaged. */
3875 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3876 	    unlikely(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) &&
3877 		(!MLX5_TXOFF_CONFIG(INLINE) ||
3878 		 unlikely((rte_pktmbuf_data_len(loc->mbuf) +
3879 			   sizeof(struct rte_vlan_hdr)) > txq->inlen_empw))) {
3880 		/*
3881 		 * eMPW does not support VLAN insertion offload,
3882 		 * we have to inline the entire packet but
3883 		 * packet is too long for inlining.
3884 		 */
3885 		return MLX5_TXCMP_CODE_SINGLE;
3886 	}
3887 	return MLX5_TXCMP_CODE_EMPW;
3888 }
3889 
3890 /**
3891  * Check the next packet attributes to match with the eMPW batch ones.
3892  * In addition, for legacy MPW the packet length is checked either.
3893  *
3894  * @param txq
3895  *   Pointer to TX queue structure.
3896  * @param es
3897  *   Pointer to Ethernet Segment of eMPW batch.
3898  * @param loc
3899  *   Pointer to burst routine local context.
3900  * @param dlen
3901  *   Length of previous packet in MPW descriptor.
3902  * @param olx
3903  *   Configured Tx offloads mask. It is fully defined at
3904  *   compile time and may be used for optimization.
3905  *
3906  * @return
3907  *  true - packet match with eMPW batch attributes.
3908  *  false - no match, eMPW should be restarted.
3909  */
3910 static __rte_always_inline bool
3911 mlx5_tx_match_empw(struct mlx5_txq_data *__rte_restrict txq,
3912 		   struct mlx5_wqe_eseg *__rte_restrict es,
3913 		   struct mlx5_txq_local *__rte_restrict loc,
3914 		   uint32_t dlen,
3915 		   unsigned int olx)
3916 {
3917 	uint8_t swp_flags = 0;
3918 
3919 	/* Compare the checksum flags, if any. */
3920 	if (MLX5_TXOFF_CONFIG(CSUM) &&
3921 	    txq_ol_cksum_to_cs(loc->mbuf) != es->cs_flags)
3922 		return false;
3923 	/* Compare the Software Parser offsets and flags. */
3924 	if (MLX5_TXOFF_CONFIG(SWP) &&
3925 	    (es->swp_offs != txq_mbuf_to_swp(loc, &swp_flags, olx) ||
3926 	     es->swp_flags != swp_flags))
3927 		return false;
3928 	/* Fill metadata field if needed. */
3929 	if (MLX5_TXOFF_CONFIG(METADATA) &&
3930 		es->metadata != (loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
3931 				 *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0))
3932 		return false;
3933 	/* Legacy MPW can send packets with the same lengt only. */
3934 	if (MLX5_TXOFF_CONFIG(MPW) &&
3935 	    dlen != rte_pktmbuf_data_len(loc->mbuf))
3936 		return false;
3937 	/* There must be no VLAN packets in eMPW loop. */
3938 	if (MLX5_TXOFF_CONFIG(VLAN))
3939 		MLX5_ASSERT(!(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT));
3940 	/* Check if the scheduling is requested. */
3941 	if (MLX5_TXOFF_CONFIG(TXPP) &&
3942 	    loc->mbuf->ol_flags & txq->ts_mask)
3943 		return false;
3944 	return true;
3945 }
3946 
3947 /*
3948  * Update send loop variables and WQE for eMPW loop
3949  * without data inlining. Number of Data Segments is
3950  * equal to the number of sent packets.
3951  *
3952  * @param txq
3953  *   Pointer to TX queue structure.
3954  * @param loc
3955  *   Pointer to burst routine local context.
3956  * @param ds
3957  *   Number of packets/Data Segments/Packets.
3958  * @param slen
3959  *   Accumulated statistics, bytes sent
3960  * @param olx
3961  *   Configured Tx offloads mask. It is fully defined at
3962  *   compile time and may be used for optimization.
3963  *
3964  * @return
3965  *  true - packet match with eMPW batch attributes.
3966  *  false - no match, eMPW should be restarted.
3967  */
3968 static __rte_always_inline void
3969 mlx5_tx_sdone_empw(struct mlx5_txq_data *__rte_restrict txq,
3970 		   struct mlx5_txq_local *__rte_restrict loc,
3971 		   unsigned int ds,
3972 		   unsigned int slen,
3973 		   unsigned int olx __rte_unused)
3974 {
3975 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
3976 #ifdef MLX5_PMD_SOFT_COUNTERS
3977 	/* Update sent data bytes counter. */
3978 	 txq->stats.obytes += slen;
3979 #else
3980 	(void)slen;
3981 #endif
3982 	loc->elts_free -= ds;
3983 	loc->pkts_sent += ds;
3984 	ds += 2;
3985 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3986 	txq->wqe_ci += (ds + 3) / 4;
3987 	loc->wqe_free -= (ds + 3) / 4;
3988 }
3989 
3990 /*
3991  * Update send loop variables and WQE for eMPW loop
3992  * with data inlining. Gets the size of pushed descriptors
3993  * and data to the WQE.
3994  *
3995  * @param txq
3996  *   Pointer to TX queue structure.
3997  * @param loc
3998  *   Pointer to burst routine local context.
3999  * @param len
4000  *   Total size of descriptor/data in bytes.
4001  * @param slen
4002  *   Accumulated statistics, data bytes sent.
4003  * @param wqem
4004  *   The base WQE for the eMPW/MPW descriptor.
4005  * @param olx
4006  *   Configured Tx offloads mask. It is fully defined at
4007  *   compile time and may be used for optimization.
4008  *
4009  * @return
4010  *  true - packet match with eMPW batch attributes.
4011  *  false - no match, eMPW should be restarted.
4012  */
4013 static __rte_always_inline void
4014 mlx5_tx_idone_empw(struct mlx5_txq_data *__rte_restrict txq,
4015 		   struct mlx5_txq_local *__rte_restrict loc,
4016 		   unsigned int len,
4017 		   unsigned int slen,
4018 		   struct mlx5_wqe *__rte_restrict wqem,
4019 		   unsigned int olx __rte_unused)
4020 {
4021 	struct mlx5_wqe_dseg *dseg = &wqem->dseg[0];
4022 
4023 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4024 #ifdef MLX5_PMD_SOFT_COUNTERS
4025 	/* Update sent data bytes counter. */
4026 	 txq->stats.obytes += slen;
4027 #else
4028 	(void)slen;
4029 #endif
4030 	if (MLX5_TXOFF_CONFIG(MPW) && dseg->bcount == RTE_BE32(0)) {
4031 		/*
4032 		 * If the legacy MPW session contains the inline packets
4033 		 * we should set the only inline data segment length
4034 		 * and align the total length to the segment size.
4035 		 */
4036 		MLX5_ASSERT(len > sizeof(dseg->bcount));
4037 		dseg->bcount = rte_cpu_to_be_32((len - sizeof(dseg->bcount)) |
4038 						MLX5_ETH_WQE_DATA_INLINE);
4039 		len = (len + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE + 2;
4040 	} else {
4041 		/*
4042 		 * The session is not legacy MPW or contains the
4043 		 * data buffer pointer segments.
4044 		 */
4045 		MLX5_ASSERT((len % MLX5_WSEG_SIZE) == 0);
4046 		len = len / MLX5_WSEG_SIZE + 2;
4047 	}
4048 	wqem->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
4049 	txq->wqe_ci += (len + 3) / 4;
4050 	loc->wqe_free -= (len + 3) / 4;
4051 	loc->wqe_last = wqem;
4052 }
4053 
4054 /**
4055  * The set of Tx burst functions for single-segment packets
4056  * without TSO and with Multi-Packet Writing feature support.
4057  * Supports all types of Tx offloads, except multi-packets
4058  * and TSO.
4059  *
4060  * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends
4061  * as many packet per WQE as it can. If eMPW is not configured
4062  * or packet can not be sent with eMPW (VLAN insertion) the
4063  * ordinary SEND opcode is used and only one packet placed
4064  * in WQE.
4065  *
4066  * Functions stop sending if it encounters the multi-segment
4067  * packet or packet with TSO requested.
4068  *
4069  * The routines are responsible for storing processed mbuf
4070  * into elts ring buffer and update elts_head if inlining
4071  * offload is requested. Otherwise the copying mbufs to elts
4072  * can be postponed and completed at the end of burst routine.
4073  *
4074  * @param txq
4075  *   Pointer to TX queue structure.
4076  * @param[in] pkts
4077  *   Packets to transmit.
4078  * @param pkts_n
4079  *   Number of packets in array.
4080  * @param loc
4081  *   Pointer to burst routine local context.
4082  * @param olx
4083  *   Configured Tx offloads mask. It is fully defined at
4084  *   compile time and may be used for optimization.
4085  *
4086  * @return
4087  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
4088  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
4089  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
4090  *   MLX5_TXCMP_CODE_TSO - TSO packet encountered.
4091  *   MLX5_TXCMP_CODE_SINGLE - used inside functions set.
4092  *   MLX5_TXCMP_CODE_EMPW - used inside functions set.
4093  *
4094  * Local context variables updated.
4095  *
4096  *
4097  * The routine sends packets with MLX5_OPCODE_EMPW
4098  * without inlining, this is dedicated optimized branch.
4099  * No VLAN insertion is supported.
4100  */
4101 static __rte_always_inline enum mlx5_txcmp_code
4102 mlx5_tx_burst_empw_simple(struct mlx5_txq_data *__rte_restrict txq,
4103 			  struct rte_mbuf **__rte_restrict pkts,
4104 			  unsigned int pkts_n,
4105 			  struct mlx5_txq_local *__rte_restrict loc,
4106 			  unsigned int olx)
4107 {
4108 	/*
4109 	 * Subroutine is the part of mlx5_tx_burst_single()
4110 	 * and sends single-segment packet with eMPW opcode
4111 	 * without data inlining.
4112 	 */
4113 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
4114 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
4115 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4116 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4117 	pkts += loc->pkts_sent + 1;
4118 	pkts_n -= loc->pkts_sent;
4119 	for (;;) {
4120 		struct mlx5_wqe_dseg *__rte_restrict dseg;
4121 		struct mlx5_wqe_eseg *__rte_restrict eseg;
4122 		enum mlx5_txcmp_code ret;
4123 		unsigned int part, loop;
4124 		unsigned int slen = 0;
4125 
4126 next_empw:
4127 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4128 		if (MLX5_TXOFF_CONFIG(TXPP)) {
4129 			enum mlx5_txcmp_code wret;
4130 
4131 			/* Generate WAIT for scheduling if requested. */
4132 			wret = mlx5_tx_schedule_send(txq, loc, olx);
4133 			if (wret == MLX5_TXCMP_CODE_EXIT)
4134 				return MLX5_TXCMP_CODE_EXIT;
4135 			if (wret == MLX5_TXCMP_CODE_ERROR)
4136 				return MLX5_TXCMP_CODE_ERROR;
4137 		}
4138 		part = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4139 				       MLX5_MPW_MAX_PACKETS :
4140 				       MLX5_EMPW_MAX_PACKETS);
4141 		if (unlikely(loc->elts_free < part)) {
4142 			/* We have no enough elts to save all mbufs. */
4143 			if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS))
4144 				return MLX5_TXCMP_CODE_EXIT;
4145 			/* But we still able to send at least minimal eMPW. */
4146 			part = loc->elts_free;
4147 		}
4148 		/* Check whether we have enough WQEs */
4149 		if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) {
4150 			if (unlikely(loc->wqe_free <
4151 				((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4152 				return MLX5_TXCMP_CODE_EXIT;
4153 			part = (loc->wqe_free * 4) - 2;
4154 		}
4155 		if (likely(part > 1))
4156 			rte_prefetch0(*pkts);
4157 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4158 		/*
4159 		 * Build eMPW title WQEBB:
4160 		 * - Control Segment, eMPW opcode
4161 		 * - Ethernet Segment, no inline
4162 		 */
4163 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2,
4164 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4165 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
4166 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4167 		eseg = &loc->wqe_last->eseg;
4168 		dseg = &loc->wqe_last->dseg[0];
4169 		loop = part;
4170 		/* Store the packet length for legacy MPW. */
4171 		if (MLX5_TXOFF_CONFIG(MPW))
4172 			eseg->mss = rte_cpu_to_be_16
4173 					(rte_pktmbuf_data_len(loc->mbuf));
4174 		for (;;) {
4175 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4176 #ifdef MLX5_PMD_SOFT_COUNTERS
4177 			/* Update sent data bytes counter. */
4178 			slen += dlen;
4179 #endif
4180 			mlx5_tx_dseg_ptr
4181 				(txq, loc, dseg,
4182 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4183 				 dlen, olx);
4184 			if (unlikely(--loop == 0))
4185 				break;
4186 			loc->mbuf = *pkts++;
4187 			if (likely(loop > 1))
4188 				rte_prefetch0(*pkts);
4189 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4190 			/*
4191 			 * Unroll the completion code to avoid
4192 			 * returning variable value - it results in
4193 			 * unoptimized sequent checking in caller.
4194 			 */
4195 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4196 				part -= loop;
4197 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4198 				if (unlikely(!loc->elts_free ||
4199 					     !loc->wqe_free))
4200 					return MLX5_TXCMP_CODE_EXIT;
4201 				return MLX5_TXCMP_CODE_MULTI;
4202 			}
4203 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4204 			if (ret == MLX5_TXCMP_CODE_TSO) {
4205 				part -= loop;
4206 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4207 				if (unlikely(!loc->elts_free ||
4208 					     !loc->wqe_free))
4209 					return MLX5_TXCMP_CODE_EXIT;
4210 				return MLX5_TXCMP_CODE_TSO;
4211 			}
4212 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4213 				part -= loop;
4214 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4215 				if (unlikely(!loc->elts_free ||
4216 					     !loc->wqe_free))
4217 					return MLX5_TXCMP_CODE_EXIT;
4218 				return MLX5_TXCMP_CODE_SINGLE;
4219 			}
4220 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4221 				MLX5_ASSERT(false);
4222 				part -= loop;
4223 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4224 				return MLX5_TXCMP_CODE_ERROR;
4225 			}
4226 			/*
4227 			 * Check whether packet parameters coincide
4228 			 * within assumed eMPW batch:
4229 			 * - check sum settings
4230 			 * - metadata value
4231 			 * - software parser settings
4232 			 * - packets length (legacy MPW only)
4233 			 * - scheduling is not required
4234 			 */
4235 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx)) {
4236 				MLX5_ASSERT(loop);
4237 				part -= loop;
4238 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4239 				if (unlikely(!loc->elts_free ||
4240 					     !loc->wqe_free))
4241 					return MLX5_TXCMP_CODE_EXIT;
4242 				pkts_n -= part;
4243 				goto next_empw;
4244 			}
4245 			/* Packet attributes match, continue the same eMPW. */
4246 			++dseg;
4247 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4248 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4249 		}
4250 		/* eMPW is built successfully, update loop parameters. */
4251 		MLX5_ASSERT(!loop);
4252 		MLX5_ASSERT(pkts_n >= part);
4253 #ifdef MLX5_PMD_SOFT_COUNTERS
4254 		/* Update sent data bytes counter. */
4255 		txq->stats.obytes += slen;
4256 #endif
4257 		loc->elts_free -= part;
4258 		loc->pkts_sent += part;
4259 		txq->wqe_ci += (2 + part + 3) / 4;
4260 		loc->wqe_free -= (2 + part + 3) / 4;
4261 		pkts_n -= part;
4262 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4263 			return MLX5_TXCMP_CODE_EXIT;
4264 		loc->mbuf = *pkts++;
4265 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4266 		if (unlikely(ret != MLX5_TXCMP_CODE_EMPW))
4267 			return ret;
4268 		/* Continue sending eMPW batches. */
4269 	}
4270 	MLX5_ASSERT(false);
4271 }
4272 
4273 /**
4274  * The routine sends packets with MLX5_OPCODE_EMPW
4275  * with inlining, optionally supports VLAN insertion.
4276  */
4277 static __rte_always_inline enum mlx5_txcmp_code
4278 mlx5_tx_burst_empw_inline(struct mlx5_txq_data *__rte_restrict txq,
4279 			  struct rte_mbuf **__rte_restrict pkts,
4280 			  unsigned int pkts_n,
4281 			  struct mlx5_txq_local *__rte_restrict loc,
4282 			  unsigned int olx)
4283 {
4284 	/*
4285 	 * Subroutine is the part of mlx5_tx_burst_single()
4286 	 * and sends single-segment packet with eMPW opcode
4287 	 * with data inlining.
4288 	 */
4289 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4290 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
4291 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4292 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4293 	pkts += loc->pkts_sent + 1;
4294 	pkts_n -= loc->pkts_sent;
4295 	for (;;) {
4296 		struct mlx5_wqe_dseg *__rte_restrict dseg;
4297 		struct mlx5_wqe *__rte_restrict wqem;
4298 		enum mlx5_txcmp_code ret;
4299 		unsigned int room, part, nlim;
4300 		unsigned int slen = 0;
4301 
4302 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4303 		if (MLX5_TXOFF_CONFIG(TXPP)) {
4304 			enum mlx5_txcmp_code wret;
4305 
4306 			/* Generate WAIT for scheduling if requested. */
4307 			wret = mlx5_tx_schedule_send(txq, loc, olx);
4308 			if (wret == MLX5_TXCMP_CODE_EXIT)
4309 				return MLX5_TXCMP_CODE_EXIT;
4310 			if (wret == MLX5_TXCMP_CODE_ERROR)
4311 				return MLX5_TXCMP_CODE_ERROR;
4312 		}
4313 		/*
4314 		 * Limits the amount of packets in one WQE
4315 		 * to improve CQE latency generation.
4316 		 */
4317 		nlim = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4318 				       MLX5_MPW_INLINE_MAX_PACKETS :
4319 				       MLX5_EMPW_MAX_PACKETS);
4320 		/* Check whether we have minimal amount WQEs */
4321 		if (unlikely(loc->wqe_free <
4322 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4323 			return MLX5_TXCMP_CODE_EXIT;
4324 		if (likely(pkts_n > 1))
4325 			rte_prefetch0(*pkts);
4326 		wqem = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4327 		/*
4328 		 * Build eMPW title WQEBB:
4329 		 * - Control Segment, eMPW opcode, zero DS
4330 		 * - Ethernet Segment, no inline
4331 		 */
4332 		mlx5_tx_cseg_init(txq, loc, wqem, 0,
4333 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4334 		mlx5_tx_eseg_none(txq, loc, wqem,
4335 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4336 		dseg = &wqem->dseg[0];
4337 		/* Store the packet length for legacy MPW. */
4338 		if (MLX5_TXOFF_CONFIG(MPW))
4339 			wqem->eseg.mss = rte_cpu_to_be_16
4340 					 (rte_pktmbuf_data_len(loc->mbuf));
4341 		room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE,
4342 			       loc->wqe_free) * MLX5_WQE_SIZE -
4343 					MLX5_WQE_CSEG_SIZE -
4344 					MLX5_WQE_ESEG_SIZE;
4345 		/* Limit the room for legacy MPW sessions for performance. */
4346 		if (MLX5_TXOFF_CONFIG(MPW))
4347 			room = RTE_MIN(room,
4348 				       RTE_MAX(txq->inlen_empw +
4349 					       sizeof(dseg->bcount) +
4350 					       (MLX5_TXOFF_CONFIG(VLAN) ?
4351 					       sizeof(struct rte_vlan_hdr) : 0),
4352 					       MLX5_MPW_INLINE_MAX_PACKETS *
4353 					       MLX5_WQE_DSEG_SIZE));
4354 		/* Build WQE till we have space, packets and resources. */
4355 		part = room;
4356 		for (;;) {
4357 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4358 			uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
4359 			unsigned int tlen;
4360 
4361 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4362 			MLX5_ASSERT((room % MLX5_WQE_DSEG_SIZE) == 0);
4363 			MLX5_ASSERT((uintptr_t)dseg < (uintptr_t)txq->wqes_end);
4364 			/*
4365 			 * Some Tx offloads may cause an error if
4366 			 * packet is not long enough, check against
4367 			 * assumed minimal length.
4368 			 */
4369 			if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) {
4370 				part -= room;
4371 				if (unlikely(!part))
4372 					return MLX5_TXCMP_CODE_ERROR;
4373 				/*
4374 				 * We have some successfully built
4375 				 * packet Data Segments to send.
4376 				 */
4377 				mlx5_tx_idone_empw(txq, loc, part,
4378 						   slen, wqem, olx);
4379 				return MLX5_TXCMP_CODE_ERROR;
4380 			}
4381 			/* Inline or not inline - that's the Question. */
4382 			if (dlen > txq->inlen_empw ||
4383 			    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE)
4384 				goto pointer_empw;
4385 			if (MLX5_TXOFF_CONFIG(MPW)) {
4386 				if (dlen > txq->inlen_send)
4387 					goto pointer_empw;
4388 				tlen = dlen;
4389 				if (part == room) {
4390 					/* Open new inline MPW session. */
4391 					tlen += sizeof(dseg->bcount);
4392 					dseg->bcount = RTE_BE32(0);
4393 					dseg = RTE_PTR_ADD
4394 						(dseg, sizeof(dseg->bcount));
4395 				} else {
4396 					/*
4397 					 * No pointer and inline descriptor
4398 					 * intermix for legacy MPW sessions.
4399 					 */
4400 					if (wqem->dseg[0].bcount)
4401 						break;
4402 				}
4403 			} else {
4404 				tlen = sizeof(dseg->bcount) + dlen;
4405 			}
4406 			/* Inline entire packet, optional VLAN insertion. */
4407 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4408 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4409 				/*
4410 				 * The packet length must be checked in
4411 				 * mlx5_tx_able_to_empw() and packet
4412 				 * fits into inline length guaranteed.
4413 				 */
4414 				MLX5_ASSERT((dlen +
4415 					     sizeof(struct rte_vlan_hdr)) <=
4416 					    txq->inlen_empw);
4417 				tlen += sizeof(struct rte_vlan_hdr);
4418 				if (room < tlen)
4419 					break;
4420 				dseg = mlx5_tx_dseg_vlan(txq, loc, dseg,
4421 							 dptr, dlen, olx);
4422 #ifdef MLX5_PMD_SOFT_COUNTERS
4423 				/* Update sent data bytes counter. */
4424 				slen +=	sizeof(struct rte_vlan_hdr);
4425 #endif
4426 			} else {
4427 				if (room < tlen)
4428 					break;
4429 				dseg = mlx5_tx_dseg_empw(txq, loc, dseg,
4430 							 dptr, dlen, olx);
4431 			}
4432 			if (!MLX5_TXOFF_CONFIG(MPW))
4433 				tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE);
4434 			MLX5_ASSERT(room >= tlen);
4435 			room -= tlen;
4436 			/*
4437 			 * Packet data are completely inline,
4438 			 * we can try to free the packet.
4439 			 */
4440 			if (likely(loc->pkts_sent == loc->mbuf_free)) {
4441 				/*
4442 				 * All the packets from the burst beginning
4443 				 * are inline, we can free mbufs directly
4444 				 * from the origin array on tx_burst exit().
4445 				 */
4446 				loc->mbuf_free++;
4447 				goto next_mbuf;
4448 			}
4449 			/*
4450 			 * In order no to call rte_pktmbuf_free_seg() here,
4451 			 * in the most inner loop (that might be very
4452 			 * expensive) we just save the mbuf in elts.
4453 			 */
4454 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
4455 			loc->elts_free--;
4456 			goto next_mbuf;
4457 pointer_empw:
4458 			/*
4459 			 * No pointer and inline descriptor
4460 			 * intermix for legacy MPW sessions.
4461 			 */
4462 			if (MLX5_TXOFF_CONFIG(MPW) &&
4463 			    part != room &&
4464 			    wqem->dseg[0].bcount == RTE_BE32(0))
4465 				break;
4466 			/*
4467 			 * Not inlinable VLAN packets are
4468 			 * proceeded outside of this routine.
4469 			 */
4470 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4471 			if (MLX5_TXOFF_CONFIG(VLAN))
4472 				MLX5_ASSERT(!(loc->mbuf->ol_flags &
4473 					    PKT_TX_VLAN_PKT));
4474 			mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
4475 			/* We have to store mbuf in elts.*/
4476 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
4477 			loc->elts_free--;
4478 			room -= MLX5_WQE_DSEG_SIZE;
4479 			/* Ring buffer wraparound is checked at the loop end.*/
4480 			++dseg;
4481 next_mbuf:
4482 #ifdef MLX5_PMD_SOFT_COUNTERS
4483 			/* Update sent data bytes counter. */
4484 			slen += dlen;
4485 #endif
4486 			loc->pkts_sent++;
4487 			pkts_n--;
4488 			if (unlikely(!pkts_n || !loc->elts_free)) {
4489 				/*
4490 				 * We have no resources/packets to
4491 				 * continue build descriptors.
4492 				 */
4493 				part -= room;
4494 				mlx5_tx_idone_empw(txq, loc, part,
4495 						   slen, wqem, olx);
4496 				return MLX5_TXCMP_CODE_EXIT;
4497 			}
4498 			loc->mbuf = *pkts++;
4499 			if (likely(pkts_n > 1))
4500 				rte_prefetch0(*pkts);
4501 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4502 			/*
4503 			 * Unroll the completion code to avoid
4504 			 * returning variable value - it results in
4505 			 * unoptimized sequent checking in caller.
4506 			 */
4507 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4508 				part -= room;
4509 				mlx5_tx_idone_empw(txq, loc, part,
4510 						   slen, wqem, olx);
4511 				if (unlikely(!loc->elts_free ||
4512 					     !loc->wqe_free))
4513 					return MLX5_TXCMP_CODE_EXIT;
4514 				return MLX5_TXCMP_CODE_MULTI;
4515 			}
4516 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4517 			if (ret == MLX5_TXCMP_CODE_TSO) {
4518 				part -= room;
4519 				mlx5_tx_idone_empw(txq, loc, part,
4520 						   slen, wqem, olx);
4521 				if (unlikely(!loc->elts_free ||
4522 					     !loc->wqe_free))
4523 					return MLX5_TXCMP_CODE_EXIT;
4524 				return MLX5_TXCMP_CODE_TSO;
4525 			}
4526 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4527 				part -= room;
4528 				mlx5_tx_idone_empw(txq, loc, part,
4529 						   slen, wqem, olx);
4530 				if (unlikely(!loc->elts_free ||
4531 					     !loc->wqe_free))
4532 					return MLX5_TXCMP_CODE_EXIT;
4533 				return MLX5_TXCMP_CODE_SINGLE;
4534 			}
4535 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4536 				MLX5_ASSERT(false);
4537 				part -= room;
4538 				mlx5_tx_idone_empw(txq, loc, part,
4539 						   slen, wqem, olx);
4540 				return MLX5_TXCMP_CODE_ERROR;
4541 			}
4542 			/* Check if we have minimal room left. */
4543 			nlim--;
4544 			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
4545 				break;
4546 			/*
4547 			 * Check whether packet parameters coincide
4548 			 * within assumed eMPW batch:
4549 			 * - check sum settings
4550 			 * - metadata value
4551 			 * - software parser settings
4552 			 * - packets length (legacy MPW only)
4553 			 * - scheduling is not required
4554 			 */
4555 			if (!mlx5_tx_match_empw(txq, &wqem->eseg,
4556 						loc, dlen, olx))
4557 				break;
4558 			/* Packet attributes match, continue the same eMPW. */
4559 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4560 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4561 		}
4562 		/*
4563 		 * We get here to close an existing eMPW
4564 		 * session and start the new one.
4565 		 */
4566 		MLX5_ASSERT(pkts_n);
4567 		part -= room;
4568 		if (unlikely(!part))
4569 			return MLX5_TXCMP_CODE_EXIT;
4570 		mlx5_tx_idone_empw(txq, loc, part, slen, wqem, olx);
4571 		if (unlikely(!loc->elts_free ||
4572 			     !loc->wqe_free))
4573 			return MLX5_TXCMP_CODE_EXIT;
4574 		/* Continue the loop with new eMPW session. */
4575 	}
4576 	MLX5_ASSERT(false);
4577 }
4578 
4579 /**
4580  * The routine sends packets with ordinary MLX5_OPCODE_SEND.
4581  * Data inlining and VLAN insertion are supported.
4582  */
4583 static __rte_always_inline enum mlx5_txcmp_code
4584 mlx5_tx_burst_single_send(struct mlx5_txq_data *__rte_restrict txq,
4585 			  struct rte_mbuf **__rte_restrict pkts,
4586 			  unsigned int pkts_n,
4587 			  struct mlx5_txq_local *__rte_restrict loc,
4588 			  unsigned int olx)
4589 {
4590 	/*
4591 	 * Subroutine is the part of mlx5_tx_burst_single()
4592 	 * and sends single-segment packet with SEND opcode.
4593 	 */
4594 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4595 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4596 	pkts += loc->pkts_sent + 1;
4597 	pkts_n -= loc->pkts_sent;
4598 	for (;;) {
4599 		struct mlx5_wqe *__rte_restrict wqe;
4600 		enum mlx5_txcmp_code ret;
4601 
4602 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4603 		if (MLX5_TXOFF_CONFIG(TXPP)) {
4604 			enum mlx5_txcmp_code wret;
4605 
4606 			/* Generate WAIT for scheduling if requested. */
4607 			wret = mlx5_tx_schedule_send(txq, loc, olx);
4608 			if (wret == MLX5_TXCMP_CODE_EXIT)
4609 				return MLX5_TXCMP_CODE_EXIT;
4610 			if (wret == MLX5_TXCMP_CODE_ERROR)
4611 				return MLX5_TXCMP_CODE_ERROR;
4612 		}
4613 		if (MLX5_TXOFF_CONFIG(INLINE)) {
4614 			unsigned int inlen, vlan = 0;
4615 
4616 			inlen = rte_pktmbuf_data_len(loc->mbuf);
4617 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4618 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4619 				vlan = sizeof(struct rte_vlan_hdr);
4620 				inlen += vlan;
4621 			}
4622 			/*
4623 			 * If inlining is enabled at configuration time
4624 			 * the limit must be not less than minimal size.
4625 			 * Otherwise we would do extra check for data
4626 			 * size to avoid crashes due to length overflow.
4627 			 */
4628 			MLX5_ASSERT(txq->inlen_send >=
4629 				    MLX5_ESEG_MIN_INLINE_SIZE);
4630 			if (inlen <= txq->inlen_send) {
4631 				unsigned int seg_n, wqe_n;
4632 
4633 				rte_prefetch0(rte_pktmbuf_mtod
4634 						(loc->mbuf, uint8_t *));
4635 				/* Check against minimal length. */
4636 				if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
4637 					return MLX5_TXCMP_CODE_ERROR;
4638 				if (loc->mbuf->ol_flags &
4639 				    PKT_TX_DYNF_NOINLINE) {
4640 					/*
4641 					 * The hint flag not to inline packet
4642 					 * data is set. Check whether we can
4643 					 * follow the hint.
4644 					 */
4645 					if ((!MLX5_TXOFF_CONFIG(EMPW) &&
4646 					      txq->inlen_mode) ||
4647 					    (MLX5_TXOFF_CONFIG(MPW) &&
4648 					     txq->inlen_mode)) {
4649 						if (inlen <= txq->inlen_send)
4650 							goto single_inline;
4651 						/*
4652 						 * The hardware requires the
4653 						 * minimal inline data header.
4654 						 */
4655 						goto single_min_inline;
4656 					}
4657 					if (MLX5_TXOFF_CONFIG(VLAN) &&
4658 					    vlan && !txq->vlan_en) {
4659 						/*
4660 						 * We must insert VLAN tag
4661 						 * by software means.
4662 						 */
4663 						goto single_part_inline;
4664 					}
4665 					goto single_no_inline;
4666 				}
4667 single_inline:
4668 				/*
4669 				 * Completely inlined packet data WQE:
4670 				 * - Control Segment, SEND opcode
4671 				 * - Ethernet Segment, no VLAN insertion
4672 				 * - Data inlined, VLAN optionally inserted
4673 				 * - Alignment to MLX5_WSEG_SIZE
4674 				 * Have to estimate amount of WQEBBs
4675 				 */
4676 				seg_n = (inlen + 3 * MLX5_WSEG_SIZE -
4677 					 MLX5_ESEG_MIN_INLINE_SIZE +
4678 					 MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4679 				/* Check if there are enough WQEBBs. */
4680 				wqe_n = (seg_n + 3) / 4;
4681 				if (wqe_n > loc->wqe_free)
4682 					return MLX5_TXCMP_CODE_EXIT;
4683 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4684 				loc->wqe_last = wqe;
4685 				mlx5_tx_cseg_init(txq, loc, wqe, seg_n,
4686 						  MLX5_OPCODE_SEND, olx);
4687 				mlx5_tx_eseg_data(txq, loc, wqe,
4688 						  vlan, inlen, 0, olx);
4689 				txq->wqe_ci += wqe_n;
4690 				loc->wqe_free -= wqe_n;
4691 				/*
4692 				 * Packet data are completely inlined,
4693 				 * free the packet immediately.
4694 				 */
4695 				rte_pktmbuf_free_seg(loc->mbuf);
4696 			} else if ((!MLX5_TXOFF_CONFIG(EMPW) ||
4697 				     MLX5_TXOFF_CONFIG(MPW)) &&
4698 					txq->inlen_mode) {
4699 				/*
4700 				 * If minimal inlining is requested the eMPW
4701 				 * feature should be disabled due to data is
4702 				 * inlined into Ethernet Segment, which can
4703 				 * not contain inlined data for eMPW due to
4704 				 * segment shared for all packets.
4705 				 */
4706 				struct mlx5_wqe_dseg *__rte_restrict dseg;
4707 				unsigned int ds;
4708 				uint8_t *dptr;
4709 
4710 				/*
4711 				 * The inline-mode settings require
4712 				 * to inline the specified amount of
4713 				 * data bytes to the Ethernet Segment.
4714 				 * We should check the free space in
4715 				 * WQE ring buffer to inline partially.
4716 				 */
4717 single_min_inline:
4718 				MLX5_ASSERT(txq->inlen_send >= txq->inlen_mode);
4719 				MLX5_ASSERT(inlen > txq->inlen_mode);
4720 				MLX5_ASSERT(txq->inlen_mode >=
4721 					    MLX5_ESEG_MIN_INLINE_SIZE);
4722 				/*
4723 				 * Check whether there are enough free WQEBBs:
4724 				 * - Control Segment
4725 				 * - Ethernet Segment
4726 				 * - First Segment of inlined Ethernet data
4727 				 * - ... data continued ...
4728 				 * - Finishing Data Segment of pointer type
4729 				 */
4730 				ds = (MLX5_WQE_CSEG_SIZE +
4731 				      MLX5_WQE_ESEG_SIZE +
4732 				      MLX5_WQE_DSEG_SIZE +
4733 				      txq->inlen_mode -
4734 				      MLX5_ESEG_MIN_INLINE_SIZE +
4735 				      MLX5_WQE_DSEG_SIZE +
4736 				      MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4737 				if (loc->wqe_free < ((ds + 3) / 4))
4738 					return MLX5_TXCMP_CODE_EXIT;
4739 				/*
4740 				 * Build the ordinary SEND WQE:
4741 				 * - Control Segment
4742 				 * - Ethernet Segment, inline inlen_mode bytes
4743 				 * - Data Segment of pointer type
4744 				 */
4745 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4746 				loc->wqe_last = wqe;
4747 				mlx5_tx_cseg_init(txq, loc, wqe, ds,
4748 						  MLX5_OPCODE_SEND, olx);
4749 				dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan,
4750 							 txq->inlen_mode,
4751 							 0, olx);
4752 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4753 				       txq->inlen_mode - vlan;
4754 				inlen -= txq->inlen_mode;
4755 				mlx5_tx_dseg_ptr(txq, loc, dseg,
4756 						 dptr, inlen, olx);
4757 				/*
4758 				 * WQE is built, update the loop parameters
4759 				 * and got to the next packet.
4760 				 */
4761 				txq->wqe_ci += (ds + 3) / 4;
4762 				loc->wqe_free -= (ds + 3) / 4;
4763 				/* We have to store mbuf in elts.*/
4764 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4765 				txq->elts[txq->elts_head++ & txq->elts_m] =
4766 						loc->mbuf;
4767 				--loc->elts_free;
4768 			} else {
4769 				uint8_t *dptr;
4770 				unsigned int dlen;
4771 
4772 				/*
4773 				 * Partially inlined packet data WQE, we have
4774 				 * some space in title WQEBB, we can fill it
4775 				 * with some packet data. It takes one WQEBB,
4776 				 * it is available, no extra space check:
4777 				 * - Control Segment, SEND opcode
4778 				 * - Ethernet Segment, no VLAN insertion
4779 				 * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data
4780 				 * - Data Segment, pointer type
4781 				 *
4782 				 * We also get here if VLAN insertion is not
4783 				 * supported by HW, the inline is enabled.
4784 				 */
4785 single_part_inline:
4786 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4787 				loc->wqe_last = wqe;
4788 				mlx5_tx_cseg_init(txq, loc, wqe, 4,
4789 						  MLX5_OPCODE_SEND, olx);
4790 				mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx);
4791 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4792 				       MLX5_ESEG_MIN_INLINE_SIZE - vlan;
4793 				/*
4794 				 * The length check is performed above, by
4795 				 * comparing with txq->inlen_send. We should
4796 				 * not get overflow here.
4797 				 */
4798 				MLX5_ASSERT(inlen > MLX5_ESEG_MIN_INLINE_SIZE);
4799 				dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE;
4800 				mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1],
4801 						 dptr, dlen, olx);
4802 				++txq->wqe_ci;
4803 				--loc->wqe_free;
4804 				/* We have to store mbuf in elts.*/
4805 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4806 				txq->elts[txq->elts_head++ & txq->elts_m] =
4807 						loc->mbuf;
4808 				--loc->elts_free;
4809 			}
4810 #ifdef MLX5_PMD_SOFT_COUNTERS
4811 			/* Update sent data bytes counter. */
4812 			txq->stats.obytes += vlan +
4813 					rte_pktmbuf_data_len(loc->mbuf);
4814 #endif
4815 		} else {
4816 			/*
4817 			 * No inline at all, it means the CPU cycles saving
4818 			 * is prioritized at configuration, we should not
4819 			 * copy any packet data to WQE.
4820 			 *
4821 			 * SEND WQE, one WQEBB:
4822 			 * - Control Segment, SEND opcode
4823 			 * - Ethernet Segment, optional VLAN, no inline
4824 			 * - Data Segment, pointer type
4825 			 */
4826 single_no_inline:
4827 			wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4828 			loc->wqe_last = wqe;
4829 			mlx5_tx_cseg_init(txq, loc, wqe, 3,
4830 					  MLX5_OPCODE_SEND, olx);
4831 			mlx5_tx_eseg_none(txq, loc, wqe, olx);
4832 			mlx5_tx_dseg_ptr
4833 				(txq, loc, &wqe->dseg[0],
4834 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4835 				 rte_pktmbuf_data_len(loc->mbuf), olx);
4836 			++txq->wqe_ci;
4837 			--loc->wqe_free;
4838 			/*
4839 			 * We should not store mbuf pointer in elts
4840 			 * if no inlining is configured, this is done
4841 			 * by calling routine in a batch copy.
4842 			 */
4843 			MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
4844 			--loc->elts_free;
4845 #ifdef MLX5_PMD_SOFT_COUNTERS
4846 			/* Update sent data bytes counter. */
4847 			txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf);
4848 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4849 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
4850 				txq->stats.obytes +=
4851 					sizeof(struct rte_vlan_hdr);
4852 #endif
4853 		}
4854 		++loc->pkts_sent;
4855 		--pkts_n;
4856 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4857 			return MLX5_TXCMP_CODE_EXIT;
4858 		loc->mbuf = *pkts++;
4859 		if (pkts_n > 1)
4860 			rte_prefetch0(*pkts);
4861 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4862 		if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE))
4863 			return ret;
4864 	}
4865 	MLX5_ASSERT(false);
4866 }
4867 
4868 static __rte_always_inline enum mlx5_txcmp_code
4869 mlx5_tx_burst_single(struct mlx5_txq_data *__rte_restrict txq,
4870 		     struct rte_mbuf **__rte_restrict pkts,
4871 		     unsigned int pkts_n,
4872 		     struct mlx5_txq_local *__rte_restrict loc,
4873 		     unsigned int olx)
4874 {
4875 	enum mlx5_txcmp_code ret;
4876 
4877 	ret = mlx5_tx_able_to_empw(txq, loc, olx, false);
4878 	if (ret == MLX5_TXCMP_CODE_SINGLE)
4879 		goto ordinary_send;
4880 	MLX5_ASSERT(ret == MLX5_TXCMP_CODE_EMPW);
4881 	for (;;) {
4882 		/* Optimize for inline/no inline eMPW send. */
4883 		ret = (MLX5_TXOFF_CONFIG(INLINE)) ?
4884 			mlx5_tx_burst_empw_inline
4885 				(txq, pkts, pkts_n, loc, olx) :
4886 			mlx5_tx_burst_empw_simple
4887 				(txq, pkts, pkts_n, loc, olx);
4888 		if (ret != MLX5_TXCMP_CODE_SINGLE)
4889 			return ret;
4890 		/* The resources to send one packet should remain. */
4891 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4892 ordinary_send:
4893 		ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx);
4894 		MLX5_ASSERT(ret != MLX5_TXCMP_CODE_SINGLE);
4895 		if (ret != MLX5_TXCMP_CODE_EMPW)
4896 			return ret;
4897 		/* The resources to send one packet should remain. */
4898 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4899 	}
4900 }
4901 
4902 /**
4903  * DPDK Tx callback template. This is configured template
4904  * used to generate routines optimized for specified offload setup.
4905  * One of this generated functions is chosen at SQ configuration
4906  * time.
4907  *
4908  * @param txq
4909  *   Generic pointer to TX queue structure.
4910  * @param[in] pkts
4911  *   Packets to transmit.
4912  * @param pkts_n
4913  *   Number of packets in array.
4914  * @param olx
4915  *   Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx
4916  *   values. Should be static to take compile time static configuration
4917  *   advantages.
4918  *
4919  * @return
4920  *   Number of packets successfully transmitted (<= pkts_n).
4921  */
4922 static __rte_always_inline uint16_t
4923 mlx5_tx_burst_tmpl(struct mlx5_txq_data *__rte_restrict txq,
4924 		   struct rte_mbuf **__rte_restrict pkts,
4925 		   uint16_t pkts_n,
4926 		   unsigned int olx)
4927 {
4928 	struct mlx5_txq_local loc;
4929 	enum mlx5_txcmp_code ret;
4930 	unsigned int part;
4931 
4932 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4933 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4934 	if (unlikely(!pkts_n))
4935 		return 0;
4936 	if (MLX5_TXOFF_CONFIG(INLINE))
4937 		loc.mbuf_free = 0;
4938 	loc.pkts_sent = 0;
4939 	loc.pkts_copy = 0;
4940 	loc.wqe_last = NULL;
4941 
4942 send_loop:
4943 	loc.pkts_loop = loc.pkts_sent;
4944 	/*
4945 	 * Check if there are some CQEs, if any:
4946 	 * - process an encountered errors
4947 	 * - process the completed WQEs
4948 	 * - free related mbufs
4949 	 * - doorbell the NIC about processed CQEs
4950 	 */
4951 	rte_prefetch0(*(pkts + loc.pkts_sent));
4952 	mlx5_tx_handle_completion(txq, olx);
4953 	/*
4954 	 * Calculate the number of available resources - elts and WQEs.
4955 	 * There are two possible different scenarios:
4956 	 * - no data inlining into WQEs, one WQEBB may contains up to
4957 	 *   four packets, in this case elts become scarce resource
4958 	 * - data inlining into WQEs, one packet may require multiple
4959 	 *   WQEBBs, the WQEs become the limiting factor.
4960 	 */
4961 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4962 	loc.elts_free = txq->elts_s -
4963 				(uint16_t)(txq->elts_head - txq->elts_tail);
4964 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4965 	loc.wqe_free = txq->wqe_s -
4966 				(uint16_t)(txq->wqe_ci - txq->wqe_pi);
4967 	if (unlikely(!loc.elts_free || !loc.wqe_free))
4968 		goto burst_exit;
4969 	for (;;) {
4970 		/*
4971 		 * Fetch the packet from array. Usually this is
4972 		 * the first packet in series of multi/single
4973 		 * segment packets.
4974 		 */
4975 		loc.mbuf = *(pkts + loc.pkts_sent);
4976 		/* Dedicated branch for multi-segment packets. */
4977 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4978 		    unlikely(NB_SEGS(loc.mbuf) > 1)) {
4979 			/*
4980 			 * Multi-segment packet encountered.
4981 			 * Hardware is able to process it only
4982 			 * with SEND/TSO opcodes, one packet
4983 			 * per WQE, do it in dedicated routine.
4984 			 */
4985 enter_send_multi:
4986 			MLX5_ASSERT(loc.pkts_sent >= loc.pkts_copy);
4987 			part = loc.pkts_sent - loc.pkts_copy;
4988 			if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4989 				/*
4990 				 * There are some single-segment mbufs not
4991 				 * stored in elts. The mbufs must be in the
4992 				 * same order as WQEs, so we must copy the
4993 				 * mbufs to elts here, before the coming
4994 				 * multi-segment packet mbufs is appended.
4995 				 */
4996 				mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy,
4997 						  part, olx);
4998 				loc.pkts_copy = loc.pkts_sent;
4999 			}
5000 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
5001 			ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx);
5002 			if (!MLX5_TXOFF_CONFIG(INLINE))
5003 				loc.pkts_copy = loc.pkts_sent;
5004 			/*
5005 			 * These returned code checks are supposed
5006 			 * to be optimized out due to routine inlining.
5007 			 */
5008 			if (ret == MLX5_TXCMP_CODE_EXIT) {
5009 				/*
5010 				 * The routine returns this code when
5011 				 * all packets are sent or there is no
5012 				 * enough resources to complete request.
5013 				 */
5014 				break;
5015 			}
5016 			if (ret == MLX5_TXCMP_CODE_ERROR) {
5017 				/*
5018 				 * The routine returns this code when
5019 				 * some error in the incoming packets
5020 				 * format occurred.
5021 				 */
5022 				txq->stats.oerrors++;
5023 				break;
5024 			}
5025 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
5026 				/*
5027 				 * The single-segment packet was encountered
5028 				 * in the array, try to send it with the
5029 				 * best optimized way, possible engaging eMPW.
5030 				 */
5031 				goto enter_send_single;
5032 			}
5033 			if (MLX5_TXOFF_CONFIG(TSO) &&
5034 			    ret == MLX5_TXCMP_CODE_TSO) {
5035 				/*
5036 				 * The single-segment TSO packet was
5037 				 * encountered in the array.
5038 				 */
5039 				goto enter_send_tso;
5040 			}
5041 			/* We must not get here. Something is going wrong. */
5042 			MLX5_ASSERT(false);
5043 			txq->stats.oerrors++;
5044 			break;
5045 		}
5046 		/* Dedicated branch for single-segment TSO packets. */
5047 		if (MLX5_TXOFF_CONFIG(TSO) &&
5048 		    unlikely(loc.mbuf->ol_flags & PKT_TX_TCP_SEG)) {
5049 			/*
5050 			 * TSO might require special way for inlining
5051 			 * (dedicated parameters) and is sent with
5052 			 * MLX5_OPCODE_TSO opcode only, provide this
5053 			 * in dedicated branch.
5054 			 */
5055 enter_send_tso:
5056 			MLX5_ASSERT(NB_SEGS(loc.mbuf) == 1);
5057 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
5058 			ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx);
5059 			/*
5060 			 * These returned code checks are supposed
5061 			 * to be optimized out due to routine inlining.
5062 			 */
5063 			if (ret == MLX5_TXCMP_CODE_EXIT)
5064 				break;
5065 			if (ret == MLX5_TXCMP_CODE_ERROR) {
5066 				txq->stats.oerrors++;
5067 				break;
5068 			}
5069 			if (ret == MLX5_TXCMP_CODE_SINGLE)
5070 				goto enter_send_single;
5071 			if (MLX5_TXOFF_CONFIG(MULTI) &&
5072 			    ret == MLX5_TXCMP_CODE_MULTI) {
5073 				/*
5074 				 * The multi-segment packet was
5075 				 * encountered in the array.
5076 				 */
5077 				goto enter_send_multi;
5078 			}
5079 			/* We must not get here. Something is going wrong. */
5080 			MLX5_ASSERT(false);
5081 			txq->stats.oerrors++;
5082 			break;
5083 		}
5084 		/*
5085 		 * The dedicated branch for the single-segment packets
5086 		 * without TSO. Often these ones can be sent using
5087 		 * MLX5_OPCODE_EMPW with multiple packets in one WQE.
5088 		 * The routine builds the WQEs till it encounters
5089 		 * the TSO or multi-segment packet (in case if these
5090 		 * offloads are requested at SQ configuration time).
5091 		 */
5092 enter_send_single:
5093 		MLX5_ASSERT(pkts_n > loc.pkts_sent);
5094 		ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx);
5095 		/*
5096 		 * These returned code checks are supposed
5097 		 * to be optimized out due to routine inlining.
5098 		 */
5099 		if (ret == MLX5_TXCMP_CODE_EXIT)
5100 			break;
5101 		if (ret == MLX5_TXCMP_CODE_ERROR) {
5102 			txq->stats.oerrors++;
5103 			break;
5104 		}
5105 		if (MLX5_TXOFF_CONFIG(MULTI) &&
5106 		    ret == MLX5_TXCMP_CODE_MULTI) {
5107 			/*
5108 			 * The multi-segment packet was
5109 			 * encountered in the array.
5110 			 */
5111 			goto enter_send_multi;
5112 		}
5113 		if (MLX5_TXOFF_CONFIG(TSO) &&
5114 		    ret == MLX5_TXCMP_CODE_TSO) {
5115 			/*
5116 			 * The single-segment TSO packet was
5117 			 * encountered in the array.
5118 			 */
5119 			goto enter_send_tso;
5120 		}
5121 		/* We must not get here. Something is going wrong. */
5122 		MLX5_ASSERT(false);
5123 		txq->stats.oerrors++;
5124 		break;
5125 	}
5126 	/*
5127 	 * Main Tx loop is completed, do the rest:
5128 	 * - set completion request if thresholds are reached
5129 	 * - doorbell the hardware
5130 	 * - copy the rest of mbufs to elts (if any)
5131 	 */
5132 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE) ||
5133 		    loc.pkts_sent >= loc.pkts_copy);
5134 	/* Take a shortcut if nothing is sent. */
5135 	if (unlikely(loc.pkts_sent == loc.pkts_loop))
5136 		goto burst_exit;
5137 	/* Request CQE generation if limits are reached. */
5138 	mlx5_tx_request_completion(txq, &loc, olx);
5139 	/*
5140 	 * Ring QP doorbell immediately after WQE building completion
5141 	 * to improve latencies. The pure software related data treatment
5142 	 * can be completed after doorbell. Tx CQEs for this SQ are
5143 	 * processed in this thread only by the polling.
5144 	 *
5145 	 * The rdma core library can map doorbell register in two ways,
5146 	 * depending on the environment variable "MLX5_SHUT_UP_BF":
5147 	 *
5148 	 * - as regular cached memory, the variable is either missing or
5149 	 *   set to zero. This type of mapping may cause the significant
5150 	 *   doorbell register writing latency and requires explicit
5151 	 *   memory write barrier to mitigate this issue and prevent
5152 	 *   write combining.
5153 	 *
5154 	 * - as non-cached memory, the variable is present and set to
5155 	 *   not "0" value. This type of mapping may cause performance
5156 	 *   impact under heavy loading conditions but the explicit write
5157 	 *   memory barrier is not required and it may improve core
5158 	 *   performance.
5159 	 *
5160 	 * - the legacy behaviour (prior 19.08 release) was to use some
5161 	 *   heuristics to decide whether write memory barrier should
5162 	 *   be performed. This behavior is supported with specifying
5163 	 *   tx_db_nc=2, write barrier is skipped if application
5164 	 *   provides the full recommended burst of packets, it
5165 	 *   supposes the next packets are coming and the write barrier
5166 	 *   will be issued on the next burst (after descriptor writing,
5167 	 *   at least).
5168 	 */
5169 	mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, !txq->db_nc &&
5170 			(!txq->db_heu || pkts_n % MLX5_TX_DEFAULT_BURST));
5171 	/* Not all of the mbufs may be stored into elts yet. */
5172 	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
5173 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
5174 		/*
5175 		 * There are some single-segment mbufs not stored in elts.
5176 		 * It can be only if the last packet was single-segment.
5177 		 * The copying is gathered into one place due to it is
5178 		 * a good opportunity to optimize that with SIMD.
5179 		 * Unfortunately if inlining is enabled the gaps in
5180 		 * pointer array may happen due to early freeing of the
5181 		 * inlined mbufs.
5182 		 */
5183 		mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx);
5184 		loc.pkts_copy = loc.pkts_sent;
5185 	}
5186 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
5187 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
5188 	if (pkts_n > loc.pkts_sent) {
5189 		/*
5190 		 * If burst size is large there might be no enough CQE
5191 		 * fetched from completion queue and no enough resources
5192 		 * freed to send all the packets.
5193 		 */
5194 		goto send_loop;
5195 	}
5196 burst_exit:
5197 #ifdef MLX5_PMD_SOFT_COUNTERS
5198 	/* Increment sent packets counter. */
5199 	txq->stats.opackets += loc.pkts_sent;
5200 #endif
5201 	if (MLX5_TXOFF_CONFIG(INLINE) && loc.mbuf_free)
5202 		__mlx5_tx_free_mbuf(txq, pkts, loc.mbuf_free, olx);
5203 	return loc.pkts_sent;
5204 }
5205 
5206 /* Generate routines with Enhanced Multi-Packet Write support. */
5207 MLX5_TXOFF_DECL(full_empw,
5208 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_EMPW)
5209 
5210 MLX5_TXOFF_DECL(none_empw,
5211 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5212 
5213 MLX5_TXOFF_DECL(md_empw,
5214 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5215 
5216 MLX5_TXOFF_DECL(mt_empw,
5217 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5218 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5219 
5220 MLX5_TXOFF_DECL(mtsc_empw,
5221 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5222 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5223 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5224 
5225 MLX5_TXOFF_DECL(mti_empw,
5226 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5227 		MLX5_TXOFF_CONFIG_INLINE |
5228 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5229 
5230 MLX5_TXOFF_DECL(mtv_empw,
5231 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5232 		MLX5_TXOFF_CONFIG_VLAN |
5233 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5234 
5235 MLX5_TXOFF_DECL(mtiv_empw,
5236 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5237 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5238 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5239 
5240 MLX5_TXOFF_DECL(sc_empw,
5241 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5242 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5243 
5244 MLX5_TXOFF_DECL(sci_empw,
5245 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5246 		MLX5_TXOFF_CONFIG_INLINE |
5247 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5248 
5249 MLX5_TXOFF_DECL(scv_empw,
5250 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5251 		MLX5_TXOFF_CONFIG_VLAN |
5252 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5253 
5254 MLX5_TXOFF_DECL(sciv_empw,
5255 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5256 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5257 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5258 
5259 MLX5_TXOFF_DECL(i_empw,
5260 		MLX5_TXOFF_CONFIG_INLINE |
5261 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5262 
5263 MLX5_TXOFF_DECL(v_empw,
5264 		MLX5_TXOFF_CONFIG_VLAN |
5265 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5266 
5267 MLX5_TXOFF_DECL(iv_empw,
5268 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5269 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5270 
5271 /* Generate routines without Enhanced Multi-Packet Write support. */
5272 MLX5_TXOFF_DECL(full,
5273 		MLX5_TXOFF_CONFIG_FULL)
5274 
5275 MLX5_TXOFF_DECL(none,
5276 		MLX5_TXOFF_CONFIG_NONE)
5277 
5278 MLX5_TXOFF_DECL(md,
5279 		MLX5_TXOFF_CONFIG_METADATA)
5280 
5281 MLX5_TXOFF_DECL(mt,
5282 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5283 		MLX5_TXOFF_CONFIG_METADATA)
5284 
5285 MLX5_TXOFF_DECL(mtsc,
5286 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5287 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5288 		MLX5_TXOFF_CONFIG_METADATA)
5289 
5290 MLX5_TXOFF_DECL(mti,
5291 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5292 		MLX5_TXOFF_CONFIG_INLINE |
5293 		MLX5_TXOFF_CONFIG_METADATA)
5294 
5295 
5296 MLX5_TXOFF_DECL(mtv,
5297 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5298 		MLX5_TXOFF_CONFIG_VLAN |
5299 		MLX5_TXOFF_CONFIG_METADATA)
5300 
5301 
5302 MLX5_TXOFF_DECL(mtiv,
5303 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5304 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5305 		MLX5_TXOFF_CONFIG_METADATA)
5306 
5307 MLX5_TXOFF_DECL(sc,
5308 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5309 		MLX5_TXOFF_CONFIG_METADATA)
5310 
5311 MLX5_TXOFF_DECL(sci,
5312 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5313 		MLX5_TXOFF_CONFIG_INLINE |
5314 		MLX5_TXOFF_CONFIG_METADATA)
5315 
5316 
5317 MLX5_TXOFF_DECL(scv,
5318 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5319 		MLX5_TXOFF_CONFIG_VLAN |
5320 		MLX5_TXOFF_CONFIG_METADATA)
5321 
5322 
5323 MLX5_TXOFF_DECL(sciv,
5324 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5325 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5326 		MLX5_TXOFF_CONFIG_METADATA)
5327 
5328 MLX5_TXOFF_DECL(i,
5329 		MLX5_TXOFF_CONFIG_INLINE |
5330 		MLX5_TXOFF_CONFIG_METADATA)
5331 
5332 MLX5_TXOFF_DECL(v,
5333 		MLX5_TXOFF_CONFIG_VLAN |
5334 		MLX5_TXOFF_CONFIG_METADATA)
5335 
5336 MLX5_TXOFF_DECL(iv,
5337 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5338 		MLX5_TXOFF_CONFIG_METADATA)
5339 
5340 /* Generate routines with timestamp scheduling. */
5341 MLX5_TXOFF_DECL(full_ts_nompw,
5342 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP)
5343 
5344 MLX5_TXOFF_DECL(full_ts_nompwi,
5345 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5346 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5347 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5348 		MLX5_TXOFF_CONFIG_TXPP)
5349 
5350 MLX5_TXOFF_DECL(full_ts,
5351 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP |
5352 		MLX5_TXOFF_CONFIG_EMPW)
5353 
5354 MLX5_TXOFF_DECL(full_ts_noi,
5355 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5356 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5357 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5358 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5359 
5360 MLX5_TXOFF_DECL(none_ts,
5361 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_TXPP |
5362 		MLX5_TXOFF_CONFIG_EMPW)
5363 
5364 MLX5_TXOFF_DECL(mdi_ts,
5365 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5366 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5367 
5368 MLX5_TXOFF_DECL(mti_ts,
5369 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5370 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5371 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5372 
5373 MLX5_TXOFF_DECL(mtiv_ts,
5374 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5375 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5376 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_TXPP |
5377 		MLX5_TXOFF_CONFIG_EMPW)
5378 
5379 /*
5380  * Generate routines with Legacy Multi-Packet Write support.
5381  * This mode is supported by ConnectX-4 Lx only and imposes
5382  * offload limitations, not supported:
5383  *   - ACL/Flows (metadata are becoming meaningless)
5384  *   - WQE Inline headers
5385  *   - SRIOV (E-Switch offloads)
5386  *   - VLAN insertion
5387  *   - tunnel encapsulation/decapsulation
5388  *   - TSO
5389  */
5390 MLX5_TXOFF_DECL(none_mpw,
5391 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5392 		MLX5_TXOFF_CONFIG_MPW)
5393 
5394 MLX5_TXOFF_DECL(mci_mpw,
5395 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5396 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5397 		MLX5_TXOFF_CONFIG_MPW)
5398 
5399 MLX5_TXOFF_DECL(mc_mpw,
5400 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5401 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5402 
5403 MLX5_TXOFF_DECL(i_mpw,
5404 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5405 		MLX5_TXOFF_CONFIG_MPW)
5406 
5407 /*
5408  * Array of declared and compiled Tx burst function and corresponding
5409  * supported offloads set. The array is used to select the Tx burst
5410  * function for specified offloads set at Tx queue configuration time.
5411  */
5412 const struct {
5413 	eth_tx_burst_t func;
5414 	unsigned int olx;
5415 } txoff_func[] = {
5416 MLX5_TXOFF_INFO(full_empw,
5417 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5418 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5419 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5420 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5421 
5422 MLX5_TXOFF_INFO(none_empw,
5423 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5424 
5425 MLX5_TXOFF_INFO(md_empw,
5426 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5427 
5428 MLX5_TXOFF_INFO(mt_empw,
5429 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5430 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5431 
5432 MLX5_TXOFF_INFO(mtsc_empw,
5433 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5434 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5435 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5436 
5437 MLX5_TXOFF_INFO(mti_empw,
5438 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5439 		MLX5_TXOFF_CONFIG_INLINE |
5440 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5441 
5442 MLX5_TXOFF_INFO(mtv_empw,
5443 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5444 		MLX5_TXOFF_CONFIG_VLAN |
5445 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5446 
5447 MLX5_TXOFF_INFO(mtiv_empw,
5448 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5449 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5450 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5451 
5452 MLX5_TXOFF_INFO(sc_empw,
5453 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5454 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5455 
5456 MLX5_TXOFF_INFO(sci_empw,
5457 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5458 		MLX5_TXOFF_CONFIG_INLINE |
5459 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5460 
5461 MLX5_TXOFF_INFO(scv_empw,
5462 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5463 		MLX5_TXOFF_CONFIG_VLAN |
5464 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5465 
5466 MLX5_TXOFF_INFO(sciv_empw,
5467 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5468 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5469 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5470 
5471 MLX5_TXOFF_INFO(i_empw,
5472 		MLX5_TXOFF_CONFIG_INLINE |
5473 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5474 
5475 MLX5_TXOFF_INFO(v_empw,
5476 		MLX5_TXOFF_CONFIG_VLAN |
5477 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5478 
5479 MLX5_TXOFF_INFO(iv_empw,
5480 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5481 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5482 
5483 MLX5_TXOFF_INFO(full_ts_nompw,
5484 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP)
5485 
5486 MLX5_TXOFF_INFO(full_ts_nompwi,
5487 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5488 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5489 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5490 		MLX5_TXOFF_CONFIG_TXPP)
5491 
5492 MLX5_TXOFF_INFO(full_ts,
5493 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP |
5494 		MLX5_TXOFF_CONFIG_EMPW)
5495 
5496 MLX5_TXOFF_INFO(full_ts_noi,
5497 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5498 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5499 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5500 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5501 
5502 MLX5_TXOFF_INFO(none_ts,
5503 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_TXPP |
5504 		MLX5_TXOFF_CONFIG_EMPW)
5505 
5506 MLX5_TXOFF_INFO(mdi_ts,
5507 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5508 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5509 
5510 MLX5_TXOFF_INFO(mti_ts,
5511 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5512 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5513 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5514 
5515 MLX5_TXOFF_INFO(mtiv_ts,
5516 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5517 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5518 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_TXPP |
5519 		MLX5_TXOFF_CONFIG_EMPW)
5520 
5521 MLX5_TXOFF_INFO(full,
5522 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5523 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5524 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5525 		MLX5_TXOFF_CONFIG_METADATA)
5526 
5527 MLX5_TXOFF_INFO(none,
5528 		MLX5_TXOFF_CONFIG_NONE)
5529 
5530 MLX5_TXOFF_INFO(md,
5531 		MLX5_TXOFF_CONFIG_METADATA)
5532 
5533 MLX5_TXOFF_INFO(mt,
5534 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5535 		MLX5_TXOFF_CONFIG_METADATA)
5536 
5537 MLX5_TXOFF_INFO(mtsc,
5538 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5539 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5540 		MLX5_TXOFF_CONFIG_METADATA)
5541 
5542 MLX5_TXOFF_INFO(mti,
5543 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5544 		MLX5_TXOFF_CONFIG_INLINE |
5545 		MLX5_TXOFF_CONFIG_METADATA)
5546 
5547 MLX5_TXOFF_INFO(mtv,
5548 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5549 		MLX5_TXOFF_CONFIG_VLAN |
5550 		MLX5_TXOFF_CONFIG_METADATA)
5551 
5552 MLX5_TXOFF_INFO(mtiv,
5553 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5554 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5555 		MLX5_TXOFF_CONFIG_METADATA)
5556 
5557 MLX5_TXOFF_INFO(sc,
5558 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5559 		MLX5_TXOFF_CONFIG_METADATA)
5560 
5561 MLX5_TXOFF_INFO(sci,
5562 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5563 		MLX5_TXOFF_CONFIG_INLINE |
5564 		MLX5_TXOFF_CONFIG_METADATA)
5565 
5566 MLX5_TXOFF_INFO(scv,
5567 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5568 		MLX5_TXOFF_CONFIG_VLAN |
5569 		MLX5_TXOFF_CONFIG_METADATA)
5570 
5571 MLX5_TXOFF_INFO(sciv,
5572 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5573 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5574 		MLX5_TXOFF_CONFIG_METADATA)
5575 
5576 MLX5_TXOFF_INFO(i,
5577 		MLX5_TXOFF_CONFIG_INLINE |
5578 		MLX5_TXOFF_CONFIG_METADATA)
5579 
5580 MLX5_TXOFF_INFO(v,
5581 		MLX5_TXOFF_CONFIG_VLAN |
5582 		MLX5_TXOFF_CONFIG_METADATA)
5583 
5584 MLX5_TXOFF_INFO(iv,
5585 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5586 		MLX5_TXOFF_CONFIG_METADATA)
5587 
5588 MLX5_TXOFF_INFO(none_mpw,
5589 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5590 		MLX5_TXOFF_CONFIG_MPW)
5591 
5592 MLX5_TXOFF_INFO(mci_mpw,
5593 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5594 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5595 		MLX5_TXOFF_CONFIG_MPW)
5596 
5597 MLX5_TXOFF_INFO(mc_mpw,
5598 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5599 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5600 
5601 MLX5_TXOFF_INFO(i_mpw,
5602 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5603 		MLX5_TXOFF_CONFIG_MPW)
5604 };
5605 
5606 /**
5607  * Configure the Tx function to use. The routine checks configured
5608  * Tx offloads for the device and selects appropriate Tx burst
5609  * routine. There are multiple Tx burst routines compiled from
5610  * the same template in the most optimal way for the dedicated
5611  * Tx offloads set.
5612  *
5613  * @param dev
5614  *   Pointer to private data structure.
5615  *
5616  * @return
5617  *   Pointer to selected Tx burst function.
5618  */
5619 eth_tx_burst_t
5620 mlx5_select_tx_function(struct rte_eth_dev *dev)
5621 {
5622 	struct mlx5_priv *priv = dev->data->dev_private;
5623 	struct mlx5_dev_config *config = &priv->config;
5624 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
5625 	unsigned int diff = 0, olx = 0, i, m;
5626 
5627 	MLX5_ASSERT(priv);
5628 	if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) {
5629 		/* We should support Multi-Segment Packets. */
5630 		olx |= MLX5_TXOFF_CONFIG_MULTI;
5631 	}
5632 	if (tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
5633 			   DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
5634 			   DEV_TX_OFFLOAD_GRE_TNL_TSO |
5635 			   DEV_TX_OFFLOAD_IP_TNL_TSO |
5636 			   DEV_TX_OFFLOAD_UDP_TNL_TSO)) {
5637 		/* We should support TCP Send Offload. */
5638 		olx |= MLX5_TXOFF_CONFIG_TSO;
5639 	}
5640 	if (tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
5641 			   DEV_TX_OFFLOAD_UDP_TNL_TSO |
5642 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5643 		/* We should support Software Parser for Tunnels. */
5644 		olx |= MLX5_TXOFF_CONFIG_SWP;
5645 	}
5646 	if (tx_offloads & (DEV_TX_OFFLOAD_IPV4_CKSUM |
5647 			   DEV_TX_OFFLOAD_UDP_CKSUM |
5648 			   DEV_TX_OFFLOAD_TCP_CKSUM |
5649 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5650 		/* We should support IP/TCP/UDP Checksums. */
5651 		olx |= MLX5_TXOFF_CONFIG_CSUM;
5652 	}
5653 	if (tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT) {
5654 		/* We should support VLAN insertion. */
5655 		olx |= MLX5_TXOFF_CONFIG_VLAN;
5656 	}
5657 	if (tx_offloads & DEV_TX_OFFLOAD_SEND_ON_TIMESTAMP &&
5658 	    rte_mbuf_dynflag_lookup
5659 			(RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL) >= 0 &&
5660 	    rte_mbuf_dynfield_lookup
5661 			(RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL) >= 0) {
5662 		/* Offload configured, dynamic entities registered. */
5663 		olx |= MLX5_TXOFF_CONFIG_TXPP;
5664 	}
5665 	if (priv->txqs_n && (*priv->txqs)[0]) {
5666 		struct mlx5_txq_data *txd = (*priv->txqs)[0];
5667 
5668 		if (txd->inlen_send) {
5669 			/*
5670 			 * Check the data inline requirements. Data inline
5671 			 * is enabled on per device basis, we can check
5672 			 * the first Tx queue only.
5673 			 *
5674 			 * If device does not support VLAN insertion in WQE
5675 			 * and some queues are requested to perform VLAN
5676 			 * insertion offload than inline must be enabled.
5677 			 */
5678 			olx |= MLX5_TXOFF_CONFIG_INLINE;
5679 		}
5680 	}
5681 	if (config->mps == MLX5_MPW_ENHANCED &&
5682 	    config->txq_inline_min <= 0) {
5683 		/*
5684 		 * The NIC supports Enhanced Multi-Packet Write
5685 		 * and does not require minimal inline data.
5686 		 */
5687 		olx |= MLX5_TXOFF_CONFIG_EMPW;
5688 	}
5689 	if (rte_flow_dynf_metadata_avail()) {
5690 		/* We should support Flow metadata. */
5691 		olx |= MLX5_TXOFF_CONFIG_METADATA;
5692 	}
5693 	if (config->mps == MLX5_MPW) {
5694 		/*
5695 		 * The NIC supports Legacy Multi-Packet Write.
5696 		 * The MLX5_TXOFF_CONFIG_MPW controls the
5697 		 * descriptor building method in combination
5698 		 * with MLX5_TXOFF_CONFIG_EMPW.
5699 		 */
5700 		if (!(olx & (MLX5_TXOFF_CONFIG_TSO |
5701 			     MLX5_TXOFF_CONFIG_SWP |
5702 			     MLX5_TXOFF_CONFIG_VLAN |
5703 			     MLX5_TXOFF_CONFIG_METADATA)))
5704 			olx |= MLX5_TXOFF_CONFIG_EMPW |
5705 			       MLX5_TXOFF_CONFIG_MPW;
5706 	}
5707 	/*
5708 	 * Scan the routines table to find the minimal
5709 	 * satisfying routine with requested offloads.
5710 	 */
5711 	m = RTE_DIM(txoff_func);
5712 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5713 		unsigned int tmp;
5714 
5715 		tmp = txoff_func[i].olx;
5716 		if (tmp == olx) {
5717 			/* Meets requested offloads exactly.*/
5718 			m = i;
5719 			break;
5720 		}
5721 		if ((tmp & olx) != olx) {
5722 			/* Does not meet requested offloads at all. */
5723 			continue;
5724 		}
5725 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_MPW)
5726 			/* Do not enable legacy MPW if not configured. */
5727 			continue;
5728 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_EMPW)
5729 			/* Do not enable eMPW if not configured. */
5730 			continue;
5731 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_INLINE)
5732 			/* Do not enable inlining if not configured. */
5733 			continue;
5734 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_TXPP)
5735 			/* Do not enable scheduling if not configured. */
5736 			continue;
5737 		/*
5738 		 * Some routine meets the requirements.
5739 		 * Check whether it has minimal amount
5740 		 * of not requested offloads.
5741 		 */
5742 		tmp = __builtin_popcountl(tmp & ~olx);
5743 		if (m >= RTE_DIM(txoff_func) || tmp < diff) {
5744 			/* First or better match, save and continue. */
5745 			m = i;
5746 			diff = tmp;
5747 			continue;
5748 		}
5749 		if (tmp == diff) {
5750 			tmp = txoff_func[i].olx ^ txoff_func[m].olx;
5751 			if (__builtin_ffsl(txoff_func[i].olx & ~tmp) <
5752 			    __builtin_ffsl(txoff_func[m].olx & ~tmp)) {
5753 				/* Lighter not requested offload. */
5754 				m = i;
5755 			}
5756 		}
5757 	}
5758 	if (m >= RTE_DIM(txoff_func)) {
5759 		DRV_LOG(DEBUG, "port %u has no selected Tx function"
5760 			       " for requested offloads %04X",
5761 				dev->data->port_id, olx);
5762 		return NULL;
5763 	}
5764 	DRV_LOG(DEBUG, "port %u has selected Tx function"
5765 		       " supporting offloads %04X/%04X",
5766 			dev->data->port_id, olx, txoff_func[m].olx);
5767 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MULTI)
5768 		DRV_LOG(DEBUG, "\tMULTI (multi segment)");
5769 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TSO)
5770 		DRV_LOG(DEBUG, "\tTSO   (TCP send offload)");
5771 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_SWP)
5772 		DRV_LOG(DEBUG, "\tSWP   (software parser)");
5773 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_CSUM)
5774 		DRV_LOG(DEBUG, "\tCSUM  (checksum offload)");
5775 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_INLINE)
5776 		DRV_LOG(DEBUG, "\tINLIN (inline data)");
5777 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_VLAN)
5778 		DRV_LOG(DEBUG, "\tVLANI (VLAN insertion)");
5779 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_METADATA)
5780 		DRV_LOG(DEBUG, "\tMETAD (tx Flow metadata)");
5781 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TXPP)
5782 		DRV_LOG(DEBUG, "\tMETAD (tx Scheduling)");
5783 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW) {
5784 		if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MPW)
5785 			DRV_LOG(DEBUG, "\tMPW   (Legacy MPW)");
5786 		else
5787 			DRV_LOG(DEBUG, "\tEMPW  (Enhanced MPW)");
5788 	}
5789 	return txoff_func[m].func;
5790 }
5791 
5792 /**
5793  * DPDK callback to get the TX queue information
5794  *
5795  * @param dev
5796  *   Pointer to the device structure.
5797  *
5798  * @param tx_queue_id
5799  *   Tx queue identificator.
5800  *
5801  * @param qinfo
5802  *   Pointer to the TX queue information structure.
5803  *
5804  * @return
5805  *   None.
5806  */
5807 
5808 void
5809 mlx5_txq_info_get(struct rte_eth_dev *dev, uint16_t tx_queue_id,
5810 		  struct rte_eth_txq_info *qinfo)
5811 {
5812 	struct mlx5_priv *priv = dev->data->dev_private;
5813 	struct mlx5_txq_data *txq = (*priv->txqs)[tx_queue_id];
5814 	struct mlx5_txq_ctrl *txq_ctrl =
5815 			container_of(txq, struct mlx5_txq_ctrl, txq);
5816 
5817 	if (!txq)
5818 		return;
5819 	qinfo->nb_desc = txq->elts_s;
5820 	qinfo->conf.tx_thresh.pthresh = 0;
5821 	qinfo->conf.tx_thresh.hthresh = 0;
5822 	qinfo->conf.tx_thresh.wthresh = 0;
5823 	qinfo->conf.tx_rs_thresh = 0;
5824 	qinfo->conf.tx_free_thresh = 0;
5825 	qinfo->conf.tx_deferred_start = txq_ctrl ? 0 : 1;
5826 	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
5827 }
5828 
5829 /**
5830  * DPDK callback to get the TX packet burst mode information
5831  *
5832  * @param dev
5833  *   Pointer to the device structure.
5834  *
5835  * @param tx_queue_id
5836  *   Tx queue identificatior.
5837  *
5838  * @param mode
5839  *   Pointer to the burts mode information.
5840  *
5841  * @return
5842  *   0 as success, -EINVAL as failure.
5843  */
5844 
5845 int
5846 mlx5_tx_burst_mode_get(struct rte_eth_dev *dev,
5847 		       uint16_t tx_queue_id,
5848 		       struct rte_eth_burst_mode *mode)
5849 {
5850 	eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
5851 	struct mlx5_priv *priv = dev->data->dev_private;
5852 	struct mlx5_txq_data *txq = (*priv->txqs)[tx_queue_id];
5853 	unsigned int i, olx;
5854 
5855 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5856 		if (pkt_burst == txoff_func[i].func) {
5857 			olx = txoff_func[i].olx;
5858 			snprintf(mode->info, sizeof(mode->info),
5859 				 "%s%s%s%s%s%s%s%s%s%s",
5860 				 (olx & MLX5_TXOFF_CONFIG_EMPW) ?
5861 				 ((olx & MLX5_TXOFF_CONFIG_MPW) ?
5862 				 "Legacy MPW" : "Enhanced MPW") : "No MPW",
5863 				 (olx & MLX5_TXOFF_CONFIG_MULTI) ?
5864 				 " + MULTI" : "",
5865 				 (olx & MLX5_TXOFF_CONFIG_TSO) ?
5866 				 " + TSO" : "",
5867 				 (olx & MLX5_TXOFF_CONFIG_SWP) ?
5868 				 " + SWP" : "",
5869 				 (olx & MLX5_TXOFF_CONFIG_CSUM) ?
5870 				 "  + CSUM" : "",
5871 				 (olx & MLX5_TXOFF_CONFIG_INLINE) ?
5872 				 " + INLINE" : "",
5873 				 (olx & MLX5_TXOFF_CONFIG_VLAN) ?
5874 				 " + VLAN" : "",
5875 				 (olx & MLX5_TXOFF_CONFIG_METADATA) ?
5876 				 " + METADATA" : "",
5877 				 (olx & MLX5_TXOFF_CONFIG_TXPP) ?
5878 				 " + TXPP" : "",
5879 				 (txq && txq->fast_free) ?
5880 				 " + Fast Free" : "");
5881 			return 0;
5882 		}
5883 	}
5884 	return -EINVAL;
5885 }
5886